7 Commits
v9.2 ... v10.1

Author SHA1 Message Date
Mahesh Doijade
1abc294982 --Update README.md for cuda 10.1 release
-- fix multidevice CG sample to check for concurrentManagedAccess as it needs systemwide atomics
2019-01-23 16:56:20 +05:30
Mahesh Doijade
b458dafcd6 Add and update samples with CUDA 10.1 support 2019-01-23 01:34:43 +05:30
Mahesh Doijade
32f0fc6111 fix the description of CUDA IPC as it is supported on Windows as well since CUDA 9.2 2018-10-24 13:35:39 +05:30
Mahesh Doijade
6bbaf9ac80 -- use DX_SDK env variable only for non-DX12 samples VS files
as DX12 does not need legacy DX SDK
2018-10-09 22:42:25 +05:30
Mahesh Doijade
fcb23487a8 -- Add simpleD3D12 sample which demonstrates DX12-CUDA interop
rendering sinewave
-- Update the documentation with description about this sample
2018-10-04 20:05:23 +05:30
Mahesh Doijade
91dc60dd96 simpleVulkan: windows.h should be included before VersionHelpers.h 2018-09-10 20:41:05 +05:30
Mahesh Doijade
21c36d3568 Add and Update samples for CUDA 10.0 2018-08-24 22:35:15 +05:30
263 changed files with 24467 additions and 1320 deletions

View File

@@ -51,457 +51,17 @@
// CUDA Runtime error messages // CUDA Runtime error messages
#ifdef __DRIVER_TYPES_H__ #ifdef __DRIVER_TYPES_H__
static const char *_cudaGetErrorEnum(cudaError_t error) { static const char *_cudaGetErrorEnum(cudaError_t error) {
switch (error) { return cudaGetErrorName(error);
case cudaSuccess:
return "cudaSuccess";
case cudaErrorMissingConfiguration:
return "cudaErrorMissingConfiguration";
case cudaErrorMemoryAllocation:
return "cudaErrorMemoryAllocation";
case cudaErrorInitializationError:
return "cudaErrorInitializationError";
case cudaErrorLaunchFailure:
return "cudaErrorLaunchFailure";
case cudaErrorPriorLaunchFailure:
return "cudaErrorPriorLaunchFailure";
case cudaErrorLaunchTimeout:
return "cudaErrorLaunchTimeout";
case cudaErrorLaunchOutOfResources:
return "cudaErrorLaunchOutOfResources";
case cudaErrorInvalidDeviceFunction:
return "cudaErrorInvalidDeviceFunction";
case cudaErrorInvalidConfiguration:
return "cudaErrorInvalidConfiguration";
case cudaErrorInvalidDevice:
return "cudaErrorInvalidDevice";
case cudaErrorInvalidValue:
return "cudaErrorInvalidValue";
case cudaErrorInvalidPitchValue:
return "cudaErrorInvalidPitchValue";
case cudaErrorInvalidSymbol:
return "cudaErrorInvalidSymbol";
case cudaErrorMapBufferObjectFailed:
return "cudaErrorMapBufferObjectFailed";
case cudaErrorUnmapBufferObjectFailed:
return "cudaErrorUnmapBufferObjectFailed";
case cudaErrorInvalidHostPointer:
return "cudaErrorInvalidHostPointer";
case cudaErrorInvalidDevicePointer:
return "cudaErrorInvalidDevicePointer";
case cudaErrorInvalidTexture:
return "cudaErrorInvalidTexture";
case cudaErrorInvalidTextureBinding:
return "cudaErrorInvalidTextureBinding";
case cudaErrorInvalidChannelDescriptor:
return "cudaErrorInvalidChannelDescriptor";
case cudaErrorInvalidMemcpyDirection:
return "cudaErrorInvalidMemcpyDirection";
case cudaErrorAddressOfConstant:
return "cudaErrorAddressOfConstant";
case cudaErrorTextureFetchFailed:
return "cudaErrorTextureFetchFailed";
case cudaErrorTextureNotBound:
return "cudaErrorTextureNotBound";
case cudaErrorSynchronizationError:
return "cudaErrorSynchronizationError";
case cudaErrorInvalidFilterSetting:
return "cudaErrorInvalidFilterSetting";
case cudaErrorInvalidNormSetting:
return "cudaErrorInvalidNormSetting";
case cudaErrorMixedDeviceExecution:
return "cudaErrorMixedDeviceExecution";
case cudaErrorCudartUnloading:
return "cudaErrorCudartUnloading";
case cudaErrorUnknown:
return "cudaErrorUnknown";
case cudaErrorNotYetImplemented:
return "cudaErrorNotYetImplemented";
case cudaErrorMemoryValueTooLarge:
return "cudaErrorMemoryValueTooLarge";
case cudaErrorInvalidResourceHandle:
return "cudaErrorInvalidResourceHandle";
case cudaErrorNotReady:
return "cudaErrorNotReady";
case cudaErrorInsufficientDriver:
return "cudaErrorInsufficientDriver";
case cudaErrorSetOnActiveProcess:
return "cudaErrorSetOnActiveProcess";
case cudaErrorInvalidSurface:
return "cudaErrorInvalidSurface";
case cudaErrorNoDevice:
return "cudaErrorNoDevice";
case cudaErrorECCUncorrectable:
return "cudaErrorECCUncorrectable";
case cudaErrorSharedObjectSymbolNotFound:
return "cudaErrorSharedObjectSymbolNotFound";
case cudaErrorSharedObjectInitFailed:
return "cudaErrorSharedObjectInitFailed";
case cudaErrorUnsupportedLimit:
return "cudaErrorUnsupportedLimit";
case cudaErrorDuplicateVariableName:
return "cudaErrorDuplicateVariableName";
case cudaErrorDuplicateTextureName:
return "cudaErrorDuplicateTextureName";
case cudaErrorDuplicateSurfaceName:
return "cudaErrorDuplicateSurfaceName";
case cudaErrorDevicesUnavailable:
return "cudaErrorDevicesUnavailable";
case cudaErrorInvalidKernelImage:
return "cudaErrorInvalidKernelImage";
case cudaErrorNoKernelImageForDevice:
return "cudaErrorNoKernelImageForDevice";
case cudaErrorIncompatibleDriverContext:
return "cudaErrorIncompatibleDriverContext";
case cudaErrorPeerAccessAlreadyEnabled:
return "cudaErrorPeerAccessAlreadyEnabled";
case cudaErrorPeerAccessNotEnabled:
return "cudaErrorPeerAccessNotEnabled";
case cudaErrorDeviceAlreadyInUse:
return "cudaErrorDeviceAlreadyInUse";
case cudaErrorProfilerDisabled:
return "cudaErrorProfilerDisabled";
case cudaErrorProfilerNotInitialized:
return "cudaErrorProfilerNotInitialized";
case cudaErrorProfilerAlreadyStarted:
return "cudaErrorProfilerAlreadyStarted";
case cudaErrorProfilerAlreadyStopped:
return "cudaErrorProfilerAlreadyStopped";
/* Since CUDA 4.0*/
case cudaErrorAssert:
return "cudaErrorAssert";
case cudaErrorTooManyPeers:
return "cudaErrorTooManyPeers";
case cudaErrorHostMemoryAlreadyRegistered:
return "cudaErrorHostMemoryAlreadyRegistered";
case cudaErrorHostMemoryNotRegistered:
return "cudaErrorHostMemoryNotRegistered";
/* Since CUDA 5.0 */
case cudaErrorOperatingSystem:
return "cudaErrorOperatingSystem";
case cudaErrorPeerAccessUnsupported:
return "cudaErrorPeerAccessUnsupported";
case cudaErrorLaunchMaxDepthExceeded:
return "cudaErrorLaunchMaxDepthExceeded";
case cudaErrorLaunchFileScopedTex:
return "cudaErrorLaunchFileScopedTex";
case cudaErrorLaunchFileScopedSurf:
return "cudaErrorLaunchFileScopedSurf";
case cudaErrorSyncDepthExceeded:
return "cudaErrorSyncDepthExceeded";
case cudaErrorLaunchPendingCountExceeded:
return "cudaErrorLaunchPendingCountExceeded";
case cudaErrorNotPermitted:
return "cudaErrorNotPermitted";
case cudaErrorNotSupported:
return "cudaErrorNotSupported";
/* Since CUDA 6.0 */
case cudaErrorHardwareStackError:
return "cudaErrorHardwareStackError";
case cudaErrorIllegalInstruction:
return "cudaErrorIllegalInstruction";
case cudaErrorMisalignedAddress:
return "cudaErrorMisalignedAddress";
case cudaErrorInvalidAddressSpace:
return "cudaErrorInvalidAddressSpace";
case cudaErrorInvalidPc:
return "cudaErrorInvalidPc";
case cudaErrorIllegalAddress:
return "cudaErrorIllegalAddress";
/* Since CUDA 6.5*/
case cudaErrorInvalidPtx:
return "cudaErrorInvalidPtx";
case cudaErrorInvalidGraphicsContext:
return "cudaErrorInvalidGraphicsContext";
case cudaErrorStartupFailure:
return "cudaErrorStartupFailure";
case cudaErrorApiFailureBase:
return "cudaErrorApiFailureBase";
/* Since CUDA 8.0*/
case cudaErrorNvlinkUncorrectable:
return "cudaErrorNvlinkUncorrectable";
/* Since CUDA 8.5*/
case cudaErrorJitCompilerNotFound:
return "cudaErrorJitCompilerNotFound";
/* Since CUDA 9.0*/
case cudaErrorCooperativeLaunchTooLarge:
return "cudaErrorCooperativeLaunchTooLarge";
}
return "<unknown>";
} }
#endif #endif
#ifdef __cuda_cuda_h__ #ifdef CUDA_DRIVER_API
// CUDA Driver API errors // CUDA Driver API errors
static const char *_cudaGetErrorEnum(CUresult error) { static const char *_cudaGetErrorEnum(CUresult error) {
switch (error) { static char unknown[] = "<unknown>";
case CUDA_SUCCESS: const char *ret = NULL;
return "CUDA_SUCCESS"; cuGetErrorName(error, &ret);
return ret ? ret : unknown;
case CUDA_ERROR_INVALID_VALUE:
return "CUDA_ERROR_INVALID_VALUE";
case CUDA_ERROR_OUT_OF_MEMORY:
return "CUDA_ERROR_OUT_OF_MEMORY";
case CUDA_ERROR_NOT_INITIALIZED:
return "CUDA_ERROR_NOT_INITIALIZED";
case CUDA_ERROR_DEINITIALIZED:
return "CUDA_ERROR_DEINITIALIZED";
case CUDA_ERROR_PROFILER_DISABLED:
return "CUDA_ERROR_PROFILER_DISABLED";
case CUDA_ERROR_PROFILER_NOT_INITIALIZED:
return "CUDA_ERROR_PROFILER_NOT_INITIALIZED";
case CUDA_ERROR_PROFILER_ALREADY_STARTED:
return "CUDA_ERROR_PROFILER_ALREADY_STARTED";
case CUDA_ERROR_PROFILER_ALREADY_STOPPED:
return "CUDA_ERROR_PROFILER_ALREADY_STOPPED";
case CUDA_ERROR_NO_DEVICE:
return "CUDA_ERROR_NO_DEVICE";
case CUDA_ERROR_INVALID_DEVICE:
return "CUDA_ERROR_INVALID_DEVICE";
case CUDA_ERROR_INVALID_IMAGE:
return "CUDA_ERROR_INVALID_IMAGE";
case CUDA_ERROR_INVALID_CONTEXT:
return "CUDA_ERROR_INVALID_CONTEXT";
case CUDA_ERROR_CONTEXT_ALREADY_CURRENT:
return "CUDA_ERROR_CONTEXT_ALREADY_CURRENT";
case CUDA_ERROR_MAP_FAILED:
return "CUDA_ERROR_MAP_FAILED";
case CUDA_ERROR_UNMAP_FAILED:
return "CUDA_ERROR_UNMAP_FAILED";
case CUDA_ERROR_ARRAY_IS_MAPPED:
return "CUDA_ERROR_ARRAY_IS_MAPPED";
case CUDA_ERROR_ALREADY_MAPPED:
return "CUDA_ERROR_ALREADY_MAPPED";
case CUDA_ERROR_NO_BINARY_FOR_GPU:
return "CUDA_ERROR_NO_BINARY_FOR_GPU";
case CUDA_ERROR_ALREADY_ACQUIRED:
return "CUDA_ERROR_ALREADY_ACQUIRED";
case CUDA_ERROR_NOT_MAPPED:
return "CUDA_ERROR_NOT_MAPPED";
case CUDA_ERROR_NOT_MAPPED_AS_ARRAY:
return "CUDA_ERROR_NOT_MAPPED_AS_ARRAY";
case CUDA_ERROR_NOT_MAPPED_AS_POINTER:
return "CUDA_ERROR_NOT_MAPPED_AS_POINTER";
case CUDA_ERROR_ECC_UNCORRECTABLE:
return "CUDA_ERROR_ECC_UNCORRECTABLE";
case CUDA_ERROR_UNSUPPORTED_LIMIT:
return "CUDA_ERROR_UNSUPPORTED_LIMIT";
case CUDA_ERROR_CONTEXT_ALREADY_IN_USE:
return "CUDA_ERROR_CONTEXT_ALREADY_IN_USE";
case CUDA_ERROR_PEER_ACCESS_UNSUPPORTED:
return "CUDA_ERROR_PEER_ACCESS_UNSUPPORTED";
case CUDA_ERROR_INVALID_PTX:
return "CUDA_ERROR_INVALID_PTX";
case CUDA_ERROR_INVALID_GRAPHICS_CONTEXT:
return "CUDA_ERROR_INVALID_GRAPHICS_CONTEXT";
case CUDA_ERROR_NVLINK_UNCORRECTABLE:
return "CUDA_ERROR_NVLINK_UNCORRECTABLE";
case CUDA_ERROR_JIT_COMPILER_NOT_FOUND:
return "CUDA_ERROR_JIT_COMPILER_NOT_FOUND";
case CUDA_ERROR_INVALID_SOURCE:
return "CUDA_ERROR_INVALID_SOURCE";
case CUDA_ERROR_FILE_NOT_FOUND:
return "CUDA_ERROR_FILE_NOT_FOUND";
case CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND:
return "CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND";
case CUDA_ERROR_SHARED_OBJECT_INIT_FAILED:
return "CUDA_ERROR_SHARED_OBJECT_INIT_FAILED";
case CUDA_ERROR_OPERATING_SYSTEM:
return "CUDA_ERROR_OPERATING_SYSTEM";
case CUDA_ERROR_INVALID_HANDLE:
return "CUDA_ERROR_INVALID_HANDLE";
case CUDA_ERROR_NOT_FOUND:
return "CUDA_ERROR_NOT_FOUND";
case CUDA_ERROR_NOT_READY:
return "CUDA_ERROR_NOT_READY";
case CUDA_ERROR_ILLEGAL_ADDRESS:
return "CUDA_ERROR_ILLEGAL_ADDRESS";
case CUDA_ERROR_LAUNCH_FAILED:
return "CUDA_ERROR_LAUNCH_FAILED";
case CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES:
return "CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES";
case CUDA_ERROR_LAUNCH_TIMEOUT:
return "CUDA_ERROR_LAUNCH_TIMEOUT";
case CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING:
return "CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING";
case CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED:
return "CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED";
case CUDA_ERROR_PEER_ACCESS_NOT_ENABLED:
return "CUDA_ERROR_PEER_ACCESS_NOT_ENABLED";
case CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE:
return "CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE";
case CUDA_ERROR_CONTEXT_IS_DESTROYED:
return "CUDA_ERROR_CONTEXT_IS_DESTROYED";
case CUDA_ERROR_ASSERT:
return "CUDA_ERROR_ASSERT";
case CUDA_ERROR_TOO_MANY_PEERS:
return "CUDA_ERROR_TOO_MANY_PEERS";
case CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED:
return "CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED";
case CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED:
return "CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED";
case CUDA_ERROR_HARDWARE_STACK_ERROR:
return "CUDA_ERROR_HARDWARE_STACK_ERROR";
case CUDA_ERROR_ILLEGAL_INSTRUCTION:
return "CUDA_ERROR_ILLEGAL_INSTRUCTION";
case CUDA_ERROR_MISALIGNED_ADDRESS:
return "CUDA_ERROR_MISALIGNED_ADDRESS";
case CUDA_ERROR_INVALID_ADDRESS_SPACE:
return "CUDA_ERROR_INVALID_ADDRESS_SPACE";
case CUDA_ERROR_INVALID_PC:
return "CUDA_ERROR_INVALID_PC";
case CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE:
return "CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE";
case CUDA_ERROR_NOT_PERMITTED:
return "CUDA_ERROR_NOT_PERMITTED";
case CUDA_ERROR_NOT_SUPPORTED:
return "CUDA_ERROR_NOT_SUPPORTED";
case CUDA_ERROR_UNKNOWN:
return "CUDA_ERROR_UNKNOWN";
}
return "<unknown>";
} }
#endif #endif
@@ -722,6 +282,42 @@ static const char *_cudaGetErrorEnum(curandStatus_t error) {
} }
#endif #endif
#ifdef NVJPEGAPI
// nvJPEG API errors
static const char *_cudaGetErrorEnum(nvjpegStatus_t error) {
switch (error) {
case NVJPEG_STATUS_SUCCESS:
return "NVJPEG_STATUS_SUCCESS";
case NVJPEG_STATUS_NOT_INITIALIZED:
return "NVJPEG_STATUS_NOT_INITIALIZED";
case NVJPEG_STATUS_INVALID_PARAMETER:
return "NVJPEG_STATUS_INVALID_PARAMETER";
case NVJPEG_STATUS_BAD_JPEG:
return "NVJPEG_STATUS_BAD_JPEG";
case NVJPEG_STATUS_JPEG_NOT_SUPPORTED:
return "NVJPEG_STATUS_JPEG_NOT_SUPPORTED";
case NVJPEG_STATUS_ALLOCATOR_FAILURE:
return "NVJPEG_STATUS_ALLOCATOR_FAILURE";
case NVJPEG_STATUS_EXECUTION_FAILED:
return "NVJPEG_STATUS_EXECUTION_FAILED";
case NVJPEG_STATUS_ARCH_MISMATCH:
return "NVJPEG_STATUS_ARCH_MISMATCH";
case NVJPEG_STATUS_INTERNAL_ERROR:
return "NVJPEG_STATUS_INTERNAL_ERROR";
}
return "<unknown>";
}
#endif
#ifdef NV_NPPIDEFS_H #ifdef NV_NPPIDEFS_H
// NPP API errors // NPP API errors
static const char *_cudaGetErrorEnum(NppStatus error) { static const char *_cudaGetErrorEnum(NppStatus error) {
@@ -1067,18 +663,19 @@ inline int _ConvertSMVer2Cores(int major, int minor) {
} sSMtoCores; } sSMtoCores;
sSMtoCores nGpuArchCoresPerSM[] = { sSMtoCores nGpuArchCoresPerSM[] = {
{0x30, 192}, // Kepler Generation (SM 3.0) GK10x class {0x30, 192},
{0x32, 192}, // Kepler Generation (SM 3.2) GK10x class {0x32, 192},
{0x35, 192}, // Kepler Generation (SM 3.5) GK11x class {0x35, 192},
{0x37, 192}, // Kepler Generation (SM 3.7) GK21x class {0x37, 192},
{0x50, 128}, // Maxwell Generation (SM 5.0) GM10x class {0x50, 128},
{0x52, 128}, // Maxwell Generation (SM 5.2) GM20x class {0x52, 128},
{0x53, 128}, // Maxwell Generation (SM 5.3) GM20x class {0x53, 128},
{0x60, 64}, // Pascal Generation (SM 6.0) GP100 class {0x60, 64},
{0x61, 128}, // Pascal Generation (SM 6.1) GP10x class {0x61, 128},
{0x62, 128}, // Pascal Generation (SM 6.2) GP10x class {0x62, 128},
{0x70, 64}, // Volta Generation (SM 7.0) GV100 class {0x70, 64},
{0x72, 64}, // Volta Generation (SM 7.2) GV11b class {0x72, 64},
{0x75, 64},
{-1, -1}}; {-1, -1}};
int index = 0; int index = 0;
@@ -1155,7 +752,7 @@ inline int gpuDeviceInit(int devID) {
inline int gpuGetMaxGflopsDeviceId() { inline int gpuGetMaxGflopsDeviceId() {
int current_device = 0, sm_per_multiproc = 0; int current_device = 0, sm_per_multiproc = 0;
int max_perf_device = 0; int max_perf_device = 0;
int device_count = 0, best_SM_arch = 0; int device_count = 0;
int devices_prohibited = 0; int devices_prohibited = 0;
uint64_t max_compute_perf = 0; uint64_t max_compute_perf = 0;
@@ -1169,30 +766,6 @@ inline int gpuGetMaxGflopsDeviceId() {
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
// Find the best major SM Architecture GPU device
while (current_device < device_count) {
cudaGetDeviceProperties(&deviceProp, current_device);
// If this GPU is not running on Compute Mode prohibited,
// then we can add it to the list
if (deviceProp.computeMode != cudaComputeModeProhibited) {
if (deviceProp.major > 0 && deviceProp.major < 9999) {
best_SM_arch = MAX(best_SM_arch, deviceProp.major);
}
} else {
devices_prohibited++;
}
current_device++;
}
if (devices_prohibited == device_count) {
fprintf(stderr,
"gpuGetMaxGflopsDeviceId() CUDA error:"
" all devices have compute mode prohibited.\n");
exit(EXIT_FAILURE);
}
// Find the best CUDA capable GPU device // Find the best CUDA capable GPU device
current_device = 0; current_device = 0;
@@ -1213,23 +786,23 @@ inline int gpuGetMaxGflopsDeviceId() {
sm_per_multiproc * deviceProp.clockRate; sm_per_multiproc * deviceProp.clockRate;
if (compute_perf > max_compute_perf) { if (compute_perf > max_compute_perf) {
// If we find GPU with SM major > 2, search only these max_compute_perf = compute_perf;
if (best_SM_arch > 2) { max_perf_device = current_device;
// If our device==dest_SM_arch, choose this, or else pass
if (deviceProp.major == best_SM_arch) {
max_compute_perf = compute_perf;
max_perf_device = current_device;
}
} else {
max_compute_perf = compute_perf;
max_perf_device = current_device;
}
} }
} else {
devices_prohibited++;
} }
++current_device; ++current_device;
} }
if (devices_prohibited == device_count) {
fprintf(stderr,
"gpuGetMaxGflopsDeviceId() CUDA error:"
" all devices have compute mode prohibited.\n");
exit(EXIT_FAILURE);
}
return max_perf_device; return max_perf_device;
} }

View File

@@ -122,18 +122,19 @@ inline int _ConvertSMVer2CoresDRV(int major, int minor) {
} sSMtoCores; } sSMtoCores;
sSMtoCores nGpuArchCoresPerSM[] = { sSMtoCores nGpuArchCoresPerSM[] = {
{0x30, 192}, // Kepler Generation (SM 3.0) GK10x class {0x30, 192},
{0x32, 192}, // Kepler Generation (SM 3.2) GK10x class {0x32, 192},
{0x35, 192}, // Kepler Generation (SM 3.5) GK11x class {0x35, 192},
{0x37, 192}, // Kepler Generation (SM 3.7) GK21x class {0x37, 192},
{0x50, 128}, // Maxwell Generation (SM 5.0) GM10x class {0x50, 128},
{0x52, 128}, // Maxwell Generation (SM 5.2) GM20x class {0x52, 128},
{0x53, 128}, // Maxwell Generation (SM 5.3) GM20x class {0x53, 128},
{0x60, 64}, // Pascal Generation (SM 6.0) GP100 class {0x60, 64},
{0x61, 128}, // Pascal Generation (SM 6.1) GP10x class {0x61, 128},
{0x62, 128}, // Pascal Generation (SM 6.2) GP10x class {0x62, 128},
{0x70, 64}, // Volta Generation (SM 7.0) GV100 class {0x70, 64},
{0x72, 64}, // Volta Generation (SM 7.2) GV11b class {0x72, 64},
{0x75, 64},
{-1, -1}}; {-1, -1}};
int index = 0; int index = 0;

View File

@@ -0,0 +1,178 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "helper_multiprocess.h"
#include <cstdlib>
#include <string>
int sharedMemoryCreate(const char *name, size_t sz, sharedMemoryInfo *info)
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
info->size = sz;
info->shmHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
NULL,
PAGE_READWRITE,
0,
(DWORD)sz,
name);
if (info->shmHandle == 0) {
return GetLastError();
}
info->addr = MapViewOfFile(info->shmHandle, FILE_MAP_ALL_ACCESS, 0, 0, sz);
if (info->addr == NULL) {
return GetLastError();
}
return 0;
#else
int status = 0;
info->size = sz;
info->shmFd = shm_open(name, O_RDWR | O_CREAT, 0777);
if (info->shmFd < 0) {
return errno;
}
status = ftruncate(info->shmFd, sz);
if (status != 0) {
return status;
}
info->addr = mmap(0, sz, PROT_READ | PROT_WRITE, MAP_SHARED, info->shmFd, 0);
if (info->addr == NULL) {
return errno;
}
return 0;
#endif
}
int sharedMemoryOpen(const char *name, size_t sz, sharedMemoryInfo *info)
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
info->size = sz;
info->shmHandle = OpenFileMapping(FILE_MAP_ALL_ACCESS, FALSE, name);
if (info->shmHandle == 0) {
return GetLastError();
}
info->addr = MapViewOfFile(info->shmHandle, FILE_MAP_ALL_ACCESS, 0, 0, sz);
if (info->addr == NULL) {
return GetLastError();
}
return 0;
#else
info->size = sz;
info->shmFd = shm_open(name, O_RDWR, 0777);
if (info->shmFd < 0) {
return errno;
}
info->addr = mmap(0, sz, PROT_READ | PROT_WRITE, MAP_SHARED, info->shmFd, 0);
if (info->addr == NULL) {
return errno;
}
return 0;
#endif
}
void sharedMemoryClose(sharedMemoryInfo *info)
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
if (info->addr) {
UnmapViewOfFile(info->addr);
}
if (info->shmHandle) {
CloseHandle(info->shmHandle);
}
#else
if (info->addr) {
munmap(info->addr, info->size);
}
if (info->shmFd) {
close(info->shmFd);
}
#endif
}
int spawnProcess(Process *process, const char *app, char * const *args)
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
STARTUPINFO si = {0};
BOOL status;
size_t arglen = 0;
size_t argIdx = 0;
std::string arg_string;
memset(process, 0, sizeof(*process));
while (*args) {
arg_string.append(*args).append(1, ' ');
args++;
}
status = CreateProcess(app, LPSTR(arg_string.c_str()), NULL, NULL, FALSE, 0, NULL, NULL, &si, process);
return status ? 0 : GetLastError();
#else
*process = fork();
if (*process == 0) {
if (0 > execvp(app, args)) {
return errno;
}
}
else if (*process < 0) {
return errno;
}
return 0;
#endif
}
int waitProcess(Process *process)
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
DWORD exitCode;
WaitForSingleObject(process->hProcess, INFINITE);
GetExitCodeProcess(process->hProcess, &exitCode);
CloseHandle(process->hProcess);
CloseHandle(process->hThread);
return (int)exitCode;
#else
int status = 0;
do {
if (0 > waitpid(*process, &status, 0)) {
return errno;
}
} while (!WIFEXITED(status));
return WEXITSTATUS(status);
#endif
}

View File

@@ -0,0 +1,71 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifndef HELPER_MULTIPROCESS_H
#define HELPER_MULTIPROCESS_H
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
#else
#include <fcntl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <errno.h>
#include <sys/wait.h>
#endif
typedef struct sharedMemoryInfo_st {
void *addr;
size_t size;
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
HANDLE shmHandle;
#else
int shmFd;
#endif
} sharedMemoryInfo;
int sharedMemoryCreate(const char *name, size_t sz, sharedMemoryInfo *info);
int sharedMemoryOpen(const char *name, size_t sz, sharedMemoryInfo *info);
void sharedMemoryClose(sharedMemoryInfo *info);
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
typedef PROCESS_INFORMATION Process;
#else
typedef pid_t Process;
#endif
int spawnProcess(Process *process, const char *app, char * const *args);
int waitProcess(Process *process);
#endif // HELPER_MULTIPROCESS_H

View File

@@ -23,3 +23,6 @@ PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
For additional information on the license terms, see the CUDA EULA at
https://docs.nvidia.com/cuda/eula/index.html

View File

@@ -1,11 +1,31 @@
# CUDA Samples # CUDA Samples
Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads). Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads).
## Release Notes ## Release Notes
This section describes the release notes for the CUDA Samples on GitHub only. This section describes the release notes for the CUDA Samples on GitHub only.
### CUDA 10.1
* Added `immaTensorCoreGemm`. Demonstrates integer GEMM computation using the Warp Matrix Multiply and Accumulate (WMMA) API for integers employing the Tensor Cores.
* Added `simpleIPC`. Demonstrates Inter Process Communication with one process per GPU for computation.
* Added `nvJPEG`. Demonstrates single and batched decoding of jpeg images using NVJPEG Library.
* Added `bandwidthTest`. It measures the memcopy bandwidth of the GPU and memcpy bandwidth across PCI-e.
* Added `reduction`. Demonstrates several important optimization strategies for Data-Parallel Algorithms like reduction.
* Update all the samples to support CUDA 10.1.
### CUDA 10.0
* Added `simpleCudaGraphs`. Demonstrates CUDA Graphs creation, instantiation and launch using Graphs APIs and Stream Capture APIs.
* Added `conjugateGradientCudaGraphs`. Demonstrates conjugate gradient solver on GPU using CUBLAS and CUSPARSE library calls captured and called using CUDA Graph APIs.
* Added `simpleVulkan`. Demonstrates Vulkan - CUDA Interop.
* Added `simpleD3D12`. Demonstrates DX12 - CUDA Interop.
* Added `UnifiedMemoryPerf`. Demonstrates performance comparision of various memory types involved in system.
* Added `p2pBandwidthLatencyTest`. Demonstrates Peer-To-Peer (P2P) data transfers between pairs of GPUs and computes latency and bandwidth.
* Added `systemWideAtomics`. Demonstrates system wide atomic instructions.
* Added `simpleCUBLASXT`. Demonstrates CUBLAS-XT library which performs GEMM operations over multiple GPUs.
* Added Windows OS support to `conjugateGradientMultiDeviceCG` sample.
* Removed support of Visual Studio 2010 from all samples.
### CUDA 9.2 ### CUDA 9.2
This is the first release of CUDA Samples on GitHub: This is the first release of CUDA Samples on GitHub:
@@ -26,7 +46,7 @@ This is the first release of CUDA Samples on GitHub:
### Prerequisites ### Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
For system requirements and installation instructions of cuda toolkit, please refer to the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/), the [Windows Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html), and the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html). For system requirements and installation instructions of cuda toolkit, please refer to the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/), the [Windows Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html), and the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html).
### Getting the CUDA Samples ### Getting the CUDA Samples
@@ -108,22 +128,29 @@ The samples makefiles can take advantage of certain options:
### Samples by OS ### Samples by OS
#### Linux #### Linux
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[shfl_scan](./Samples/shfl_scan)** | **[deviceQuery](./Samples/deviceQuery)** | **[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[simpleIPC](./Samples/simpleIPC)** | **[shfl_scan](./Samples/shfl_scan)** |
---|---|---|---| ---|---|---|---|
**[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[nvJPEG](./Samples/nvJPEG)** | **[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** |
**[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[matrixMul](./Samples/matrixMul)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** | **[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** |
**[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** | **[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** |
**[bandwidthTest](./Samples/bandwidthTest)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[simpleVulkan](./Samples/simpleVulkan)** |
**[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** | **[systemWideAtomics](./Samples/systemWideAtomics)** |
#### Windows #### Windows
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[shfl_scan](./Samples/shfl_scan)** | **[deviceQuery](./Samples/deviceQuery)** | **[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[simpleIPC](./Samples/simpleIPC)** | **[shfl_scan](./Samples/shfl_scan)** |
---|---|---|---| ---|---|---|---|
**[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[simpleD3D12](./Samples/simpleD3D12)** | **[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** |
**[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[matrixMul](./Samples/matrixMul)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** | **[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** |
**[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** | **[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** |
**[bandwidthTest](./Samples/bandwidthTest)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[simpleVulkan](./Samples/simpleVulkan)** |
**[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** |
#### Mac OSX #### Mac OSX
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[shfl_scan](./Samples/shfl_scan)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[shfl_scan](./Samples/shfl_scan)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[simpleCudaGraphs](./Samples/simpleCudaGraphs)** |
---|---|---|---| ---|---|---|---|
**[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** | **[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** |
**[matrixMul](./Samples/matrixMul)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** | **[bandwidthTest](./Samples/bandwidthTest)** |
**[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** |
## Dependencies ## Dependencies
@@ -153,6 +180,10 @@ Some samples can only be run on a 64-bit operating system.
DirectX is a collection of APIs designed to allow development of multimedia applications on Microsoft platforms. For Microsoft platforms, NVIDIA's CUDA Driver supports DirectX. Several CUDA Samples for Windows demonstrates CUDA-DirectX Interoperability, for building such samples one needs to install [Direct X SDK (June 2010 or newer)](http://www.microsoft.com/en-us/download/details.aspx?id=6812) , this is required to be installed on Windows 7, Windows 10 and Windows Server 2008, Other Windows OSes do not need to explicitly install the DirectX SDK. DirectX is a collection of APIs designed to allow development of multimedia applications on Microsoft platforms. For Microsoft platforms, NVIDIA's CUDA Driver supports DirectX. Several CUDA Samples for Windows demonstrates CUDA-DirectX Interoperability, for building such samples one needs to install [Direct X SDK (June 2010 or newer)](http://www.microsoft.com/en-us/download/details.aspx?id=6812) , this is required to be installed on Windows 7, Windows 10 and Windows Server 2008, Other Windows OSes do not need to explicitly install the DirectX SDK.
#### DirectX12
DirectX 12 is a collection of advanced low-level programming APIs which can reduce driver overhead, designed to allow development of multimedia applications on Microsoft platforms starting with Windows 10 OS onwards. For Microsoft platforms, NVIDIA's CUDA Driver supports DirectX. Few CUDA Samples for Windows demonstrates CUDA-DirectX12 Interoperability, for building such samples one needs to install [Windows 10 SDK or higher](https://developer.microsoft.com/en-us/windows/downloads/windows-10-sdk), with VS 2015 or VS 2017.
#### OpenGL #### OpenGL
OpenGL is a graphics library used for 2D and 3D rendering. On systems which support OpenGL, NVIDIA's OpenGL implementation is provided with the CUDA Driver. OpenGL is a graphics library used for 2D and 3D rendering. On systems which support OpenGL, NVIDIA's OpenGL implementation is provided with the CUDA Driver.
@@ -161,6 +192,10 @@ OpenGL is a graphics library used for 2D and 3D rendering. On systems which supp
OpenGL ES is an embedded systems graphics library used for 2D and 3D rendering. On systems which support OpenGL ES, NVIDIA's OpenGL ES implementation is provided with the CUDA Driver. OpenGL ES is an embedded systems graphics library used for 2D and 3D rendering. On systems which support OpenGL ES, NVIDIA's OpenGL ES implementation is provided with the CUDA Driver.
#### Vulkan
Vulkan is a low-overhead, cross-platform 3D graphics and compute API. Vulkan targets high-performance realtime 3D graphics applications such as video games and interactive media across all platforms. On systems which support Vulkan, NVIDIA's Vulkan implementation is provided with the CUDA Driver. For building and running Vulkan applications one needs to install the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/).
#### OpenMP #### OpenMP
OpenMP is an API for multiprocessing programming. OpenMP can be installed using your Linux distribution's package manager system. It usually comes preinstalled with GCC. It can also be found at the [OpenMP website](http://openmp.org/). OpenMP is an API for multiprocessing programming. OpenMP can be installed using your Linux distribution's package manager system. It usually comes preinstalled with GCC. It can also be found at the [OpenMP website](http://openmp.org/).
@@ -211,7 +246,7 @@ CUBLAS (CUDA Basic Linear Algebra Subroutines) is a GPU-accelerated version of t
#### CUDA Interprocess Communication #### CUDA Interprocess Communication
IPC (Interprocess Communication) allows processes to share device pointers. IPC is only available on Linux x86_64 and ppc64le systems. IPC (Interprocess Communication) allows processes to share device pointers.
#### CUFFT #### CUFFT
@@ -237,6 +272,10 @@ NPP (NVIDIA Performance Primitives) provides GPU-accelerated image, video, and s
NVGRAPH is a GPU-accelerated graph analytics library. NVGRAPH is a GPU-accelerated graph analytics library.
#### NVJPEG
NVJPEG library provides high-performance, GPU accelerated JPEG decoding functionality for image formats commonly used in deep learning and hyperscale multimedia applications.
#### NVRTC #### NVRTC
NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++. NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++.

View File

@@ -0,0 +1,316 @@
################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
################################################################################
#
# Makefile project only supported on Mac OS X and Linux Platforms)
#
################################################################################
# Location of the CUDA Toolkit
CUDA_PATH ?= /usr/local/cuda
##############################
# start deprecated interface #
##############################
ifeq ($(x86_64),1)
$(info WARNING - x86_64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=x86_64 instead)
TARGET_ARCH ?= x86_64
endif
ifeq ($(ARMv7),1)
$(info WARNING - ARMv7 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=armv7l instead)
TARGET_ARCH ?= armv7l
endif
ifeq ($(aarch64),1)
$(info WARNING - aarch64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=aarch64 instead)
TARGET_ARCH ?= aarch64
endif
ifeq ($(ppc64le),1)
$(info WARNING - ppc64le variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=ppc64le instead)
TARGET_ARCH ?= ppc64le
endif
ifneq ($(GCC),)
$(info WARNING - GCC variable has been deprecated)
$(info WARNING - please use HOST_COMPILER=$(GCC) instead)
HOST_COMPILER ?= $(GCC)
endif
ifneq ($(abi),)
$(error ERROR - abi variable has been removed)
endif
############################
# end deprecated interface #
############################
# architecture
HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif
endif
# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
TARGET_ARCH = armv7l
endif
# operating system
HOST_OS := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
TARGET_OS ?= $(HOST_OS)
ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
$(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
endif
# host compiler
ifeq ($(TARGET_OS),darwin)
ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
HOST_COMPILER ?= clang++
endif
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
ifeq ($(TARGET_OS),linux)
HOST_COMPILER ?= arm-linux-gnueabihf-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
else ifeq ($(TARGET_OS),android)
HOST_COMPILER ?= arm-linux-androideabi-g++
endif
else ifeq ($(TARGET_ARCH),aarch64)
ifeq ($(TARGET_OS), linux)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++
endif
else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif
endif
HOST_COMPILER ?= g++
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
# internal flags
NVCCFLAGS := -m${TARGET_SIZE}
CCFLAGS :=
LDFLAGS :=
# build flags
ifeq ($(TARGET_OS),darwin)
LDFLAGS += -rpath $(CUDA_PATH)/lib
CCFLAGS += -arch $(HOST_ARCH)
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
CCFLAGS += -mfloat-abi=hard
else ifeq ($(TARGET_OS),android)
LDFLAGS += -pie
CCFLAGS += -fpie -fpic -fexceptions
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
endif
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
else ifeq ($(TARGET_ARCH),ppc64le)
CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
endif
# Debug build flags
ifeq ($(dbg),1)
NVCCFLAGS += -g -G
BUILD_TYPE := debug
else
BUILD_TYPE := release
endif
ALL_CCFLAGS :=
ALL_CCFLAGS += $(NVCCFLAGS)
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
SAMPLE_ENABLED := 1
# This sample is not supported on QNX
ifeq ($(TARGET_OS),qnx)
$(info >>> WARNING - UnifiedMemoryPerf is not supported on QNX - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
# Common includes and paths for CUDA
INCLUDES := -I../../Common
LIBRARIES :=
################################################################################
# Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75
else
SMS ?= 30 35 37 50 52 60 61 70 75
endif
ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ifeq ($(GENCODE_FLAGS),)
# Generate SASS code for each SM architecture listed in $(SMS)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
HIGHEST_SM := $(lastword $(sort $(SMS)))
ifneq ($(HIGHEST_SM),)
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: UnifiedMemoryPerf
check.deps:
ifeq ($(SAMPLE_ENABLED),0)
@echo "Sample will be waived due to the above missing dependencies"
else
@echo "Sample is ready - all dependencies have been met"
endif
commonKernels.o:commonKernels.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
helperFunctions.o:helperFunctions.cpp
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
matrixMultiplyPerf.o:matrixMultiplyPerf.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
UnifiedMemoryPerf: commonKernels.o helperFunctions.o matrixMultiplyPerf.o
$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
run: build
$(EXEC) ./UnifiedMemoryPerf
clean:
rm -f UnifiedMemoryPerf commonKernels.o helperFunctions.o matrixMultiplyPerf.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/UnifiedMemoryPerf
clobber: clean

View File

@@ -0,0 +1,85 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>UnifiedMemoryPerf</name>
<cuda_api_list>
<toolkit>cudaMallocManaged</toolkit>
<toolkit>cudaStreamAttachMemAsync</toolkit>
<toolkit>cudaMemcpyAsync</toolkit>
<toolkit>cudaMallocHost</toolkit>
<toolkit>cudaMalloc</toolkit>
</cuda_api_list>
<description><![CDATA[This sample demonstrates the performance comparision using matrix multiplication kernel of Unified Memory with/without hints and other types of memory like zero copy buffers, pageable, pagelocked memory performing synchronous and Asynchronous transfers on a single GPU.]]></description>
<devicecompilation>whole</devicecompilation>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">CUDA Systems Integration</concept>
<concept level="basic">Unified Memory</concept>
<concept level="basic">CUDA Streams and Events</concept>
<concept level="basic">Pinned System Paged Memory</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>Unified Memory</keyword>
<keyword>Pinned Memory</keyword>
<keyword>Zero copy buffer</keyword>
<keyword>UVM</keyword>
<keyword>Streams</keyword>
</keywords>
<libraries>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>matrixMultiplyPerf.cu</primary_file>
<required_dependencies>
<dependency>UVM</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Basic Topics</scope>
<scope>1:CUDA Systems Integration</scope>
<scope>1:Unified Memory</scope>
</scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch>
<sm-arch>sm52</sm-arch>
<sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<arch>x86_64</arch>
<platform>macosx</platform>
</env>
<env>
<platform>windows7</platform>
</env>
<env>
<arch>arm</arch>
</env>
<env>
<arch>aarch64</arch>
</env>
<env>
<arch>ppc64le</arch>
<platform>linux</platform>
</env>
</supported_envs>
<supported_sm_architectures>
<from>3.0</from>
</supported_sm_architectures>
<title>Unified and other CUDA Memories Performance</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,98 @@
# UnifiedMemoryPerf - Unified and other CUDA Memories Performance
## Description
This sample demonstrates the performance comparision using matrix multiplication kernel of Unified Memory with/without hints and other types of memory like zero copy buffers, pageable, pagelocked memory performing synchronous and Asynchronous transfers on a single GPU.
## Key Concepts
CUDA Systems Integration, Unified Memory, CUDA Streams and Events, Pinned System Paged Memory
## Supported SM Architectures
[SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows, MacOSX
## Supported CPU Architecture
x86_64, ppc64le, armv7l, aarch64
## CUDA APIs involved
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaMallocManaged, cudaStreamAttachMemAsync, cudaMemcpyAsync, cudaMallocHost, cudaMalloc
## Dependencies needed to build/run
[UVM](../../README.md#uvm)
## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
## Build and Run
### Windows
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
```
*_vs<version>.sln - for Visual Studio <version>
```
Each individual sample has its own set of solution files in its directory:
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
### Linux
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, armv7l, aarch64.
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=armv7l` <br/> `$ make TARGET_ARCH=aarch64` <br/>
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
```
$ make SMS="50 60"
```
* **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=g++
```
### Mac
The Mac samples are built using makefiles. To use the makefiles, change directory into the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where "A B ..." is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use SMS="50 60".
```
$ make SMS="A B ..."
```
* **HOST_COMPILER=<host_compiler>** - override the default clang host compiler. See the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=clang
```
## References (for more details)

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,110 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>UnifiedMemoryPerf_vs2012</RootNamespace>
<ProjectName>UnifiedMemoryPerf</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v110</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="commonKernels.cu" />
<ClCompile Include="helperFunctions.cpp" />
<CudaCompile Include="matrixMultiplyPerf.cu" />
<ClInclude Include="commonDefs.hpp" />
<ClInclude Include="commonKernels.hpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,110 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>UnifiedMemoryPerf_vs2013</RootNamespace>
<ProjectName>UnifiedMemoryPerf</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="commonKernels.cu" />
<ClCompile Include="helperFunctions.cpp" />
<CudaCompile Include="matrixMultiplyPerf.cu" />
<ClInclude Include="commonDefs.hpp" />
<ClInclude Include="commonKernels.hpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,110 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>UnifiedMemoryPerf_vs2015</RootNamespace>
<ProjectName>UnifiedMemoryPerf</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v140</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="commonKernels.cu" />
<ClCompile Include="helperFunctions.cpp" />
<CudaCompile Include="matrixMultiplyPerf.cu" />
<ClInclude Include="commonDefs.hpp" />
<ClInclude Include="commonKernels.hpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,111 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>UnifiedMemoryPerf_vs2017</RootNamespace>
<ProjectName>UnifiedMemoryPerf</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</PlatformToolset>
<WindowsTargetPlatformVersion>10.0.15063.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="commonKernels.cu" />
<ClCompile Include="helperFunctions.cpp" />
<CudaCompile Include="matrixMultiplyPerf.cu" />
<ClInclude Include="commonDefs.hpp" />
<ClInclude Include="commonKernels.hpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,88 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifndef _COMMON_DEFS_
#define _COMMON_DEFS_
#include <cuda.h>
#define ONE_KB 1024
#define ONE_MB (ONE_KB * ONE_KB)
extern size_t maxSampleSizeInMb;
extern int numKernelRuns;
extern int verboseResults;
extern unsigned int findNumSizesToTest(unsigned int minSize,
unsigned int maxSize,
unsigned int multiplier);
// For Tracking the different memory allocation types
typedef enum memAllocType_enum {
MEMALLOC_TYPE_START,
USE_MANAGED_MEMORY_WITH_HINTS = MEMALLOC_TYPE_START,
USE_MANAGED_MEMORY_WITH_HINTS_ASYNC,
USE_MANAGED_MEMORY,
USE_ZERO_COPY,
USE_HOST_PAGEABLE_AND_DEVICE_MEMORY,
USE_HOST_PAGEABLE_AND_DEVICE_MEMORY_ASYNC,
USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY,
USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY_ASYNC,
MEMALLOC_TYPE_END = USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY_ASYNC,
MEMALLOC_TYPE_INVALID,
MEMALLOC_TYPE_COUNT = MEMALLOC_TYPE_INVALID
} MemAllocType;
typedef enum bandwidthType_enum {
READ_BANDWIDTH,
WRITE_BANDWIDTH
} BandwidthType;
extern const char *memAllocTypeStr[];
extern const char *memAllocTypeShortStr[];
struct resultsData;
struct testResults;
void createAndInitTestResults(struct testResults **results,
const char *testName,
unsigned int numMeasurements,
unsigned int numSizesToTest);
unsigned long *getPtrSizesToTest(struct testResults *results);
void freeTestResultsAndAllResultsData(struct testResults *results);
void createResultDataAndAddToTestResults(struct resultsData **ptrData,
struct testResults *results,
const char *resultsName,
bool printOnlyInVerbose,
bool reportAsBandwidth);
double *getPtrRunTimesInMs(struct resultsData *data, int allocType,
int sizeIndex);
void printResults(struct testResults *results,
bool print_launch_transfer_results, bool print_std_deviation);
#endif

View File

@@ -0,0 +1,33 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "commonKernels.hpp"
__global__ void spinWhileLessThanOne(volatile unsigned int *latch) {
while (latch[0] < 1)
;
}

View File

@@ -0,0 +1,28 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
__global__ void spinWhileLessThanOne(volatile unsigned int *latch);

View File

@@ -0,0 +1,303 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <stdio.h>
#include <string.h>
#include "commonDefs.hpp"
#define CU_INIT_UUID
#include <cmath>
#define UNITS_Time "ms"
#define UNITS_BW "MB/s"
#define KB_str "KB"
#define MB_str "MB"
struct resultsData {
char resultsName[64];
struct testResults *results;
// this has MEMALLOC_TYPE_COUNT * results->numSizesToTest *
// results->numMeasurements elements
double **runTimesInMs[MEMALLOC_TYPE_COUNT];
double *averageRunTimesInMs[MEMALLOC_TYPE_COUNT];
double *stdDevRunTimesInMs[MEMALLOC_TYPE_COUNT];
double *stdDevBandwidthInMBps[MEMALLOC_TYPE_COUNT];
bool printOnlyInVerbose;
bool reportAsBandwidth;
struct resultsData *next;
};
struct testResults {
char testName[64];
unsigned int numMeasurements;
unsigned long *sizesToTest;
unsigned int numSizesToTest;
struct resultsData *resultsDataHead;
struct resultsData *resultsDataTail;
};
unsigned int findNumSizesToTest(unsigned int minSize, unsigned int maxSize,
unsigned int multiplier) {
unsigned int numSizesToTest = 0;
while (minSize <= maxSize) {
numSizesToTest++;
minSize *= multiplier;
}
return numSizesToTest;
}
int compareDoubles(const void *ptr1, const void *ptr2) {
return (*(double *)ptr1 > *(double *)ptr2) ? 1 : -1;
}
static inline double getTimeOrBandwidth(double runTimeInMs, unsigned long size,
bool getBandwidth) {
return (getBandwidth) ? (1000 * (size / runTimeInMs)) / ONE_MB : runTimeInMs;
}
void createAndInitTestResults(struct testResults **ptrResults,
const char *testName,
unsigned int numMeasurements,
unsigned int numSizesToTest) {
unsigned int i;
struct testResults *results;
results = (struct testResults *)malloc(sizeof(struct testResults));
memset(results, 0, sizeof(struct testResults));
strcpy(results->testName, testName);
results->numMeasurements = numMeasurements;
results->numSizesToTest = numSizesToTest;
results->sizesToTest =
(unsigned long *)malloc(numSizesToTest * sizeof(unsigned long));
results->resultsDataHead = NULL;
results->resultsDataTail = NULL;
*ptrResults = results;
}
unsigned long *getPtrSizesToTest(struct testResults *results) {
return results->sizesToTest;
}
void createResultDataAndAddToTestResults(struct resultsData **ptrData,
struct testResults *results,
const char *resultsName,
bool printOnlyInVerbose,
bool reportAsBandwidth) {
unsigned int i, j;
struct resultsData *data;
data = (struct resultsData *)malloc(sizeof(struct resultsData));
memset(data, 0, sizeof(struct resultsData));
strcpy(data->resultsName, resultsName);
data->results = results;
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
data->runTimesInMs[i] =
(double **)malloc(results->numSizesToTest * sizeof(double *));
for (j = 0; j < results->numSizesToTest; j++) {
data->runTimesInMs[i][j] =
(double *)malloc(results->numMeasurements * sizeof(double));
}
data->averageRunTimesInMs[i] =
(double *)malloc(results->numSizesToTest * sizeof(double));
data->stdDevRunTimesInMs[i] =
(double *)malloc(results->numSizesToTest * sizeof(double));
data->stdDevBandwidthInMBps[i] =
(double *)malloc(results->numSizesToTest * sizeof(double));
}
data->printOnlyInVerbose = printOnlyInVerbose;
data->reportAsBandwidth = reportAsBandwidth;
data->next = NULL;
*ptrData = data;
if (results->resultsDataHead == NULL) {
results->resultsDataHead = data;
results->resultsDataTail = data;
} else {
results->resultsDataTail->next = data;
results->resultsDataTail = data;
}
}
double *getPtrRunTimesInMs(struct resultsData *data, int allocType,
int sizeIndex) {
return data->runTimesInMs[allocType][sizeIndex];
}
void freeTestResultsAndAllResultsData(struct testResults *results) {
struct resultsData *data, *dataToFree;
unsigned int i, j;
for (data = results->resultsDataHead; data != NULL;) {
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
for (j = 0; j < results->numSizesToTest; j++) {
free(data->runTimesInMs[i][j]);
}
free(data->runTimesInMs[i]);
free(data->averageRunTimesInMs[i]);
free(data->stdDevRunTimesInMs[i]);
free(data->stdDevBandwidthInMBps[i]);
}
dataToFree = data;
data = data->next;
free(dataToFree);
}
free(results->sizesToTest);
free(results);
}
void calculateAverageAndStdDev(double *pAverage, double *pStdDev,
double *allResults, unsigned int count) {
unsigned int i;
double average = 0.0;
double stdDev = 0.0;
for (i = 0; i < count; i++) {
average += allResults[i];
}
average /= count;
for (i = 0; i < count; i++) {
stdDev += (allResults[i] - average) * (allResults[i] - average);
}
stdDev /= count;
stdDev = sqrt(stdDev);
*pAverage = average;
*pStdDev = (average == 0.0) ? 0.0 : ((100.0 * stdDev) / average);
}
void calculateStdDevBandwidth(double *pStdDev, double *allResults,
unsigned int count, unsigned long size) {
unsigned int i;
double bandwidth;
double average = 0.0;
double stdDev = 0.0;
for (i = 0; i < count; i++) {
bandwidth = (1000 * (size / allResults[i])) / ONE_MB;
average += bandwidth;
}
average /= count;
for (i = 0; i < count; i++) {
bandwidth = (1000 * (size / allResults[i])) / ONE_MB;
stdDev += (bandwidth - average) * (bandwidth - average);
}
stdDev /= count;
stdDev = sqrt(stdDev);
*pStdDev = (average == 0.0) ? 0.0 : ((100.0 * stdDev) / average);
}
void printTimesInTableFormat(struct testResults *results,
struct resultsData *data, bool printAverage,
bool printStdDev) {
unsigned int i, j;
bool printStdDevBandwidth = printStdDev && data->reportAsBandwidth;
printf("Size_KB");
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
printf("\t%7s", memAllocTypeShortStr[i]);
}
printf("\n");
for (j = 0; j < results->numSizesToTest; j++) {
printf("%lu", results->sizesToTest[j] / ONE_KB);
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
printf(data->reportAsBandwidth ? "\t%7.2lf" : "\t%7.3lf",
printStdDevBandwidth
? data->stdDevBandwidthInMBps[i][j]
: getTimeOrBandwidth(
printAverage ? data->averageRunTimesInMs[i][j]
: data->stdDevRunTimesInMs[i][j],
results->sizesToTest[j], data->reportAsBandwidth));
}
printf("\n");
}
}
void printAllResultsInVerboseMode(struct testResults *results,
struct resultsData *data) {
unsigned int i, j, k;
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
printf("Verbose mode, printing all results for %s\n", memAllocTypeStr[i]);
printf("Instance");
for (j = 0; j < results->numSizesToTest; j++) {
printf("\t%lu", results->sizesToTest[j] / ONE_KB);
}
printf("\n");
for (k = 0; k < results->numMeasurements; k++) {
printf("%u", k);
for (j = 0; j < results->numSizesToTest; j++) {
printf(data->reportAsBandwidth ? "\t%7.2lf" : "\t%7.3lf",
getTimeOrBandwidth(data->runTimesInMs[i][j][k],
results->sizesToTest[j],
data->reportAsBandwidth));
}
printf("\n");
}
}
}
void printResults(struct testResults *results,
bool print_launch_transfer_results,
bool print_std_deviation) {
char vulcanPrint[256];
char resultNameNoSpaces[64];
unsigned int i, j, k;
struct resultsData *resultsIter;
bool sizeGreaterThan1MB;
for (resultsIter = results->resultsDataHead; resultsIter != NULL;
resultsIter = resultsIter->next) {
if (!verboseResults && resultsIter->printOnlyInVerbose) {
continue;
}
if (!print_launch_transfer_results) {
if (!(strcmp(resultsIter->resultsName, "Overall Time") == 0)) {
continue;
}
}
// regular print
printf("\n%s For %s ", resultsIter->resultsName, results->testName);
printf("\n");
for (j = 0; j < results->numSizesToTest; j++) {
for (i = 0; i < MEMALLOC_TYPE_COUNT; i++) {
calculateAverageAndStdDev(&resultsIter->averageRunTimesInMs[i][j],
&resultsIter->stdDevRunTimesInMs[i][j],
resultsIter->runTimesInMs[i][j],
results->numMeasurements);
if (resultsIter->reportAsBandwidth) {
calculateStdDevBandwidth(&resultsIter->stdDevBandwidthInMBps[i][j],
resultsIter->runTimesInMs[i][j],
results->numMeasurements,
results->sizesToTest[j]);
}
}
}
printf("\nPrinting Average of %u measurements in (%s)\n",
results->numMeasurements,
resultsIter->reportAsBandwidth ? UNITS_BW : UNITS_Time);
printTimesInTableFormat(results, resultsIter, true, false);
if (print_std_deviation) {
printf(
"\nPrinting Standard Deviation as %% of average of %u measurements\n",
results->numMeasurements);
printTimesInTableFormat(results, resultsIter, false, true);
}
if (verboseResults) {
printAllResultsInVerboseMode(results, resultsIter);
}
}
}

View File

@@ -0,0 +1,697 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <helper_cuda.h>
#include <helper_timer.h>
#include "commonDefs.hpp"
#include "commonKernels.hpp"
#define VERIFY_GPU_CORRECTNESS 0
size_t maxSampleSizeInMb = 64;
int numKernelRuns = 100;
int verboseResults = 0;
const char *memAllocTypeStr[MEMALLOC_TYPE_COUNT] = {
"Managed_Memory_With_Hints",
"Managed_Memory_With_Hints_FullyAsync",
"Managed_Memory_NoHints",
"Zero_Copy",
"Memcpy_HostMalloc_DeviceCudaMalloc",
"MemcpyAsync_HostMalloc_DeviceCudaMalloc",
"Memcpy_HostCudaHostAlloc_DeviceCudaMalloc",
"MemcpyAsync_HostCudaHostAlloc_DeviceCudaMalloc"};
const char *memAllocTypeShortStr[MEMALLOC_TYPE_COUNT] = {
"UMhint", // Managed Memory With Hints
"UMhntAs", // Managed Memory With_Hints Async
"UMeasy", // Managed_Memory with No Hints
"0Copy", // Zero Copy
"MemCopy", // USE HOST PAGEABLE AND DEVICE_MEMORY
"CpAsync", // USE HOST PAGEABLE AND DEVICE_MEMORY ASYNC
"CpHpglk", // USE HOST PAGELOCKED AND DEVICE MEMORY
"CpPglAs" // USE HOST PAGELOCKED AND DEVICE MEMORY ASYNC
};
static float RandFloat(float low, float high) {
float t = (float)rand() / (float)RAND_MAX;
return (1.0f - t) * low + t * high;
}
void fillMatrixWithRandomValues(float *matrix, unsigned int matrixDim) {
unsigned int i, j;
for (i = 0; i < matrixDim; ++i) {
for (j = 0; j < matrixDim; ++j) {
matrix[j + i * matrixDim] = RandFloat(0.0f, 10.0f);
}
}
}
#if VERIFY_GPU_CORRECTNESS
void verifyMatrixMultiplyCorrectness(float *C, float *A, float *B,
unsigned int matrixDim) {
unsigned int i, j, k, numErrors = 0;
for (i = 0; i < matrixDim; ++i) {
for (j = 0; j < matrixDim; ++j) {
float result = 0.0f;
for (k = 0; k < matrixDim; ++k) {
result += A[k + i * matrixDim] * B[j + k * matrixDim];
}
if (fabs(C[j + i * matrixDim] - result) > 0.001 * matrixDim) {
printf("At [%u, %u]: Expected %f, Found %f\n", i, j, result,
C[j + i * matrixDim]);
++numErrors;
}
}
}
if (numErrors != 0) {
printf("%d value mismatches occured\n", numErrors);
fflush(stdout);
exit(EXIT_FAILURE); // exit since value mismatches occured
}
}
#endif
void copyMatrix(float *dstMatrix, float *srcMatrix, unsigned int matrixDim) {
size_t size = matrixDim * matrixDim * sizeof(float);
memcpy(dstMatrix, srcMatrix, size);
}
void verifyMatrixData(float *expectedData, float *observedData,
unsigned int matrixDim) {
unsigned int i, j, numErrors = 0;
for (i = 0; i < matrixDim; ++i) {
for (j = 0; j < matrixDim; ++j) {
if (expectedData[j + i * matrixDim] != observedData[j + i * matrixDim]) {
++numErrors;
if (verboseResults) {
printf("At [%u, %u]: Expected %f, Found %f\n", i, j,
expectedData[j + i * matrixDim],
observedData[j + i * matrixDim]);
}
}
}
}
if (numErrors != 0) {
printf("%d value mismatches occured\n", numErrors);
fflush(stdout);
exit(EXIT_FAILURE); // exit since value mismatches occured
}
}
#define BLOCK_SIZE 32
__global__ void matrixMultiplyKernel(float *C, float *A, float *B,
unsigned int matrixDim) {
// Block index
int bx = blockIdx.x;
int by = blockIdx.y;
// Thread index
int tx = threadIdx.x;
int ty = threadIdx.y;
unsigned int wA = matrixDim;
unsigned int wB = matrixDim;
// Index of the first sub-matrix of A processed by the block
int aBegin = matrixDim * BLOCK_SIZE * by;
// Index of the last sub-matrix of A processed by the block
int aEnd = aBegin + wA - 1;
// Step size used to iterate through the sub-matrices of A
int aStep = BLOCK_SIZE;
// Index of the first sub-matrix of B processed by the block
int bBegin = BLOCK_SIZE * bx;
// Step size used to iterate through the sub-matrices of B
int bStep = BLOCK_SIZE * wB;
// Csub is used to store the element of the block sub-matrix
// that is computed by the thread
float Csub = 0;
// Loop over all the sub-matrices of A and B
// required to compute the block sub-matrix
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
// Declaration of the shared memory array As used to
// store the sub-matrix of A
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
// Declaration of the shared memory array Bs used to
// store the sub-matrix of B
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
// Load the matrices from device memory
// to shared memory; each thread loads
// one element of each matrix
As[ty][tx] = A[a + wA * ty + tx];
Bs[ty][tx] = B[b + wB * ty + tx];
// Synchronize to make sure the matrices are loaded
__syncthreads();
// Multiply the two matrices together;
// each thread computes one element
// of the block sub-matrix
#pragma unroll
for (int k = 0; k < BLOCK_SIZE; ++k) {
Csub += As[ty][k] * Bs[k][tx];
}
// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
__syncthreads();
}
// Write the block sub-matrix to device memory;
// each thread writes one element
int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
C[c + wB * ty + tx] = Csub;
}
void runMatrixMultiplyKernel(unsigned int matrixDim, int allocType,
unsigned int numLoops, double *gpuLaunchCallsTimes,
double *gpuTransferToCallsTimes,
double *gpuTransferFromCallsTimes,
double *gpuLaunchAndTransferCallsTimes,
double *gpuLaunchTransferSyncTimes,
double *cpuAccessTimes, double *overallTimes,
int device_id) {
float *dptrA = NULL, *hptrA = NULL;
float *dptrB = NULL, *hptrB = NULL;
float *dptrC = NULL, *hptrC = NULL;
float *randValuesX = NULL, *randValuesY = NULL;
float *randValuesVerifyXmulY = NULL, *randValuesVerifyYmulX = NULL;
bool copyRequired = false, hintsRequired = false;
bool someTransferOpRequired;
bool isAsync = false;
cudaStream_t streamToRunOn;
unsigned int *latch;
size_t size = matrixDim * matrixDim * sizeof(float);
dim3 threads(32, 32);
dim3 grid(matrixDim / threads.x, matrixDim / threads.y);
StopWatchInterface *gpuLaunchCallsTimer = 0, *gpuTransferCallsTimer = 0;
StopWatchInterface *gpuSyncTimer = 0, *cpuAccessTimer = 0;
sdkCreateTimer(&gpuLaunchCallsTimer);
sdkCreateTimer(&gpuTransferCallsTimer);
sdkCreateTimer(&gpuSyncTimer);
sdkCreateTimer(&cpuAccessTimer);
unsigned int i;
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, device_id));
checkCudaErrors(cudaStreamCreate(&streamToRunOn));
randValuesX = (float *)malloc(size);
if (!randValuesX) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
randValuesY = (float *)malloc(size);
if (!randValuesY) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
randValuesVerifyXmulY = (float *)malloc(size);
if (!randValuesVerifyXmulY) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
randValuesVerifyYmulX = (float *)malloc(size);
if (!randValuesVerifyYmulX) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
checkCudaErrors(cudaMalloc(&dptrA, size));
checkCudaErrors(cudaMalloc(&dptrB, size));
checkCudaErrors(cudaMalloc(&dptrC, size));
fillMatrixWithRandomValues(randValuesX, matrixDim);
fillMatrixWithRandomValues(randValuesY, matrixDim);
checkCudaErrors(
cudaMemcpyAsync(dptrA, randValuesX, size, cudaMemcpyHostToDevice));
checkCudaErrors(
cudaMemcpyAsync(dptrB, randValuesY, size, cudaMemcpyHostToDevice));
matrixMultiplyKernel<<<grid, threads>>>(dptrC, dptrA, dptrB, matrixDim);
checkCudaErrors(cudaMemcpyAsync(randValuesVerifyXmulY, dptrC, size,
cudaMemcpyDeviceToHost));
checkCudaErrors(cudaStreamSynchronize(NULL));
matrixMultiplyKernel<<<grid, threads>>>(dptrC, dptrB, dptrA, matrixDim);
checkCudaErrors(cudaMemcpyAsync(randValuesVerifyYmulX, dptrC, size,
cudaMemcpyDeviceToHost));
checkCudaErrors(cudaStreamSynchronize(NULL));
#if VERIFY_GPU_CORRECTNESS
verifyMatrixMultiplyCorrectness(randValuesVerifyXmulY, randValuesX,
randValuesY, matrixDim);
verifyMatrixMultiplyCorrectness(randValuesVerifyYmulX, randValuesY,
randValuesX, matrixDim);
#endif
checkCudaErrors(cudaFree(dptrA));
checkCudaErrors(cudaFree(dptrB));
checkCudaErrors(cudaFree(dptrC));
checkCudaErrors(cudaMallocHost(&latch, sizeof(unsigned int)));
switch (allocType) {
case USE_HOST_PAGEABLE_AND_DEVICE_MEMORY:
case USE_HOST_PAGEABLE_AND_DEVICE_MEMORY_ASYNC:
hptrA = (float *)malloc(size);
if (!hptrA) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
hptrB = (float *)malloc(size);
if (!hptrB) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
hptrC = (float *)malloc(size);
if (!hptrC) {
exit(EXIT_FAILURE); // exit since memory allocation error
}
checkCudaErrors(cudaMalloc(&dptrA, size));
checkCudaErrors(cudaMalloc(&dptrB, size));
checkCudaErrors(cudaMalloc(&dptrC, size));
copyRequired = true;
break;
case USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY:
case USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY_ASYNC:
checkCudaErrors(cudaMallocHost(&hptrA, size));
checkCudaErrors(cudaMallocHost(&hptrB, size));
checkCudaErrors(cudaMallocHost(&hptrC, size));
checkCudaErrors(cudaMalloc(&dptrA, size));
checkCudaErrors(cudaMalloc(&dptrB, size));
checkCudaErrors(cudaMalloc(&dptrC, size));
copyRequired = true;
break;
case USE_ZERO_COPY:
checkCudaErrors(cudaMallocHost(&hptrA, size));
checkCudaErrors(cudaMallocHost(&hptrB, size));
checkCudaErrors(cudaMallocHost(&hptrC, size));
checkCudaErrors(cudaHostGetDevicePointer(&dptrA, hptrA, 0));
checkCudaErrors(cudaHostGetDevicePointer(&dptrB, hptrB, 0));
checkCudaErrors(cudaHostGetDevicePointer(&dptrC, hptrC, 0));
break;
case USE_MANAGED_MEMORY:
checkCudaErrors(cudaMallocManaged(&dptrA, size));
checkCudaErrors(cudaMallocManaged(&dptrB, size));
checkCudaErrors(cudaMallocManaged(&dptrC, size));
hptrA = dptrA;
hptrB = dptrB;
hptrC = dptrC;
break;
case USE_MANAGED_MEMORY_WITH_HINTS:
case USE_MANAGED_MEMORY_WITH_HINTS_ASYNC:
if (deviceProp.concurrentManagedAccess) {
checkCudaErrors(cudaMallocManaged(&dptrA, size));
checkCudaErrors(cudaMallocManaged(&dptrB, size));
checkCudaErrors(cudaMallocManaged(&dptrC, size));
checkCudaErrors(cudaMemPrefetchAsync(dptrA, size, cudaCpuDeviceId));
checkCudaErrors(cudaMemPrefetchAsync(dptrB, size, cudaCpuDeviceId));
checkCudaErrors(cudaMemPrefetchAsync(dptrC, size, cudaCpuDeviceId));
} else {
checkCudaErrors(cudaMallocManaged(&dptrA, size, cudaMemAttachHost));
checkCudaErrors(cudaMallocManaged(&dptrB, size, cudaMemAttachHost));
checkCudaErrors(cudaMallocManaged(&dptrC, size, cudaMemAttachHost));
}
hptrA = dptrA;
hptrB = dptrB;
hptrC = dptrC;
hintsRequired = true;
break;
default:
exit(EXIT_FAILURE); // exit with error
}
if (allocType == USE_HOST_PAGEABLE_AND_DEVICE_MEMORY_ASYNC ||
allocType == USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY_ASYNC ||
allocType == USE_MANAGED_MEMORY_WITH_HINTS_ASYNC) {
isAsync = true;
}
someTransferOpRequired = copyRequired || hintsRequired;
// fill buffers with 0 to avoid any first access page-fault overheads.
memset(hptrA, 0, size);
memset(hptrB, 0, size);
memset(hptrC, 0, size);
for (i = 0; i < numLoops; i++) {
cpuAccessTimes[i] = 0.0;
gpuLaunchCallsTimes[i] = 0.0;
gpuTransferToCallsTimes[i] = 0.0;
gpuTransferFromCallsTimes[i] = 0.0;
sdkStartTimer(&cpuAccessTimer);
{
copyMatrix(hptrA, (i & 0x1 == 0) ? randValuesX : randValuesY, matrixDim);
copyMatrix(hptrB, (i & 0x1 == 0) ? randValuesY : randValuesX, matrixDim);
}
sdkStopTimer(&cpuAccessTimer);
cpuAccessTimes[i] += sdkGetAverageTimerValue(&cpuAccessTimer);
sdkResetTimer(&cpuAccessTimer);
if (isAsync && hintsRequired) {
*latch = 0;
// Prevent any work on stream from starting until all work is pushed
spinWhileLessThanOne<<<1, 1, 0, streamToRunOn>>>(latch);
}
if (someTransferOpRequired) {
sdkStartTimer(&gpuTransferCallsTimer);
if (copyRequired) {
if (isAsync) {
checkCudaErrors(cudaMemcpyAsync(
dptrA, hptrA, size, cudaMemcpyHostToDevice, streamToRunOn));
checkCudaErrors(cudaMemcpyAsync(
dptrB, hptrB, size, cudaMemcpyHostToDevice, streamToRunOn));
} else {
checkCudaErrors(
cudaMemcpy(dptrA, hptrA, size, cudaMemcpyHostToDevice));
checkCudaErrors(
cudaMemcpy(dptrB, hptrB, size, cudaMemcpyHostToDevice));
}
}
if (hintsRequired) {
if (deviceProp.concurrentManagedAccess) {
checkCudaErrors(
cudaMemPrefetchAsync(dptrA, size, device_id, streamToRunOn));
checkCudaErrors(
cudaMemPrefetchAsync(dptrB, size, device_id, streamToRunOn));
checkCudaErrors(
cudaMemPrefetchAsync(dptrC, size, device_id, streamToRunOn));
} else {
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrA, 0,
cudaMemAttachGlobal));
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrB, 0,
cudaMemAttachGlobal));
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrC, 0,
cudaMemAttachGlobal));
}
if (!isAsync) {
checkCudaErrors(cudaStreamSynchronize(streamToRunOn));
}
}
sdkStopTimer(&gpuTransferCallsTimer);
gpuTransferToCallsTimes[i] +=
sdkGetAverageTimerValue(&gpuTransferCallsTimer);
sdkResetTimer(&gpuTransferCallsTimer);
}
sdkStartTimer(&gpuLaunchCallsTimer);
{
matrixMultiplyKernel<<<grid, threads, 0, streamToRunOn>>>(
dptrC, dptrA, dptrB, matrixDim);
if (!isAsync) {
checkCudaErrors(cudaStreamSynchronize(streamToRunOn));
}
}
sdkStopTimer(&gpuLaunchCallsTimer);
gpuLaunchCallsTimes[i] += sdkGetAverageTimerValue(&gpuLaunchCallsTimer);
sdkResetTimer(&gpuLaunchCallsTimer);
if (someTransferOpRequired) {
sdkStartTimer(&gpuTransferCallsTimer);
if (hintsRequired) {
if (deviceProp.concurrentManagedAccess) {
checkCudaErrors(cudaMemPrefetchAsync(dptrA, size, cudaCpuDeviceId));
checkCudaErrors(cudaMemPrefetchAsync(dptrB, size, cudaCpuDeviceId));
checkCudaErrors(cudaMemPrefetchAsync(dptrC, size, cudaCpuDeviceId));
} else {
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrA, 0,
cudaMemAttachHost));
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrB, 0,
cudaMemAttachHost));
checkCudaErrors(cudaStreamAttachMemAsync(streamToRunOn, dptrC, 0,
cudaMemAttachHost));
}
if (!isAsync) {
checkCudaErrors(cudaStreamSynchronize(streamToRunOn));
}
}
if (copyRequired) {
if (isAsync) {
checkCudaErrors(cudaMemcpyAsync(
hptrC, dptrC, size, cudaMemcpyDeviceToHost, streamToRunOn));
} else {
checkCudaErrors(
cudaMemcpy(hptrC, dptrC, size, cudaMemcpyDeviceToHost));
}
}
sdkStopTimer(&gpuTransferCallsTimer);
gpuTransferFromCallsTimes[i] +=
sdkGetAverageTimerValue(&gpuTransferCallsTimer);
sdkResetTimer(&gpuTransferCallsTimer);
}
gpuLaunchAndTransferCallsTimes[i] = gpuLaunchCallsTimes[i] +
gpuTransferToCallsTimes[i] +
gpuTransferFromCallsTimes[i];
gpuLaunchTransferSyncTimes[i] = gpuLaunchAndTransferCallsTimes[i];
if (isAsync) {
sdkStartTimer(&gpuSyncTimer);
{
if (hintsRequired) {
*latch = 1;
}
checkCudaErrors(cudaStreamSynchronize(streamToRunOn));
}
sdkStopTimer(&gpuSyncTimer);
gpuLaunchTransferSyncTimes[i] += sdkGetAverageTimerValue(&gpuSyncTimer);
sdkResetTimer(&gpuSyncTimer);
}
sdkStartTimer(&cpuAccessTimer);
{
verifyMatrixData(
(i & 0x1 == 0) ? randValuesVerifyXmulY : randValuesVerifyYmulX, hptrC,
matrixDim);
}
sdkStopTimer(&cpuAccessTimer);
cpuAccessTimes[i] += sdkGetAverageTimerValue(&cpuAccessTimer);
sdkResetTimer(&cpuAccessTimer);
overallTimes[i] = cpuAccessTimes[i] + gpuLaunchTransferSyncTimes[i];
}
switch (allocType) {
case USE_HOST_PAGEABLE_AND_DEVICE_MEMORY:
case USE_HOST_PAGEABLE_AND_DEVICE_MEMORY_ASYNC:
free(hptrA);
free(hptrB);
free(hptrC);
checkCudaErrors(cudaFree(dptrA));
checkCudaErrors(cudaFree(dptrB));
checkCudaErrors(cudaFree(dptrC));
break;
case USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY:
case USE_HOST_PAGELOCKED_AND_DEVICE_MEMORY_ASYNC:
checkCudaErrors(cudaFreeHost(hptrA));
checkCudaErrors(cudaFreeHost(hptrB));
checkCudaErrors(cudaFreeHost(hptrC));
checkCudaErrors(cudaFree(dptrA));
checkCudaErrors(cudaFree(dptrB));
checkCudaErrors(cudaFree(dptrC));
break;
case USE_ZERO_COPY:
checkCudaErrors(cudaFreeHost(hptrA));
checkCudaErrors(cudaFreeHost(hptrB));
checkCudaErrors(cudaFreeHost(hptrC));
break;
case USE_MANAGED_MEMORY:
case USE_MANAGED_MEMORY_WITH_HINTS:
case USE_MANAGED_MEMORY_WITH_HINTS_ASYNC:
checkCudaErrors(cudaFree(dptrA));
checkCudaErrors(cudaFree(dptrB));
checkCudaErrors(cudaFree(dptrC));
break;
default:
exit(EXIT_FAILURE); // exit due to error
}
checkCudaErrors(cudaStreamDestroy(streamToRunOn));
checkCudaErrors(cudaFreeHost(latch));
free(randValuesX);
free(randValuesY);
free(randValuesVerifyXmulY);
free(randValuesVerifyYmulX);
sdkDeleteTimer(&gpuLaunchCallsTimer);
sdkDeleteTimer(&gpuTransferCallsTimer);
sdkDeleteTimer(&gpuSyncTimer);
sdkDeleteTimer(&cpuAccessTimer);
}
void matrixMultiplyPerfRunner(bool reportAsBandwidth,
bool print_launch_transfer_results,
bool print_std_deviation, int device_id) {
int i;
unsigned int minMatrixDim = 32;
unsigned int multiplierDim = 2;
unsigned int matrixDim;
unsigned int minSize = minMatrixDim * minMatrixDim * sizeof(float);
unsigned int maxSize =
(maxSampleSizeInMb * ONE_MB) /
4; // 3 buffers are used, but dividing by 4 (power of 2)
unsigned int multiplier = multiplierDim * multiplierDim;
unsigned int numSizesToTest;
struct testResults *results;
struct resultsData *gpuLaunchCallsTimes;
struct resultsData *gpuTransferToCallsTimes;
struct resultsData *gpuTransferFromCallsTimes;
struct resultsData *gpuLaunchAndTransferCallsTimes;
struct resultsData *gpuLaunchTransferSyncTimes;
struct resultsData *cpuAccessTimes;
struct resultsData *overallTimes;
unsigned long *sizesToTest;
unsigned int j;
numSizesToTest = findNumSizesToTest(minSize, maxSize, multiplier);
createAndInitTestResults(&results, "matrixMultiplyPerf", numKernelRuns,
numSizesToTest);
sizesToTest = getPtrSizesToTest(results);
createResultDataAndAddToTestResults(&gpuLaunchCallsTimes, results,
"GPU Kernel Launch Call Time", false,
reportAsBandwidth);
createResultDataAndAddToTestResults(&gpuTransferToCallsTimes, results,
"CPU to GPU Transfer Calls Time", false,
reportAsBandwidth);
createResultDataAndAddToTestResults(&gpuTransferFromCallsTimes, results,
"GPU to CPU Transfer Calls Time", false,
reportAsBandwidth);
createResultDataAndAddToTestResults(&gpuLaunchAndTransferCallsTimes, results,
"GPU Launch and Transfer Calls Time",
false, reportAsBandwidth);
createResultDataAndAddToTestResults(&gpuLaunchTransferSyncTimes, results,
"GPU Launch Transfer and Sync Time",
false, reportAsBandwidth);
createResultDataAndAddToTestResults(
&cpuAccessTimes, results, "CPU Access Time", false, reportAsBandwidth);
createResultDataAndAddToTestResults(&overallTimes, results, "Overall Time",
false, reportAsBandwidth);
printf("Running ");
for (matrixDim = minMatrixDim, j = 0;
matrixDim * matrixDim <= maxSize / sizeof(float);
matrixDim *= multiplierDim, ++j) {
sizesToTest[j] = matrixDim * matrixDim * sizeof(float);
for (i = MEMALLOC_TYPE_START; i <= MEMALLOC_TYPE_END; i++) {
printf(".");
fflush(stdout);
runMatrixMultiplyKernel(
matrixDim, i, numKernelRuns,
getPtrRunTimesInMs(gpuLaunchCallsTimes, i, j),
getPtrRunTimesInMs(gpuTransferToCallsTimes, i, j),
getPtrRunTimesInMs(gpuTransferFromCallsTimes, i, j),
getPtrRunTimesInMs(gpuLaunchAndTransferCallsTimes, i, j),
getPtrRunTimesInMs(gpuLaunchTransferSyncTimes, i, j),
getPtrRunTimesInMs(cpuAccessTimes, i, j),
getPtrRunTimesInMs(overallTimes, i, j), device_id);
}
}
printf("\n");
printResults(results, print_launch_transfer_results, print_std_deviation);
freeTestResultsAndAllResultsData(results);
}
static void usage() {
printf(
"./cudaMemoryTypesPerf [-device=<device_id>] [-reportAsBandwidth] "
"[-print-launch-transfer-results] [-print-std-deviation] [-verbose]\n");
printf("Options:\n");
printf(
"-reportAsBandwidth: By default time taken is printed, this "
"option allows to instead print bandwidth.\n");
printf(
"-print-launch-transfer-results: By default overall results are printed, "
"this option allows to print data transfers and kernel time as well.\n");
printf(
"-print-std-deviation: Prints std deviation of the results.\n");
printf(
"-kernel-iterations=<num>: Number of times the kernel tests should "
"be run[default is 100 iterations].\n");
printf(
"-device=<device_id>: Allows to pass GPU Device ID on which "
"the tests will be run.\n");
printf("-verbose: Prints highly verbose output.\n");
}
int main(int argc, char **argv) {
bool reportAsBandwidth = false;
bool print_launch_transfer_results = false;
bool print_std_deviation = false;
if (checkCmdLineFlag(argc, (const char **)argv, "help") ||
checkCmdLineFlag(argc, (const char **)argv, "h")) {
usage();
printf("&&&& %s WAIVED\n", argv[0]);
exit(EXIT_WAIVED);
}
if (checkCmdLineFlag(argc, (const char **)argv, "reportAsBandwidth")) {
reportAsBandwidth = true;
}
if (checkCmdLineFlag(argc, (const char **)argv,
"print-launch-transfer-results")) {
print_launch_transfer_results = true;
}
if (checkCmdLineFlag(argc, (const char **)argv, "print-std-deviation")) {
print_std_deviation = true;
}
if (checkCmdLineFlag(argc, (const char **)argv, "kernel-iterations")) {
numKernelRuns =
getCmdLineArgumentInt(argc, (const char **)argv, "kernel-iterations");
}
if (checkCmdLineFlag(argc, (const char **)argv, "verbose")) {
verboseResults = 1;
}
int device_id = findCudaDevice(argc, (const char **)argv);
matrixMultiplyPerfRunner(reportAsBandwidth, print_launch_transfer_results,
print_std_deviation, device_id);
printf(
"\nNOTE: The CUDA Samples are not meant for performance measurements. "
"Results may vary when GPU Boost is enabled.\n");
exit(EXIT_SUCCESS);
}

View File

@@ -0,0 +1,304 @@
################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
################################################################################
#
# Makefile project only supported on Mac OS X and Linux Platforms)
#
################################################################################
# Location of the CUDA Toolkit
CUDA_PATH ?= /usr/local/cuda
##############################
# start deprecated interface #
##############################
ifeq ($(x86_64),1)
$(info WARNING - x86_64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=x86_64 instead)
TARGET_ARCH ?= x86_64
endif
ifeq ($(ARMv7),1)
$(info WARNING - ARMv7 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=armv7l instead)
TARGET_ARCH ?= armv7l
endif
ifeq ($(aarch64),1)
$(info WARNING - aarch64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=aarch64 instead)
TARGET_ARCH ?= aarch64
endif
ifeq ($(ppc64le),1)
$(info WARNING - ppc64le variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=ppc64le instead)
TARGET_ARCH ?= ppc64le
endif
ifneq ($(GCC),)
$(info WARNING - GCC variable has been deprecated)
$(info WARNING - please use HOST_COMPILER=$(GCC) instead)
HOST_COMPILER ?= $(GCC)
endif
ifneq ($(abi),)
$(error ERROR - abi variable has been removed)
endif
############################
# end deprecated interface #
############################
# architecture
HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif
endif
# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
TARGET_ARCH = armv7l
endif
# operating system
HOST_OS := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
TARGET_OS ?= $(HOST_OS)
ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
$(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
endif
# host compiler
ifeq ($(TARGET_OS),darwin)
ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
HOST_COMPILER ?= clang++
endif
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
ifeq ($(TARGET_OS),linux)
HOST_COMPILER ?= arm-linux-gnueabihf-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
else ifeq ($(TARGET_OS),android)
HOST_COMPILER ?= arm-linux-androideabi-g++
endif
else ifeq ($(TARGET_ARCH),aarch64)
ifeq ($(TARGET_OS), linux)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++
endif
else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif
endif
HOST_COMPILER ?= g++
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
# internal flags
NVCCFLAGS := -m${TARGET_SIZE}
CCFLAGS :=
LDFLAGS :=
# build flags
ifeq ($(TARGET_OS),darwin)
LDFLAGS += -rpath $(CUDA_PATH)/lib
CCFLAGS += -arch $(HOST_ARCH)
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
CCFLAGS += -mfloat-abi=hard
else ifeq ($(TARGET_OS),android)
LDFLAGS += -pie
CCFLAGS += -fpie -fpic -fexceptions
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
endif
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
else ifeq ($(TARGET_ARCH),ppc64le)
CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
endif
# Debug build flags
ifeq ($(dbg),1)
NVCCFLAGS += -g -G
BUILD_TYPE := debug
else
BUILD_TYPE := release
endif
ALL_CCFLAGS :=
ALL_CCFLAGS += $(NVCCFLAGS)
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
SAMPLE_ENABLED := 1
ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
# Common includes and paths for CUDA
INCLUDES := -I../../Common
LIBRARIES :=
################################################################################
# Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75
else
SMS ?= 30 35 37 50 52 60 61 70 75
endif
ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ifeq ($(GENCODE_FLAGS),)
# Generate SASS code for each SM architecture listed in $(SMS)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
HIGHEST_SM := $(lastword $(sort $(SMS)))
ifneq ($(HIGHEST_SM),)
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: bandwidthTest
check.deps:
ifeq ($(SAMPLE_ENABLED),0)
@echo "Sample will be waived due to the above missing dependencies"
else
@echo "Sample is ready - all dependencies have been met"
endif
bandwidthTest.o:bandwidthTest.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
bandwidthTest: bandwidthTest.o
$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
run: build
$(EXEC) ./bandwidthTest
clean:
rm -f bandwidthTest bandwidthTest.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/bandwidthTest
clobber: clean

View File

@@ -0,0 +1,79 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>bandwidthTest</name>
<cuda_api_list>
<toolkit>cudaSetDevice</toolkit>
<toolkit>cudaHostAlloc</toolkit>
<toolkit>cudaFree</toolkit>
<toolkit>cudaMallocHost</toolkit>
<toolkit>cudaFreeHost</toolkit>
<toolkit>cudaMemcpy</toolkit>
<toolkit>cudaMemcpyAsync</toolkit>
<toolkit>cudaEventCreate</toolkit>
<toolkit>cudaEventRecord</toolkit>
<toolkit>cudaEventDestroy</toolkit>
<toolkit>cudaDeviceSynchronize</toolkit>
<toolkit>cudaEventElapsedTime</toolkit>
</cuda_api_list>
<description><![CDATA[This is a simple test program to measure the memcopy bandwidth of the GPU and memcpy bandwidth across PCI-e. This test application is capable of measuring device to device copy bandwidth, host to device copy bandwidth for pageable and page-locked memory, and device to host copy bandwidth for pageable and page-locked memory.]]></description>
<devicecompilation>whole</devicecompilation>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">CUDA Streams and Events</concept>
<concept level="basic">Performance Strategies</concept>
</keyconcepts>
<keywords>
<keyword>GPGPU</keyword>
<keyword>bandwidth</keyword>
</keywords>
<libraries>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>bandwidthTest.cu</primary_file>
<scopes>
<scope>1:CUDA Basic Topics</scope>
<scope>1:Performance Strategies</scope>
</scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch>
<sm-arch>sm52</sm-arch>
<sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<platform>windows7</platform>
</env>
<env>
<arch>x86_64</arch>
<platform>macosx</platform>
</env>
<env>
<arch>arm</arch>
</env>
<env>
<arch>ppc64le</arch>
<platform>linux</platform>
</env>
</supported_envs>
<supported_sm_architectures>
<include>all</include>
</supported_sm_architectures>
<title>Bandwidth Test</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,94 @@
# bandwidthTest - Bandwidth Test
## Description
This is a simple test program to measure the memcopy bandwidth of the GPU and memcpy bandwidth across PCI-e. This test application is capable of measuring device to device copy bandwidth, host to device copy bandwidth for pageable and page-locked memory, and device to host copy bandwidth for pageable and page-locked memory.
## Key Concepts
CUDA Streams and Events, Performance Strategies
## Supported SM Architectures
[SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows, MacOSX
## Supported CPU Architecture
x86_64, ppc64le, armv7l
## CUDA APIs involved
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaSetDevice, cudaHostAlloc, cudaFree, cudaMallocHost, cudaFreeHost, cudaMemcpy, cudaMemcpyAsync, cudaEventCreate, cudaEventRecord, cudaEventDestroy, cudaDeviceSynchronize, cudaEventElapsedTime
## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run
### Windows
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
```
*_vs<version>.sln - for Visual Studio <version>
```
Each individual sample has its own set of solution files in its directory:
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
### Linux
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, armv7l.
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=armv7l` <br/>
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
```
$ make SMS="50 60"
```
* **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=g++
```
### Mac
The Mac samples are built using makefiles. To use the makefiles, change directory into the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where "A B ..." is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use SMS="50 60".
```
$ make SMS="A B ..."
```
* **HOST_COMPILER=<host_compiler>** - override the default clang host compiler. See the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=clang
```
## References (for more details)

View File

@@ -0,0 +1,969 @@
/* Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
/*
* This is a simple test program to measure the memcopy bandwidth of the GPU.
* It can measure device to device copy bandwidth, host to device copy bandwidth
* for pageable and pinned memory, and device to host copy bandwidth for
* pageable and pinned memory.
*
* Usage:
* ./bandwidthTest [option]...
*/
// CUDA runtime
#include <cuda_runtime.h>
// includes
#include <helper_cuda.h> // helper functions for CUDA error checking and initialization
#include <helper_functions.h> // helper for shared functions common to CUDA Samples
#include <cuda.h>
#include <cassert>
#include <iostream>
#include <memory>
static const char *sSDKsample = "CUDA Bandwidth Test";
// defines, project
#define MEMCOPY_ITERATIONS 100
#define DEFAULT_SIZE (32 * (1e6)) // 32 M
#define DEFAULT_INCREMENT (4 * (1e6)) // 4 M
#define CACHE_CLEAR_SIZE (16 * (1e6)) // 16 M
// shmoo mode defines
#define SHMOO_MEMSIZE_MAX (64 * (1e6)) // 64 M
#define SHMOO_MEMSIZE_START (1e3) // 1 KB
#define SHMOO_INCREMENT_1KB (1e3) // 1 KB
#define SHMOO_INCREMENT_2KB (2 * 1e3) // 2 KB
#define SHMOO_INCREMENT_10KB (10 * (1e3)) // 10KB
#define SHMOO_INCREMENT_100KB (100 * (1e3)) // 100 KB
#define SHMOO_INCREMENT_1MB (1e6) // 1 MB
#define SHMOO_INCREMENT_2MB (2 * 1e6) // 2 MB
#define SHMOO_INCREMENT_4MB (4 * 1e6) // 4 MB
#define SHMOO_LIMIT_20KB (20 * (1e3)) // 20 KB
#define SHMOO_LIMIT_50KB (50 * (1e3)) // 50 KB
#define SHMOO_LIMIT_100KB (100 * (1e3)) // 100 KB
#define SHMOO_LIMIT_1MB (1e6) // 1 MB
#define SHMOO_LIMIT_16MB (16 * 1e6) // 16 MB
#define SHMOO_LIMIT_32MB (32 * 1e6) // 32 MB
// CPU cache flush
#define FLUSH_SIZE (256 * 1024 * 1024)
char *flush_buf;
// enums, project
enum testMode { QUICK_MODE, RANGE_MODE, SHMOO_MODE };
enum memcpyKind { DEVICE_TO_HOST, HOST_TO_DEVICE, DEVICE_TO_DEVICE };
enum printMode { USER_READABLE, CSV };
enum memoryMode { PINNED, PAGEABLE };
const char *sMemoryCopyKind[] = {"Device to Host", "Host to Device",
"Device to Device", NULL};
const char *sMemoryMode[] = {"PINNED", "PAGEABLE", NULL};
// if true, use CPU based timing for everything
static bool bDontUseGPUTiming;
int *pArgc = NULL;
char **pArgv = NULL;
////////////////////////////////////////////////////////////////////////////////
// declaration, forward
int runTest(const int argc, const char **argv);
void testBandwidth(unsigned int start, unsigned int end, unsigned int increment,
testMode mode, memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice, bool wc);
void testBandwidthQuick(unsigned int size, memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice,
bool wc);
void testBandwidthRange(unsigned int start, unsigned int end,
unsigned int increment, memcpyKind kind,
printMode printmode, memoryMode memMode,
int startDevice, int endDevice, bool wc);
void testBandwidthShmoo(memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice,
bool wc);
float testDeviceToHostTransfer(unsigned int memSize, memoryMode memMode,
bool wc);
float testHostToDeviceTransfer(unsigned int memSize, memoryMode memMode,
bool wc);
float testDeviceToDeviceTransfer(unsigned int memSize);
void printResultsReadable(unsigned int *memSizes, double *bandwidths,
unsigned int count, memcpyKind kind,
memoryMode memMode, int iNumDevs, bool wc);
void printResultsCSV(unsigned int *memSizes, double *bandwidths,
unsigned int count, memcpyKind kind, memoryMode memMode,
int iNumDevs, bool wc);
void printHelp(void);
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv) {
pArgc = &argc;
pArgv = argv;
flush_buf = (char *)malloc(FLUSH_SIZE);
// set logfile name and start logs
printf("[%s] - Starting...\n", sSDKsample);
int iRetVal = runTest(argc, (const char **)argv);
if (iRetVal < 0) {
checkCudaErrors(cudaSetDevice(0));
}
// finish
printf("%s\n", (iRetVal == 0) ? "Result = PASS" : "Result = FAIL");
printf(
"\nNOTE: The CUDA Samples are not meant for performance measurements. "
"Results may vary when GPU Boost is enabled.\n");
free(flush_buf);
exit((iRetVal == 0) ? EXIT_SUCCESS : EXIT_FAILURE);
}
///////////////////////////////////////////////////////////////////////////////
// Parse args, run the appropriate tests
///////////////////////////////////////////////////////////////////////////////
int runTest(const int argc, const char **argv) {
int start = DEFAULT_SIZE;
int end = DEFAULT_SIZE;
int startDevice = 0;
int endDevice = 0;
int increment = DEFAULT_INCREMENT;
testMode mode = QUICK_MODE;
bool htod = false;
bool dtoh = false;
bool dtod = false;
bool wc = false;
char *modeStr;
char *device = NULL;
printMode printmode = USER_READABLE;
char *memModeStr = NULL;
memoryMode memMode = PINNED;
// process command line args
if (checkCmdLineFlag(argc, argv, "help")) {
printHelp();
return 0;
}
if (checkCmdLineFlag(argc, argv, "csv")) {
printmode = CSV;
}
if (getCmdLineArgumentString(argc, argv, "memory", &memModeStr)) {
if (strcmp(memModeStr, "pageable") == 0) {
memMode = PAGEABLE;
} else if (strcmp(memModeStr, "pinned") == 0) {
memMode = PINNED;
} else {
printf("Invalid memory mode - valid modes are pageable or pinned\n");
printf("See --help for more information\n");
return -1000;
}
} else {
// default - pinned memory
memMode = PINNED;
}
if (getCmdLineArgumentString(argc, argv, "device", &device)) {
int deviceCount;
cudaError_t error_id = cudaGetDeviceCount(&deviceCount);
if (error_id != cudaSuccess) {
printf("cudaGetDeviceCount returned %d\n-> %s\n", (int)error_id,
cudaGetErrorString(error_id));
exit(EXIT_FAILURE);
}
if (deviceCount == 0) {
printf("!!!!!No devices found!!!!!\n");
return -2000;
}
if (strcmp(device, "all") == 0) {
printf(
"\n!!!!!Cumulative Bandwidth to be computed from all the devices "
"!!!!!!\n\n");
startDevice = 0;
endDevice = deviceCount - 1;
} else {
startDevice = endDevice = atoi(device);
if (startDevice >= deviceCount || startDevice < 0) {
printf(
"\n!!!!!Invalid GPU number %d given hence default gpu %d will be "
"used !!!!!\n",
startDevice, 0);
startDevice = endDevice = 0;
}
}
}
printf("Running on...\n\n");
for (int currentDevice = startDevice; currentDevice <= endDevice;
currentDevice++) {
cudaDeviceProp deviceProp;
cudaError_t error_id = cudaGetDeviceProperties(&deviceProp, currentDevice);
if (error_id == cudaSuccess) {
printf(" Device %d: %s\n", currentDevice, deviceProp.name);
if (deviceProp.computeMode == cudaComputeModeProhibited) {
fprintf(stderr,
"Error: device is running in <Compute Mode Prohibited>, no "
"threads can use ::cudaSetDevice().\n");
checkCudaErrors(cudaSetDevice(currentDevice));
exit(EXIT_FAILURE);
}
} else {
printf("cudaGetDeviceProperties returned %d\n-> %s\n", (int)error_id,
cudaGetErrorString(error_id));
checkCudaErrors(cudaSetDevice(currentDevice));
exit(EXIT_FAILURE);
}
}
if (getCmdLineArgumentString(argc, argv, "mode", &modeStr)) {
// figure out the mode
if (strcmp(modeStr, "quick") == 0) {
printf(" Quick Mode\n\n");
mode = QUICK_MODE;
} else if (strcmp(modeStr, "shmoo") == 0) {
printf(" Shmoo Mode\n\n");
mode = SHMOO_MODE;
} else if (strcmp(modeStr, "range") == 0) {
printf(" Range Mode\n\n");
mode = RANGE_MODE;
} else {
printf("Invalid mode - valid modes are quick, range, or shmoo\n");
printf("See --help for more information\n");
return -3000;
}
} else {
// default mode - quick
printf(" Quick Mode\n\n");
mode = QUICK_MODE;
}
if (checkCmdLineFlag(argc, argv, "htod")) {
htod = true;
}
if (checkCmdLineFlag(argc, argv, "dtoh")) {
dtoh = true;
}
if (checkCmdLineFlag(argc, argv, "dtod")) {
dtod = true;
}
#if CUDART_VERSION >= 2020
if (checkCmdLineFlag(argc, argv, "wc")) {
wc = true;
}
#endif
if (checkCmdLineFlag(argc, argv, "cputiming")) {
bDontUseGPUTiming = true;
}
if (!htod && !dtoh && !dtod) {
// default: All
htod = true;
dtoh = true;
dtod = true;
}
if (RANGE_MODE == mode) {
if (checkCmdLineFlag(argc, (const char **)argv, "start")) {
start = getCmdLineArgumentInt(argc, argv, "start");
if (start <= 0) {
printf("Illegal argument - start must be greater than zero\n");
return -4000;
}
} else {
printf("Must specify a starting size in range mode\n");
printf("See --help for more information\n");
return -5000;
}
if (checkCmdLineFlag(argc, (const char **)argv, "end")) {
end = getCmdLineArgumentInt(argc, argv, "end");
if (end <= 0) {
printf("Illegal argument - end must be greater than zero\n");
return -6000;
}
if (start > end) {
printf("Illegal argument - start is greater than end\n");
return -7000;
}
} else {
printf("Must specify an end size in range mode.\n");
printf("See --help for more information\n");
return -8000;
}
if (checkCmdLineFlag(argc, argv, "increment")) {
increment = getCmdLineArgumentInt(argc, argv, "increment");
if (increment <= 0) {
printf("Illegal argument - increment must be greater than zero\n");
return -9000;
}
} else {
printf("Must specify an increment in user mode\n");
printf("See --help for more information\n");
return -10000;
}
}
if (htod) {
testBandwidth((unsigned int)start, (unsigned int)end,
(unsigned int)increment, mode, HOST_TO_DEVICE, printmode,
memMode, startDevice, endDevice, wc);
}
if (dtoh) {
testBandwidth((unsigned int)start, (unsigned int)end,
(unsigned int)increment, mode, DEVICE_TO_HOST, printmode,
memMode, startDevice, endDevice, wc);
}
if (dtod) {
testBandwidth((unsigned int)start, (unsigned int)end,
(unsigned int)increment, mode, DEVICE_TO_DEVICE, printmode,
memMode, startDevice, endDevice, wc);
}
// Ensure that we reset all CUDA Devices in question
for (int nDevice = startDevice; nDevice <= endDevice; nDevice++) {
cudaSetDevice(nDevice);
}
return 0;
}
///////////////////////////////////////////////////////////////////////////////
// Run a bandwidth test
///////////////////////////////////////////////////////////////////////////////
void testBandwidth(unsigned int start, unsigned int end, unsigned int increment,
testMode mode, memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice,
bool wc) {
switch (mode) {
case QUICK_MODE:
testBandwidthQuick(DEFAULT_SIZE, kind, printmode, memMode, startDevice,
endDevice, wc);
break;
case RANGE_MODE:
testBandwidthRange(start, end, increment, kind, printmode, memMode,
startDevice, endDevice, wc);
break;
case SHMOO_MODE:
testBandwidthShmoo(kind, printmode, memMode, startDevice, endDevice, wc);
break;
default:
break;
}
}
//////////////////////////////////////////////////////////////////////
// Run a quick mode bandwidth test
//////////////////////////////////////////////////////////////////////
void testBandwidthQuick(unsigned int size, memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice,
bool wc) {
testBandwidthRange(size, size, DEFAULT_INCREMENT, kind, printmode, memMode,
startDevice, endDevice, wc);
}
///////////////////////////////////////////////////////////////////////
// Run a range mode bandwidth test
//////////////////////////////////////////////////////////////////////
void testBandwidthRange(unsigned int start, unsigned int end,
unsigned int increment, memcpyKind kind,
printMode printmode, memoryMode memMode,
int startDevice, int endDevice, bool wc) {
// count the number of copies we're going to run
unsigned int count = 1 + ((end - start) / increment);
unsigned int *memSizes = (unsigned int *)malloc(count * sizeof(unsigned int));
double *bandwidths = (double *)malloc(count * sizeof(double));
// Before calculating the cumulative bandwidth, initialize bandwidths array to
// NULL
for (unsigned int i = 0; i < count; i++) {
bandwidths[i] = 0.0;
}
// Use the device asked by the user
for (int currentDevice = startDevice; currentDevice <= endDevice;
currentDevice++) {
cudaSetDevice(currentDevice);
// run each of the copies
for (unsigned int i = 0; i < count; i++) {
memSizes[i] = start + i * increment;
switch (kind) {
case DEVICE_TO_HOST:
bandwidths[i] += testDeviceToHostTransfer(memSizes[i], memMode, wc);
break;
case HOST_TO_DEVICE:
bandwidths[i] += testHostToDeviceTransfer(memSizes[i], memMode, wc);
break;
case DEVICE_TO_DEVICE:
bandwidths[i] += testDeviceToDeviceTransfer(memSizes[i]);
break;
}
}
} // Complete the bandwidth computation on all the devices
// print results
if (printmode == CSV) {
printResultsCSV(memSizes, bandwidths, count, kind, memMode,
(1 + endDevice - startDevice), wc);
} else {
printResultsReadable(memSizes, bandwidths, count, kind, memMode,
(1 + endDevice - startDevice), wc);
}
// clean up
free(memSizes);
free(bandwidths);
}
//////////////////////////////////////////////////////////////////////////////
// Intense shmoo mode - covers a large range of values with varying increments
//////////////////////////////////////////////////////////////////////////////
void testBandwidthShmoo(memcpyKind kind, printMode printmode,
memoryMode memMode, int startDevice, int endDevice,
bool wc) {
// count the number of copies to make
unsigned int count =
1 + (SHMOO_LIMIT_20KB / SHMOO_INCREMENT_1KB) +
((SHMOO_LIMIT_50KB - SHMOO_LIMIT_20KB) / SHMOO_INCREMENT_2KB) +
((SHMOO_LIMIT_100KB - SHMOO_LIMIT_50KB) / SHMOO_INCREMENT_10KB) +
((SHMOO_LIMIT_1MB - SHMOO_LIMIT_100KB) / SHMOO_INCREMENT_100KB) +
((SHMOO_LIMIT_16MB - SHMOO_LIMIT_1MB) / SHMOO_INCREMENT_1MB) +
((SHMOO_LIMIT_32MB - SHMOO_LIMIT_16MB) / SHMOO_INCREMENT_2MB) +
((SHMOO_MEMSIZE_MAX - SHMOO_LIMIT_32MB) / SHMOO_INCREMENT_4MB);
unsigned int *memSizes = (unsigned int *)malloc(count * sizeof(unsigned int));
double *bandwidths = (double *)malloc(count * sizeof(double));
// Before calculating the cumulative bandwidth, initialize bandwidths array to
// NULL
for (unsigned int i = 0; i < count; i++) {
bandwidths[i] = 0.0;
}
// Use the device asked by the user
for (int currentDevice = startDevice; currentDevice <= endDevice;
currentDevice++) {
cudaSetDevice(currentDevice);
// Run the shmoo
int iteration = 0;
unsigned int memSize = 0;
while (memSize <= SHMOO_MEMSIZE_MAX) {
if (memSize < SHMOO_LIMIT_20KB) {
memSize += SHMOO_INCREMENT_1KB;
} else if (memSize < SHMOO_LIMIT_50KB) {
memSize += SHMOO_INCREMENT_2KB;
} else if (memSize < SHMOO_LIMIT_100KB) {
memSize += SHMOO_INCREMENT_10KB;
} else if (memSize < SHMOO_LIMIT_1MB) {
memSize += SHMOO_INCREMENT_100KB;
} else if (memSize < SHMOO_LIMIT_16MB) {
memSize += SHMOO_INCREMENT_1MB;
} else if (memSize < SHMOO_LIMIT_32MB) {
memSize += SHMOO_INCREMENT_2MB;
} else {
memSize += SHMOO_INCREMENT_4MB;
}
memSizes[iteration] = memSize;
switch (kind) {
case DEVICE_TO_HOST:
bandwidths[iteration] +=
testDeviceToHostTransfer(memSizes[iteration], memMode, wc);
break;
case HOST_TO_DEVICE:
bandwidths[iteration] +=
testHostToDeviceTransfer(memSizes[iteration], memMode, wc);
break;
case DEVICE_TO_DEVICE:
bandwidths[iteration] +=
testDeviceToDeviceTransfer(memSizes[iteration]);
break;
}
iteration++;
printf(".");
fflush(0);
}
} // Complete the bandwidth computation on all the devices
// print results
printf("\n");
if (CSV == printmode) {
printResultsCSV(memSizes, bandwidths, count, kind, memMode,
(1 + endDevice - startDevice), wc);
} else {
printResultsReadable(memSizes, bandwidths, count, kind, memMode,
(1 + endDevice - startDevice), wc);
}
// clean up
free(memSizes);
free(bandwidths);
}
///////////////////////////////////////////////////////////////////////////////
// test the bandwidth of a device to host memcopy of a specific size
///////////////////////////////////////////////////////////////////////////////
float testDeviceToHostTransfer(unsigned int memSize, memoryMode memMode,
bool wc) {
StopWatchInterface *timer = NULL;
float elapsedTimeInMs = 0.0f;
float bandwidthInGBs = 0.0f;
unsigned char *h_idata = NULL;
unsigned char *h_odata = NULL;
cudaEvent_t start, stop;
sdkCreateTimer(&timer);
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
// allocate host memory
if (PINNED == memMode) {
// pinned memory mode - use special function to get OS-pinned memory
#if CUDART_VERSION >= 2020
checkCudaErrors(cudaHostAlloc((void **)&h_idata, memSize,
(wc) ? cudaHostAllocWriteCombined : 0));
checkCudaErrors(cudaHostAlloc((void **)&h_odata, memSize,
(wc) ? cudaHostAllocWriteCombined : 0));
#else
checkCudaErrors(cudaMallocHost((void **)&h_idata, memSize));
checkCudaErrors(cudaMallocHost((void **)&h_odata, memSize));
#endif
} else {
// pageable memory mode - use malloc
h_idata = (unsigned char *)malloc(memSize);
h_odata = (unsigned char *)malloc(memSize);
if (h_idata == 0 || h_odata == 0) {
fprintf(stderr, "Not enough memory avaialable on host to run test!\n");
exit(EXIT_FAILURE);
}
}
// initialize the memory
for (unsigned int i = 0; i < memSize / sizeof(unsigned char); i++) {
h_idata[i] = (unsigned char)(i & 0xff);
}
// allocate device memory
unsigned char *d_idata;
checkCudaErrors(cudaMalloc((void **)&d_idata, memSize));
// initialize the device memory
checkCudaErrors(
cudaMemcpy(d_idata, h_idata, memSize, cudaMemcpyHostToDevice));
// copy data from GPU to Host
if (PINNED == memMode) {
if (bDontUseGPUTiming) sdkStartTimer(&timer);
checkCudaErrors(cudaEventRecord(start, 0));
for (unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++) {
checkCudaErrors(cudaMemcpyAsync(h_odata, d_idata, memSize,
cudaMemcpyDeviceToHost, 0));
}
checkCudaErrors(cudaEventRecord(stop, 0));
checkCudaErrors(cudaDeviceSynchronize());
checkCudaErrors(cudaEventElapsedTime(&elapsedTimeInMs, start, stop));
if (bDontUseGPUTiming) {
sdkStopTimer(&timer);
elapsedTimeInMs = sdkGetTimerValue(&timer);
sdkResetTimer(&timer);
}
} else {
elapsedTimeInMs = 0;
for (unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++) {
sdkStartTimer(&timer);
checkCudaErrors(
cudaMemcpy(h_odata, d_idata, memSize, cudaMemcpyDeviceToHost));
sdkStopTimer(&timer);
elapsedTimeInMs += sdkGetTimerValue(&timer);
sdkResetTimer(&timer);
memset(flush_buf, i, FLUSH_SIZE);
}
}
// calculate bandwidth in GB/s
double time_s = elapsedTimeInMs / 1e3;
bandwidthInGBs = (memSize * (float)MEMCOPY_ITERATIONS) / (double)1e9;
bandwidthInGBs = bandwidthInGBs / time_s;
// clean up memory
checkCudaErrors(cudaEventDestroy(stop));
checkCudaErrors(cudaEventDestroy(start));
sdkDeleteTimer(&timer);
if (PINNED == memMode) {
checkCudaErrors(cudaFreeHost(h_idata));
checkCudaErrors(cudaFreeHost(h_odata));
} else {
free(h_idata);
free(h_odata);
}
checkCudaErrors(cudaFree(d_idata));
return bandwidthInGBs;
}
///////////////////////////////////////////////////////////////////////////////
//! test the bandwidth of a host to device memcopy of a specific size
///////////////////////////////////////////////////////////////////////////////
float testHostToDeviceTransfer(unsigned int memSize, memoryMode memMode,
bool wc) {
StopWatchInterface *timer = NULL;
float elapsedTimeInMs = 0.0f;
float bandwidthInGBs = 0.0f;
cudaEvent_t start, stop;
sdkCreateTimer(&timer);
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
// allocate host memory
unsigned char *h_odata = NULL;
if (PINNED == memMode) {
#if CUDART_VERSION >= 2020
// pinned memory mode - use special function to get OS-pinned memory
checkCudaErrors(cudaHostAlloc((void **)&h_odata, memSize,
(wc) ? cudaHostAllocWriteCombined : 0));
#else
// pinned memory mode - use special function to get OS-pinned memory
checkCudaErrors(cudaMallocHost((void **)&h_odata, memSize));
#endif
} else {
// pageable memory mode - use malloc
h_odata = (unsigned char *)malloc(memSize);
if (h_odata == 0) {
fprintf(stderr, "Not enough memory available on host to run test!\n");
exit(EXIT_FAILURE);
}
}
unsigned char *h_cacheClear1 = (unsigned char *)malloc(CACHE_CLEAR_SIZE);
unsigned char *h_cacheClear2 = (unsigned char *)malloc(CACHE_CLEAR_SIZE);
if (h_cacheClear1 == 0 || h_cacheClear2 == 0) {
fprintf(stderr, "Not enough memory available on host to run test!\n");
exit(EXIT_FAILURE);
}
// initialize the memory
for (unsigned int i = 0; i < memSize / sizeof(unsigned char); i++) {
h_odata[i] = (unsigned char)(i & 0xff);
}
for (unsigned int i = 0; i < CACHE_CLEAR_SIZE / sizeof(unsigned char); i++) {
h_cacheClear1[i] = (unsigned char)(i & 0xff);
h_cacheClear2[i] = (unsigned char)(0xff - (i & 0xff));
}
// allocate device memory
unsigned char *d_idata;
checkCudaErrors(cudaMalloc((void **)&d_idata, memSize));
// copy host memory to device memory
if (PINNED == memMode) {
if (bDontUseGPUTiming) sdkStartTimer(&timer);
checkCudaErrors(cudaEventRecord(start, 0));
for (unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++) {
checkCudaErrors(cudaMemcpyAsync(d_idata, h_odata, memSize,
cudaMemcpyHostToDevice, 0));
}
checkCudaErrors(cudaEventRecord(stop, 0));
checkCudaErrors(cudaDeviceSynchronize());
checkCudaErrors(cudaEventElapsedTime(&elapsedTimeInMs, start, stop));
if (bDontUseGPUTiming) {
sdkStopTimer(&timer);
elapsedTimeInMs = sdkGetTimerValue(&timer);
sdkResetTimer(&timer);
}
} else {
elapsedTimeInMs = 0;
for (unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++) {
sdkStartTimer(&timer);
checkCudaErrors(
cudaMemcpy(d_idata, h_odata, memSize, cudaMemcpyHostToDevice));
sdkStopTimer(&timer);
elapsedTimeInMs += sdkGetTimerValue(&timer);
sdkResetTimer(&timer);
memset(flush_buf, i, FLUSH_SIZE);
}
}
// calculate bandwidth in GB/s
double time_s = elapsedTimeInMs / 1e3;
bandwidthInGBs = (memSize * (float)MEMCOPY_ITERATIONS) / (double)1e9;
bandwidthInGBs = bandwidthInGBs / time_s;
// clean up memory
checkCudaErrors(cudaEventDestroy(stop));
checkCudaErrors(cudaEventDestroy(start));
sdkDeleteTimer(&timer);
if (PINNED == memMode) {
checkCudaErrors(cudaFreeHost(h_odata));
} else {
free(h_odata);
}
free(h_cacheClear1);
free(h_cacheClear2);
checkCudaErrors(cudaFree(d_idata));
return bandwidthInGBs;
}
///////////////////////////////////////////////////////////////////////////////
//! test the bandwidth of a device to device memcopy of a specific size
///////////////////////////////////////////////////////////////////////////////
float testDeviceToDeviceTransfer(unsigned int memSize) {
StopWatchInterface *timer = NULL;
float elapsedTimeInMs = 0.0f;
float bandwidthInGBs = 0.0f;
cudaEvent_t start, stop;
sdkCreateTimer(&timer);
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
// allocate host memory
unsigned char *h_idata = (unsigned char *)malloc(memSize);
if (h_idata == 0) {
fprintf(stderr, "Not enough memory avaialable on host to run test!\n");
exit(EXIT_FAILURE);
}
// initialize the host memory
for (unsigned int i = 0; i < memSize / sizeof(unsigned char); i++) {
h_idata[i] = (unsigned char)(i & 0xff);
}
// allocate device memory
unsigned char *d_idata;
checkCudaErrors(cudaMalloc((void **)&d_idata, memSize));
unsigned char *d_odata;
checkCudaErrors(cudaMalloc((void **)&d_odata, memSize));
// initialize memory
checkCudaErrors(
cudaMemcpy(d_idata, h_idata, memSize, cudaMemcpyHostToDevice));
// run the memcopy
sdkStartTimer(&timer);
checkCudaErrors(cudaEventRecord(start, 0));
for (unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++) {
checkCudaErrors(
cudaMemcpy(d_odata, d_idata, memSize, cudaMemcpyDeviceToDevice));
}
checkCudaErrors(cudaEventRecord(stop, 0));
// Since device to device memory copies are non-blocking,
// cudaDeviceSynchronize() is required in order to get
// proper timing.
checkCudaErrors(cudaDeviceSynchronize());
// get the total elapsed time in ms
sdkStopTimer(&timer);
checkCudaErrors(cudaEventElapsedTime(&elapsedTimeInMs, start, stop));
if (bDontUseGPUTiming) {
elapsedTimeInMs = sdkGetTimerValue(&timer);
}
// calculate bandwidth in GB/s
double time_s = elapsedTimeInMs / 1e3;
bandwidthInGBs = (2.0f * memSize * (float)MEMCOPY_ITERATIONS) / (double)1e9;
bandwidthInGBs = bandwidthInGBs / time_s;
// clean up memory
sdkDeleteTimer(&timer);
free(h_idata);
checkCudaErrors(cudaEventDestroy(stop));
checkCudaErrors(cudaEventDestroy(start));
checkCudaErrors(cudaFree(d_idata));
checkCudaErrors(cudaFree(d_odata));
return bandwidthInGBs;
}
/////////////////////////////////////////////////////////
// print results in an easily read format
////////////////////////////////////////////////////////
void printResultsReadable(unsigned int *memSizes, double *bandwidths,
unsigned int count, memcpyKind kind,
memoryMode memMode, int iNumDevs, bool wc) {
printf(" %s Bandwidth, %i Device(s)\n", sMemoryCopyKind[kind], iNumDevs);
printf(" %s Memory Transfers\n", sMemoryMode[memMode]);
if (wc) {
printf(" Write-Combined Memory Writes are Enabled");
}
printf(" Transfer Size (Bytes)\tBandwidth(GB/s)\n");
unsigned int i;
for (i = 0; i < (count - 1); i++) {
printf(" %u\t\t\t%s%.1f\n", memSizes[i],
(memSizes[i] < 10000) ? "\t" : "", bandwidths[i]);
}
printf(" %u\t\t\t%s%.1f\n\n", memSizes[i],
(memSizes[i] < 10000) ? "\t" : "", bandwidths[i]);
}
///////////////////////////////////////////////////////////////////////////
// print results in a database format
///////////////////////////////////////////////////////////////////////////
void printResultsCSV(unsigned int *memSizes, double *bandwidths,
unsigned int count, memcpyKind kind, memoryMode memMode,
int iNumDevs, bool wc) {
std::string sConfig;
// log config information
if (kind == DEVICE_TO_DEVICE) {
sConfig += "D2D";
} else {
if (kind == DEVICE_TO_HOST) {
sConfig += "D2H";
} else if (kind == HOST_TO_DEVICE) {
sConfig += "H2D";
}
if (memMode == PAGEABLE) {
sConfig += "-Paged";
} else if (memMode == PINNED) {
sConfig += "-Pinned";
if (wc) {
sConfig += "-WriteCombined";
}
}
}
unsigned int i;
double dSeconds = 0.0;
for (i = 0; i < count; i++) {
dSeconds = (double)memSizes[i] / (bandwidths[i] * (double)(1 << 20));
printf(
"bandwidthTest-%s, Bandwidth = %.1f GB/s, Time = %.5f s, Size = %u "
"bytes, NumDevsUsed = %d\n",
sConfig.c_str(), bandwidths[i], dSeconds, memSizes[i], iNumDevs);
}
}
///////////////////////////////////////////////////////////////////////////
// Print help screen
///////////////////////////////////////////////////////////////////////////
void printHelp(void) {
printf("Usage: bandwidthTest [OPTION]...\n");
printf(
"Test the bandwidth for device to host, host to device, and device to "
"device transfers\n");
printf("\n");
printf(
"Example: measure the bandwidth of device to host pinned memory copies "
"in the range 1024 Bytes to 102400 Bytes in 1024 Byte increments\n");
printf(
"./bandwidthTest --memory=pinned --mode=range --start=1024 --end=102400 "
"--increment=1024 --dtoh\n");
printf("\n");
printf("Options:\n");
printf("--help\tDisplay this help menu\n");
printf("--csv\tPrint results as a CSV\n");
printf("--device=[deviceno]\tSpecify the device device to be used\n");
printf(" all - compute cumulative bandwidth on all the devices\n");
printf(" 0,1,2,...,n - Specify any particular device to be used\n");
printf("--memory=[MEMMODE]\tSpecify which memory mode to use\n");
printf(" pageable - pageable memory\n");
printf(" pinned - non-pageable system memory\n");
printf("--mode=[MODE]\tSpecify the mode to use\n");
printf(" quick - performs a quick measurement\n");
printf(" range - measures a user-specified range of values\n");
printf(" shmoo - performs an intense shmoo of a large range of values\n");
printf("--htod\tMeasure host to device transfers\n");
printf("--dtoh\tMeasure device to host transfers\n");
printf("--dtod\tMeasure device to device transfers\n");
#if CUDART_VERSION >= 2020
printf("--wc\tAllocate pinned memory as write-combined\n");
#endif
printf("--cputiming\tForce CPU-based timing always\n");
printf("Range mode options\n");
printf("--start=[SIZE]\tStarting transfer size in bytes\n");
printf("--end=[SIZE]\tEnding transfer size in bytes\n");
printf("--increment=[SIZE]\tIncrement size in bytes\n");
}

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bandwidthTest", "bandwidthTest_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>deviceQuery_vs2010</RootNamespace> <RootNamespace>bandwidthTest_vs2012</RootNamespace>
<ProjectName>deviceQuery</ProjectName> <ProjectName>bandwidthTest</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v110</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -58,10 +59,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/deviceQuery.exe</OutputFile> <OutputFile>$(OutDir)/bandwidthTest.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -96,11 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<ClCompile Include="deviceQuery.cpp" /> <CudaCompile Include="bandwidthTest.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bandwidthTest", "bandwidthTest_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>matrixMul_vs2010</RootNamespace> <RootNamespace>bandwidthTest_vs2013</RootNamespace>
<ProjectName>matrixMul</ProjectName> <ProjectName>bandwidthTest</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -58,10 +59,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/matrixMul.exe</OutputFile> <OutputFile>$(OutDir)/bandwidthTest.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -96,11 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="matrixMul.cu" /> <CudaCompile Include="bandwidthTest.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bandwidthTest", "bandwidthTest_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>warpAggregatedAtomicsCG_vs2010</RootNamespace> <RootNamespace>bandwidthTest_vs2015</RootNamespace>
<ProjectName>warpAggregatedAtomicsCG</ProjectName> <ProjectName>bandwidthTest</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v140</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -58,10 +59,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/warpAggregatedAtomicsCG.exe</OutputFile> <OutputFile>$(OutDir)/bandwidthTest.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -96,11 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="warpAggregatedAtomicsCG.cu" /> <CudaCompile Include="bandwidthTest.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bandwidthTest", "bandwidthTest_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,108 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>bandwidthTest_vs2017</RootNamespace>
<ProjectName>bandwidthTest</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</PlatformToolset>
<WindowsTargetPlatformVersion>10.0.15063.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/bandwidthTest.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="bandwidthTest.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,306 @@
################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
################################################################################
#
# Makefile project only supported on Mac OS X and Linux Platforms)
#
################################################################################
# Location of the CUDA Toolkit
CUDA_PATH ?= /usr/local/cuda
##############################
# start deprecated interface #
##############################
ifeq ($(x86_64),1)
$(info WARNING - x86_64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=x86_64 instead)
TARGET_ARCH ?= x86_64
endif
ifeq ($(ARMv7),1)
$(info WARNING - ARMv7 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=armv7l instead)
TARGET_ARCH ?= armv7l
endif
ifeq ($(aarch64),1)
$(info WARNING - aarch64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=aarch64 instead)
TARGET_ARCH ?= aarch64
endif
ifeq ($(ppc64le),1)
$(info WARNING - ppc64le variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=ppc64le instead)
TARGET_ARCH ?= ppc64le
endif
ifneq ($(GCC),)
$(info WARNING - GCC variable has been deprecated)
$(info WARNING - please use HOST_COMPILER=$(GCC) instead)
HOST_COMPILER ?= $(GCC)
endif
ifneq ($(abi),)
$(error ERROR - abi variable has been removed)
endif
############################
# end deprecated interface #
############################
# architecture
HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif
endif
# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
TARGET_ARCH = armv7l
endif
# operating system
HOST_OS := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
TARGET_OS ?= $(HOST_OS)
ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
$(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
endif
# host compiler
ifeq ($(TARGET_OS),darwin)
ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
HOST_COMPILER ?= clang++
endif
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
ifeq ($(TARGET_OS),linux)
HOST_COMPILER ?= arm-linux-gnueabihf-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
else ifeq ($(TARGET_OS),android)
HOST_COMPILER ?= arm-linux-androideabi-g++
endif
else ifeq ($(TARGET_ARCH),aarch64)
ifeq ($(TARGET_OS), linux)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++
endif
else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif
endif
HOST_COMPILER ?= g++
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
# internal flags
NVCCFLAGS := -m${TARGET_SIZE}
CCFLAGS :=
LDFLAGS :=
# build flags
ifeq ($(TARGET_OS),darwin)
LDFLAGS += -rpath $(CUDA_PATH)/lib
CCFLAGS += -arch $(HOST_ARCH)
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
CCFLAGS += -mfloat-abi=hard
else ifeq ($(TARGET_OS),android)
LDFLAGS += -pie
CCFLAGS += -fpie -fpic -fexceptions
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
endif
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
else ifeq ($(TARGET_ARCH),ppc64le)
CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
endif
# Debug build flags
ifeq ($(dbg),1)
NVCCFLAGS += -g -G
BUILD_TYPE := debug
else
BUILD_TYPE := release
endif
ALL_CCFLAGS :=
ALL_CCFLAGS += $(NVCCFLAGS)
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
SAMPLE_ENABLED := 1
ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
# Common includes and paths for CUDA
INCLUDES := -I../../Common
LIBRARIES :=
################################################################################
# Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75
else
SMS ?= 30 35 37 50 52 60 61 70 75
endif
ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ifeq ($(GENCODE_FLAGS),)
# Generate SASS code for each SM architecture listed in $(SMS)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
HIGHEST_SM := $(lastword $(sort $(SMS)))
ifneq ($(HIGHEST_SM),)
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif
LIBRARIES += -lcublas_static -lcublasLt_static -lcusparse_static -lculibos
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: conjugateGradientCudaGraphs
check.deps:
ifeq ($(SAMPLE_ENABLED),0)
@echo "Sample will be waived due to the above missing dependencies"
else
@echo "Sample is ready - all dependencies have been met"
endif
conjugateGradientCudaGraphs.o:conjugateGradientCudaGraphs.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
conjugateGradientCudaGraphs: conjugateGradientCudaGraphs.o
$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
run: build
$(EXEC) ./conjugateGradientCudaGraphs
clean:
rm -f conjugateGradientCudaGraphs conjugateGradientCudaGraphs.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/conjugateGradientCudaGraphs
clobber: clean

View File

@@ -0,0 +1,86 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>conjugateGradientCudaGraphs</name>
<cuda_api_list>
<toolkit>cudaStreamBeginCapture</toolkit>
<toolkit>cudaStreamEndCapture</toolkit>
<toolkit>cudaGraphCreate</toolkit>
<toolkit>cudaGraphLaunch</toolkit>
<toolkit>cudaGraphInstantiate</toolkit>
<toolkit>cudaGraphExecDestroy</toolkit>
<toolkit>cudaGraphDestroy</toolkit>
</cuda_api_list>
<description><![CDATA[This sample implements a conjugate gradient solver on GPU using CUBLAS and CUSPARSE library calls captured and called using CUDA Graph APIs.]]></description>
<devicecompilation>whole</devicecompilation>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="advanced">Linear Algebra</concept>
<concept level="advanced">CUBLAS Library</concept>
<concept level="advanced">CUSPARSE Library</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>CUBLAS</keyword>
<keyword>CUSPARSE</keyword>
<keyword>Sparse Matrix</keyword>
</keywords>
<libraries>
<library>cublas_static</library>
<library>cublasLt_static</library>
<library>cusparse_static</library>
<library>culibos</library>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>conjugateGradientCudaGraphs.cu</primary_file>
<required_dependencies>
<dependency>CUBLAS</dependency>
<dependency>CUSPARSE</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Advanced Topics</scope>
<scope>3:Linear Algebra</scope>
<scope>1:CUDA Graphs</scope>
</scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch>
<sm-arch>sm52</sm-arch>
<sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<platform>windows7</platform>
</env>
<env>
<arch>x86_64</arch>
<platform>macosx</platform>
</env>
<env>
<arch>arm</arch>
</env>
<env>
<arch>ppc64le</arch>
<platform>linux</platform>
</env>
</supported_envs>
<supported_sm_architectures>
<include>all</include>
</supported_sm_architectures>
<title>Conjugate Gradient using Cuda Graphs</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,98 @@
# conjugateGradientCudaGraphs - Conjugate Gradient using Cuda Graphs
## Description
This sample implements a conjugate gradient solver on GPU using CUBLAS and CUSPARSE library calls captured and called using CUDA Graph APIs.
## Key Concepts
Linear Algebra, CUBLAS Library, CUSPARSE Library
## Supported SM Architectures
[SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows, MacOSX
## Supported CPU Architecture
x86_64, ppc64le, armv7l
## CUDA APIs involved
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaStreamBeginCapture, cudaStreamEndCapture, cudaGraphCreate, cudaGraphLaunch, cudaGraphInstantiate, cudaGraphExecDestroy, cudaGraphDestroy
## Dependencies needed to build/run
[CUBLAS](../../README.md#cublas), [CUSPARSE](../../README.md#cusparse)
## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
## Build and Run
### Windows
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
```
*_vs<version>.sln - for Visual Studio <version>
```
Each individual sample has its own set of solution files in its directory:
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
### Linux
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, armv7l.
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=armv7l` <br/>
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
```
$ make SMS="50 60"
```
* **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=g++
```
### Mac
The Mac samples are built using makefiles. To use the makefiles, change directory into the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where "A B ..." is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use SMS="50 60".
```
$ make SMS="A B ..."
```
* **HOST_COMPILER=<host_compiler>** - override the default clang host compiler. See the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=clang
```
## References (for more details)

View File

@@ -0,0 +1,466 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
/*
* This sample implements a conjugate gradient solver on GPU
* using CUBLAS and CUSPARSE with CUDA Graphs
*
*/
// includes, system
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
/* Using updated (v2) interfaces to cublas */
#include <cublas_v2.h>
#include <cuda_runtime.h>
#include <cusparse.h>
#include <cooperative_groups.h>
// Utilities and system includes
#include <helper_cuda.h> // helper function CUDA error checking and initialization
#include <helper_functions.h> // helper for shared functions common to CUDA Samples
namespace cg = cooperative_groups;
const char *sSDKname = "conjugateGradientCudaGraphs";
#ifndef WITH_GRAPH
#define WITH_GRAPH 1
#endif
/* genTridiag: generate a random tridiagonal symmetric matrix */
void genTridiag(int *I, int *J, float *val, int N, int nz) {
I[0] = 0, J[0] = 0, J[1] = 1;
val[0] = (float)rand() / RAND_MAX + 10.0f;
val[1] = (float)rand() / RAND_MAX;
int start;
for (int i = 1; i < N; i++) {
if (i > 1) {
I[i] = I[i - 1] + 3;
} else {
I[1] = 2;
}
start = (i - 1) * 3 + 2;
J[start] = i - 1;
J[start + 1] = i;
if (i < N - 1) {
J[start + 2] = i + 1;
}
val[start] = val[start - 1];
val[start + 1] = (float)rand() / RAND_MAX + 10.0f;
if (i < N - 1) {
val[start + 2] = (float)rand() / RAND_MAX;
}
}
I[N] = nz;
}
__global__ void initVectors(float *rhs, float *x, int N) {
size_t gid = blockIdx.x * blockDim.x + threadIdx.x;
for (size_t i = gid; i < N; i += gridDim.x * blockDim.x) {
rhs[i] = 1.0;
x[i] = 0.0;
}
}
__global__ void gpuDotProduct(float *vecA, float *vecB, float *result,
int size) {
cg::thread_block cta = cg::this_thread_block();
int gid = blockIdx.x * blockDim.x + threadIdx.x;
extern __shared__ double tmp[];
double temp_sum = 0.0;
for (int i = gid; i < size; i += gridDim.x * blockDim.x) {
temp_sum += (double)(vecA[i] * vecB[i]);
}
tmp[cta.thread_rank()] = temp_sum;
cg::sync(cta);
cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
double beta = temp_sum;
double temp;
for (int i = tile32.size() / 2; i > 0; i >>= 1) {
if (tile32.thread_rank() < i) {
temp = tmp[cta.thread_rank() + i];
beta += temp;
tmp[cta.thread_rank()] = beta;
}
cg::sync(tile32);
}
cg::sync(cta);
if (cta.thread_rank() == 0) {
beta = 0.0;
for (int i = 0; i < cta.size(); i += tile32.size()) {
beta += tmp[i];
}
atomicAdd(result, (float)beta);
}
}
__global__ void gpuSpMV(int *I, int *J, float *val, int nnz, int num_rows,
float alpha, float *inputVecX, float *outputVecY) {
size_t gid = blockIdx.x * blockDim.x + threadIdx.x;
for (size_t i = gid; i < num_rows; i += blockDim.x * gridDim.x) {
int row_elem = I[i];
int next_row_elem = I[i + 1];
int num_elems_this_row = next_row_elem - row_elem;
float output = 0.0;
for (int j = 0; j < num_elems_this_row; j++) {
output += alpha * val[row_elem + j] * inputVecX[J[row_elem + j]];
}
outputVecY[i] = output;
}
}
__global__ void r1_div_x(float *r1, float *r0, float *b) {
int gid = blockIdx.x * blockDim.x + threadIdx.x;
if (gid == 0) {
b[0] = r1[0] / r0[0];
}
}
__global__ void a_minus(float *a, float *na) {
int gid = blockIdx.x * blockDim.x + threadIdx.x;
if (gid == 0) {
na[0] = -(a[0]);
}
}
int main(int argc, char **argv) {
int N = 0, nz = 0, *I = NULL, *J = NULL;
float *val = NULL;
const float tol = 1e-5f;
const int max_iter = 10000;
float *x;
float *rhs;
float r1;
int *d_col, *d_row;
float *d_val, *d_x;
float *d_r, *d_p, *d_Ax;
int k;
float alpha, beta, alpham1;
cudaStream_t stream1, streamForGraph;
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
if (devID < 0) {
printf("exiting...\n");
exit(EXIT_SUCCESS);
}
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
// Statistics about the GPU device
printf(
"> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
/* Generate a random tridiagonal symmetric matrix in CSR format */
N = 1048576;
nz = (N - 2) * 3 + 4;
I = (int *)malloc(sizeof(int) * (N + 1));
J = (int *)malloc(sizeof(int) * nz);
val = (float *)malloc(sizeof(float) * nz);
genTridiag(I, J, val, N, nz);
x = (float *)malloc(sizeof(float) * N);
rhs = (float *)malloc(sizeof(float) * N);
for (int i = 0; i < N; i++) {
rhs[i] = 1.0;
x[i] = 0.0;
}
/* Get handle to the CUBLAS context */
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
/* Get handle to the CUSPARSE context */
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
checkCudaErrors(cudaStreamCreate(&stream1));
checkCudaErrors(cudaMalloc((void **)&d_col, nz * sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_row, (N + 1) * sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_val, nz * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_x, N * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_r, N * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_p, N * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_Ax, N * sizeof(float)));
float *d_r1, *d_r0, *d_dot, *d_a, *d_na, *d_b;
checkCudaErrors(cudaMalloc((void **)&d_r1, sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_r0, sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_dot, sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_a, sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_na, sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_b, sizeof(float)));
cusparseMatDescr_t descr = 0;
checkCudaErrors(cusparseCreateMatDescr(&descr));
checkCudaErrors(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_GENERAL));
checkCudaErrors(cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO));
int numBlocks = 0, blockSize = 0, numBlocks2 = 0, blockSize2 = 0;
checkCudaErrors(
cudaOccupancyMaxPotentialBlockSize(&numBlocks, &blockSize, initVectors));
checkCudaErrors(cudaMemcpyAsync(d_col, J, nz * sizeof(int),
cudaMemcpyHostToDevice, stream1));
checkCudaErrors(cudaMemcpyAsync(d_row, I, (N + 1) * sizeof(int),
cudaMemcpyHostToDevice, stream1));
checkCudaErrors(cudaMemcpyAsync(d_val, val, nz * sizeof(float),
cudaMemcpyHostToDevice, stream1));
initVectors<<<numBlocks, blockSize, 0, stream1>>>(d_r, d_x, N);
checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(&numBlocks2, &blockSize2,
gpuSpMV));
checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(&numBlocks, &blockSize,
gpuDotProduct));
alpha = 1.0;
alpham1 = -1.0;
beta = 0.0;
checkCudaErrors(cusparseSetStream(cusparseHandle, stream1));
checkCudaErrors(
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz,
&alpha, descr, d_val, d_row, d_col, d_x, &beta, d_Ax));
checkCudaErrors(cublasSetStream(cublasHandle, stream1));
checkCudaErrors(cublasSaxpy(cublasHandle, N, &alpham1, d_Ax, 1, d_r, 1));
checkCudaErrors(
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_DEVICE));
checkCudaErrors(cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, d_r1));
k = 1;
// First Iteration when k=1 starts
checkCudaErrors(cublasScopy(cublasHandle, N, d_r, 1, d_p, 1));
checkCudaErrors(
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz,
&alpha, descr, d_val, d_row, d_col, d_p, &beta, d_Ax));
checkCudaErrors(cublasSdot(cublasHandle, N, d_p, 1, d_Ax, 1, d_dot));
r1_div_x<<<1, 1, 0, stream1>>>(d_r1, d_dot, d_a);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_a, d_p, 1, d_x, 1));
a_minus<<<1, 1, 0, stream1>>>(d_a, d_na);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_na, d_Ax, 1, d_r, 1));
checkCudaErrors(cudaMemcpyAsync(d_r0, d_r1, sizeof(float),
cudaMemcpyDeviceToDevice, stream1));
checkCudaErrors(cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, d_r1));
checkCudaErrors(cudaMemcpyAsync(&r1, d_r1, sizeof(float),
cudaMemcpyDeviceToHost, stream1));
checkCudaErrors(cudaStreamSynchronize(stream1));
printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
// First Iteration when k=1 ends
k++;
#if WITH_GRAPH
cudaGraph_t initGraph;
checkCudaErrors(cudaStreamCreate(&streamForGraph));
checkCudaErrors(cublasSetStream(cublasHandle, stream1));
checkCudaErrors(cusparseSetStream(cusparseHandle, stream1));
checkCudaErrors(cudaStreamBeginCapture(stream1, cudaStreamCaptureModeGlobal));
r1_div_x<<<1, 1, 0, stream1>>>(d_r1, d_r0, d_b);
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_DEVICE);
checkCudaErrors(cublasSscal(cublasHandle, N, d_b, d_p, 1));
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_HOST);
checkCudaErrors(cublasSaxpy(cublasHandle, N, &alpha, d_r, 1, d_p, 1));
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_DEVICE);
#if 0 // Use cusparseScsrmv API when it is cuda graph compliant
checkCudaErrors(
cusparseSetPointerMode(cusparseHandle, CUSPARSE_POINTER_MODE_HOST));
checkCudaErrors(
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz,
&alpha, descr, d_val, d_row, d_col, d_p, &beta, d_Ax));
#else
gpuSpMV<<<numBlocks2, blockSize2, 0, stream1>>>(d_row, d_col, d_val, nz,
N, alpha, d_p, d_Ax);
#endif
checkCudaErrors(cudaMemsetAsync(d_dot, 0, sizeof(float), stream1));
// Use cublasSdot API when it is cuda graph compliant.
// checkCudaErrors(cublasSdot(cublasHandle, N, d_p, 1, d_Ax, 1, d_dot));
gpuDotProduct<<<numBlocks, blockSize, blockSize * sizeof(double), stream1>>>(
d_p, d_Ax, d_dot, N);
r1_div_x<<<1, 1, 0, stream1>>>(d_r1, d_dot, d_a);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_a, d_p, 1, d_x, 1));
a_minus<<<1, 1, 0, stream1>>>(d_a, d_na);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_na, d_Ax, 1, d_r, 1));
checkCudaErrors(cudaMemcpyAsync(d_r0, d_r1, sizeof(float),
cudaMemcpyDeviceToDevice, stream1));
checkCudaErrors(cudaMemsetAsync(d_r1, 0, sizeof(float), stream1));
// Use cublasSdot API when it is cuda graph compliant.
// checkCudaErrors(cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, d_r1));
gpuDotProduct<<<numBlocks, blockSize, blockSize * sizeof(double), stream1>>>(
d_r, d_r, d_r1, N);
checkCudaErrors(cudaMemcpyAsync((float *)&r1, d_r1, sizeof(float),
cudaMemcpyDeviceToHost, stream1));
checkCudaErrors(cudaStreamEndCapture(stream1, &initGraph));
cudaGraphExec_t graphExec;
checkCudaErrors(cudaGraphInstantiate(&graphExec, initGraph, NULL, NULL, 0));
#endif
checkCudaErrors(cublasSetStream(cublasHandle, stream1));
checkCudaErrors(cusparseSetStream(cusparseHandle, stream1));
while (r1 > tol * tol && k <= max_iter) {
#if WITH_GRAPH
checkCudaErrors(cudaGraphLaunch(graphExec, streamForGraph));
checkCudaErrors(cudaStreamSynchronize(streamForGraph));
#else
r1_div_x<<<1, 1, 0, stream1>>>(d_r1, d_r0, d_b);
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_DEVICE);
checkCudaErrors(cublasSscal(cublasHandle, N, d_b, d_p, 1));
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_HOST);
checkCudaErrors(cublasSaxpy(cublasHandle, N, &alpha, d_r, 1, d_p, 1));
checkCudaErrors(cusparseScsrmv(
cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha,
descr, d_val, d_row, d_col, d_p, &beta, d_Ax));
cublasSetPointerMode(cublasHandle, CUBLAS_POINTER_MODE_DEVICE);
checkCudaErrors(cublasSdot(cublasHandle, N, d_p, 1, d_Ax, 1, d_dot));
r1_div_x<<<1, 1, 0, stream1>>>(d_r1, d_dot, d_a);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_a, d_p, 1, d_x, 1));
a_minus<<<1, 1, 0, stream1>>>(d_a, d_na);
checkCudaErrors(cublasSaxpy(cublasHandle, N, d_na, d_Ax, 1, d_r, 1));
checkCudaErrors(cudaMemcpyAsync(d_r0, d_r1, sizeof(float),
cudaMemcpyDeviceToDevice, stream1));
checkCudaErrors(cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, d_r1));
checkCudaErrors(cudaMemcpyAsync((float *)&r1, d_r1, sizeof(float),
cudaMemcpyDeviceToHost, stream1));
checkCudaErrors(cudaStreamSynchronize(stream1));
#endif
printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
k++;
}
#if WITH_GRAPH
checkCudaErrors(cudaMemcpyAsync(x, d_x, N * sizeof(float),
cudaMemcpyDeviceToHost, streamForGraph));
checkCudaErrors(cudaStreamSynchronize(streamForGraph));
#else
checkCudaErrors(cudaMemcpyAsync(x, d_x, N * sizeof(float),
cudaMemcpyDeviceToHost, stream1));
checkCudaErrors(cudaStreamSynchronize(stream1));
#endif
float rsum, diff, err = 0.0;
for (int i = 0; i < N; i++) {
rsum = 0.0;
for (int j = I[i]; j < I[i + 1]; j++) {
rsum += val[j] * x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err) {
err = diff;
}
}
#if WITH_GRAPH
checkCudaErrors(cudaGraphExecDestroy(graphExec));
checkCudaErrors(cudaGraphDestroy(initGraph));
checkCudaErrors(cudaStreamDestroy(streamForGraph));
#endif
checkCudaErrors(cudaStreamDestroy(stream1));
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
free(I);
free(J);
free(val);
free(x);
free(rhs);
cudaFree(d_col);
cudaFree(d_row);
cudaFree(d_val);
cudaFree(d_x);
cudaFree(d_r);
cudaFree(d_p);
cudaFree(d_Ax);
printf("Test Summary: Error amount = %f\n", err);
exit((k <= max_iter) ? 0 : 1);
}

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 11.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2010 # Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientMultiBlockCG", "conjugateGradientMultiBlockCG_vs2010.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientCudaGraphs", "conjugateGradientCudaGraphs_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>simpleCUFFT_vs2010</RootNamespace> <RootNamespace>conjugateGradientCudaGraphs_vs2012</RootNamespace>
<ProjectName>simpleCUFFT</ProjectName> <ProjectName>conjugateGradientCudaGraphs</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v110</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -56,12 +57,12 @@
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cufft.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cublas.lib;cusparse.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/simpleCUFFT.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientCudaGraphs.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -96,11 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="simpleCUFFT.cu" /> <CudaCompile Include="conjugateGradientCudaGraphs.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientCudaGraphs", "conjugateGradientCudaGraphs_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,107 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientCudaGraphs_vs2013</RootNamespace>
<ProjectName>conjugateGradientCudaGraphs</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cublas.lib;cusparse.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientCudaGraphs.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientCudaGraphs.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientCudaGraphs", "conjugateGradientCudaGraphs_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,107 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientCudaGraphs_vs2015</RootNamespace>
<ProjectName>conjugateGradientCudaGraphs</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v140</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cublas.lib;cusparse.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientCudaGraphs.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientCudaGraphs.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientCudaGraphs", "conjugateGradientCudaGraphs_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,108 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientCudaGraphs_vs2017</RootNamespace>
<ProjectName>conjugateGradientCudaGraphs</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</PlatformToolset>
<WindowsTargetPlatformVersion>10.0.15063.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cublas.lib;cusparse.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientCudaGraphs.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientCudaGraphs.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
@@ -266,7 +264,11 @@ LIBRARIES :=
################################################################################ ################################################################################
# Gencode arguments # Gencode arguments
SMS ?= 60 61 70 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 60 61 70 72 75
else
SMS ?= 60 61 70 75
endif
ifeq ($(SMS),) ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<) $(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)

View File

@@ -42,6 +42,8 @@
<sm-arch>sm60</sm-arch> <sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch> <sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,7 +10,7 @@ Unified Memory, Linear Algebra, Cooperative Groups, MultiBlock Cooperative Group
## Supported SM Architectures ## Supported SM Architectures
[SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
@@ -27,7 +27,7 @@ x86_64, ppc64le
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Make sure the dependencies mentioned in [Dependencies]() section above are installed. Make sure the dependencies mentioned in [Dependencies]() section above are installed.
## Build and Run ## Build and Run

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -34,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -63,7 +63,7 @@
<OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -104,6 +104,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
@@ -266,7 +264,11 @@ LIBRARIES :=
################################################################################ ################################################################################
# Gencode arguments # Gencode arguments
SMS ?= 60 61 70 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 60 61 70 72 75
else
SMS ?= 60 61 70 75
endif
ifeq ($(SMS),) ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<) $(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)

View File

@@ -49,6 +49,8 @@
<sm-arch>sm60</sm-arch> <sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch> <sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>
@@ -58,6 +60,9 @@
<arch>ppc64le</arch> <arch>ppc64le</arch>
<platform>linux</platform> <platform>linux</platform>
</env> </env>
<env>
<platform>windows</platform>
</env>
</supported_envs> </supported_envs>
<supported_sm_architectures> <supported_sm_architectures>
<from>6.0</from> <from>6.0</from>

View File

@@ -10,11 +10,11 @@ Unified Memory, Linear Algebra, Cooperative Groups, MultiDevice Cooperative Grou
## Supported SM Architectures ## Supported SM Architectures
[SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -30,11 +30,21 @@ cudaMemAdvise, cudaMemPrefetchAsync, cudaLaunchCooperativeKernelMultiDevice, cud
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Make sure the dependencies mentioned in [Dependencies]() section above are installed. Make sure the dependencies mentioned in [Dependencies]() section above are installed.
## Build and Run ## Build and Run
### Windows
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
```
*_vs<version>.sln - for Visual Studio <version>
```
Each individual sample has its own set of solution files in its directory:
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
### Linux ### Linux
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make: The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
``` ```

View File

@@ -414,7 +414,8 @@ void getIdenticalGPUs(int num_of_gpus, std::set<int> &identicalGPUs) {
deviceProp.minor != maxMajorMinor[1]) { deviceProp.minor != maxMajorMinor[1]) {
identicalGPUs.erase(it); identicalGPUs.erase(it);
} }
if (!deviceProp.cooperativeMultiDeviceLaunch) { if (!deviceProp.cooperativeMultiDeviceLaunch ||
!deviceProp.concurrentManagedAccess) {
identicalGPUs.erase(it); identicalGPUs.erase(it);
} }
it++; it++;
@@ -449,7 +450,7 @@ int main(int argc, char **argv) {
if (identicalGPUs.size() <= 1) { if (identicalGPUs.size() <= 1) {
printf( printf(
"No Two or more GPUs with same architecture capable of " "No Two or more GPUs with same architecture capable of "
"cooperativeMultiDeviceLaunch found. \nWaiving the sample\n"); "cooperativeMultiDeviceLaunch & concurrentManagedAccess found. \nWaiving the sample\n");
exit(EXIT_WAIVED); exit(EXIT_WAIVED);
} }
@@ -617,9 +618,12 @@ int main(int argc, char **argv) {
cudaCooperativeLaunchMultiDeviceNoPreSync | cudaCooperativeLaunchMultiDeviceNoPreSync |
cudaCooperativeLaunchMultiDeviceNoPostSync)); cudaCooperativeLaunchMultiDeviceNoPostSync));
checkCudaErrors(cudaMemPrefetchAsync(x, sizeof(float) * N, cudaCpuDeviceId)); if (deviceProp.concurrentManagedAccess) {
checkCudaErrors( checkCudaErrors(
cudaMemPrefetchAsync(dot_result, sizeof(double), cudaCpuDeviceId)); cudaMemPrefetchAsync(x, sizeof(float) * N, cudaCpuDeviceId));
checkCudaErrors(
cudaMemPrefetchAsync(dot_result, sizeof(double), cudaCpuDeviceId));
}
deviceId = identicalGPUs.begin(); deviceId = identicalGPUs.begin();
device_count = 0; device_count = 0;

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientMultiDeviceCG", "conjugateGradientMultiDeviceCG_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientMultiBlockCG_vs2010</RootNamespace> <RootNamespace>conjugateGradientMultiDeviceCG_vs2012</RootNamespace>
<ProjectName>conjugateGradientMultiBlockCG</ProjectName> <ProjectName>conjugateGradientMultiDeviceCG</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v110</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -58,10 +59,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudadevrt.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudadevrt.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientMultiBlockCG.exe</OutputFile> <OutputFile>$(OutDir)/conjugateGradientMultiDeviceCG.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -97,11 +98,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="conjugateGradientMultiBlockCG.cu" /> <CudaCompile Include="conjugateGradientMultiDeviceCG.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientMultiDeviceCG", "conjugateGradientMultiDeviceCG_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,108 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientMultiDeviceCG_vs2013</RootNamespace>
<ProjectName>conjugateGradientMultiDeviceCG</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudadevrt.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientMultiDeviceCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
<GenerateRelocatableDeviceCode>true</GenerateRelocatableDeviceCode>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientMultiDeviceCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientMultiDeviceCG", "conjugateGradientMultiDeviceCG_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,108 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientMultiDeviceCG_vs2015</RootNamespace>
<ProjectName>conjugateGradientMultiDeviceCG</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v140</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudadevrt.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientMultiDeviceCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
<GenerateRelocatableDeviceCode>true</GenerateRelocatableDeviceCode>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientMultiDeviceCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "conjugateGradientMultiDeviceCG", "conjugateGradientMultiDeviceCG_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,109 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientMultiDeviceCG_vs2017</RootNamespace>
<ProjectName>conjugateGradientMultiDeviceCG</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</PlatformToolset>
<WindowsTargetPlatformVersion>10.0.15063.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudadevrt.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/conjugateGradientMultiDeviceCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
<GenerateRelocatableDeviceCode>true</GenerateRelocatableDeviceCode>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="conjugateGradientMultiDeviceCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
@@ -248,12 +246,6 @@ ifeq ($(TARGET_ARCH),armv7l)
SAMPLE_ENABLED := 0 SAMPLE_ENABLED := 0
endif endif
# This sample is not supported on aarch64
ifeq ($(TARGET_ARCH),aarch64)
$(info >>> WARNING - cudaTensorCoreGemm is not supported on aarch64 - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ALL_LDFLAGS := ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS) ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS)) ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
@@ -266,7 +258,11 @@ LIBRARIES :=
################################################################################ ################################################################################
# Gencode arguments # Gencode arguments
SMS ?= 70 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 70 72 75
else
SMS ?= 70 75
endif
ifeq ($(SMS),) ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<) $(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)

View File

@@ -43,11 +43,16 @@ In addition to that, it demonstrates the use of the new CUDA function attribute
<scope>1:CUDA Basic Topics</scope> <scope>1:CUDA Basic Topics</scope>
</scopes> </scopes>
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>
<platform>linux</platform> <platform>linux</platform>
</env> </env>
<env>
<arch>aarch64</arch>
</env>
<env> <env>
<platform>windows7</platform> <platform>windows7</platform>
</env> </env>

View File

@@ -14,7 +14,7 @@ Matrix Multiply, WMMA, Tensor Cores
## Supported SM Architectures ## Supported SM Architectures
[SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
@@ -22,7 +22,7 @@ Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
x86_64, ppc64le x86_64, ppc64le, aarch64
## CUDA APIs involved ## CUDA APIs involved
@@ -31,7 +31,7 @@ cudaMallocManaged, cudaDeviceSynchronize, cudaFuncSetAttribute, cudaEventCreate,
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run ## Build and Run
@@ -52,9 +52,9 @@ $ cd <sample_dir>
$ make $ make
``` ```
The samples makefiles can take advantage of certain options: The samples makefiles can take advantage of certain options:
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le. * **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, aarch64.
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/> By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=aarch64` <br/>
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details. See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
* **dbg=1** - build with debug symbols * **dbg=1** - build with debug symbols
``` ```

View File

@@ -72,6 +72,21 @@
#include <helper_cuda.h> #include <helper_cuda.h>
#include <helper_functions.h> #include <helper_functions.h>
// Externally configurable parameters.
#ifndef CPU_DEBUG
// Set this to 1 to verify the correctness of the GPU-computed matrix.
#define CPU_DEBUG 0
#endif
#ifndef SHARED_MEMORY_LIMIT_64K
// Set this to 0 to use more than 64 Kb of shared memory to cache data, to
// improve the performance of the computations on GPU.
// Note that you need a GPU that can have more than 64 Kb of shared memory
// per multiprocessor.
#define SHARED_MEMORY_LIMIT_64K 1
#endif
// GPU configuration. // GPU configuration.
#define WARP_SIZE 32 #define WARP_SIZE 32
@@ -82,6 +97,10 @@
#define N 16 #define N 16
#define K 16 #define K 16
#define WMMA_M 16
#define WMMA_N 16
#define WMMA_K 16
// GEMM configuration. // GEMM configuration.
#define M_TILES 256 #define M_TILES 256
@@ -99,7 +118,24 @@
#define WARPS_PER_BLOCK 8 #define WARPS_PER_BLOCK 8
#define THREADS_PER_BLOCK (WARP_SIZE * WARPS_PER_BLOCK) #define THREADS_PER_BLOCK (WARP_SIZE * WARPS_PER_BLOCK)
#if SHARED_MEMORY_LIMIT_64K
// With only 64 Kb shared memory available, we can fit two 8-tile chunks of
// the A and B matrix data, that are 16 * 16 * 8 * 8 * 2 = 32 Kb each
// (i.e. two 8x8 arrays of tiles of 16x16 half-typed elements per CTA).
// But we cannot account the 8 Kb total skew overhead, without which the
// performance would be severely impacted. So we choose to reduce the chunk size
// in half, i.e. the amount of A and B matrix data we cache in shared memory.
// Accordingly, this doubles the number of outer iterations across the global K
// dimension, which only slightly impacts the performance.
#define CHUNK_K 4
#else
#define CHUNK_K 8 #define CHUNK_K 8
#endif
#define CHUNK_LINE_BYTES (CHUNK_K * K * sizeof(half))
#define WARP_COPY_BYTES (WARP_SIZE * sizeof(int4))
#define CHUNK_COPY_LINES_PER_WARP (WARP_COPY_BYTES / CHUNK_LINE_BYTES)
#define CHUNK_COPY_LINE_LANES (WARP_SIZE / CHUNK_COPY_LINES_PER_WARP)
#define BLOCK_ROW_WARPS 2 #define BLOCK_ROW_WARPS 2
#define BLOCK_COL_WARPS 4 #define BLOCK_COL_WARPS 4
@@ -144,16 +180,16 @@
using namespace nvcuda; using namespace nvcuda;
__host__ void init_host_matrices(float *a, float *b, float *c) { __host__ void init_host_matrices(half *a, half *b, float *c) {
for (int i = 0; i < M_GLOBAL; i++) { for (int i = 0; i < M_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) { for (int j = 0; j < K_GLOBAL; j++) {
a[i * K_GLOBAL + j] = static_cast<float>(rand() % 3); a[i * K_GLOBAL + j] = (half)(rand() % 3);
} }
} }
for (int i = 0; i < N_GLOBAL; i++) { for (int i = 0; i < N_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) { for (int j = 0; j < K_GLOBAL; j++) {
b[i * K_GLOBAL + j] = static_cast<float>(rand() % 3); b[i * K_GLOBAL + j] = (half)(rand() % 3);
} }
} }
@@ -162,26 +198,6 @@ __host__ void init_host_matrices(float *a, float *b, float *c) {
} }
} }
__global__ void init_device_matrices(const float *A_h, const float *B_h,
const float *C_h, half *A, half *B,
float *C, float *D) {
for (int i = blockDim.x * blockIdx.x + threadIdx.x; i < M_GLOBAL * K_GLOBAL;
i += gridDim.x * blockDim.x)
A[i] = __float2half(A_h[i]);
for (int i = blockDim.x * blockIdx.x + threadIdx.x; i < N_GLOBAL * K_GLOBAL;
i += gridDim.x * blockDim.x)
B[i] = __float2half(B_h[i]);
for (int i = blockDim.x * blockIdx.x + threadIdx.x; i < M_GLOBAL * N_GLOBAL;
i += gridDim.x * blockDim.x)
C[i] = C_h[i];
for (int i = blockDim.x * blockIdx.x + threadIdx.x; i < M_GLOBAL * N_GLOBAL;
i += gridDim.x * blockDim.x)
D[i] = 0;
}
__global__ void compute_gemm(const half *A, const half *B, const float *C, __global__ void compute_gemm(const half *A, const half *B, const float *C,
float *D, float alpha, float beta) { float *D, float alpha, float beta) {
extern __shared__ half shmem[][CHUNK_K * K + SKEW_HALF]; extern __shared__ half shmem[][CHUNK_K * K + SKEW_HALF];
@@ -194,14 +210,14 @@ __global__ void compute_gemm(const half *A, const half *B, const float *C,
const size_t shmem_idx_b_off = BLOCK_COL_TILES * M; const size_t shmem_idx_b_off = BLOCK_COL_TILES * M;
// This pointer is used to access the C and D matrix tiles this warp computes. // This pointer is used to access the C and D matrix tiles this warp computes.
float *shmem_warp_tile_ptr = reinterpret_cast<float *>( float *shmem_warp_tile_ptr = (float *)&shmem[0][0] +
&shmem[0][0] + (warpId / 2) * SHMEM_STRIDE * K * 2 + (warpId / 2) * SHMEM_STRIDE * K * 2 +
(warpId % 2) * SHMEM_OFFSET); (warpId % 2) * SHMEM_OFFSET;
// This pointer is used to stream the C and D matrices block-wide tile to and // This pointer is used to stream the C and D matrices block-wide tile to and
// from shared memory. // from shared memory.
float *shmem_warp_stream_ptr = float *shmem_warp_stream_ptr =
reinterpret_cast<float *>(&shmem[0][0] + warpId * SHMEM_STRIDE * K); (float *)&shmem[0][0] + warpId * SHMEM_STRIDE * K;
// Adjust the beta scaler, as it'll be multiplied by alpha at the end of // Adjust the beta scaler, as it'll be multiplied by alpha at the end of
// each tile computation. Technically this is not generally correct (may // each tile computation. Technically this is not generally correct (may
@@ -292,23 +308,24 @@ __global__ void compute_gemm(const half *A, const half *B, const float *C,
// First half of the warp copies the first row / column of the matrix, // First half of the warp copies the first row / column of the matrix,
// the second half of the warp copies the next. // the second half of the warp copies the next.
int4 *lane_ptr = (int4 *)(warp_ptr + tile_k * K + int4 *lane_ptr = (int4 *)(warp_ptr + tile_k * K +
(laneId / (WARP_SIZE / 2)) * K_GLOBAL) + (laneId / CHUNK_COPY_LINE_LANES) * K_GLOBAL) +
(laneId % (WARP_SIZE / 2)); (laneId % CHUNK_COPY_LINE_LANES);
// Shift the second half of the warp to the next row / column in the // Shift the second half of the warp to the next row / column in the
// shared memory. // shared memory.
shmem_idx += laneId / (WARP_SIZE / 2); shmem_idx += laneId / CHUNK_COPY_LINE_LANES;
#pragma unroll #pragma unroll
for (int i = 0; i < (WARP_SIZE / 2); i++) { for (int i = 0; i < ((WARP_SIZE / 2) / CHUNK_COPY_LINES_PER_WARP) * 2;
i++) {
// Copy 16 bytes at once in each lane. // Copy 16 bytes at once in each lane.
*((int4 *)&shmem[shmem_idx][0] + (laneId % (WARP_SIZE / 2))) = *((int4 *)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES)) =
*lane_ptr; *lane_ptr;
// Advance the global memory pointer and the shared memory index. // Advance the global memory pointer and the shared memory index.
lane_ptr = reinterpret_cast<int4 *>( lane_ptr =
reinterpret_cast<half *>(lane_ptr + K_GLOBAL * 2)); (int4 *)((half *)lane_ptr + K_GLOBAL * CHUNK_COPY_LINES_PER_WARP);
shmem_idx += 2; shmem_idx += CHUNK_COPY_LINES_PER_WARP;
} }
__syncthreads(); __syncthreads();
@@ -374,17 +391,98 @@ __global__ void compute_gemm(const half *A, const half *B, const float *C,
#pragma unroll #pragma unroll
for (int i = 0; i < K; i++) { for (int i = 0; i < K; i++) {
*(reinterpret_cast<int4 *>(dst_gmem_warp_stream_ptr + *((int4 *)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
GLOBAL_MEM_STRIDE * i) + *((int4 *)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
laneId) =
*(reinterpret_cast<int4 *>(shmem_warp_stream_ptr + SHMEM_STRIDE * i) +
laneId);
} }
__syncthreads(); __syncthreads();
} }
} }
// Performs an MxNxK GEMM (C=alpha*A*B + beta*C) assuming:
// 1) Matrices are packed in memory.
// 2) M, N and K are multiples of 16.
// 3) Neither A nor B are transposed.
// Note: This is a less performant version of the compute_gemm kernel. It is
// designed for
// demonstration purposes only to show the CUDA WMMA API use without
// relying on availability of the shared memory.
__global__ void simple_wmma_gemm(half *a, half *b, float *c, float *d, int m_ld,
int n_ld, int k_ld, float alpha, float beta) {
// Leading dimensions. Packed with no transpositions.
int lda = m_ld;
int ldb = k_ld;
int ldc = n_ld;
// Tile using a 2D grid
int warpM = (blockIdx.x * blockDim.x + threadIdx.x) / warpSize;
int warpN = (blockIdx.y * blockDim.y + threadIdx.y);
// Declare the fragments
wmma::fragment<wmma::matrix_a, WMMA_M, WMMA_N, WMMA_K, half, wmma::row_major>
a_frag;
wmma::fragment<wmma::matrix_b, WMMA_M, WMMA_N, WMMA_K, half, wmma::col_major>
b_frag;
wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, float> acc_frag;
wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, float> c_frag;
wmma::fill_fragment(acc_frag, 0.0f);
// Loop over k
for (int i = 0; i < k_ld; i += WMMA_K) {
int aCol = i;
int aRow = warpM * WMMA_M;
int bCol = i;
int bRow = warpN * WMMA_N;
// Bounds checking
if (aRow < m_ld && aCol < k_ld && bRow < k_ld && bCol < n_ld) {
// Load the inputs
wmma::load_matrix_sync(a_frag, a + aCol + aRow * lda, lda);
wmma::load_matrix_sync(b_frag, b + bCol + bRow * ldb, ldb);
// Perform the matrix multiplication
wmma::mma_sync(acc_frag, a_frag, b_frag, acc_frag);
}
}
// Load in the current value of c, scale it by beta, and add this our result
// scaled by alpha
int cCol = warpN * WMMA_N;
int cRow = warpM * WMMA_M;
if (cRow < m_ld && cCol < n_ld) {
wmma::load_matrix_sync(c_frag, c + cCol + cRow * ldc, ldc,
wmma::mem_row_major);
for (int i = 0; i < c_frag.num_elements; i++) {
c_frag.x[i] = alpha * acc_frag.x[i] + beta * c_frag.x[i];
}
// Store the output
wmma::store_matrix_sync(d + cCol + cRow * ldc, c_frag, ldc,
wmma::mem_row_major);
}
}
__host__ void matMultiplyOnHost(half *A, half *B, float *C, float alpha,
float beta, int numARows, int numAColumns,
int numBRows, int numBColumns, int numCRows,
int numCColumns) {
for (int i = 0; i < numCRows; i++) {
for (int j = 0; j < numCColumns; j++) {
float temp = 0.0;
for (int k = 0; k < numAColumns; k++) {
temp += (float)A[i * numAColumns + k] * (float)B[j * numBRows + k];
}
C[i * numCColumns + j] = temp * alpha + beta * C[i * numCColumns + j];
}
}
}
int main(int argc, char **argv) { int main(int argc, char **argv) {
printf("Initializing...\n"); printf("Initializing...\n");
@@ -396,7 +494,7 @@ int main(int argc, char **argv) {
// Tensor cores require a GPU of Volta (SM7X) architecture or higher. // Tensor cores require a GPU of Volta (SM7X) architecture or higher.
if (deviceProp.major < 7) { if (deviceProp.major < 7) {
printf( printf(
"cudaTensorCoreGemm requires requires SM 7.0 or higher to use Tensor " "cudaTensorCoreGemm requires SM 7.0 or higher to use Tensor "
"Cores. Exiting...\n"); "Cores. Exiting...\n");
exit(EXIT_WAIVED); exit(EXIT_WAIVED);
} }
@@ -405,16 +503,21 @@ int main(int argc, char **argv) {
printf("N: %d (%d x %d)\n", N_GLOBAL, N, N_TILES); printf("N: %d (%d x %d)\n", N_GLOBAL, N, N_TILES);
printf("K: %d (%d x %d)\n", K_GLOBAL, K, K_TILES); printf("K: %d (%d x %d)\n", K_GLOBAL, K, K_TILES);
float *A_h = NULL; half *A_h = NULL;
float *B_h = NULL; half *B_h = NULL;
float *C_h = NULL; float *C_h = NULL;
#if CPU_DEBUG
float *result_hD = NULL;
float *result_host = NULL;
#endif
checkCudaErrors(cudaMallocManaged(reinterpret_cast<void **>(&A_h), A_h = (half *)malloc(sizeof(half) * M_GLOBAL * K_GLOBAL);
sizeof(float) * M_GLOBAL * K_GLOBAL)); B_h = (half *)malloc(sizeof(half) * K_GLOBAL * N_GLOBAL);
checkCudaErrors(cudaMallocManaged(reinterpret_cast<void **>(&B_h), C_h = (float *)malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
sizeof(float) * K_GLOBAL * N_GLOBAL)); #if CPU_DEBUG
checkCudaErrors(cudaMallocManaged(reinterpret_cast<void **>(&C_h), result_hD = (float *)malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
sizeof(float) * M_GLOBAL * N_GLOBAL)); result_host = (float *)malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
#endif
half *A = NULL; half *A = NULL;
half *B = NULL; half *B = NULL;
@@ -439,23 +542,31 @@ int main(int argc, char **argv) {
printf("Preparing data for GPU...\n"); printf("Preparing data for GPU...\n");
checkKernelErrors( checkCudaErrors(cudaMemcpy(A, A_h, sizeof(half) * M_GLOBAL * K_GLOBAL,
(init_device_matrices<<<deviceProp.multiProcessorCount, cudaMemcpyHostToDevice));
THREADS_PER_BLOCK>>>(A_h, B_h, C_h, A, B, C, D))); checkCudaErrors(cudaMemcpy(B, B_h, sizeof(half) * N_GLOBAL * K_GLOBAL,
cudaMemcpyHostToDevice));
checkCudaErrors(cudaDeviceSynchronize()); checkCudaErrors(cudaMemcpy(C, C_h, sizeof(float) * M_GLOBAL * N_GLOBAL,
cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemset(D, 0, sizeof(float) * M_GLOBAL * N_GLOBAL));
enum { enum {
SHMEM_SZ = // Compute the right amount of shared memory to request.
sizeof(half) * (BLOCK_COL_TILES * M) * (CHUNK_K * K + SKEW_HALF) * 2 // We need shared memory to hold per-CTA C and D matrix tiles, and to cache
// per-CTA chunks
// of the A and B matrices. Therefore, the right amount to request is the
// maximum of those
// two numbers.
SHMEM_SZ = MAX(
sizeof(half) * (BLOCK_COL_TILES * M) * (CHUNK_K * K + SKEW_HALF) * 2,
M * (BLOCK_ROW_WARPS * WARP_ROW_TILES) * N *
(BLOCK_COL_WARPS * WARP_COL_TILES) * sizeof(float))
}; };
printf("Required shared memory size: %lu Kb\n", SHMEM_SZ / 1024UL); printf("Required shared memory size: %lu Kb\n", SHMEM_SZ / 1024UL);
checkCudaErrors(cudaFuncSetAttribute( const float alpha = 1.1f;
compute_gemm, cudaFuncAttributeMaxDynamicSharedMemorySize, SHMEM_SZ)); const float beta = 1.2f;
printf("Computing...\n");
cudaEvent_t start, stop; cudaEvent_t start, stop;
@@ -463,16 +574,63 @@ int main(int argc, char **argv) {
checkCudaErrors(cudaEventCreate(&stop)); checkCudaErrors(cudaEventCreate(&stop));
checkCudaErrors(cudaEventRecord(start)); checkCudaErrors(cudaEventRecord(start));
const float alpha = 1.1f; // If enough shared memory available on the GPU use high performant kernel
const float beta = 1.2f; if (deviceProp.sharedMemPerMultiprocessor >= SHMEM_SZ) {
printf("Computing... using high performance kernel compute_gemm \n");
checkKernelErrors( checkCudaErrors(cudaFuncSetAttribute(
(compute_gemm<<<deviceProp.multiProcessorCount, THREADS_PER_BLOCK, compute_gemm, cudaFuncAttributeMaxDynamicSharedMemorySize, SHMEM_SZ));
SHMEM_SZ>>>(A, B, C, D, alpha, beta))); checkKernelErrors(
(compute_gemm<<<deviceProp.multiProcessorCount, THREADS_PER_BLOCK,
SHMEM_SZ>>>(A, B, C, D, alpha, beta)));
#if CPU_DEBUG
checkCudaErrors(cudaMemcpy(result_hD, D,
sizeof(float) * M_GLOBAL * N_GLOBAL,
cudaMemcpyDeviceToHost));
#endif
} else {
dim3 gridDim;
dim3 blockDim;
// blockDim.x must be a multple of warpSize
// 128x4 means we have 16 warps and a block computes a 64x64 output tile
blockDim.x = 128;
blockDim.y = 4;
gridDim.x = (M_GLOBAL + (WMMA_M * blockDim.x / 32 - 1)) /
(WMMA_M * blockDim.x / 32);
gridDim.y = (N_GLOBAL + WMMA_N * blockDim.y - 1) / (WMMA_N * blockDim.y);
printf("Computing... using simple_wmma_gemm kernel\n");
simple_wmma_gemm<<<gridDim, blockDim>>>(A, B, C, D, M_GLOBAL, N_GLOBAL,
K_GLOBAL, alpha, beta);
#if CPU_DEBUG
checkCudaErrors(cudaMemcpy(result_hD, D,
sizeof(float) * M_GLOBAL * N_GLOBAL,
cudaMemcpyDeviceToHost));
#endif
}
checkCudaErrors(cudaEventRecord(stop)); checkCudaErrors(cudaEventRecord(stop));
checkCudaErrors(cudaEventSynchronize(stop)); checkCudaErrors(cudaEventSynchronize(stop));
#if CPU_DEBUG
printf("Verifying correctness of the computations...\n");
memcpy(result_host, C_h, sizeof(float) * M_GLOBAL * N_GLOBAL);
matMultiplyOnHost(A_h, B_h, result_host, alpha, beta, M_GLOBAL, K_GLOBAL,
K_GLOBAL, N_GLOBAL, M_GLOBAL, N_GLOBAL);
for (int i = 0; i < N_GLOBAL * M_GLOBAL; i++) {
if (fabs(result_hD[i] - result_host[i]) > 0.1f)
printf("mismatch i=%d result_hD=%f result_host=%f\n", i, result_hD[i],
result_host[i]);
}
free(result_hD);
free(result_host);
#endif
float milliseconds = 0; float milliseconds = 0;
checkCudaErrors(cudaEventElapsedTime(&milliseconds, start, stop)); checkCudaErrors(cudaEventElapsedTime(&milliseconds, start, stop));
@@ -483,9 +641,9 @@ int main(int argc, char **argv) {
(milliseconds / 1000.)) / (milliseconds / 1000.)) /
1e12); 1e12);
checkCudaErrors(cudaFree(reinterpret_cast<void *>(A_h))); free(A_h);
checkCudaErrors(cudaFree(reinterpret_cast<void *>(B_h))); free(B_h);
checkCudaErrors(cudaFree(reinterpret_cast<void *>(C_h))); free(C_h);
checkCudaErrors(cudaFree(reinterpret_cast<void *>(A))); checkCudaErrors(cudaFree(reinterpret_cast<void *>(A)));
checkCudaErrors(cudaFree(reinterpret_cast<void *>(B))); checkCudaErrors(cudaFree(reinterpret_cast<void *>(B)));
checkCudaErrors(cudaFree(reinterpret_cast<void *>(C))); checkCudaErrors(cudaFree(reinterpret_cast<void *>(C)));

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile> <OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile> <OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile> <OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -34,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -63,7 +63,7 @@
<OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile> <OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
@@ -248,7 +246,11 @@ LIBRARIES :=
################################################################################ ################################################################################
# Gencode arguments # Gencode arguments
SMS ?= 30 35 37 50 52 60 61 70 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75
else
SMS ?= 30 35 37 50 52 60 61 70 75
endif
ifeq ($(SMS),) ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<) $(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)

View File

@@ -39,6 +39,8 @@
<sm-arch>sm60</sm-arch> <sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch> <sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,7 +10,7 @@ CUDA Runtime API, Device Query
## Supported SM Architectures ## Supported SM Architectures
[SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
@@ -27,7 +27,7 @@ cudaSetDevice, cudaGetDeviceCount, cudaGetDeviceProperties, cudaDriverGetVersion
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run ## Build and Run

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/deviceQuery.exe</OutputFile> <OutputFile>$(OutDir)/deviceQuery.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/deviceQuery.exe</OutputFile> <OutputFile>$(OutDir)/deviceQuery.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/deviceQuery.exe</OutputFile> <OutputFile>$(OutDir)/deviceQuery.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -34,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -63,7 +63,7 @@
<OutputFile>$(OutDir)/deviceQuery.exe</OutputFile> <OutputFile>$(OutDir)/deviceQuery.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,318 @@
################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
################################################################################
#
# Makefile project only supported on Mac OS X and Linux Platforms)
#
################################################################################
# Location of the CUDA Toolkit
CUDA_PATH ?= /usr/local/cuda
##############################
# start deprecated interface #
##############################
ifeq ($(x86_64),1)
$(info WARNING - x86_64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=x86_64 instead)
TARGET_ARCH ?= x86_64
endif
ifeq ($(ARMv7),1)
$(info WARNING - ARMv7 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=armv7l instead)
TARGET_ARCH ?= armv7l
endif
ifeq ($(aarch64),1)
$(info WARNING - aarch64 variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=aarch64 instead)
TARGET_ARCH ?= aarch64
endif
ifeq ($(ppc64le),1)
$(info WARNING - ppc64le variable has been deprecated)
$(info WARNING - please use TARGET_ARCH=ppc64le instead)
TARGET_ARCH ?= ppc64le
endif
ifneq ($(GCC),)
$(info WARNING - GCC variable has been deprecated)
$(info WARNING - please use HOST_COMPILER=$(GCC) instead)
HOST_COMPILER ?= $(GCC)
endif
ifneq ($(abi),)
$(error ERROR - abi variable has been removed)
endif
############################
# end deprecated interface #
############################
# architecture
HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif
endif
# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
TARGET_ARCH = armv7l
endif
# operating system
HOST_OS := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
TARGET_OS ?= $(HOST_OS)
ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
$(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
endif
# host compiler
ifeq ($(TARGET_OS),darwin)
ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
HOST_COMPILER ?= clang++
endif
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
ifeq ($(TARGET_OS),linux)
HOST_COMPILER ?= arm-linux-gnueabihf-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
else ifeq ($(TARGET_OS),android)
HOST_COMPILER ?= arm-linux-androideabi-g++
endif
else ifeq ($(TARGET_ARCH),aarch64)
ifeq ($(TARGET_OS), linux)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_OS),qnx)
ifeq ($(QNX_HOST),)
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
endif
ifeq ($(QNX_TARGET),)
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
endif
export QNX_HOST
export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++
endif
else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif
endif
HOST_COMPILER ?= g++
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
# internal flags
NVCCFLAGS := -m${TARGET_SIZE}
CCFLAGS :=
LDFLAGS :=
# build flags
ifeq ($(TARGET_OS),darwin)
LDFLAGS += -rpath $(CUDA_PATH)/lib
CCFLAGS += -arch $(HOST_ARCH)
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
CCFLAGS += -mfloat-abi=hard
else ifeq ($(TARGET_OS),android)
LDFLAGS += -pie
CCFLAGS += -fpie -fpic -fexceptions
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
ifneq ($(TARGET_FS),)
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
ifeq ($(GCCVERSIONLTEQ46),1)
CCFLAGS += --sysroot=$(TARGET_FS)
endif
LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
endif
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
else ifeq ($(TARGET_ARCH),ppc64le)
CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
endif
# Debug build flags
ifeq ($(dbg),1)
NVCCFLAGS += -g -G
BUILD_TYPE := debug
else
BUILD_TYPE := release
endif
ALL_CCFLAGS :=
ALL_CCFLAGS += $(NVCCFLAGS)
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
SAMPLE_ENABLED := 1
# This sample is not supported on Mac OSX
ifeq ($(TARGET_OS),darwin)
$(info >>> WARNING - immaTensorCoreGemm is not supported on Mac OSX - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
# This sample is not supported on ARMv7
ifeq ($(TARGET_ARCH),armv7l)
$(info >>> WARNING - immaTensorCoreGemm is not supported on ARMv7 - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ALL_LDFLAGS :=
ALL_LDFLAGS += $(ALL_CCFLAGS)
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
# Common includes and paths for CUDA
INCLUDES := -I../../Common
LIBRARIES :=
################################################################################
# Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 72 75
else
SMS ?= 75
endif
ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
ifeq ($(GENCODE_FLAGS),)
# Generate SASS code for each SM architecture listed in $(SMS)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
HIGHEST_SM := $(lastword $(sort $(SMS)))
ifneq ($(HIGHEST_SM),)
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif
ALL_CCFLAGS += -maxrregcount=255
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: immaTensorCoreGemm
check.deps:
ifeq ($(SAMPLE_ENABLED),0)
@echo "Sample will be waived due to the above missing dependencies"
else
@echo "Sample is ready - all dependencies have been met"
endif
immaTensorCoreGemm.o:immaTensorCoreGemm.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
immaTensorCoreGemm: immaTensorCoreGemm.o
$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
run: build
$(EXEC) ./immaTensorCoreGemm
clean:
rm -f immaTensorCoreGemm immaTensorCoreGemm.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/immaTensorCoreGemm
clobber: clean

View File

@@ -0,0 +1,64 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>immaTensorCoreGemm</name>
<cflags>
<flag>-maxrregcount=255</flag>
</cflags>
<cuda_api_list>
<toolkit>cudaMallocManaged</toolkit>
<toolkit>cudaDeviceSynchronize</toolkit>
<toolkit>cudaFuncSetAttribute</toolkit>
<toolkit>cudaEventCreate</toolkit>
<toolkit>cudaEventRecord</toolkit>
<toolkit>cudaEventSynchronize</toolkit>
<toolkit>cudaEventElapsedTime</toolkit>
<toolkit>cudaFree</toolkit>
</cuda_api_list>
<description><![CDATA[CUDA sample demonstrating a integer GEMM computation using the Warp Matrix Multiply and Accumulate (WMMA) API for integer introduced in CUDA 10. This sample demonstrates the use of the CUDA WMMA API employing the Tensor Cores introduced in the Volta chip family for faster matrix operations. In addition to that, it demonstrates the use of the new CUDA function attribute cudaFuncAttributeMaxDynamicSharedMemorySize that allows the application to reserve an extended amount of shared memory than it is available by default.]]></description>
<devicecompilation>whole</devicecompilation>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">Matrix Multiply</concept>
<concept level="advanced">WMMA</concept>
<concept level="advanced">Tensor Cores</concept>
</keyconcepts>
<keywords>
</keywords>
<libraries>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>immaTensorCoreGemm.cu</primary_file>
<scopes>
<scope>1:CUDA Basic Topics</scope>
</scopes>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<arch>aarch64</arch>
</env>
<env>
<platform>windows7</platform>
</env>
<env>
<arch>ppc64le</arch>
<platform>linux</platform>
</env>
</supported_envs>
<supported_sm_architectures>
<from>7.2</from>
</supported_sm_architectures>
<title>Tensor Core GEMM Integer MMA</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,70 @@
# immaTensorCoreGemm - Tensor Core GEMM Integer MMA
## Description
CUDA sample demonstrating a integer GEMM computation using the Warp Matrix Multiply and Accumulate (WMMA) API for integer introduced in CUDA 10. This sample demonstrates the use of the CUDA WMMA API employing the Tensor Cores introduced in the Volta chip family for faster matrix operations. In addition to that, it demonstrates the use of the new CUDA function attribute cudaFuncAttributeMaxDynamicSharedMemorySize that allows the application to reserve an extended amount of shared memory than it is available by default.
## Key Concepts
Matrix Multiply, WMMA, Tensor Cores
## Supported SM Architectures
[SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, ppc64le, aarch64
## CUDA APIs involved
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaMallocManaged, cudaDeviceSynchronize, cudaFuncSetAttribute, cudaEventCreate, cudaEventRecord, cudaEventSynchronize, cudaEventElapsedTime, cudaFree
## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run
### Windows
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
```
*_vs<version>.sln - for Visual Studio <version>
```
Each individual sample has its own set of solution files in its directory:
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
### Linux
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
```
$ cd <sample_dir>
$ make
```
The samples makefiles can take advantage of certain options:
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, aarch64.
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=aarch64` <br/>
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
* **dbg=1** - build with debug symbols
```
$ make dbg=1
```
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
```
$ make SMS="50 60"
```
* **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
```
$ make HOST_COMPILER=g++
```
## References (for more details)

View File

@@ -0,0 +1,655 @@
/* Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
// CUDA sample demonstrating a integer GEMM computation using the Warp Matrix
// Multiply and Accumulate API.
// In this program, the compute_gemm kernel computes the result of a matrix
// multiplication and addition: D = alpha * A * B + beta * C. The dimensions of
// both C and D matrices are M_GLOBAL x N_GLOBAL. The A matrix is M_GLOBAL x
// K_GLOBAL (row-major), the B matrix is K_GLOBAL x N_GLOBAL (column-major). In
// that kernel, each CTA computes one 128 x 128 tile of the resulting matrix per
// iteration. When the tile is computed, the CTA stores it to the global memory
// and begins a new iteration, selecting a new 128 x 128 tile to compute.
// Each CTA consists of eight warps. For the 128 x 128 tile, each warp computes
// eight 16 x 16 subtiles, organized in a 2 x 4 two-dimensional array. Warps
// compute the 16 x 16 subtiles using nvcuda::wmma::mma_sync operations by
// moving through the K_GLOBAL dimension of the A and B matrices and
// accumulating the intermediate result in the local thread state.
// There are a number of simple optimizations used in the algorithm:
// - The CTA copies the 128 x 128 tile of the C matrix from the global memory to
// shared memory. After that is done, each warp loads the C matrix fragments
// from shared memory, thus avoiding a random global memory access.
// - On each internal iteration, the CTA copies a portion of the A and B
// matrices from
// global memory to shared memory. After that, all warps in the CTA reuse the
// A and B data from shared memory, thus reducing the number of data copies
// from global memory.
// - The portions of the A and B matrices are stored in shared memory with an
// additional
// padding (skew) to reduce the number of shared memory access bank conflicts.
// (See a detailed explanation near the SKEW_HALF macro definition.)
// - When the CTA finishes computing the tiles of the resulting matrix, each
// warp stores
// its subtiles to shared memory. The CTA then copies the shared memory
// contents to global memory, again avoiding redundant random global memory
// accesses.
// - Note that the CTA tile size is chosen to maximize the GPU register
// utilization,
// but carefully enough to avoid local memory use.
#include <assert.h>
#include <cuda.h>
#include <mma.h>
#include <stdio.h>
// helper functions and utilities to work with CUDA
#include <helper_cuda.h>
#include <helper_functions.h>
// Externally configurable parameters.
#ifndef CPU_DEBUG
// Set this to 1 to verify the correctness of the GPU-computed matrix.
#define CPU_DEBUG 0
#endif
#ifndef SHARED_MEMORY_LIMIT_64K
// Set this to 0 to use more than 64 Kb of shared memory to cache data, to
// improve the performance of the computations on GPU.
// Note that you need a GPU that can have more than 64 Kb of shared memory
// per multiprocessor.
#define SHARED_MEMORY_LIMIT_64K 1
#endif
// GPU configuration.
#define WARP_SIZE 32
// MMA matrix tile dimensions.
#define M 16
#define N 16
#define K 16
#define WMMA_M 16
#define WMMA_N 16
#define WMMA_K 16
// GEMM configuration.
#define M_TILES 256
#define N_TILES 256
#define K_TILES 256
#define M_GLOBAL (M * M_TILES)
#define N_GLOBAL (N * N_TILES)
#define K_GLOBAL (K * K_TILES)
#define C_LAYOUT wmma::mem_row_major
// Implementation constants.
#define WARPS_PER_BLOCK 8
#define THREADS_PER_BLOCK (WARP_SIZE * WARPS_PER_BLOCK)
#if SHARED_MEMORY_LIMIT_64K
// With only 64 Kb shared memory available, we can fit two 8-tile chunks of
// the A and B matrix data, that are 16 * 16 * 8 * 8 * 2 = 32 Kb each
// (i.e. two 8x8 arrays of tiles of 16x16 uint8_t-typed elements per CTA).
// But we cannot account the 8 Kb total skew overhead, without which the
// performance would be severely impacted. So we choose to reduce the chunk size
// in half, i.e. the amount of A and B matrix data we cache in shared memory.
// Accordingly, this doubles the number of outer iterations across the global K
// dimension, which only slightly impacts the performance.
#define CHUNK_K 8
#else
#define CHUNK_K 16
#endif
#define CHUNK_LINE_BYTES (CHUNK_K * K * sizeof(uint8_t))
#define WARP_COPY_BYTES (WARP_SIZE * sizeof(int4))
#define CHUNK_COPY_LINES_PER_WARP (WARP_COPY_BYTES / CHUNK_LINE_BYTES)
#define CHUNK_COPY_LINE_LANES (WARP_SIZE / CHUNK_COPY_LINES_PER_WARP)
#define BLOCK_ROW_WARPS 2
#define BLOCK_COL_WARPS 4
#define WARP_ROW_TILES 4
#define WARP_COL_TILES 2
#define BLOCK_ROW_TILES (WARP_ROW_TILES * BLOCK_ROW_WARPS)
#define BLOCK_COL_TILES (WARP_COL_TILES * BLOCK_COL_WARPS)
#define GLOBAL_MEM_STRIDE N_GLOBAL
#define SHMEM_STRIDE (N * BLOCK_ROW_TILES)
#define SHMEM_OFFSET (N * WARP_ROW_TILES)
// The macro below is used to shift rows of the A matrix and columns of the B
// matrix in shared memory to minimize possible bank conflicts. Before
// performing the nvcuda::wmma::mma_sync operation, the warp must load the
// matrix data using the nvcuda::wmma::load_matrix_sync operation. Although the
// memory access pattern is not specified for that function, each lane in the
// warp can read one or multiple matrix elements from different matrix rows or
// columns. For shared memory, such access can result in bank conflicts if
// different rows / columns of the matrix map to the same bank. By shifting each
// row and column by a few bytes, we make sure that they map to different banks,
// thus reducing the number of possible bank conflicts. The number of 16
// one-byte "uint8_t" elements is chosen as the minimum possible shift because
// we must keep each row and column 128-bit aligned, as required by
// nvcuda::wmma::load_matrix_sync.
#define SKEW_UINT8 16
#define checkKernelErrors(expr) \
do { \
expr; \
\
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
printf("Line %d: '%s' failed: %s\n", __LINE__, #expr, \
cudaGetErrorString(__err)); \
abort(); \
} \
} while (0)
using namespace nvcuda;
__host__ void init_host_matrices(uint8_t *a, uint8_t *b, int *c) {
for (int i = 0; i < M_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) {
a[i * K_GLOBAL + j] = (uint8_t)(rand() % 3);
}
}
for (int i = 0; i < N_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) {
b[i * K_GLOBAL + j] = (uint8_t)(rand() % 3);
}
}
for (int t = 0; t < M_GLOBAL * N_GLOBAL; t++) {
c[t] = (rand() % 3);
}
}
__global__ void compute_gemm_imma(const uint8_t *A, const uint8_t *B,
const int *C, int *D, int alpha, int beta) {
extern __shared__ uint8_t shmem[][CHUNK_K * K + SKEW_UINT8];
// Warp and lane identification.
const unsigned int warpId = threadIdx.x / WARP_SIZE;
const unsigned int laneId = threadIdx.x % WARP_SIZE;
// Offset in shared memory from which the B matrix is stored.
const size_t shmem_idx_b_off = BLOCK_COL_TILES * M;
// This pointer is used to access the C and D matrix tiles this warp computes.
int *shmem_warp_tile_ptr = (int *)&shmem[0][0] +
(warpId / 2) * SHMEM_STRIDE * K * 2 +
(warpId % 2) * SHMEM_OFFSET;
// This pointer is used to stream the C and D matrices block-wide tile to and
// from shared memory.
int *shmem_warp_stream_ptr = (int *)&shmem[0][0] + warpId * SHMEM_STRIDE * K;
// Adjust the beta scaler, as it'll be multiplied by alpha at the end of
// each tile computation. Technically this is not generally correct (may
// result in a loss of precision). Zero still needs to be specially handled
// though.
beta /= alpha;
// Each CTA slides along the 128 x 128 tiles from the top left corner of the
// matrix to the right and down, and selects the next tile to compute. Once
// there's no such tile, all warps in this CTA exit.
for (unsigned int block_pos = blockIdx.x;; block_pos += gridDim.x) {
const unsigned int block_tile_i =
((block_pos * BLOCK_ROW_TILES) / N_TILES) * (BLOCK_COL_TILES);
const unsigned int block_tile_j = (block_pos * BLOCK_COL_TILES) % N_TILES;
// Stop when there are no more D matrix tiles to compute in this CTA.
if (block_tile_i >= M_TILES) {
break;
}
// This warp's pointer to the C matrix data to copy memory from to shared
// memory.
const size_t gmem_idx =
(block_tile_i + warpId) * M * GLOBAL_MEM_STRIDE + block_tile_j * N;
const int *src_gmem_warp_stream_ptr = &C[gmem_idx];
// Stream multiple C tiles to shared memory.
#pragma unroll
for (int i = 0; i < K; i++) {
typedef int4 copy_t;
*((copy_t *)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId) =
*((copy_t *)(src_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) +
laneId);
}
__syncthreads();
// These fragments will accumulate the result of A and B matrix fragment
// multiplications along the K_GLOBAL dimension.
wmma::fragment<wmma::accumulator, M, N, K, int> c[WARP_COL_TILES]
[WARP_ROW_TILES];
// Load the C matrix tiles into fragments from shared memory.
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
#pragma unroll
for (int j = 0; j < WARP_ROW_TILES; j++) {
const int *tile_ptr =
shmem_warp_tile_ptr + i * SHMEM_STRIDE * K + j * N;
wmma::load_matrix_sync(c[i][j], tile_ptr, SHMEM_STRIDE, C_LAYOUT);
}
}
__syncthreads();
// Scale the C matrix.
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
#pragma unroll
for (int j = 0; j < WARP_ROW_TILES; j++) {
#pragma unroll
for (int t = 0; t < c[i][j].num_elements; t++) {
c[i][j].x[t] *= beta;
}
}
}
// Select what warp copies what matrix to shared memory.
// Warps 0-3 copy the A matrix, warps 4-7 copy the B matrix.
const uint8_t *warp_ptr = (warpId < 4) ? (&A[block_tile_i * M * K_GLOBAL] +
M * K_GLOBAL * (warpId % 4) * 2)
: (&B[block_tile_j * N * K_GLOBAL] +
N * K_GLOBAL * (warpId % 4) * 2);
// Go through the global K dimension by a fixed step at a time.
#pragma unroll
for (int tile_k = 0; tile_k < K_TILES; tile_k += CHUNK_K) {
// Copy slices of the A and B matrices to shared memory.
// The first half of the warps in the CTA copy the A matrix, the rest copy
// the B matrix.
size_t shmem_idx =
warpId < (WARPS_PER_BLOCK / 2)
? (M * (warpId % (WARPS_PER_BLOCK / 2)) * 2)
: (N * (warpId % (WARPS_PER_BLOCK / 2)) * 2 + shmem_idx_b_off);
// First half of the warp copies the first row / column of the matrix,
// the second half of the warp copies the next.
int4 *lane_ptr = (int4 *)(warp_ptr + tile_k * K +
(laneId / CHUNK_COPY_LINE_LANES) * K_GLOBAL) +
(laneId % CHUNK_COPY_LINE_LANES);
// Shift the second half of the warp to the next row / column in the
// shared memory.
shmem_idx += laneId / CHUNK_COPY_LINE_LANES;
#pragma unroll
for (int i = 0; i < ((WARP_SIZE / 2) / CHUNK_COPY_LINES_PER_WARP) * 2;
i++) {
// Copy 16 bytes at once in each lane.
*((int4 *)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES)) =
*lane_ptr;
// Advance the global memory pointer and the shared memory index.
lane_ptr = (int4 *)((uint8_t *)lane_ptr +
K_GLOBAL * CHUNK_COPY_LINES_PER_WARP);
shmem_idx += CHUNK_COPY_LINES_PER_WARP;
}
__syncthreads();
// Compute a grid of C matrix tiles in each warp.
#pragma unroll
for (int k_step = 0; k_step < CHUNK_K; k_step++) {
wmma::fragment<wmma::matrix_a, M, N, K, uint8_t, wmma::row_major>
a[WARP_COL_TILES];
wmma::fragment<wmma::matrix_b, M, N, K, uint8_t, wmma::col_major>
b[WARP_ROW_TILES];
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
size_t shmem_idx_a = (warpId / 2) * M * 2 + (i * M);
const uint8_t *tile_ptr = &shmem[shmem_idx_a][k_step * K];
wmma::load_matrix_sync(a[i], tile_ptr, K * CHUNK_K + SKEW_UINT8);
#pragma unroll
for (int j = 0; j < WARP_ROW_TILES; j++) {
if (i == 0) {
// Load the B matrix fragment once, because it is going to be
// reused against the other A matrix fragments.
size_t shmem_idx_b = shmem_idx_b_off +
(WARP_ROW_TILES * N) * (warpId % 2) +
(j * N);
const uint8_t *tile_ptr = &shmem[shmem_idx_b][k_step * K];
wmma::load_matrix_sync(b[j], tile_ptr, K * CHUNK_K + SKEW_UINT8);
}
wmma::mma_sync(c[i][j], a[i], b[j], c[i][j]);
}
}
}
__syncthreads();
}
// Store the D fragments to shared memory.
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
#pragma unroll
for (int j = 0; j < WARP_ROW_TILES; j++) {
#pragma unroll
// Uniform, point-wise transformations of ALL fragment elements by ALL
// threads in the warp are well-defined even though element indices
// within fragment storage are not defined.
for (int t = 0; t < c[i][j].num_elements; t++) c[i][j].x[t] *= alpha;
int *tile_ptr = shmem_warp_tile_ptr + i * SHMEM_STRIDE * K + j * N;
wmma::store_matrix_sync(tile_ptr, c[i][j], SHMEM_STRIDE, C_LAYOUT);
}
}
__syncthreads();
// Now that shared memory contains all the D tiles, stream them to global
// memory.
int *dst_gmem_warp_stream_ptr = &D[gmem_idx];
#pragma unroll
for (int i = 0; i < K; i++) {
*((int4 *)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
*((int4 *)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
}
__syncthreads();
}
}
// Performs an MxNxK GEMM (C=alpha*A*B + beta*C) assuming:
// 1) Matrices are packed in memory.
// 2) M, N and K are multiples of 16.
// 3) Neither A nor B are transposed.
// Note: This is a less performant version of the compute_gemm_imma kernel. It
// is designed for
// demonstration purposes only to show the CUDA WMMA API use without
// relying on availability of the shared memory.
__global__ void simple_wmma_gemm_imma(const uint8_t *a, const uint8_t *b,
const int *c, int *d, int m_ld, int n_ld,
int k_ld, int alpha, int beta) {
// Leading dimensions. Packed with no transpositions.
int lda = m_ld;
int ldb = k_ld;
int ldc = n_ld;
// Tile using a 2D grid
int warpM = (blockIdx.x * blockDim.x + threadIdx.x) / warpSize;
int warpN = (blockIdx.y * blockDim.y + threadIdx.y);
// Declare the fragments
wmma::fragment<wmma::matrix_a, WMMA_M, WMMA_N, WMMA_K, uint8_t,
wmma::row_major>
a_frag;
wmma::fragment<wmma::matrix_b, WMMA_M, WMMA_N, WMMA_K, uint8_t,
wmma::col_major>
b_frag;
wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, int> acc_frag;
wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, int> c_frag;
wmma::fill_fragment(acc_frag, 0.0f);
// Loop over k
for (int i = 0; i < k_ld; i += WMMA_K) {
int aCol = i;
int aRow = warpM * WMMA_M;
int bCol = i;
int bRow = warpN * WMMA_N;
// Bounds checking
if (aRow < m_ld && aCol < k_ld && bRow < k_ld && bCol < n_ld) {
// Load the inputs
wmma::load_matrix_sync(a_frag, a + aCol + aRow * lda, lda);
wmma::load_matrix_sync(b_frag, b + bCol + bRow * ldb, ldb);
// Perform the matrix multiplication
wmma::mma_sync(acc_frag, a_frag, b_frag, acc_frag);
}
}
// Load in the current value of c, scale it by beta, and add this our result
// scaled by alpha
int cCol = warpN * WMMA_N;
int cRow = warpM * WMMA_M;
if (cRow < m_ld && cCol < n_ld) {
wmma::load_matrix_sync(c_frag, c + cCol + cRow * ldc, ldc,
wmma::mem_row_major);
for (int i = 0; i < c_frag.num_elements; i++) {
c_frag.x[i] = alpha * acc_frag.x[i] + beta * c_frag.x[i];
}
// Store the output
wmma::store_matrix_sync(d + cCol + cRow * ldc, c_frag, ldc,
wmma::mem_row_major);
}
}
__host__ void matMultiplyOnHost(uint8_t *A, uint8_t *B, int *C, int alpha,
int beta, int numARows, int numAColumns,
int numBRows, int numBColumns, int numCRows,
int numCColumns) {
for (int i = 0; i < numCRows; i++) {
for (int j = 0; j < numCColumns; j++) {
int temp = 0;
for (int k = 0; k < numAColumns; k++) {
temp += A[i * numAColumns + k] * B[j * numBRows + k];
}
C[i * numCColumns + j] = temp * alpha + beta * C[i * numCColumns + j];
}
}
}
int main(int argc, char **argv) {
printf("Initializing...\n");
int dev = findCudaDevice(argc, (const char **)argv);
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
// Tensor cores require a GPU of Volta (SM72) architecture or higher.
if (deviceProp.major < 7 || (deviceProp.major <= 7 && deviceProp.minor < 2)) {
printf(
"immaTensorCoreGemm requires SM 7.2 or higher to use Tensor Cores. "
"Exiting...\n");
exit(EXIT_WAIVED);
}
printf("M: %d (%d x %d)\n", M_GLOBAL, M, M_TILES);
printf("N: %d (%d x %d)\n", N_GLOBAL, N, N_TILES);
printf("K: %d (%d x %d)\n", K_GLOBAL, K, K_TILES);
uint8_t *A_h = NULL;
uint8_t *B_h = NULL;
int *C_h = NULL;
#if CPU_DEBUG
int *result_hD = NULL;
int *result_host = NULL;
#endif
A_h = (uint8_t *)malloc(sizeof(uint8_t) * M_GLOBAL * K_GLOBAL);
B_h = (uint8_t *)malloc(sizeof(uint8_t) * K_GLOBAL * N_GLOBAL);
C_h = (int *)malloc(sizeof(int) * M_GLOBAL * N_GLOBAL);
#if CPU_DEBUG
result_hD = (int *)malloc(sizeof(int) * M_GLOBAL * N_GLOBAL);
result_host = (int *)malloc(sizeof(int) * M_GLOBAL * N_GLOBAL);
#endif
uint8_t *A = NULL;
uint8_t *B = NULL;
int *C = NULL;
int *D = NULL;
checkCudaErrors(
cudaMalloc(reinterpret_cast<void **>(&A), sizeof(uint8_t) * M_GLOBAL * K_GLOBAL));
checkCudaErrors(
cudaMalloc(reinterpret_cast<void **>(&B), sizeof(uint8_t) * N_GLOBAL * K_GLOBAL));
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&C), sizeof(int) * M_GLOBAL * N_GLOBAL));
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&D), sizeof(int) * M_GLOBAL * N_GLOBAL));
assert(((unsigned long long)A) % 128 == 0);
assert(((unsigned long long)B) % 128 == 0);
assert(((unsigned long long)C) % 128 == 0);
assert(((unsigned long long)D) % 128 == 0);
init_host_matrices(A_h, B_h, C_h);
checkCudaErrors(cudaMemcpy(A, A_h, sizeof(uint8_t) * M_GLOBAL * K_GLOBAL,
cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(B, B_h, sizeof(uint8_t) * N_GLOBAL * K_GLOBAL,
cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(C, C_h, sizeof(int) * M_GLOBAL * N_GLOBAL,
cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemset(D, 0, sizeof(int) * M_GLOBAL * N_GLOBAL));
printf("Preparing data for GPU...\n");
assert(((unsigned long long)A) % 128 == 0);
assert(((unsigned long long)B) % 128 == 0);
assert(((unsigned long long)C) % 128 == 0);
assert(((unsigned long long)D) % 128 == 0);
enum {
// Compute the right amount of shared memory to request.
// We need shared memory to hold per-CTA C and D matrix tiles, and to cache
// per-CTA chunks
// of the A and B matrices. Therefore, the right amount to request is the
// maximum of those
// two numbers.
SHMEM_SZ = MAX(sizeof(uint8_t) * (BLOCK_COL_TILES * M) *
(CHUNK_K * K + SKEW_UINT8) * 2,
M * (BLOCK_ROW_WARPS * WARP_ROW_TILES) * N *
(BLOCK_COL_WARPS * WARP_COL_TILES) * sizeof(int))
};
printf("Required shared memory size: %lu Kb\n", SHMEM_SZ / 1024UL);
int alpha = 1;
int beta = 1;
cudaEvent_t start, stop;
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
checkCudaErrors(cudaEventRecord(start));
// If enough shared memory available on the GPU use high performant kernel
if (deviceProp.sharedMemPerMultiprocessor >= SHMEM_SZ) {
printf("Computing... using high performance kernel compute_gemm_imma \n");
checkCudaErrors(cudaFuncSetAttribute(
compute_gemm_imma, cudaFuncAttributeMaxDynamicSharedMemorySize,
SHMEM_SZ));
checkKernelErrors(
(compute_gemm_imma<<<deviceProp.multiProcessorCount, THREADS_PER_BLOCK,
SHMEM_SZ>>>(A, B, C, D, alpha, beta)));
#if CPU_DEBUG
checkCudaErrors(cudaMemcpy(result_hD, D, sizeof(int) * M_GLOBAL * N_GLOBAL,
cudaMemcpyDeviceToHost));
#endif
} else {
dim3 gridDim;
dim3 blockDim;
// blockDim.x must be a multiple of warpSize
// 128x4 means we have 16 warps and a block computes a 64x64 output tile
blockDim.x = 128;
blockDim.y = 4;
gridDim.x = (M_GLOBAL + (WMMA_M * blockDim.x / 32 - 1)) /
(WMMA_M * blockDim.x / 32);
gridDim.y = (N_GLOBAL + WMMA_N * blockDim.y - 1) / (WMMA_N * blockDim.y);
printf("Computing... using simple_wmma_gemm_imma kernel\n");
simple_wmma_gemm_imma<<<gridDim, blockDim>>>(A, B, C, D, M_GLOBAL, N_GLOBAL,
K_GLOBAL, alpha, beta);
#if CPU_DEBUG
checkCudaErrors(cudaMemcpy(result_hD, D, sizeof(int) * M_GLOBAL * N_GLOBAL,
cudaMemcpyDeviceToHost));
#endif
}
checkCudaErrors(cudaEventRecord(stop));
checkCudaErrors(cudaEventSynchronize(stop));
#if CPU_DEBUG
printf("Verifying correctness of the computations...\n");
memcpy(result_host, C_h, sizeof(int) * M_GLOBAL * N_GLOBAL);
matMultiplyOnHost(A_h, B_h, result_host, alpha, beta, M_GLOBAL, K_GLOBAL,
K_GLOBAL, N_GLOBAL, M_GLOBAL, N_GLOBAL);
for (int i = 0; i < N_GLOBAL * M_GLOBAL; i++) {
if (abs(result_hD[i] - result_host[i]) > 0) {
printf("mismatch i=%d result_hD=%d result_host=%d\n", i, result_hD[i],
result_host[i]);
}
}
free(result_host);
free(result_hD);
#endif
float milliseconds = 0;
checkCudaErrors(cudaEventElapsedTime(&milliseconds, start, stop));
printf("Time: %f ms\n", milliseconds);
printf("TOPS: %.2f\n", (((double)M_GLOBAL * N_GLOBAL * K_GLOBAL * 2)/(milliseconds/1000.)) / 1e12);
free(A_h);
free(B_h);
free(C_h);
checkCudaErrors(cudaFree(reinterpret_cast<void *>(A)));
checkCudaErrors(cudaFree(reinterpret_cast<void *>(B)));
checkCudaErrors(cudaFree(reinterpret_cast<void *>(C)));
checkCudaErrors(cudaFree(reinterpret_cast<void *>(D)));
return EXIT_SUCCESS;
}

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 11.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2010 # Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "cudaTensorCoreGemm", "cudaTensorCoreGemm_vs2010.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "immaTensorCoreGemm", "immaTensorCoreGemm_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,14 +15,15 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>cudaTensorCoreGemm_vs2010</RootNamespace> <RootNamespace>immaTensorCoreGemm_vs2012</RootNamespace>
<ProjectName>cudaTensorCoreGemm</ProjectName> <ProjectName>immaTensorCoreGemm</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v110</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -32,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -58,10 +59,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/cudaTensorCoreGemm.exe</OutputFile> <OutputFile>$(OutDir)/immaTensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -96,11 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="cudaTensorCoreGemm.cu" /> <CudaCompile Include="immaTensorCoreGemm.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "immaTensorCoreGemm", "immaTensorCoreGemm_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,107 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>immaTensorCoreGemm_vs2013</RootNamespace>
<ProjectName>immaTensorCoreGemm</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/immaTensorCoreGemm.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="immaTensorCoreGemm.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "immaTensorCoreGemm", "immaTensorCoreGemm_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,107 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>immaTensorCoreGemm_vs2015</RootNamespace>
<ProjectName>immaTensorCoreGemm</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v140</PlatformToolset>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/immaTensorCoreGemm.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="immaTensorCoreGemm.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,20 @@

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "immaTensorCoreGemm", "immaTensorCoreGemm_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject
Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution
Debug|x64 = Debug|x64
Release|x64 = Release|x64
EndGlobalSection
GlobalSection(ProjectConfigurationPlatforms) = postSolution
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
EndGlobalSection
GlobalSection(SolutionProperties) = preSolution
HideSolutionNode = FALSE
EndGlobalSection
EndGlobal

View File

@@ -0,0 +1,108 @@
<?xml version="1.0" encoding="utf-8"?>
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
<PropertyGroup>
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
</PropertyGroup>
<ItemGroup Label="ProjectConfigurations">
<ProjectConfiguration Include="Debug|x64">
<Configuration>Debug</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
<ProjectConfiguration Include="Release|x64">
<Configuration>Release</Configuration>
<Platform>x64</Platform>
</ProjectConfiguration>
</ItemGroup>
<PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>immaTensorCoreGemm_vs2017</RootNamespace>
<ProjectName>immaTensorCoreGemm</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</PlatformToolset>
<WindowsTargetPlatformVersion>10.0.15063.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries>
</PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Release'">
<WholeProgramOptimization>true</WholeProgramOptimization>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
</ImportGroup>
<PropertyGroup Label="UserMacros" />
<PropertyGroup>
<IntDir>$(Platform)/$(Configuration)/</IntDir>
<IncludePath>$(IncludePath)</IncludePath>
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
<CodeAnalysisRules />
<CodeAnalysisRuleAssemblies />
</PropertyGroup>
<PropertyGroup Condition="'$(Platform)'=='x64'">
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
</PropertyGroup>
<ItemDefinitionGroup>
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/immaTensorCoreGemm.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
<ClCompile>
<Optimization>Disabled</Optimization>
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>true</GenerateDebugInformation>
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MTd</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="immaTensorCoreGemm.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
@@ -248,7 +246,11 @@ LIBRARIES :=
################################################################################ ################################################################################
# Gencode arguments # Gencode arguments
SMS ?= 30 35 37 50 52 60 61 70 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75
else
SMS ?= 30 35 37 50 52 60 61 70 75
endif
ifeq ($(SMS),) ifeq ($(SMS),)
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<) $(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)

View File

@@ -46,6 +46,8 @@
<sm-arch>sm60</sm-arch> <sm-arch>sm60</sm-arch>
<sm-arch>sm61</sm-arch> <sm-arch>sm61</sm-arch>
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,7 +10,7 @@ CUDA Runtime API, Linear Algebra
## Supported SM Architectures ## Supported SM Architectures
[SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.0 ](https://developer.nvidia.com/cuda-gpus) [SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
@@ -27,7 +27,7 @@ cudaEventCreate, cudaEventRecord, cudaEventQuery, cudaEventDestroy, cudaEventEla
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 9.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run ## Build and Run

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/matrixMul.exe</OutputFile> <OutputFile>$(OutDir)/matrixMul.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/matrixMul.exe</OutputFile> <OutputFile>$(OutDir)/matrixMul.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -62,7 +62,7 @@
<OutputFile>$(OutDir)/matrixMul.exe</OutputFile> <OutputFile>$(OutDir)/matrixMul.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -102,6 +102,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -34,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.props" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" /> <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -63,7 +63,7 @@
<OutputFile>$(OutDir)/matrixMul.exe</OutputFile> <OutputFile>$(OutDir)/matrixMul.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;</CodeGeneration> <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -103,6 +103,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 9.2.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,31 +1,29 @@
################################################################################ ################################################################################
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
# #
# Copyright 1993-2015 NVIDIA Corporation. All rights reserved. # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
# #
# NOTICE TO USER: # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# # EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# This source code is subject to NVIDIA ownership rights under U.S. and # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# international Copyright laws. # PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# # CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# NVIDIA MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS SOURCE # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# CODE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESS OR # PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# IMPLIED WARRANTY OF ANY KIND. NVIDIA DISCLAIMS ALL WARRANTIES WITH # PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# REGARD TO THIS SOURCE CODE, INCLUDING ALL IMPLIED WARRANTIES OF # OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# MERCHANTABILITY, NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL, # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# OR CONSEQUENTIAL DAMAGES, OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
# OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
# OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE
# OR PERFORMANCE OF THIS SOURCE CODE.
#
# U.S. Government End Users. This source code is a "commercial item" as
# that term is defined at 48 C.F.R. 2.101 (OCT 1995), consisting of
# "commercial computer software" and "commercial computer software
# documentation" as such terms are used in 48 C.F.R. 12.212 (SEPT 1995)
# and is provided to the U.S. Government only as a commercial end item.
# Consistent with 48 C.F.R.12.212 and 48 C.F.R. 227.7202-1 through
# 227.7202-4 (JUNE 1995), all U.S. Government End Users acquire the
# source code with only those rights set forth herein.
# #
################################################################################ ################################################################################
# #
@@ -141,7 +139,7 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
export QNX_TARGET export QNX_TARGET
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++ HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-g++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++

Some files were not shown because too many files have changed in this diff Show More