13 Commits

Author SHA1 Message Date
Mahesh Doijade
c4e2869a2b add multi-warp cooperative groups based reduction kernel in reduction sample 2020-09-24 16:49:58 +05:30
Mahesh Doijade
dd2dba3489 helper_string: fix the missing "/" at end of 3 up path in sdkFindFilePath() 2020-09-17 20:45:23 +05:30
Mahesh Doijade
44052982a7 add ewp tags to VS project of conjugateGradientMultiDeviceCG 2020-09-17 20:01:41 +05:30
Mahesh Doijade
cd76533c3f Add and update samples for cuda 11.1 support 2020-09-15 23:45:56 +05:30
Mahesh Doijade
e6ce58fef4 [p2pBandwidthLatency] increase default buffer size and add support to specific buffer size through command line 2020-06-30 19:05:55 +05:30
Mahesh Doijade
0ec4bd58e5 cudaCompressibleMemory: when unable to allocate compressible memory waive the execution 2020-06-01 20:22:22 +05:30
Mahesh Doijade
908dddb207 cudaCompressibleMemory: refactor and refine, and use only cuMemMap API for non-compressible allocation.
hence, remove use of cudaMalloc.
2020-05-27 19:00:33 +05:30
Mahesh Doijade
b60c149b17 Correct the runtime API section in documentation when sample has both driver as well as runtime API 2020-05-22 22:33:02 +05:30
Mahesh Doijade
56b5b4c767 -- add cudaCompressibleMemory sample
-- fix simpleD3D12 build issue
-- fix simpleVulkanMMAP shader file locating issue
2020-05-22 22:07:30 +05:30
Mahesh Doijade
53e9e413a8 Update VS project files to use cuda 11.0 props/targets 2020-05-19 18:16:05 +05:30
Mahesh Doijade
4f6e02970b Add and update samples for cuda 11.0 support 2020-05-18 22:22:06 +05:30
Mahesh Doijade
6be514679b Add and update samples for cuda 10.2 support 2019-10-23 19:00:39 +05:30
Mahesh Doijade
489d9f7b1f conjugateGradientCudaGraphs: Add cusparseScsrmv & cublasSdot instead of
gpuSpMV & gpuDotProduct as they are graph compliant
2019-08-19 15:47:24 +05:30
531 changed files with 101273 additions and 8102 deletions

Binary file not shown.

Binary file not shown.

Binary file not shown.

File diff suppressed because one or more lines are too long

Binary file not shown.

Binary file not shown.

Binary file not shown.

File diff suppressed because one or more lines are too long

Binary file not shown.

160
Common/dynlink_d3d11.h Normal file
View File

@@ -0,0 +1,160 @@
/* 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.
*/
//--------------------------------------------------------------------------------------
// File: dynlink_d3d11.h
//
// Shortcut macros and functions for using DX objects
//
// Copyright (c) Microsoft Corporation. All rights reserved
//--------------------------------------------------------------------------------------
#ifndef _DYNLINK_D3D11_H_
#define _DYNLINK_D3D11_H_
// Standard Windows includes
#include <windows.h>
#include <initguid.h>
#include <assert.h>
#include <wchar.h>
#include <mmsystem.h>
#include <commctrl.h> // for InitCommonControls()
#include <shellapi.h> // for ExtractIcon()
#include <new.h> // for placement new
#include <shlobj.h>
#include <math.h>
#include <limits.h>
#include <stdio.h>
// CRT's memory leak detection
#if defined(DEBUG) || defined(_DEBUG)
#include <crtdbg.h>
#endif
// Direct3D10 includes
#include <dxgi.h>
#include <d3d11.h>
// #include <..\Samples\C++\Effects11\Inc\d3dx11effect.h>
// XInput includes
#include <xinput.h>
// strsafe.h deprecates old unsecure string functions. If you
// really do not want to it to (not recommended), then uncomment the next line
//#define STRSAFE_NO_DEPRECATE
#ifndef STRSAFE_NO_DEPRECATE
#pragma deprecated("strncpy")
#pragma deprecated("wcsncpy")
#pragma deprecated("_tcsncpy")
#pragma deprecated("wcsncat")
#pragma deprecated("strncat")
#pragma deprecated("_tcsncat")
#endif
#pragma warning( disable : 4996 ) // disable deprecated warning
#include <strsafe.h>
#pragma warning( default : 4996 )
typedef HRESULT(WINAPI *LPCREATEDXGIFACTORY)(REFIID, void **);
typedef HRESULT(WINAPI *LPD3D11CREATEDEVICEANDSWAPCHAIN)(__in_opt IDXGIAdapter *pAdapter, D3D_DRIVER_TYPE DriverType, HMODULE Software, UINT Flags, __in_ecount_opt(FeatureLevels) CONST D3D_FEATURE_LEVEL *pFeatureLevels, UINT FeatureLevels, UINT SDKVersion, __in_opt CONST DXGI_SWAP_CHAIN_DESC *pSwapChainDesc, __out_opt IDXGISwapChain **ppSwapChain, __out_opt ID3D11Device **ppDevice, __out_opt D3D_FEATURE_LEVEL *pFeatureLevel, __out_opt ID3D11DeviceContext **ppImmediateContext);
typedef HRESULT(WINAPI *LPD3D11CREATEDEVICE)(IDXGIAdapter *, D3D_DRIVER_TYPE, HMODULE, UINT32, D3D_FEATURE_LEVEL *, UINT, UINT32, ID3D11Device **, D3D_FEATURE_LEVEL *, ID3D11DeviceContext **);
static HMODULE s_hModDXGI = NULL;
static LPCREATEDXGIFACTORY sFnPtr_CreateDXGIFactory = NULL;
static HMODULE s_hModD3D11 = NULL;
static LPD3D11CREATEDEVICE sFnPtr_D3D11CreateDevice = NULL;
static LPD3D11CREATEDEVICEANDSWAPCHAIN sFnPtr_D3D11CreateDeviceAndSwapChain = NULL;
// unload the D3D10 DLLs
static bool dynlinkUnloadD3D11API(void)
{
if (s_hModDXGI)
{
FreeLibrary(s_hModDXGI);
s_hModDXGI = NULL;
}
if (s_hModD3D11)
{
FreeLibrary(s_hModD3D11);
s_hModD3D11 = NULL;
}
return true;
}
// Dynamically load the D3D11 DLLs loaded and map the function pointers
static bool dynlinkLoadD3D11API(void)
{
// If both modules are non-NULL, this function has already been called. Note
// that this doesn't guarantee that all ProcAddresses were found.
if (s_hModD3D11 != NULL && s_hModDXGI != NULL)
{
return true;
}
#if 1
// This may fail if Direct3D 11 isn't installed
s_hModD3D11 = LoadLibrary("d3d11.dll");
if (s_hModD3D11 != NULL)
{
sFnPtr_D3D11CreateDevice = (LPD3D11CREATEDEVICE)GetProcAddress(s_hModD3D11, "D3D11CreateDevice");
sFnPtr_D3D11CreateDeviceAndSwapChain = (LPD3D11CREATEDEVICEANDSWAPCHAIN)GetProcAddress(s_hModD3D11, "D3D11CreateDeviceAndSwapChain");
}
else
{
printf("\nLoad d3d11.dll failed\n");
fflush(0);
}
if (!sFnPtr_CreateDXGIFactory)
{
s_hModDXGI = LoadLibrary("dxgi.dll");
if (s_hModDXGI)
{
sFnPtr_CreateDXGIFactory = (LPCREATEDXGIFACTORY)GetProcAddress(s_hModDXGI, "CreateDXGIFactory1");
}
return (s_hModDXGI != NULL) && (s_hModD3D11 != NULL);
}
return (s_hModD3D11 != NULL);
#else
sFnPtr_D3D11CreateDevice = (LPD3D11CREATEDEVICE)D3D11CreateDeviceAndSwapChain;
sFnPtr_D3D11CreateDeviceAndSwapChain = (LPD3D11CREATEDEVICEANDSWAPCHAIN)D3D11CreateDeviceAndSwapChain;
//sFnPtr_D3DX11CreateEffectFromMemory = ( LPD3DX11CREATEEFFECTFROMMEMORY )D3DX11CreateEffectFromMemory;
sFnPtr_D3DX11CompileFromMemory = (LPD3DX11COMPILEFROMMEMORY)D3DX11CompileFromMemory;
sFnPtr_CreateDXGIFactory = (LPCREATEDXGIFACTORY)CreateDXGIFactory;
return true;
#endif
return true;
}
#endif

View File

@@ -579,24 +579,12 @@ static const char *_cudaGetErrorEnum(NppStatus error) {
} }
#endif #endif
#ifdef __DRIVER_TYPES_H__
#ifndef DEVICE_RESET
#define DEVICE_RESET cudaDeviceReset();
#endif
#else
#ifndef DEVICE_RESET
#define DEVICE_RESET
#endif
#endif
template <typename T> template <typename T>
void check(T result, char const *const func, const char *const file, void check(T result, char const *const func, const char *const file,
int const line) { int const line) {
if (result) { if (result) {
fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \"%s\" \n", file, line, fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \"%s\" \n", file, line,
static_cast<unsigned int>(result), _cudaGetErrorEnum(result), func); static_cast<unsigned int>(result), _cudaGetErrorEnum(result), func);
DEVICE_RESET
// Make sure we call CUDA Device Reset before exiting
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
} }
@@ -619,7 +607,6 @@ inline void __getLastCudaError(const char *errorMessage, const char *file,
" %s : (%d) %s.\n", " %s : (%d) %s.\n",
file, line, errorMessage, static_cast<int>(err), file, line, errorMessage, static_cast<int>(err),
cudaGetErrorString(err)); cudaGetErrorString(err));
DEVICE_RESET
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
} }
@@ -676,6 +663,8 @@ inline int _ConvertSMVer2Cores(int major, int minor) {
{0x70, 64}, {0x70, 64},
{0x72, 64}, {0x72, 64},
{0x75, 64}, {0x75, 64},
{0x80, 64},
{0x86, 128},
{-1, -1}}; {-1, -1}};
int index = 0; int index = 0;
@@ -696,6 +685,52 @@ inline int _ConvertSMVer2Cores(int major, int minor) {
major, minor, nGpuArchCoresPerSM[index - 1].Cores); major, minor, nGpuArchCoresPerSM[index - 1].Cores);
return nGpuArchCoresPerSM[index - 1].Cores; return nGpuArchCoresPerSM[index - 1].Cores;
} }
inline const char* _ConvertSMVer2ArchName(int major, int minor) {
// Defines for GPU Architecture types (using the SM version to determine
// the GPU Arch name)
typedef struct {
int SM; // 0xMm (hexidecimal notation), M = SM Major version,
// and m = SM minor version
const char* name;
} sSMtoArchName;
sSMtoArchName nGpuArchNameSM[] = {
{0x30, "Kepler"},
{0x32, "Kepler"},
{0x35, "Kepler"},
{0x37, "Kepler"},
{0x50, "Maxwell"},
{0x52, "Maxwell"},
{0x53, "Maxwell"},
{0x60, "Pascal"},
{0x61, "Pascal"},
{0x62, "Pascal"},
{0x70, "Volta"},
{0x72, "Xavier"},
{0x75, "Turing"},
{0x80, "Ampere"},
{0x86, "Ampere"},
{-1, "Graphics Device"}};
int index = 0;
while (nGpuArchNameSM[index].SM != -1) {
if (nGpuArchNameSM[index].SM == ((major << 4) + minor)) {
return nGpuArchNameSM[index].name;
}
index++;
}
// If we don't find the values, we default use the previous one
// to run properly
printf(
"MapSMtoArchName for SM %d.%d is undefined."
" Default to use %s\n",
major, minor, nGpuArchNameSM[index - 1].name);
return nGpuArchNameSM[index - 1].name;
}
// end of GPU Architecture definitions // end of GPU Architecture definitions
#ifdef __CUDA_RUNTIME_H__ #ifdef __CUDA_RUNTIME_H__
@@ -727,23 +762,24 @@ inline int gpuDeviceInit(int devID) {
return -devID; return -devID;
} }
cudaDeviceProp deviceProp; int computeMode = -1, major = 0, minor = 0;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID)); checkCudaErrors(cudaDeviceGetAttribute(&computeMode, cudaDevAttrComputeMode, devID));
checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, devID));
if (deviceProp.computeMode == cudaComputeModeProhibited) { checkCudaErrors(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, devID));
if (computeMode == cudaComputeModeProhibited) {
fprintf(stderr, fprintf(stderr,
"Error: device is running in <Compute Mode " "Error: device is running in <Compute Mode "
"Prohibited>, no threads can use cudaSetDevice().\n"); "Prohibited>, no threads can use cudaSetDevice().\n");
return -1; return -1;
} }
if (deviceProp.major < 1) { if (major < 1) {
fprintf(stderr, "gpuDeviceInit(): GPU device does not support CUDA.\n"); fprintf(stderr, "gpuDeviceInit(): GPU device does not support CUDA.\n");
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
checkCudaErrors(cudaSetDevice(devID)); checkCudaErrors(cudaSetDevice(devID));
printf("gpuDeviceInit() CUDA Device [%d]: \"%s\n", devID, deviceProp.name); printf("gpuDeviceInit() CUDA Device [%d]: \"%s\n", devID, _ConvertSMVer2ArchName(major, minor));
return devID; return devID;
} }
@@ -756,7 +792,6 @@ inline int gpuGetMaxGflopsDeviceId() {
int devices_prohibited = 0; int devices_prohibited = 0;
uint64_t max_compute_perf = 0; uint64_t max_compute_perf = 0;
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceCount(&device_count)); checkCudaErrors(cudaGetDeviceCount(&device_count));
if (device_count == 0) { if (device_count == 0) {
@@ -770,20 +805,36 @@ inline int gpuGetMaxGflopsDeviceId() {
current_device = 0; current_device = 0;
while (current_device < device_count) { while (current_device < device_count) {
cudaGetDeviceProperties(&deviceProp, current_device); int computeMode = -1, major = 0, minor = 0;
checkCudaErrors(cudaDeviceGetAttribute(&computeMode, cudaDevAttrComputeMode, current_device));
checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, current_device));
checkCudaErrors(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, current_device));
// If this GPU is not running on Compute Mode prohibited, // If this GPU is not running on Compute Mode prohibited,
// then we can add it to the list // then we can add it to the list
if (deviceProp.computeMode != cudaComputeModeProhibited) { if (computeMode != cudaComputeModeProhibited) {
if (deviceProp.major == 9999 && deviceProp.minor == 9999) { if (major == 9999 && minor == 9999) {
sm_per_multiproc = 1; sm_per_multiproc = 1;
} else { } else {
sm_per_multiproc = sm_per_multiproc =
_ConvertSMVer2Cores(deviceProp.major, deviceProp.minor); _ConvertSMVer2Cores(major, minor);
} }
int multiProcessorCount = 0, clockRate = 0;
uint64_t compute_perf = (uint64_t)deviceProp.multiProcessorCount * checkCudaErrors(cudaDeviceGetAttribute(&multiProcessorCount, cudaDevAttrMultiProcessorCount, current_device));
sm_per_multiproc * deviceProp.clockRate; cudaError_t result = cudaDeviceGetAttribute(&clockRate, cudaDevAttrClockRate, current_device);
if (result != cudaSuccess) {
// If cudaDevAttrClockRate attribute is not supported we
// set clockRate as 1, to consider GPU with most SMs and CUDA Cores.
if(result == cudaErrorInvalidValue) {
clockRate = 1;
}
else {
fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \n", __FILE__, __LINE__,
static_cast<unsigned int>(result), _cudaGetErrorEnum(result));
exit(EXIT_FAILURE);
}
}
uint64_t compute_perf = (uint64_t)multiProcessorCount * sm_per_multiproc * clockRate;
if (compute_perf > max_compute_perf) { if (compute_perf > max_compute_perf) {
max_compute_perf = compute_perf; max_compute_perf = compute_perf;
@@ -808,7 +859,6 @@ inline int gpuGetMaxGflopsDeviceId() {
// Initialization code to find the best CUDA Device // Initialization code to find the best CUDA Device
inline int findCudaDevice(int argc, const char **argv) { inline int findCudaDevice(int argc, const char **argv) {
cudaDeviceProp deviceProp;
int devID = 0; int devID = 0;
// If the command-line has a device number specified, use it // If the command-line has a device number specified, use it
@@ -830,9 +880,12 @@ inline int findCudaDevice(int argc, const char **argv) {
// Otherwise pick the device with highest Gflops/s // Otherwise pick the device with highest Gflops/s
devID = gpuGetMaxGflopsDeviceId(); devID = gpuGetMaxGflopsDeviceId();
checkCudaErrors(cudaSetDevice(devID)); checkCudaErrors(cudaSetDevice(devID));
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID)); int major = 0, minor = 0;
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", devID, checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, devID));
deviceProp.name, deviceProp.major, deviceProp.minor); checkCudaErrors(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, devID));
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n",
devID, _ConvertSMVer2ArchName(major, minor), major, minor);
} }
return devID; return devID;
@@ -843,7 +896,6 @@ inline int findIntegratedGPU() {
int device_count = 0; int device_count = 0;
int devices_prohibited = 0; int devices_prohibited = 0;
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceCount(&device_count)); checkCudaErrors(cudaGetDeviceCount(&device_count));
if (device_count == 0) { if (device_count == 0) {
@@ -853,17 +905,19 @@ inline int findIntegratedGPU() {
// Find the integrated GPU which is compute capable // Find the integrated GPU which is compute capable
while (current_device < device_count) { while (current_device < device_count) {
cudaGetDeviceProperties(&deviceProp, current_device); int computeMode = -1, integrated = -1;
checkCudaErrors(cudaDeviceGetAttribute(&computeMode, cudaDevAttrComputeMode, current_device));
checkCudaErrors(cudaDeviceGetAttribute(&integrated, cudaDevAttrIntegrated, current_device));
// If GPU is integrated and is not running on Compute Mode prohibited, // If GPU is integrated and is not running on Compute Mode prohibited,
// then cuda can map to GLES resource // then cuda can map to GLES resource
if (deviceProp.integrated && if (integrated && (computeMode != cudaComputeModeProhibited)) {
(deviceProp.computeMode != cudaComputeModeProhibited)) {
checkCudaErrors(cudaSetDevice(current_device)); checkCudaErrors(cudaSetDevice(current_device));
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, current_device));
int major = 0, minor = 0;
checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, current_device));
checkCudaErrors(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, current_device));
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n",
current_device, deviceProp.name, deviceProp.major, current_device, _ConvertSMVer2ArchName(major, minor), major, minor);
deviceProp.minor);
return current_device; return current_device;
} else { } else {
@@ -885,19 +939,18 @@ inline int findIntegratedGPU() {
// General check for CUDA GPU SM Capabilities // General check for CUDA GPU SM Capabilities
inline bool checkCudaCapabilities(int major_version, int minor_version) { inline bool checkCudaCapabilities(int major_version, int minor_version) {
cudaDeviceProp deviceProp;
deviceProp.major = 0;
deviceProp.minor = 0;
int dev; int dev;
int major = 0, minor = 0;
checkCudaErrors(cudaGetDevice(&dev)); checkCudaErrors(cudaGetDevice(&dev));
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev)); checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, dev));
checkCudaErrors(cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, dev));
if ((deviceProp.major > major_version) || if ((major > major_version) ||
(deviceProp.major == major_version && (major == major_version &&
deviceProp.minor >= minor_version)) { minor >= minor_version)) {
printf(" Device %d: <%16s >, Compute SM %d.%d detected\n", dev, printf(" Device %d: <%16s >, Compute SM %d.%d detected\n", dev,
deviceProp.name, deviceProp.major, deviceProp.minor); _ConvertSMVer2ArchName(major, minor), major, minor);
return true; return true;
} else { } else {
printf( printf(

View File

@@ -33,8 +33,10 @@
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
#include <string.h> #include <string.h>
#include <iostream>
#include <cstring>
#include <sstream>
#include <drvapi_error_string.h>
#include <helper_string.h> #include <helper_string.h>
#ifndef MAX #ifndef MAX
@@ -66,48 +68,22 @@ inline int ftoi(float value) {
// These are the inline versions for all of the SDK helper functions // These are the inline versions for all of the SDK helper functions
inline void __checkCudaErrors(CUresult err, const char *file, const int line) { inline void __checkCudaErrors(CUresult err, const char *file, const int line) {
if (CUDA_SUCCESS != err) { if (CUDA_SUCCESS != err) {
const char *errorStr = NULL;
cuGetErrorString(err, &errorStr);
fprintf(stderr, fprintf(stderr,
"checkCudaErrors() Driver API error = %04d \"%s\" from file <%s>, " "checkCudaErrors() Driver API error = %04d \"%s\" from file <%s>, "
"line %i.\n", "line %i.\n",
err, getCudaDrvErrorString(err), file, line); err, errorStr, file, line);
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
} }
#endif #endif
#ifdef getLastCudaDrvErrorMsg
#undef getLastCudaDrvErrorMsg
#endif
#define getLastCudaDrvErrorMsg(msg) \
__getLastCudaDrvErrorMsg(msg, __FILE__, __LINE__)
inline void __getLastCudaDrvErrorMsg(const char *msg, const char *file,
const int line) {
CUresult err = cuCtxSynchronize();
if (CUDA_SUCCESS != err) {
fprintf(stderr, "getLastCudaDrvErrorMsg -> %s", msg);
fprintf(stderr,
"getLastCudaDrvErrorMsg -> cuCtxSynchronize API error = %04d "
"\"%s\" in file <%s>, line %i.\n",
err, getCudaDrvErrorString(err), file, line);
exit(EXIT_FAILURE);
}
}
// This function wraps the CUDA Driver API into a template function // This function wraps the CUDA Driver API into a template function
template <class T> template <class T>
inline void getCudaAttribute(T *attribute, CUdevice_attribute device_attribute, inline void getCudaAttribute(T *attribute, CUdevice_attribute device_attribute,
int device) { int device) {
CUresult error_result = checkCudaErrors(cuDeviceGetAttribute(attribute, device_attribute, device));
cuDeviceGetAttribute(attribute, device_attribute, device);
if (error_result != CUDA_SUCCESS) {
printf("cuDeviceGetAttribute returned %d\n-> %s\n",
static_cast<int>(error_result), getCudaDrvErrorString(error_result));
exit(EXIT_SUCCESS);
}
} }
#endif #endif
@@ -135,6 +111,8 @@ inline int _ConvertSMVer2CoresDRV(int major, int minor) {
{0x70, 64}, {0x70, 64},
{0x72, 64}, {0x72, 64},
{0x75, 64}, {0x75, 64},
{0x80, 64},
{0x86, 128},
{-1, -1}}; {-1, -1}};
int index = 0; int index = 0;
@@ -161,11 +139,9 @@ inline int _ConvertSMVer2CoresDRV(int major, int minor) {
inline int gpuDeviceInitDRV(int ARGC, const char **ARGV) { inline int gpuDeviceInitDRV(int ARGC, const char **ARGV) {
int cuDevice = 0; int cuDevice = 0;
int deviceCount = 0; int deviceCount = 0;
CUresult err = cuInit(0); checkCudaErrors(cuInit(0));
if (CUDA_SUCCESS == err) {
checkCudaErrors(cuDeviceGetCount(&deviceCount)); checkCudaErrors(cuDeviceGetCount(&deviceCount));
}
if (deviceCount == 0) { if (deviceCount == 0) {
fprintf(stderr, "cudaDeviceInit error: no devices supporting CUDA\n"); fprintf(stderr, "cudaDeviceInit error: no devices supporting CUDA\n");
@@ -192,7 +168,7 @@ inline int gpuDeviceInitDRV(int ARGC, const char **ARGV) {
checkCudaErrors(cuDeviceGet(&cuDevice, dev)); checkCudaErrors(cuDeviceGet(&cuDevice, dev));
char name[100]; char name[100];
cuDeviceGetName(name, 100, cuDevice); checkCudaErrors(cuDeviceGetName(name, 100, cuDevice));
int computeMode; int computeMode;
getCudaAttribute<int>(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, dev); getCudaAttribute<int>(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, dev);
@@ -218,7 +194,6 @@ inline int gpuGetMaxGflopsDeviceIdDRV() {
int device_count = 0; int device_count = 0;
int sm_per_multiproc = 0; int sm_per_multiproc = 0;
unsigned long long max_compute_perf = 0; unsigned long long max_compute_perf = 0;
int best_SM_arch = 0;
int major = 0; int major = 0;
int minor = 0; int minor = 0;
int multiProcessorCount; int multiProcessorCount;
@@ -234,19 +209,6 @@ inline int gpuGetMaxGflopsDeviceIdDRV() {
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
// Find the best major SM Architecture GPU device
while (current_device < device_count) {
checkCudaErrors(cuDeviceGetAttribute(
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, current_device));
checkCudaErrors(cuDeviceGetAttribute(
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, current_device));
if (major > 0 && major < 9999) {
best_SM_arch = MAX(best_SM_arch, major);
}
current_device++;
}
// Find the best CUDA capable GPU device // Find the best CUDA capable GPU device
current_device = 0; current_device = 0;
@@ -277,18 +239,9 @@ inline int gpuGetMaxGflopsDeviceIdDRV() {
clockRate); clockRate);
if (compute_perf > max_compute_perf) { if (compute_perf > max_compute_perf) {
// If we find GPU with SM major > 2, search only these
if (best_SM_arch > 2) {
// If our device==dest_SM_arch, choose this, or else pass
if (major == best_SM_arch) {
max_compute_perf = compute_perf; max_compute_perf = compute_perf;
max_perf_device = current_device; max_perf_device = current_device;
} }
} else {
max_compute_perf = compute_perf;
max_perf_device = current_device;
}
}
} else { } else {
devices_prohibited++; devices_prohibited++;
} }
@@ -414,7 +367,39 @@ inline bool checkCudaCapabilitiesDRV(int major_version, int minor_version,
} }
} }
#endif #endif
bool inline findFatbinPath(const char *module_file, std::string &module_path, char **argv, std::ostringstream &ostrm)
{
char *actual_path = sdkFindFilePath(module_file, argv[0]);
// end of CUDA Helper Functions if (actual_path)
{
module_path = actual_path;
}
else
{
printf("> findModulePath file not found: <%s> \n", module_file);
return false;
}
if (module_path.empty())
{
printf("> findModulePath could not find file: <%s> \n", module_file);
return false;
}
else
{
printf("> findModulePath found file at <%s>\n", module_path.c_str());
if (module_path.rfind("fatbin") != std::string::npos)
{
std::ifstream fileIn(module_path.c_str(), std::ios::binary);
ostrm << fileIn.rdbuf();
fileIn.close();
}
return true;
}
}
// end of CUDA Helper Functions
#endif // COMMON_HELPER_CUDA_DRVAPI_H_ #endif // COMMON_HELPER_CUDA_DRVAPI_H_

View File

@@ -29,16 +29,11 @@
#include <cstdlib> #include <cstdlib>
#include <string> #include <string>
int sharedMemoryCreate(const char *name, size_t sz, sharedMemoryInfo *info) int sharedMemoryCreate(const char *name, size_t sz, sharedMemoryInfo *info) {
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
info->size = sz; info->size = sz;
info->shmHandle = CreateFileMapping(INVALID_HANDLE_VALUE, info->shmHandle = CreateFileMapping(INVALID_HANDLE_VALUE, NULL,
NULL, PAGE_READWRITE, 0, (DWORD)sz, name);
PAGE_READWRITE,
0,
(DWORD)sz,
name);
if (info->shmHandle == 0) { if (info->shmHandle == 0) {
return GetLastError(); return GetLastError();
} }
@@ -73,8 +68,7 @@ int sharedMemoryCreate(const char *name, size_t sz, sharedMemoryInfo *info)
#endif #endif
} }
int sharedMemoryOpen(const char *name, size_t sz, sharedMemoryInfo *info) int sharedMemoryOpen(const char *name, size_t sz, sharedMemoryInfo *info) {
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
info->size = sz; info->size = sz;
@@ -106,8 +100,7 @@ int sharedMemoryOpen(const char *name, size_t sz, sharedMemoryInfo *info)
#endif #endif
} }
void sharedMemoryClose(sharedMemoryInfo *info) void sharedMemoryClose(sharedMemoryInfo *info) {
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
if (info->addr) { if (info->addr) {
UnmapViewOfFile(info->addr); UnmapViewOfFile(info->addr);
@@ -125,8 +118,7 @@ void sharedMemoryClose(sharedMemoryInfo *info)
#endif #endif
} }
int spawnProcess(Process *process, const char *app, char * const *args) int spawnProcess(Process *process, const char *app, char *const *args) {
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
STARTUPINFO si = {0}; STARTUPINFO si = {0};
BOOL status; BOOL status;
@@ -140,7 +132,8 @@ int spawnProcess(Process *process, const char *app, char * const *args)
args++; args++;
} }
status = CreateProcess(app, LPSTR(arg_string.c_str()), NULL, NULL, FALSE, 0, NULL, NULL, &si, process); status = CreateProcess(app, LPSTR(arg_string.c_str()), NULL, NULL, FALSE, 0,
NULL, NULL, &si, process);
return status ? 0 : GetLastError(); return status ? 0 : GetLastError();
#else #else
@@ -149,16 +142,14 @@ int spawnProcess(Process *process, const char *app, char * const *args)
if (0 > execvp(app, args)) { if (0 > execvp(app, args)) {
return errno; return errno;
} }
} } else if (*process < 0) {
else if (*process < 0) {
return errno; return errno;
} }
return 0; return 0;
#endif #endif
} }
int waitProcess(Process *process) int waitProcess(Process *process) {
{
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
DWORD exitCode; DWORD exitCode;
WaitForSingleObject(process->hProcess, INFINITE); WaitForSingleObject(process->hProcess, INFINITE);
@@ -176,3 +167,377 @@ int waitProcess(Process *process)
return WEXITSTATUS(status); return WEXITSTATUS(status);
#endif #endif
} }
#if defined(__linux__)
int ipcCreateSocket(ipcHandle *&handle, const char *name,
const std::vector<Process> &processes) {
int server_fd;
struct sockaddr_un servaddr;
handle = new ipcHandle;
memset(handle, 0, sizeof(*handle));
handle->socket = -1;
handle->socketName = NULL;
// Creating socket file descriptor
if ((server_fd = socket(AF_UNIX, SOCK_DGRAM, 0)) == 0) {
perror("IPC failure: Socket creation failed");
return -1;
}
unlink(name);
bzero(&servaddr, sizeof(servaddr));
servaddr.sun_family = AF_UNIX;
size_t len = strlen(name);
if (len > (sizeof(servaddr.sun_path) - 1)) {
perror("IPC failure: Cannot bind provided name to socket. Name too large");
return -1;
}
strncpy(servaddr.sun_path, name, len);
if (bind(server_fd, (struct sockaddr *)&servaddr, SUN_LEN(&servaddr)) < 0) {
perror("IPC failure: Binding socket failed");
return -1;
}
handle->socketName = new char[strlen(name) + 1];
strcpy(handle->socketName, name);
handle->socket = server_fd;
return 0;
}
int ipcOpenSocket(ipcHandle *&handle) {
int sock = 0;
struct sockaddr_un cliaddr;
handle = new ipcHandle;
memset(handle, 0, sizeof(*handle));
if ((sock = socket(AF_UNIX, SOCK_DGRAM, 0)) < 0) {
perror("IPC failure:Socket creation error");
return -1;
}
bzero(&cliaddr, sizeof(cliaddr));
cliaddr.sun_family = AF_UNIX;
char temp[10];
// Create unique name for the socket.
sprintf(temp, "%u", getpid());
strcpy(cliaddr.sun_path, temp);
if (bind(sock, (struct sockaddr *)&cliaddr, sizeof(cliaddr)) < 0) {
perror("IPC failure: Binding socket failed");
return -1;
}
handle->socket = sock;
handle->socketName = new char[strlen(temp) + 1];
strcpy(handle->socketName, temp);
return 0;
}
int ipcCloseSocket(ipcHandle *handle) {
if (!handle) {
return -1;
}
if (handle->socketName) {
unlink(handle->socketName);
delete[] handle->socketName;
}
close(handle->socket);
delete handle;
return 0;
}
int ipcRecvShareableHandle(ipcHandle *handle, ShareableHandle *shHandle) {
struct msghdr msg = {0};
struct iovec iov[1];
struct cmsghdr cm;
// Union to guarantee alignment requirements for control array
union {
struct cmsghdr cm;
char control[CMSG_SPACE(sizeof(int))];
} control_un;
struct cmsghdr *cmptr;
ssize_t n;
int receivedfd;
char dummy_buffer[1];
ssize_t sendResult;
msg.msg_control = control_un.control;
msg.msg_controllen = sizeof(control_un.control);
iov[0].iov_base = (void *)dummy_buffer;
iov[0].iov_len = sizeof(dummy_buffer);
msg.msg_iov = iov;
msg.msg_iovlen = 1;
if ((n = recvmsg(handle->socket, &msg, 0)) <= 0) {
perror("IPC failure: Receiving data over socket failed");
return -1;
}
if (((cmptr = CMSG_FIRSTHDR(&msg)) != NULL) &&
(cmptr->cmsg_len == CMSG_LEN(sizeof(int)))) {
if ((cmptr->cmsg_level != SOL_SOCKET) || (cmptr->cmsg_type != SCM_RIGHTS)) {
return -1;
}
memmove(&receivedfd, CMSG_DATA(cmptr), sizeof(receivedfd));
*(int *)shHandle = receivedfd;
} else {
return -1;
}
return 0;
}
int ipcRecvDataFromClient(ipcHandle *serverHandle, void *data, size_t size) {
ssize_t readResult;
struct sockaddr_un cliaddr;
socklen_t len = sizeof(cliaddr);
readResult = recvfrom(serverHandle->socket, data, size, 0,
(struct sockaddr *)&cliaddr, &len);
if (readResult == -1) {
perror("IPC failure: Receiving data over socket failed");
return -1;
}
return 0;
}
int ipcSendDataToServer(ipcHandle *handle, const char *serverName,
const void *data, size_t size) {
ssize_t sendResult;
struct sockaddr_un serveraddr;
bzero(&serveraddr, sizeof(serveraddr));
serveraddr.sun_family = AF_UNIX;
strncpy(serveraddr.sun_path, serverName, sizeof(serveraddr.sun_path) - 1);
sendResult = sendto(handle->socket, data, size, 0,
(struct sockaddr *)&serveraddr, sizeof(serveraddr));
if (sendResult <= 0) {
perror("IPC failure: Sending data over socket failed");
}
return 0;
}
int ipcSendShareableHandle(ipcHandle *handle,
const std::vector<ShareableHandle> &shareableHandles,
Process process, int data) {
struct msghdr msg;
struct iovec iov[1];
union {
struct cmsghdr cm;
char control[CMSG_SPACE(sizeof(int))];
} control_un;
struct cmsghdr *cmptr;
ssize_t readResult;
struct sockaddr_un cliaddr;
socklen_t len = sizeof(cliaddr);
// Construct client address to send this SHareable handle to
bzero(&cliaddr, sizeof(cliaddr));
cliaddr.sun_family = AF_UNIX;
char temp[10];
sprintf(temp, "%u", process);
strcpy(cliaddr.sun_path, temp);
len = sizeof(cliaddr);
// Send corresponding shareable handle to the client
int sendfd = (int)shareableHandles[data];
msg.msg_control = control_un.control;
msg.msg_controllen = sizeof(control_un.control);
cmptr = CMSG_FIRSTHDR(&msg);
cmptr->cmsg_len = CMSG_LEN(sizeof(int));
cmptr->cmsg_level = SOL_SOCKET;
cmptr->cmsg_type = SCM_RIGHTS;
memmove(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
msg.msg_name = (void *)&cliaddr;
msg.msg_namelen = sizeof(struct sockaddr_un);
iov[0].iov_base = (void *)"";
iov[0].iov_len = 1;
msg.msg_iov = iov;
msg.msg_iovlen = 1;
ssize_t sendResult = sendmsg(handle->socket, &msg, 0);
if (sendResult <= 0) {
perror("IPC failure: Sending data over socket failed");
return -1;
}
return 0;
}
int ipcSendShareableHandles(
ipcHandle *handle, const std::vector<ShareableHandle> &shareableHandles,
const std::vector<Process> &processes) {
// Send all shareable handles to every single process.
for (int i = 0; i < shareableHandles.size(); i++) {
for (int j = 0; j < processes.size(); j++) {
checkIpcErrors(
ipcSendShareableHandle(handle, shareableHandles, processes[j], i));
}
}
return 0;
}
int ipcRecvShareableHandles(ipcHandle *handle,
std::vector<ShareableHandle> &shareableHandles) {
for (int i = 0; i < shareableHandles.size(); i++) {
checkIpcErrors(ipcRecvShareableHandle(handle, &shareableHandles[i]));
}
return 0;
}
int ipcCloseShareableHandle(ShareableHandle shHandle) {
return close(shHandle);
}
#elif defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
// Generic name to build individual Mailslot names by appending process ids.
LPTSTR SlotName = (LPTSTR)TEXT("\\\\.\\mailslot\\sample_mailslot_");
int ipcCreateSocket(ipcHandle *&handle, const char *name,
const std::vector<Process> &processes) {
handle = new ipcHandle;
handle->hMailslot.resize(processes.size());
// Open Mailslots of all clients and store respective handles.
for (int i = 0; i < handle->hMailslot.size(); ++i) {
std::basic_string<TCHAR> childSlotName(SlotName);
char tempBuf[20];
_itoa_s(processes[i].dwProcessId, tempBuf, 10);
childSlotName += TEXT(tempBuf);
HANDLE hFile =
CreateFile(TEXT(childSlotName.c_str()), GENERIC_WRITE, FILE_SHARE_READ,
(LPSECURITY_ATTRIBUTES)NULL, OPEN_EXISTING,
FILE_ATTRIBUTE_NORMAL, (HANDLE)NULL);
if (hFile == INVALID_HANDLE_VALUE) {
printf("IPC failure: Opening Mailslot by CreateFile failed with %d\n",
GetLastError());
return -1;
}
handle->hMailslot[i] = hFile;
}
return 0;
}
int ipcOpenSocket(ipcHandle *&handle) {
handle = new ipcHandle;
HANDLE hSlot;
std::basic_string<TCHAR> clientSlotName(SlotName);
char tempBuf[20];
_itoa_s(GetCurrentProcessId(), tempBuf, 10);
clientSlotName += TEXT(tempBuf);
hSlot = CreateMailslot((LPSTR)clientSlotName.c_str(), 0,
MAILSLOT_WAIT_FOREVER, (LPSECURITY_ATTRIBUTES)NULL);
if (hSlot == INVALID_HANDLE_VALUE) {
printf("IPC failure: CreateMailslot failed for client with %d\n",
GetLastError());
return -1;
}
handle->hMailslot.push_back(hSlot);
return 0;
}
int ipcSendData(HANDLE mailslot, const void *data, size_t sz) {
BOOL result;
DWORD cbWritten;
result = WriteFile(mailslot, data, (DWORD)sz, &cbWritten, (LPOVERLAPPED)NULL);
if (!result) {
printf("IPC failure: WriteFile failed with %d.\n", GetLastError());
return -1;
}
return 0;
}
int ipcRecvData(ipcHandle *handle, void *data, size_t sz) {
DWORD cbRead = 0;
if (!ReadFile(handle->hMailslot[0], data, (DWORD)sz, &cbRead, NULL)) {
printf("IPC failure: ReadFile failed with %d.\n", GetLastError());
return -1;
}
if (sz != (size_t)cbRead) {
printf(
"IPC failure: ReadFile didn't receive the expected number of bytes\n");
return -1;
}
return 0;
}
int ipcSendShareableHandles(
ipcHandle *handle, const std::vector<ShareableHandle> &shareableHandles,
const std::vector<Process> &processes) {
// Send all shareable handles to every single process.
for (int i = 0; i < processes.size(); i++) {
HANDLE hProcess =
OpenProcess(PROCESS_DUP_HANDLE, FALSE, processes[i].dwProcessId);
if (hProcess == INVALID_HANDLE_VALUE) {
printf("IPC failure: OpenProcess failed (%d)\n", GetLastError());
return -1;
}
for (int j = 0; j < shareableHandles.size(); j++) {
HANDLE hDup = INVALID_HANDLE_VALUE;
// Duplicate the handle into the target process's space
if (!DuplicateHandle(GetCurrentProcess(), shareableHandles[j], hProcess,
&hDup, 0, FALSE, DUPLICATE_SAME_ACCESS)) {
printf("IPC failure: DuplicateHandle failed (%d)\n", GetLastError());
return -1;
}
checkIpcErrors(ipcSendData(handle->hMailslot[i], &hDup, sizeof(hDup)));
}
CloseHandle(hProcess);
}
return 0;
}
int ipcRecvShareableHandles(ipcHandle *handle,
std::vector<ShareableHandle> &shareableHandles) {
for (int i = 0; i < shareableHandles.size(); i++) {
checkIpcErrors(
ipcRecvData(handle, &shareableHandles[i], sizeof(shareableHandles[i])));
}
return 0;
}
int ipcCloseSocket(ipcHandle *handle) {
for (int i = 0; i < handle->hMailslot.size(); i++) {
CloseHandle(handle->hMailslot[i]);
}
delete handle;
return 0;
}
int ipcCloseShareableHandle(ShareableHandle shHandle) {
CloseHandle(shHandle);
return 0;
}
#endif

View File

@@ -33,13 +33,26 @@
#define WIN32_LEAN_AND_MEAN #define WIN32_LEAN_AND_MEAN
#endif #endif
#include <windows.h> #include <windows.h>
#include <iostream>
#include <stdio.h>
#include <tchar.h>
#include <strsafe.h>
#include <sddl.h>
#include <aclapi.h>
#include <winternl.h>
#else #else
#include <stdio.h>
#include <fcntl.h> #include <fcntl.h>
#include <sys/mman.h> #include <sys/mman.h>
#include <unistd.h> #include <unistd.h>
#include <errno.h> #include <errno.h>
#include <sys/wait.h> #include <sys/wait.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <memory.h>
#include <sys/un.h>
#endif #endif
#include <vector>
typedef struct sharedMemoryInfo_st { typedef struct sharedMemoryInfo_st {
void *addr; void *addr;
@@ -68,4 +81,40 @@ int spawnProcess(Process *process, const char *app, char * const *args);
int waitProcess(Process *process); int waitProcess(Process *process);
#define checkIpcErrors(ipcFuncResult) \
if (ipcFuncResult == -1) { fprintf(stderr, "Failure at %u %s\n", __LINE__, __FILE__); exit(EXIT_FAILURE); }
#if defined(__linux__)
struct ipcHandle_st {
int socket;
char *socketName;
};
typedef int ShareableHandle;
#elif defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
struct ipcHandle_st {
std::vector<HANDLE> hMailslot; // 1 Handle in case of child and `num children` Handles for parent.
};
typedef HANDLE ShareableHandle;
#endif
typedef struct ipcHandle_st ipcHandle;
int
ipcCreateSocket(ipcHandle *&handle, const char *name, const std::vector<Process>& processes);
int
ipcOpenSocket(ipcHandle *&handle);
int
ipcCloseSocket(ipcHandle *handle);
int
ipcRecvShareableHandles(ipcHandle *handle, std::vector<ShareableHandle>& shareableHandles);
int
ipcSendShareableHandles(ipcHandle *handle, const std::vector<ShareableHandle>& shareableHandles, const std::vector<Process>& processes);
int
ipcCloseShareableHandle(ShareableHandle shHandle);
#endif // HELPER_MULTIPROCESS_H #endif // HELPER_MULTIPROCESS_H

View File

@@ -297,6 +297,19 @@ int writeBMPi(const char *filename, const unsigned char *d_RGB, int pitch,
return 0; return 0;
} }
int inputDirExists(const char *pathname) {
struct stat info;
if (stat(pathname, &info) != 0) {
return 0; // Directory does not exists
} else if (info.st_mode & S_IFDIR) {
// is a directory
return 1;
} else {
// is not a directory
return 0;
}
}
int readInput(const std::string &sInputPath, int readInput(const std::string &sInputPath,
std::vector<std::string> &filelist) { std::vector<std::string> &filelist) {
int error_code = 1; int error_code = 1;
@@ -315,15 +328,18 @@ int readInput(const std::string &sInputPath,
if (dir_handle) { if (dir_handle) {
error_code = 0; error_code = 0;
while ((dir = readdir(dir_handle)) != NULL) { while ((dir = readdir(dir_handle)) != NULL) {
if (dir->d_type == DT_REG) {
std::string sFileName = sInputPath + dir->d_name; std::string sFileName = sInputPath + dir->d_name;
filelist.push_back(sFileName); if (inputDirExists(sFileName.c_str()))
} else if (dir->d_type == DT_DIR) { {
std::string sname = dir->d_name; std::string sname = dir->d_name;
if (sname != "." && sname != "..") { if (sname != "." && sname != "..") {
readInput(sInputPath + sname + "/", filelist); readInput(sInputPath + sname + "/", filelist);
} }
} }
else
{
filelist.push_back(sFileName);
}
} }
closedir(dir_handle); closedir(dir_handle);
} else { } else {
@@ -360,18 +376,6 @@ int readInput(const std::string &sInputPath,
return 0; return 0;
} }
int inputDirExists(const char *pathname) {
struct stat info;
if (stat(pathname, &info) != 0) {
return 0; // Directory does not exists
} else if (info.st_mode & S_IFDIR) {
// is a directory
return 1;
} else {
// is not a directory
return 0;
}
}
int getInputDir(std::string &input_dir, const char *executable_path) { int getInputDir(std::string &input_dir, const char *executable_path) {
int found = 0; int found = 0;
@@ -399,6 +403,7 @@ int getInputDir(std::string &input_dir, const char *executable_path) {
"../../../Samples/<executable_name>/images", "../../../Samples/<executable_name>/images",
"../../Samples/<executable_name>/images"}; "../../Samples/<executable_name>/images"};
for (unsigned int i = 0; i < sizeof(searchPath) / sizeof(char *); ++i) { for (unsigned int i = 0; i < sizeof(searchPath) / sizeof(char *); ++i) {
std::string pathname(searchPath[i]); std::string pathname(searchPath[i]);
size_t executable_name_pos = pathname.find("<executable_name>"); size_t executable_name_pos = pathname.find("<executable_name>");

View File

@@ -293,6 +293,9 @@ inline char *sdkFindFilePath(const char *filename,
"../../../../Samples/<executable_name>/data/", // up 4 in tree "../../../../Samples/<executable_name>/data/", // up 4 in tree
"../../../Samples/<executable_name>/data/", // up 3 in tree "../../../Samples/<executable_name>/data/", // up 3 in tree
"../../Samples/<executable_name>/data/", // up 2 in tree "../../Samples/<executable_name>/data/", // up 2 in tree
"../../../../Common/data/", // up 4 in tree
"../../../Common/data/", // up 3 in tree
"../../Common/data/" // up 2 in tree
}; };
// Extract the executable name // Extract the executable name

View File

@@ -47,8 +47,8 @@
} \ } \
} while (0) } while (0)
void compileFileToPTX(char *filename, int argc, char **argv, char **ptxResult, void compileFileToCUBIN(char *filename, int argc, char **argv, char **cubinResult,
size_t *ptxResultSize, int requiresCGheaders) { size_t *cubinResultSize, int requiresCGheaders) {
std::ifstream inputFile(filename, std::ifstream inputFile(filename,
std::ios::in | std::ios::binary | std::ios::ate); std::ios::in | std::ios::binary | std::ios::ate);
@@ -68,7 +68,37 @@ void compileFileToPTX(char *filename, int argc, char **argv, char **ptxResult,
int numCompileOptions = 0; int numCompileOptions = 0;
char *compileParams[1]; char *compileParams[2];
int major = 0, minor = 0;
char deviceName[256];
// Picks the best CUDA device available
CUdevice cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
// get compute capabilities and the devicename
checkCudaErrors(cuDeviceGetAttribute(
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice));
checkCudaErrors(cuDeviceGetAttribute(
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice));
{
// Compile cubin for the GPU arch on which are going to run cuda kernel.
std::string compileOptions;
compileOptions = "--gpu-architecture=sm_";
compileParams[numCompileOptions] = reinterpret_cast<char *>(
malloc(sizeof(char) * (compileOptions.length() + 10)));
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
sprintf_s(compileParams[numCompileOptions], sizeof(char) * (compileOptions.length() + 10),
"%s%d%d", compileOptions.c_str(), major, minor);
#else
snprintf(compileParams[numCompileOptions], compileOptions.size() + 10, "%s%d%d",
compileOptions.c_str(), major, minor);
#endif
}
numCompileOptions++;
if (requiresCGheaders) { if (requiresCGheaders) {
std::string compileOptions; std::string compileOptions;
@@ -92,13 +122,13 @@ void compileFileToPTX(char *filename, int argc, char **argv, char **ptxResult,
argv[0]); argv[0]);
} }
compileOptions += path.c_str(); compileOptions += path.c_str();
compileParams[0] = reinterpret_cast<char *>( compileParams[numCompileOptions] = reinterpret_cast<char *>(
malloc(sizeof(char) * (compileOptions.length() + 1))); malloc(sizeof(char) * (compileOptions.length() + 1)));
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64) #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
sprintf_s(compileParams[0], sizeof(char) * (compileOptions.length() + 1), sprintf_s(compileParams[numCompileOptions], sizeof(char) * (compileOptions.length() + 1),
"%s", compileOptions.c_str()); "%s", compileOptions.c_str());
#else #else
snprintf(compileParams[0], compileOptions.size(), "%s", snprintf(compileParams[numCompileOptions], compileOptions.size(), "%s",
compileOptions.c_str()); compileOptions.c_str());
#endif #endif
numCompileOptions++; numCompileOptions++;
@@ -128,19 +158,20 @@ void compileFileToPTX(char *filename, int argc, char **argv, char **ptxResult,
free(log); free(log);
NVRTC_SAFE_CALL("nvrtcCompileProgram", res); NVRTC_SAFE_CALL("nvrtcCompileProgram", res);
// fetch PTX
size_t ptxSize;
NVRTC_SAFE_CALL("nvrtcGetPTXSize", nvrtcGetPTXSize(prog, &ptxSize));
char *ptx = reinterpret_cast<char *>(malloc(sizeof(char) * ptxSize));
NVRTC_SAFE_CALL("nvrtcGetPTX", nvrtcGetPTX(prog, ptx));
NVRTC_SAFE_CALL("nvrtcDestroyProgram", nvrtcDestroyProgram(&prog));
*ptxResult = ptx;
*ptxResultSize = ptxSize;
if (requiresCGheaders) free(compileParams[0]); size_t codeSize;
NVRTC_SAFE_CALL("nvrtcGetCUBINSize", nvrtcGetCUBINSize(prog, &codeSize));
char *code = new char[codeSize];
NVRTC_SAFE_CALL("nvrtcGetCUBIN", nvrtcGetCUBIN(prog, code));
*cubinResult = code;
*cubinResultSize = codeSize;
for (int i = 0; i < numCompileOptions; i++) {
free(compileParams[i]);
}
} }
CUmodule loadPTX(char *ptx, int argc, char **argv) { CUmodule loadCUBIN(char *cubin, int argc, char **argv) {
CUmodule module; CUmodule module;
CUcontext context; CUcontext context;
int major = 0, minor = 0; int major = 0, minor = 0;
@@ -158,11 +189,10 @@ CUmodule loadPTX(char *ptx, int argc, char **argv) {
printf("> GPU Device has SM %d.%d compute capability\n", major, minor); printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
checkCudaErrors(cuInit(0)); checkCudaErrors(cuInit(0));
checkCudaErrors(cuDeviceGet(&cuDevice, 0));
checkCudaErrors(cuCtxCreate(&context, 0, cuDevice)); checkCudaErrors(cuCtxCreate(&context, 0, cuDevice));
checkCudaErrors(cuModuleLoadDataEx(&module, ptx, 0, 0, 0)); checkCudaErrors(cuModuleLoadData(&module, cubin));
free(ptx); free(cubin);
return module; return module;
} }

View File

@@ -0,0 +1,124 @@
/* 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.
*/
////////////////////////////////////////////////////////////////////////////////
//
// Utility funcs to wrap up saving a surface or the back buffer as a PPM file
// In addition, wraps up a threshold comparision of two PPMs.
//
// These functions are designed to be used to implement an automated QA testing for SDK samples.
//
// Author: Bryan Dudash
// Email: sdkfeedback@nvidia.com
//
// Copyright (c) NVIDIA Corporation. All rights reserved.
////////////////////////////////////////////////////////////////////////////////
#include <helper_functions.h>
#include <rendercheck_d3d11.h>
HRESULT CheckRenderD3D11::ActiveRenderTargetToPPM(ID3D11Device *pDevice, const char *zFileName)
{
ID3D11DeviceContext *pDeviceCtxt;
pDevice->GetImmediateContext(&pDeviceCtxt);
ID3D11RenderTargetView *pRTV = NULL;
pDeviceCtxt->OMGetRenderTargets(1,&pRTV,NULL);
ID3D11Resource *pSourceResource = NULL;
pRTV->GetResource(&pSourceResource);
return ResourceToPPM(pDevice,pSourceResource,zFileName);
}
HRESULT CheckRenderD3D11::ResourceToPPM(ID3D11Device *pDevice, ID3D11Resource *pResource, const char *zFileName)
{
ID3D11DeviceContext *pDeviceCtxt;
pDevice->GetImmediateContext(&pDeviceCtxt);
D3D11_RESOURCE_DIMENSION rType;
pResource->GetType(&rType);
if (rType != D3D11_RESOURCE_DIMENSION_TEXTURE2D)
{
printf("SurfaceToPPM: pResource is not a 2D texture! Aborting...\n");
return E_FAIL;
}
ID3D11Texture2D *pSourceTexture = (ID3D11Texture2D *)pResource;
ID3D11Texture2D *pTargetTexture = NULL;
D3D11_TEXTURE2D_DESC desc;
pSourceTexture->GetDesc(&desc);
desc.BindFlags = 0;
desc.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
desc.Usage = D3D11_USAGE_STAGING;
if (FAILED(pDevice->CreateTexture2D(&desc,NULL,&pTargetTexture)))
{
printf("SurfaceToPPM: Unable to create target Texture resoruce! Aborting... \n");
return E_FAIL;
}
pDeviceCtxt->CopyResource(pTargetTexture,pSourceTexture);
D3D11_MAPPED_SUBRESOURCE mappedTex2D;
pDeviceCtxt->Map(pTargetTexture, 0, D3D11_MAP_READ,0,&mappedTex2D);
// Need to convert from dx pitch to pitch=width
unsigned char *pPPMData = new unsigned char[desc.Width*desc.Height*4];
for (unsigned int iHeight = 0; iHeight<desc.Height; iHeight++)
{
memcpy(&(pPPMData[iHeight*desc.Width*4]),(unsigned char *)(mappedTex2D.pData)+iHeight*mappedTex2D.RowPitch,desc.Width*4);
}
pDeviceCtxt->Unmap(pTargetTexture, 0);
// Prepends the PPM header info and bumps byte data afterwards
sdkSavePPM4ub(zFileName, pPPMData, desc.Width, desc.Height);
delete [] pPPMData;
pTargetTexture->Release();
return S_OK;
}
bool CheckRenderD3D11::PPMvsPPM(const char *src_file, const char *ref_file, const char *exec_path,
const float epsilon, const float threshold)
{
char *ref_file_path = sdkFindFilePath(ref_file, exec_path);
if (ref_file_path == NULL)
{
printf("CheckRenderD3D11::PPMvsPPM unable to find <%s> in <%s> Aborting comparison!\n", ref_file, exec_path);
printf(">>> Check info.xml and [project//data] folder <%s> <<<\n", ref_file);
printf("Aborting comparison!\n");
printf(" FAILURE!\n");
return false;
}
return sdkComparePPM(src_file,ref_file_path,epsilon,threshold,true) == true;
}

View File

@@ -0,0 +1,52 @@
/* 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.
*/
#pragma once
#ifndef _RENDERCHECK_D3D11_H_
#define _RENDERCHECK_D3D11_H_
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <d3d11.h>
class CheckRenderD3D11
{
public:
CheckRenderD3D11() {}
static HRESULT ActiveRenderTargetToPPM(ID3D11Device *pDevice, const char *zFileName);
static HRESULT ResourceToPPM(ID3D11Device *pDevice, ID3D11Resource *pResource, const char *zFileName);
static bool PPMvsPPM(const char *src_file, const char *ref_file, const char *exec_path,
const float epsilon, const float threshold = 0.0f);
};
#endif

116
README.md
View File

@@ -1,11 +1,45 @@
# CUDA Samples # CUDA Samples
Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 10.1 Update 2](https://developer.nvidia.com/cuda-downloads). Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 11.1](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 11.1
* Added `watershedSegmentationNPP`. Demonstrates how to use the NPP watershed segmentation function.
* Added `batchedLabelMarkersAndLabelCompressionNPP`. Demonstrates how to use the NPP label markers generation and label compression functions based on a Union Find (UF) algorithm including both single image and batched image versions.
* Dropped Visual Studio 2012, 2013 support from all the windows supported samples.
* Added kernel performing warp aggregated atomic max in multi buckets using cg::labeled_partition & cg::reduce in `warpAggregatedAtomicsCG`.
* Added extended CG shuffle mechanics to `shfl_scan` sample.
* Added `cudaOpenMP`. Demonstrates how to use OpenMP API to write an application for multiple GPUs.
* Added `simpleZeroCopy`. Demonstrates how to use zero copy, kernels can read and write directly to pinned system memory.
### CUDA 11.0
* Added `dmmaTensorCoreGemm`. Demonstrates double precision GEMM computation using the Double precision Warp Matrix Multiply and Accumulate (WMMA) API introduced with CUDA 11 in Ampere chip family tensor cores.
* Added `bf16TensorCoreGemm`. Demonstrates __nv_bfloat16 (e8m7) GEMM computation using the __nv_bfloat16 WMMA API introduced with CUDA 11 in Ampere chip family tensor cores.
* Added `tf32TensorCoreGemm`. Demonstrates tf32 (e8m10) GEMM computation using the tf32 WMMA API introduced with CUDA 11 in Ampere chip family tensor cores.
* Added `globalToShmemAsyncCopy`. Demonstrates async copy of data from global to shared memory when on compute capability 8.0 or higher. Also demonstrates arrive-wait barrier for synchronization.
* Added `simpleAWBarrier`. Demonstrates arrive wait barriers.
* Added `simpleAttributes`. Demonstrates the stream attributes that affect L2 locality.
* Added warp aggregated atomic multi bucket increments kernel using labeled_partition cooperative groups in `warpAggregatedAtomicsCG` which can be used on compute capability 7.0 and above GPU architectures.
* Added `binaryPartitionCG`. Demonstrates binary partition cooperative groups and reduction within the thread block.
* Added two new reduction kernels in `reduction` one which demonstrates reduce_add_sync intrinstic supported on compute capability 8.0 and another which uses cooperative_groups::reduce function which does thread_block_tile level reduction introduced from CUDA 11.0.
* Added `cudaCompressibleMemory`. Demonstrates compressible memory allocation using cuMemMap API.
* Added `simpleVulkanMMAP`. Demonstrates Vulkan CUDA Interop via cuMemMap APIs.
* Added `concurrentKernels`. Demonstrates the use of CUDA streams for concurrent execution of several kernels on a GPU.
* Dropped Mac OSX support from all samples.
### CUDA 10.2
* Added `simpleD3D11`. Demonstrates CUDA-D3D11 External Resource Interoperability APIs for updating D3D11 buffers from CUDA and synchronization between D3D11 and CUDA with Keyed Mutexes.
* Added `simpleDrvRuntime`. Demonstrates CUDA Driver and Runtime APIs working together to load fatbinary of a CUDA kernel.
* Added `vectorAddMMAP`. Demonstrates how cuMemMap API allows the user to specify the physical properties of their memory while retaining the contiguous nature of their access.
* Added `memMapIPCDrv`. Demonstrates Inter Process Communication using cuMemMap APIs.
* Added `cudaNvSci`. Demonstrates CUDA-NvSciBuf/NvSciSync Interop.
* Added `jacobiCudaGraphs`. Demonstrates Instantiated CUDA Graph Update with Jacobi Iterative Method using different approaches.
* Added `cuSolverSp_LinearSolver`. Demonstrates cuSolverSP's LU, QR and Cholesky factorization.
* Added `MersenneTwisterGP11213`. Demonstrates the Mersenne Twister random number generator GP11213 in cuRAND.
### CUDA 10.1 Update 2 ### CUDA 10.1 Update 2
* Added `vulkanImageCUDA`. Demonstrates how to perform Vulkan image - CUDA Interop. * Added `vulkanImageCUDA`. Demonstrates how to perform Vulkan image - CUDA Interop.
* Added `nvJPEG_encoder`. Demonstrates encoding of jpeg images using NVJPEG Library. * Added `nvJPEG_encoder`. Demonstrates encoding of jpeg images using NVJPEG Library.
@@ -59,8 +93,8 @@ This is the first release of CUDA Samples on GitHub:
### Prerequisites ### Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](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/), and the [Windows Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html).
### Getting the CUDA Samples ### Getting the CUDA Samples
@@ -85,7 +119,6 @@ Each individual sample has its own set of solution files at:
`<CUDA_SAMPLES_REPO>\Samples\<sample_dir>\` `<CUDA_SAMPLES_REPO>\Samples\<sample_dir>\`
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. 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](#directx) 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:
@@ -112,63 +145,41 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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
```
## Samples list ## Samples list
### Samples by OS ### Samples by OS
#### Linux #### Linux
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[boxFilterNPP](./Samples/boxFilterNPP)** | **[EGLStream_CUDA_Interop](./Samples/EGLStream_CUDA_Interop)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[boxFilterNPP](./Samples/boxFilterNPP)** | **[binaryPartitionCG](./Samples/binaryPartitionCG)** | **[dmmaTensorCoreGemm](./Samples/dmmaTensorCoreGemm)** |
---|---|---|---| ---|---|---|---|
**[simpleIPC](./Samples/simpleIPC)** | **[shfl_scan](./Samples/shfl_scan)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[nvJPEG](./Samples/nvJPEG)** | **[EGLStream_CUDA_Interop](./Samples/EGLStream_CUDA_Interop)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[simpleIPC](./Samples/simpleIPC)** | **[memMapIPCDrv](./Samples/memMapIPCDrv)** |
**[vectorAddMMAP](./Samples/vectorAddMMAP)** | **[shfl_scan](./Samples/shfl_scan)** | **[simpleZeroCopy](./Samples/simpleZeroCopy)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** |
**[globalToShmemAsyncCopy](./Samples/globalToShmemAsyncCopy)** | **[nvJPEG](./Samples/nvJPEG)** | **[batchedLabelMarkersAndLabelCompressionNPP](./Samples/batchedLabelMarkersAndLabelCompressionNPP)** | **[watershedSegmentationNPP](./Samples/watershedSegmentationNPP)** |
**[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** | **[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** |
**[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[cuSolverDn_LinearSolver](./Samples/cuSolverDn_LinearSolver)** | **[vulkanImageCUDA](./Samples/vulkanImageCUDA)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** | **[simpleAttributes](./Samples/simpleAttributes)** | **[cudaNvSci](./Samples/cudaNvSci)** | **[tf32TensorCoreGemm](./Samples/tf32TensorCoreGemm)** | **[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** |
**[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** | **[nvJPEG_encoder](./Samples/nvJPEG_encoder)** | **[cudaCompressibleMemory](./Samples/cudaCompressibleMemory)** | **[bf16TensorCoreGemm](./Samples/bf16TensorCoreGemm)** | **[cuSolverDn_LinearSolver](./Samples/cuSolverDn_LinearSolver)** | **[vulkanImageCUDA](./Samples/vulkanImageCUDA)** |
**[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** | **[bandwidthTest](./Samples/bandwidthTest)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[NV12toBGRandResize](./Samples/NV12toBGRandResize)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[cuSolverSp_LinearSolver](./Samples/cuSolverSp_LinearSolver)** | **[simpleCUFFT](./Samples/simpleCUFFT)** |
**[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[simpleVulkan](./Samples/simpleVulkan)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[cannyEdgeDetectorNPP](./Samples/cannyEdgeDetectorNPP)** | **[reduction](./Samples/reduction)** | **[nvJPEG_encoder](./Samples/nvJPEG_encoder)** | **[simpleDrvRuntime](./Samples/simpleDrvRuntime)** | **[MersenneTwisterGP11213](./Samples/MersenneTwisterGP11213)** |
**[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** | **[systemWideAtomics](./Samples/systemWideAtomics)** | **[simpleAWBarrier](./Samples/simpleAWBarrier)** | **[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** | **[bandwidthTest](./Samples/bandwidthTest)** | **[concurrentKernels](./Samples/concurrentKernels)** |
**[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[NV12toBGRandResize](./Samples/NV12toBGRandResize)** | **[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[jacobiCudaGraphs](./Samples/jacobiCudaGraphs)** |
**[simpleVulkan](./Samples/simpleVulkan)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[cannyEdgeDetectorNPP](./Samples/cannyEdgeDetectorNPP)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** |
**[simpleVulkanMMAP](./Samples/simpleVulkanMMAP)** | **[cudaOpenMP](./Samples/cudaOpenMP)** | **[matrixMul](./Samples/matrixMul)** | **[systemWideAtomics](./Samples/systemWideAtomics)** |
#### Windows #### Windows
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[boxFilterNPP](./Samples/boxFilterNPP)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[simpleIPC](./Samples/simpleIPC)** | **[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[boxFilterNPP](./Samples/boxFilterNPP)** | **[binaryPartitionCG](./Samples/binaryPartitionCG)** | **[dmmaTensorCoreGemm](./Samples/dmmaTensorCoreGemm)** |
---|---|---|---| ---|---|---|---|
**[shfl_scan](./Samples/shfl_scan)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[nvJPEG](./Samples/nvJPEG)** | **[simpleD3D12](./Samples/simpleD3D12)** | **[conjugateGradientMultiBlockCG](./Samples/conjugateGradientMultiBlockCG)** | **[simpleIPC](./Samples/simpleIPC)** | **[memMapIPCDrv](./Samples/memMapIPCDrv)** | **[vectorAddMMAP](./Samples/vectorAddMMAP)** |
**[shfl_scan](./Samples/shfl_scan)** | **[simpleZeroCopy](./Samples/simpleZeroCopy)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[globalToShmemAsyncCopy](./Samples/globalToShmemAsyncCopy)** |
**[nvJPEG](./Samples/nvJPEG)** | **[batchedLabelMarkersAndLabelCompressionNPP](./Samples/batchedLabelMarkersAndLabelCompressionNPP)** | **[simpleD3D12](./Samples/simpleD3D12)** | **[watershedSegmentationNPP](./Samples/watershedSegmentationNPP)** |
**[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** | **[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** |
**[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[cuSolverDn_LinearSolver](./Samples/cuSolverDn_LinearSolver)** | **[vulkanImageCUDA](./Samples/vulkanImageCUDA)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** | **[simpleAttributes](./Samples/simpleAttributes)** | **[tf32TensorCoreGemm](./Samples/tf32TensorCoreGemm)** | **[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[cudaCompressibleMemory](./Samples/cudaCompressibleMemory)** |
**[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** | **[nvJPEG_encoder](./Samples/nvJPEG_encoder)** | **[bf16TensorCoreGemm](./Samples/bf16TensorCoreGemm)** | **[cuSolverDn_LinearSolver](./Samples/cuSolverDn_LinearSolver)** | **[vulkanImageCUDA](./Samples/vulkanImageCUDA)** | **[conjugateGradientMultiDeviceCG](./Samples/conjugateGradientMultiDeviceCG)** |
**[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** | **[bandwidthTest](./Samples/bandwidthTest)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[NV12toBGRandResize](./Samples/NV12toBGRandResize)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[cuSolverSp_LinearSolver](./Samples/cuSolverSp_LinearSolver)** | **[simpleCUFFT](./Samples/simpleCUFFT)** | **[reduction](./Samples/reduction)** |
**[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[simpleVulkan](./Samples/simpleVulkan)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[cannyEdgeDetectorNPP](./Samples/cannyEdgeDetectorNPP)** | **[nvJPEG_encoder](./Samples/nvJPEG_encoder)** | **[simpleDrvRuntime](./Samples/simpleDrvRuntime)** | **[simpleD3D11](./Samples/simpleD3D11)** | **[MersenneTwisterGP11213](./Samples/MersenneTwisterGP11213)** |
**[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** | **[simpleAWBarrier](./Samples/simpleAWBarrier)** | **[immaTensorCoreGemm](./Samples/immaTensorCoreGemm)** | **[bandwidthTest](./Samples/bandwidthTest)** | **[concurrentKernels](./Samples/concurrentKernels)** |
**[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[NV12toBGRandResize](./Samples/NV12toBGRandResize)** | **[cudaTensorCoreGemm](./Samples/cudaTensorCoreGemm)** | **[jacobiCudaGraphs](./Samples/jacobiCudaGraphs)** |
#### Mac OSX **[simpleVulkan](./Samples/simpleVulkan)** | **[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[cannyEdgeDetectorNPP](./Samples/cannyEdgeDetectorNPP)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** |
**[warpAggregatedAtomicsCG](./Samples/warpAggregatedAtomicsCG)** | **[boxFilterNPP](./Samples/boxFilterNPP)** | **[shfl_scan](./Samples/shfl_scan)** | **[conjugateGradientCudaGraphs](./Samples/conjugateGradientCudaGraphs)** | **[simpleVulkanMMAP](./Samples/simpleVulkanMMAP)** | **[cudaOpenMP](./Samples/cudaOpenMP)** | **[matrixMul](./Samples/matrixMul)** |
---|---|---|---|
**[simpleCudaGraphs](./Samples/simpleCudaGraphs)** | **[deviceQuery](./Samples/deviceQuery)** | **[simpleVoteIntrinsics](./Samples/simpleVoteIntrinsics)** | **[simpleCUBLASXT](./Samples/simpleCUBLASXT)** |
**[UnifiedMemoryPerf](./Samples/UnifiedMemoryPerf)** | **[cuSolverDn_LinearSolver](./Samples/cuSolverDn_LinearSolver)** | **[matrixMulDrv](./Samples/matrixMulDrv)** | **[simpleCUFFT](./Samples/simpleCUFFT)** |
**[reduction](./Samples/reduction)** | **[bandwidthTest](./Samples/bandwidthTest)** | **[simpleCUBLAS](./Samples/simpleCUBLAS)** | **[NV12toBGRandResize](./Samples/NV12toBGRandResize)** |
**[vectorAdd_nvrtc](./Samples/vectorAdd_nvrtc)** | **[cannyEdgeDetectorNPP](./Samples/cannyEdgeDetectorNPP)** | **[p2pBandwidthLatencyTest](./Samples/p2pBandwidthLatencyTest)** | **[matrixMul](./Samples/matrixMul)** |
## Dependencies ## Dependencies
@@ -196,7 +207,7 @@ Some samples can only be run on a 64-bit operating system.
#### DirectX #### DirectX
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 Microsoft Visual Studio 2012 or higher which provides Microsoft Windows SDK for Windows 8.
#### DirectX12 #### DirectX12
@@ -238,6 +249,9 @@ EGLOutput is a set of EGL extensions which allow EGL to render directly to the d
EGLSync is a set of EGL extensions which provides sync objects that are synchronization primitive, representing events whose completion can be tested or waited upon. EGLSync is a set of EGL extensions which provides sync objects that are synchronization primitive, representing events whose completion can be tested or waited upon.
#### NVSCI
NvSci is a set of communication interface libraries out of which CUDA interops with NvSciBuf and NvSciSync. NvSciBuf allows applications to allocate and exchange buffers in memory. NvSciSync allows applications to manage synchronization objects which coordinate when sequences of operations begin and end.
### CUDA Features ### CUDA Features
These CUDA features are needed by some CUDA samples. They are provided by either the CUDA Toolkit or CUDA Driver. Some features may not be available on your system. These CUDA features are needed by some CUDA samples. They are provided by either the CUDA Toolkit or CUDA Driver. Some features may not be available on your system.

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -289,6 +310,10 @@ else
CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/aarch64-linux/lib/stubs CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/aarch64-linux/lib/stubs
endif endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/sbsa-linux/lib/stubs
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/armv7-linux-androideabi/lib/stubs CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/armv7-linux-androideabi/lib/stubs
endif endif
@@ -303,12 +328,19 @@ else
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx) ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/aarch64-qnx/lib/stubs CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/aarch64-qnx/lib/stubs
ifdef TARGET_OVERRIDE
CUDA_SEARCH_PATH := $(CUDA_PATH)/targets/$(TARGET_OVERRIDE)/lib/stubs
endif
endif endif
ifeq ($(TARGET_ARCH),ppc64le) ifeq ($(TARGET_ARCH),ppc64le)
CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/ppc64le-linux/lib/stubs CUDA_SEARCH_PATH ?= $(CUDA_PATH)/targets/ppc64le-linux/lib/stubs
endif endif
ifeq ($(HOST_ARCH),ppc64le)
CUDA_SEARCH_PATH += $(CUDA_PATH)/lib64/stubs
endif
CUDALIB ?= $(shell find -L $(CUDA_SEARCH_PATH) -maxdepth 1 -name libcuda.so 2> /dev/null) CUDALIB ?= $(shell find -L $(CUDA_SEARCH_PATH) -maxdepth 1 -name libcuda.so 2> /dev/null)
ifeq ("$(CUDALIB)","") ifeq ("$(CUDALIB)","")
$(info >>> WARNING - libcuda.so not found, CUDA Driver is not installed. Please re-install the driver. <<<) $(info >>> WARNING - libcuda.so not found, CUDA Driver is not installed. Please re-install the driver. <<<)

View File

@@ -49,7 +49,6 @@
<scope>1:CUDA Basic Topics</scope> <scope>1:CUDA Basic Topics</scope>
<scope>2:Graphics Interop</scope> <scope>2:Graphics Interop</scope>
</scopes> </scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch> <sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch> <sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch> <sm-arch>sm50</sm-arch>
@@ -59,6 +58,8 @@
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch> <sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch> <sm-arch>sm75</sm-arch>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,7 +10,7 @@ EGLStreams Interop
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
@@ -30,7 +30,7 @@ cuDeviceGet, cuDeviceGetAttribute, cuDeviceComputeCapability, cuDeviceGetCount,
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](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

@@ -245,12 +245,30 @@ CUresult cudaProducerTest(test_cuda_producer_s *cudaProducer, char *file) {
cudaEgl.eglColorFormat = eglColorFormat; cudaEgl.eglColorFormat = eglColorFormat;
cudaEgl.cuFormat = CU_AD_FORMAT_UNSIGNED_INT8; cudaEgl.cuFormat = CU_AD_FORMAT_UNSIGNED_INT8;
static int numFramesPresented = 0;
// If there is a frame presented before we check if consumer
// is done with it using cuEGLStreamProducerReturnFrame.
while (numFramesPresented) {
CUeglFrame returnedCudaEgl;
cuStatus = cuEGLStreamProducerReturnFrame(&cudaProducer->cudaConn,
&returnedCudaEgl, NULL);
if (cuStatus == CUDA_ERROR_LAUNCH_TIMEOUT) {
continue;
} else if (cuStatus != CUDA_SUCCESS) {
printf("cuda Producer return frame FAILED with custatus= %d\n", cuStatus);
return cuStatus;
} else {
numFramesPresented--;
}
}
cuStatus = cuStatus =
cuEGLStreamProducerPresentFrame(&cudaProducer->cudaConn, cudaEgl, NULL); cuEGLStreamProducerPresentFrame(&cudaProducer->cudaConn, cudaEgl, NULL);
if (cuStatus != CUDA_SUCCESS) { if (cuStatus != CUDA_SUCCESS) {
printf("cuda Producer present frame FAILED with custatus= %d\n", cuStatus); printf("cuda Producer present frame FAILED with custatus= %d\n", cuStatus);
goto done; goto done;
} }
numFramesPresented++;
done: done:
if (file_p) { if (file_p) {
@@ -281,6 +299,13 @@ CUresult cudaDeviceCreateProducer(test_cuda_producer_s *cudaProducer,
"%d.%d\n\n", "%d.%d\n\n",
device, deviceName, major, minor); device, deviceName, major, minor);
if (major < 6) {
printf(
"EGLStreams_CUDA_Interop requires SM 6.0 or higher arch GPU. "
"Exiting...\n");
exit(2); // EXIT_WAIVED
}
if (CUDA_SUCCESS != if (CUDA_SUCCESS !=
(status = cuCtxCreate(&cudaProducer->context, 0, device))) { (status = cuCtxCreate(&cudaProducer->context, 0, device))) {
printf("failed to create CUDA context\n"); printf("failed to create CUDA context\n");

View File

@@ -133,10 +133,14 @@ ifeq ("$(TARGET_OS)","linux")
else else
endif endif
ifeq ("$(TARGET_OS)","qnx")
HOST_CCFLAGS := -V5.4.0,gcc_ntoaarch64le
endif
# Attempt to compile a minimal EGL application and run to check if EGL_SUPPORT_REUSE_NV is supported in the EGL headers available. # Attempt to compile a minimal EGL application and run to check if EGL_SUPPORT_REUSE_NV is supported in the EGL headers available.
ifneq ($(SAMPLE_ENABLED), 0) ifneq ($(SAMPLE_ENABLED), 0)
$(shell printf "#include <EGL/egl.h>\n#include <EGL/eglext.h>\nint main() {\n#ifdef EGL_SUPPORT_REUSE_NV \n #error \"Compatible EGL header found\" \n return 0;\n#endif \n return 1;\n}" > test.c; ) $(shell printf "#include <EGL/egl.h>\n#include <EGL/eglext.h>\nint main() {\n#ifdef EGL_SUPPORT_REUSE_NV \n #error \"Compatible EGL header found\" \n return 0;\n#endif \n return 1;\n}" > test.c; )
EGL_DEFINES := $(shell $(HOST_COMPILER) $(CCFLAGS) $(EXTRA_CCFLAGS) -lEGL test.c -c 2>&1 | grep -ic "Compatible EGL header found";) EGL_DEFINES := $(shell $(HOST_COMPILER) $(HOST_CCFLAGS) $(CCFLAGS) $(EXTRA_CCFLAGS) -lEGL test.c -c 2>&1 | grep -ic "Compatible EGL header found";)
SHOULD_WAIVE := 0 SHOULD_WAIVE := 0
ifeq ($(EGL_DEFINES),0) ifeq ($(EGL_DEFINES),0)
SHOULD_WAIVE := 1 SHOULD_WAIVE := 1

View File

@@ -0,0 +1,310 @@
################################################################################
# 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.
#
################################################################################
#
# 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 $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa 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 ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(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)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(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 -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(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)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(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))
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
SMS ?=
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_$(sm)))
ifeq ($(SMS),)
# Generate PTX code from SM 35
GENCODE_FLAGS += -gencode arch=compute_35,code=compute_35
endif
# 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 += -lcurand_static -lculibos
################################################################################
# Target rules
all: build
build: MersenneTwisterGP11213
MersenneTwister.o:MersenneTwister.cpp
$(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
MersenneTwisterGP11213: MersenneTwister.o
$(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
run: build
./MersenneTwisterGP11213
clean:
rm -f MersenneTwisterGP11213 MersenneTwister.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/MersenneTwisterGP11213
clobber: clean

View File

@@ -0,0 +1,184 @@
/* 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 sample demonstrates the use of CURAND to generate
* random numbers on GPU and CPU.
*/
// Utilities and system includes
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <curand.h>
// Utilities and system includes
#include <helper_functions.h>
#include <helper_cuda.h>
#include <cuda_runtime.h>
#include <curand.h>
float compareResults(int rand_n, float *h_RandGPU, float *h_RandCPU);
const int DEFAULT_RAND_N = 2400000;
const unsigned int DEFAULT_SEED = 777;
///////////////////////////////////////////////////////////////////////////////
// Main program
///////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
// Start logs
printf("%s Starting...\n\n", argv[0]);
// initialize the GPU, either identified by --device
// or by picking the device with highest flop rate.
int devID = findCudaDevice(argc, (const char **)argv);
// parsing the number of random numbers to generate
int rand_n = DEFAULT_RAND_N;
if (checkCmdLineFlag(argc, (const char **) argv, "count"))
{
rand_n = getCmdLineArgumentInt(argc, (const char **) argv, "count");
}
printf("Allocating data for %i samples...\n", rand_n);
// parsing the seed
int seed = DEFAULT_SEED;
if (checkCmdLineFlag(argc, (const char **) argv, "seed"))
{
seed = getCmdLineArgumentInt(argc, (const char **) argv, "seed");
}
printf("Seeding with %i ...\n", seed);
cudaStream_t stream;
checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
float *d_Rand;
checkCudaErrors(cudaMalloc((void **)&d_Rand, rand_n * sizeof(float)));
curandGenerator_t prngGPU;
checkCudaErrors(curandCreateGenerator(&prngGPU, CURAND_RNG_PSEUDO_MTGP32));
checkCudaErrors(curandSetStream(prngGPU, stream));
checkCudaErrors(curandSetPseudoRandomGeneratorSeed(prngGPU, seed));
curandGenerator_t prngCPU;
checkCudaErrors(curandCreateGeneratorHost(&prngCPU, CURAND_RNG_PSEUDO_MTGP32));
checkCudaErrors(curandSetPseudoRandomGeneratorSeed(prngCPU, seed));
//
// Example 1: Compare random numbers generated on GPU and CPU
float *h_RandGPU = (float *)malloc(rand_n * sizeof(float));
printf("Generating random numbers on GPU...\n\n");
checkCudaErrors(curandGenerateUniform(prngGPU, (float *) d_Rand, rand_n));
printf("\nReading back the results...\n");
checkCudaErrors(cudaMemcpyAsync(h_RandGPU, d_Rand, rand_n * sizeof(float), cudaMemcpyDeviceToHost, stream));
float *h_RandCPU = (float *)malloc(rand_n * sizeof(float));
printf("Generating random numbers on CPU...\n\n");
checkCudaErrors(curandGenerateUniform(prngCPU, (float *) h_RandCPU, rand_n));
checkCudaErrors(cudaStreamSynchronize(stream));
printf("Comparing CPU/GPU random numbers...\n\n");
float L1norm = compareResults(rand_n, h_RandGPU, h_RandCPU);
//
// Example 2: Timing of random number generation on GPU
const int numIterations = 10;
int i;
StopWatchInterface *hTimer;
sdkCreateTimer(&hTimer);
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (i = 0; i < numIterations; i++)
{
checkCudaErrors(curandGenerateUniform(prngGPU, (float *) d_Rand, rand_n));
}
checkCudaErrors(cudaStreamSynchronize(stream));
sdkStopTimer(&hTimer);
double gpuTime = 1.0e-3 * sdkGetTimerValue(&hTimer)/(double)numIterations;
printf("MersenneTwisterGP11213, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers\n",
1.0e-9 * rand_n / gpuTime, gpuTime, rand_n);
printf("Shutting down...\n");
checkCudaErrors(curandDestroyGenerator(prngGPU));
checkCudaErrors(curandDestroyGenerator(prngCPU));
checkCudaErrors(cudaStreamDestroy(stream));
checkCudaErrors(cudaFree(d_Rand));
sdkDeleteTimer(&hTimer);
free(h_RandGPU);
free(h_RandCPU);
exit(L1norm < 1e-6 ? EXIT_SUCCESS : EXIT_FAILURE);
}
float compareResults(int rand_n, float *h_RandGPU, float *h_RandCPU)
{
int i;
float rCPU, rGPU, delta;
float max_delta = 0.;
float sum_delta = 0.;
float sum_ref = 0.;
for (i = 0; i < rand_n; i++)
{
rCPU = h_RandCPU[i];
rGPU = h_RandGPU[i];
delta = fabs(rCPU - rGPU);
sum_delta += delta;
sum_ref += fabs(rCPU);
if (delta >= max_delta)
{
max_delta = delta;
}
}
float L1norm = (float)(sum_delta / sum_ref);
printf("Max absolute error: %E\n", max_delta);
printf("L1 norm: %E\n\n", L1norm);
return L1norm;
}

View File

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

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "MersenneTwisterGP11213", "MersenneTwisterGP11213_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

@@ -15,15 +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>nvJPEG_vs2012</RootNamespace> <RootNamespace>MersenneTwisterGP11213_vs2015</RootNamespace>
<ProjectName>nvJPEG</ProjectName> <ProjectName>MersenneTwisterGP11213</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> <PlatformToolset>v140</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -57,12 +57,12 @@
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>nvjpeg.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>curand.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>$(OutDir)/nvJPEG.exe</OutputFile> <OutputFile>$(OutDir)/MersenneTwisterGP11213.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration></CodeGeneration> <CodeGeneration>compute_35,compute_35;</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 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<ClCompile Include="nvJPEG.cpp" /> <ClCompile Include="MersenneTwister.cpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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}") = "MersenneTwisterGP11213", "MersenneTwisterGP11213_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 +1,112 @@
<?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>MersenneTwisterGP11213_vs2017</RootNamespace>
<ProjectName>MersenneTwisterGP11213</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
<LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
<WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
<TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</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 11.1.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;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>curand.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/MersenneTwisterGP11213.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_35,compute_35;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<ClCompile Include="MersenneTwister.cpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

View File

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

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2019
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "MersenneTwisterGP11213", "MersenneTwisterGP11213_vs2019.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

@@ -15,15 +15,16 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>nvJPEG_encoder_vs2012</RootNamespace> <RootNamespace>MersenneTwisterGP11213_vs2019</RootNamespace>
<ProjectName>nvJPEG_encoder</ProjectName> <ProjectName>MersenneTwisterGP11213</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> <PlatformToolset>v142</PlatformToolset>
<WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -33,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -57,12 +58,12 @@
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>nvjpeg.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>curand.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>$(OutDir)/nvJPEG_encoder.exe</OutputFile> <OutputFile>$(OutDir)/MersenneTwisterGP11213.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration></CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
<ClCompile Include="nvJPEG_encoder.cpp" /> <ClCompile Include="MersenneTwister.cpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,74 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>MersenneTwisterGP11213</name>
<description><![CDATA[This sample demonstrates the Mersenne Twister random number generator GP11213 in cuRAND.]]></description>
<devicecompilation>whole</devicecompilation>
<fallback_min_ptx>true</fallback_min_ptx>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">CURAND Library</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>CURAND</keyword>
<keyword>Monte-Carlo</keyword>
<keyword>random number generation</keyword>
<keyword>GSFR</keyword>
<keyword>Mersenne Twister</keyword>
</keywords>
<libraries>
<library>curand_static</library>
<library>culibos</library>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>MersenneTwister.cpp</primary_file>
<required_dependencies>
<dependency>CURAND</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Advanced Topics</scope>
</scopes>
<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>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</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>MersenneTwisterGP11213</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,71 @@
# MersenneTwisterGP11213 - MersenneTwisterGP11213
## Description
This sample demonstrates the Mersenne Twister random number generator GP11213 in cuRAND.
## Key Concepts
CURAND Library
## Supported SM Architectures
[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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, ppc64le, armv7l
## CUDA APIs involved
## Dependencies needed to build/run
[CURAND](../../README.md#curand)
## Prerequisites
Download and install the [CUDA Toolkit 11.1](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++
```
## References (for more details)

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -253,9 +274,9 @@ LIBRARIES :=
# Gencode arguments # Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64)) ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75 SMS ?= 35 37 50 52 60 61 70 72 75 80 86
else else
SMS ?= 30 35 37 50 52 60 61 70 75 SMS ?= 35 37 50 52 60 61 70 75 80 86
endif endif
ifeq ($(SMS),) ifeq ($(SMS),)

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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/NV12toBGRandResize.exe</OutputFile> <OutputFile>$(OutDir)/NV12toBGRandResize.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -107,6 +107,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -38,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -67,7 +67,7 @@
<OutputFile>$(OutDir)/NV12toBGRandResize.exe</OutputFile> <OutputFile>$(OutDir)/NV12toBGRandResize.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -112,6 +112,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/NV12toBGRandResize.exe</OutputFile> <OutputFile>$(OutDir)/NV12toBGRandResize.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -108,6 +108,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -32,7 +32,6 @@
<scope>2:Image Processing</scope> <scope>2:Image Processing</scope>
<scope>2:Computer Vision</scope> <scope>2:Computer Vision</scope>
</scopes> </scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch> <sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch> <sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch> <sm-arch>sm50</sm-arch>
@@ -42,6 +41,8 @@
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch> <sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch> <sm-arch>sm75</sm-arch>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,11 +10,11 @@ Graphics Interop, Image Processing, Video Processing
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux, Windows, MacOSX Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -27,7 +27,7 @@ cudaMemcpy2D, cudaMallocManaged
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run ## Build and Run
@@ -66,29 +66,5 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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) ## References (for more details)

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -253,9 +274,9 @@ LIBRARIES :=
# Gencode arguments # Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64)) ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75 SMS ?= 35 37 50 52 60 61 70 72 75 80 86
else else
SMS ?= 30 35 37 50 52 60 61 70 75 SMS ?= 35 37 50 52 60 61 70 75 80 86
endif endif
ifeq ($(SMS),) ifeq ($(SMS),)

View File

@@ -44,16 +44,6 @@
<scope>1:CUDA Systems Integration</scope> <scope>1:CUDA Systems Integration</scope>
<scope>1:Unified Memory</scope> <scope>1:Unified Memory</scope>
</scopes> </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> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,11 +10,9 @@ CUDA Systems Integration, Unified Memory, CUDA Streams and Events, Pinned System
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux, Windows, MacOSX Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -30,7 +28,7 @@ cudaMallocManaged, cudaStreamAttachMemAsync, cudaMemcpyAsync, cudaMallocHost, cu
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](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
@@ -70,29 +68,5 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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) ## References (for more details)

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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile> <OutputFile>$(OutDir)/UnifiedMemoryPerf.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -105,6 +105,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -38,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -67,7 +67,7 @@
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile> <OutputFile>$(OutDir)/UnifiedMemoryPerf.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -110,6 +110,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile> <OutputFile>$(OutDir)/UnifiedMemoryPerf.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -106,6 +106,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -33,7 +33,7 @@
#define VERIFY_GPU_CORRECTNESS 0 #define VERIFY_GPU_CORRECTNESS 0
size_t maxSampleSizeInMb = 64; size_t maxSampleSizeInMb = 64;
int numKernelRuns = 100; int numKernelRuns = 20;
int verboseResults = 0; int verboseResults = 0;
const char *memAllocTypeStr[MEMALLOC_TYPE_COUNT] = { const char *memAllocTypeStr[MEMALLOC_TYPE_COUNT] = {

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -247,9 +268,9 @@ LIBRARIES :=
# Gencode arguments # Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64)) ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 30 35 37 50 52 60 61 70 72 75 SMS ?= 35 37 50 52 60 61 70 72 75 80 86
else else
SMS ?= 30 35 37 50 52 60 61 70 75 SMS ?= 35 37 50 52 60 61 70 75 80 86
endif endif
ifeq ($(SMS),) ifeq ($(SMS),)

View File

@@ -41,7 +41,6 @@
<scope>1:CUDA Basic Topics</scope> <scope>1:CUDA Basic Topics</scope>
<scope>1:Performance Strategies</scope> <scope>1:Performance Strategies</scope>
</scopes> </scopes>
<sm-arch>sm30</sm-arch>
<sm-arch>sm35</sm-arch> <sm-arch>sm35</sm-arch>
<sm-arch>sm37</sm-arch> <sm-arch>sm37</sm-arch>
<sm-arch>sm50</sm-arch> <sm-arch>sm50</sm-arch>
@@ -51,6 +50,8 @@
<sm-arch>sm70</sm-arch> <sm-arch>sm70</sm-arch>
<sm-arch>sm72</sm-arch> <sm-arch>sm72</sm-arch>
<sm-arch>sm75</sm-arch> <sm-arch>sm75</sm-arch>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</sm-arch>
<supported_envs> <supported_envs>
<env> <env>
<arch>x86_64</arch> <arch>x86_64</arch>

View File

@@ -10,11 +10,11 @@ CUDA Streams and Events, Performance Strategies
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux, Windows, MacOSX Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -27,7 +27,7 @@ cudaSetDevice, cudaHostAlloc, cudaFree, cudaMallocHost, cudaFreeHost, cudaMemcpy
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run ## Build and Run
@@ -66,29 +66,5 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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) ## References (for more details)

View File

@@ -915,7 +915,7 @@ void printResultsCSV(unsigned int *memSizes, double *bandwidths,
double dSeconds = 0.0; double dSeconds = 0.0;
for (i = 0; i < count; i++) { for (i = 0; i < count; i++) {
dSeconds = (double)memSizes[i] / (bandwidths[i] * (double)(1 << 20)); dSeconds = (double)memSizes[i] / (bandwidths[i] * (double)(1e9));
printf( printf(
"bandwidthTest-%s, Bandwidth = %.1f GB/s, Time = %.5f s, Size = %u " "bandwidthTest-%s, Bandwidth = %.1f GB/s, Time = %.5f s, Size = %u "
"bytes, NumDevsUsed = %d\n", "bytes, NumDevsUsed = %d\n",

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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/bandwidthTest.exe</OutputFile> <OutputFile>$(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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -38,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -67,7 +67,7 @@
<OutputFile>$(OutDir)/bandwidthTest.exe</OutputFile> <OutputFile>$(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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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>
@@ -107,6 +107,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/bandwidthTest.exe</OutputFile> <OutputFile>$(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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,329 @@
################################################################################
# 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.
#
################################################################################
#
# 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 $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa 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 ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(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)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(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 -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(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)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(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 - batchedLabelMarkersAndLabelCompressionNPP is not supported on Mac OSX - 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
SMS ?=
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_$(sm)))
ifeq ($(SMS),)
# Generate PTX code from SM 35
GENCODE_FLAGS += -gencode arch=compute_35,code=compute_35
endif
# 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 += -lnppisu_static -lnppif_static -lnppc_static -lculibos
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: batchedLabelMarkersAndLabelCompressionNPP
check.deps:
ifeq ($(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
batchedLabelMarkersAndLabelCompressionNPP.o:batchedLabelMarkersAndLabelCompressionNPP.cpp
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
batchedLabelMarkersAndLabelCompressionNPP: batchedLabelMarkersAndLabelCompressionNPP.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
$(EXEC) ./batchedLabelMarkersAndLabelCompressionNPP
clean:
rm -f batchedLabelMarkersAndLabelCompressionNPP batchedLabelMarkersAndLabelCompressionNPP.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/batchedLabelMarkersAndLabelCompressionNPP
clobber: clean

View File

@@ -0,0 +1,76 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>batchedLabelMarkersAndLabelCompressionNPP</name>
<description><![CDATA[An NPP CUDA Sample that demonstrates how to use the NPP label markers generation and label compression functions based on a Union Find (UF) algorithm including both
single image and batched image versions.]]></description>
<devicecompilation>whole</devicecompilation>
<fallback_min_ptx>true</fallback_min_ptx>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">Performance Strategies</concept>
<concept level="basic">Image Processing</concept>
<concept level="basic">NPP Library</concept>
<concept level="basic">Using NPP Batch Functions</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>NPP</keyword>
<keyword>Image Processing</keyword>
</keywords>
<libraries>
<library>nppisu_static</library>
<library>nppif_static</library>
<library>nppc_static</library>
<library>culibos</library>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>batchedLabelMarkersAndLabelCompressionNPP.cpp</primary_file>
<required_dependencies>
<dependency>NPP</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Basic Topics</scope>
<scope>1:Performance Strategies</scope>
<scope>2:Image Processing</scope>
<scope>2:Computer Vision</scope>
</scopes>
<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>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<platform>windows7</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>Batched Label Markers And Label Compression NPP</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,72 @@
# batchedLabelMarkersAndLabelCompressionNPP - Batched Label Markers And Label Compression NPP
## Description
An NPP CUDA Sample that demonstrates how to use the NPP label markers generation and label compression functions based on a Union Find (UF) algorithm including both
single image and batched image versions.
## Key Concepts
Performance Strategies, Image Processing, NPP Library, Using NPP Batch Functions
## Supported SM Architectures
[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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, ppc64le, armv7l
## CUDA APIs involved
## Dependencies needed to build/run
[NPP](../../README.md#npp)
## Prerequisites
Download and install the [CUDA Toolkit 11.1](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++
```
## References (for more details)

View File

@@ -0,0 +1,798 @@
/* Copyright (c) 2020, 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.
*/
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
#define WINDOWS_LEAN_AND_MEAN
#define NOMINMAX
#include <windows.h>
#pragma warning(disable : 4819)
#endif
#include <stdio.h>
#include <string.h>
#include <fstream>
#include <helper_cuda.h>
#include <npp.h>
// Note: If you want to view these images we HIGHLY recommend using imagej
// which is free on the internet and works on most platforms
// because it is one of the few image viewing apps that can display 32
// bit integer image data. While it normalizes the data to floating
// point values for viewing it still provides a good representation of
// the relative brightness of each label value. Note that label
// compression output results in smaller differences between label values
// making it visually more difficult to detect differences in labeled
// regions. If you have an editor that can display hex values you can
// see what the exact values of each label is, every 4 bytes represents 1
// 32 bit integer label value.
//
// The files read and written by this sample app use RAW image format,
// that is, only the image data itself exists in the files with no image
// format information. When viewing RAW files with imagej just enter
// the image size and bit depth values that are part of the file name
// when requested by imagej.
//
// This sample app works in 2 stages, first it processes all of the
// images individually then it processes them all again in 1 batch using
// the Batch_Advanced versions of the NPP batch functions which allow
// each image to have it's own ROI. The 2 stages are completely
// separable but in this sample the second stage takes advantage of some
// of the data that has already been initialized.
//
// Note that there is a small amount of variability in the number of
// unique label markers generated from one run to the next by the UF
// algorithm.
//
// Performance of ALL NPP image batch functions is limited by the maximum
// ROI height in the list of images.
// Batched label compression support is only available on NPP versions > 11.0,
// comment out if using NPP 11.0
#define USE_BATCHED_LABEL_COMPRESSION 1
#define NUMBER_OF_IMAGES 5
Npp8u *pInputImageDev[NUMBER_OF_IMAGES];
Npp8u *pInputImageHost[NUMBER_OF_IMAGES];
Npp8u *pUFGenerateLabelsScratchBufferDev[NUMBER_OF_IMAGES];
Npp8u *pUFCompressedLabelsScratchBufferDev[NUMBER_OF_IMAGES];
Npp32u *pUFLabelDev[NUMBER_OF_IMAGES];
Npp32u *pUFLabelHost[NUMBER_OF_IMAGES];
NppiImageDescriptor *pUFBatchSrcImageListDev = 0;
NppiImageDescriptor *pUFBatchSrcDstImageListDev = 0;
NppiImageDescriptor *pUFBatchSrcImageListHost = 0;
NppiImageDescriptor *pUFBatchSrcDstImageListHost = 0;
NppiBufferDescriptor *pUFBatchSrcDstScratchBufferListDev =
0; // from nppi_filtering_functions.h
NppiBufferDescriptor *pUFBatchSrcDstScratchBufferListHost = 0;
Npp32u *pUFBatchPerImageCompressedCountListDev = 0;
Npp32u *pUFBatchPerImageCompressedCountListHost = 0;
void tearDown() // Clean up and tear down
{
if (pUFBatchPerImageCompressedCountListDev != 0)
cudaFree(pUFBatchPerImageCompressedCountListDev);
if (pUFBatchSrcDstScratchBufferListDev != 0)
cudaFree(pUFBatchSrcDstScratchBufferListDev);
if (pUFBatchSrcDstImageListDev != 0) cudaFree(pUFBatchSrcDstImageListDev);
if (pUFBatchSrcImageListDev != 0) cudaFree(pUFBatchSrcImageListDev);
if (pUFBatchPerImageCompressedCountListHost != 0)
free(pUFBatchPerImageCompressedCountListHost);
if (pUFBatchSrcDstScratchBufferListHost != 0)
free(pUFBatchSrcDstScratchBufferListHost);
if (pUFBatchSrcDstImageListHost != 0) free(pUFBatchSrcDstImageListHost);
if (pUFBatchSrcImageListHost != 0) free(pUFBatchSrcImageListHost);
for (int j = 0; j < NUMBER_OF_IMAGES; j++) {
if (pUFCompressedLabelsScratchBufferDev[j] != 0)
cudaFree(pUFCompressedLabelsScratchBufferDev[j]);
if (pUFGenerateLabelsScratchBufferDev[j] != 0)
cudaFree(pUFGenerateLabelsScratchBufferDev[j]);
if (pUFLabelDev[j] != 0) cudaFree(pUFLabelDev[j]);
if (pInputImageDev[j] != 0) cudaFree(pInputImageDev[j]);
if (pUFLabelHost[j] != 0) free(pUFLabelHost[j]);
if (pInputImageHost[j] != 0) free(pInputImageHost[j]);
}
}
const std::string &LabelMarkersOutputFile0 =
"Lena_LabelMarkersUF_8Way_512x512_32u.raw";
const std::string &LabelMarkersOutputFile1 =
"CT_skull_LabelMarkersUF_8Way_512x512_32u.raw";
const std::string &LabelMarkersOutputFile2 =
"PCB_METAL_LabelMarkersUF_8Way_509x335_32u.raw";
const std::string &LabelMarkersOutputFile3 =
"PCB2_LabelMarkersUF_8Way_1024x683_32u.raw";
const std::string &LabelMarkersOutputFile4 =
"PCB_LabelMarkersUF_8Way_1280x720_32u.raw";
const std::string &CompressedMarkerLabelsOutputFile0 =
"Lena_CompressedMarkerLabelsUF_8Way_512x512_32u.raw";
const std::string &CompressedMarkerLabelsOutputFile1 =
"CT_skull_CompressedMarkerLabelsUF_8Way_512x512_32u.raw";
const std::string &CompressedMarkerLabelsOutputFile2 =
"PCB_METAL_CompressedMarkerLabelsUF_8Way_509x335_32u.raw";
const std::string &CompressedMarkerLabelsOutputFile3 =
"PCB2_CompressedMarkerLabelsUF_8Way_1024x683_32u.raw";
const std::string &CompressedMarkerLabelsOutputFile4 =
"PCB_CompressedMarkerLabelsUF_8Way_1280x720_32u.raw";
const std::string &LabelMarkersBatchOutputFile0 =
"Lena_LabelMarkersUFBatch_8Way_512x512_32u.raw";
const std::string &LabelMarkersBatchOutputFile1 =
"CT_skull_LabelMarkersUFBatch_8Way_512x512_32u.raw";
const std::string &LabelMarkersBatchOutputFile2 =
"PCB_METAL_LabelMarkersUFBatch_8Way_509x335_32u.raw";
const std::string &LabelMarkersBatchOutputFile3 =
"PCB2_LabelMarkersUFBatch_8Way_1024x683_32u.raw";
const std::string &LabelMarkersBatchOutputFile4 =
"PCB_LabelMarkersUFBatch_8Way_1280x720_32u.raw";
const std::string &CompressedMarkerLabelsBatchOutputFile0 =
"Lena_CompressedMarkerLabelsUFBatch_8Way_512x512_32u.raw";
const std::string &CompressedMarkerLabelsBatchOutputFile1 =
"CT_skull_CompressedMarkerLabelsUFBatch_8Way_512x512_32u.raw";
const std::string &CompressedMarkerLabelsBatchOutputFile2 =
"PCB_METAL_CompressedMarkerLabelsUFBatch_8Way_509x335_32u.raw";
const std::string &CompressedMarkerLabelsBatchOutputFile3 =
"PCB2_CompressedMarkerLabelsUFBatch_8Way_1024x683_32u.raw";
const std::string &CompressedMarkerLabelsBatchOutputFile4 =
"PCB_CompressedMarkerLabelsUFBatch_8Way_1280x720_32u.raw";
int loadRaw8BitImage(Npp8u *pImage, int nWidth, int nHeight, int nImage) {
FILE *bmpFile;
size_t nSize;
if (nImage == 0) {
if (nWidth != 512 || nHeight != 512) return -1;
const char *fileName = "lena_512x512_8u.raw";
const char *InputFile = sdkFindFilePath(fileName, ".");
if (InputFile == NULL) {
printf("%s file not found.. exiting\n", fileName);
exit(EXIT_WAIVED);
}
bmpFile = fopen(InputFile, "rb");
} else if (nImage == 1) {
if (nWidth != 512 || nHeight != 512) return -1;
const char *fileName = "CT_skull_512x512_8u.raw";
const char *InputFile = sdkFindFilePath(fileName, ".");
if (InputFile == NULL) {
printf("%s file not found.. exiting\n", fileName);
exit(EXIT_WAIVED);
}
bmpFile = fopen(InputFile, "rb");
} else if (nImage == 2) {
if (nWidth != 509 || nHeight != 335) return -1;
const char *fileName = "PCB_METAL_509x335_8u.raw";
const char *InputFile = sdkFindFilePath(fileName, ".");
if (InputFile == NULL) {
printf("%s file not found.. exiting\n", fileName);
exit(EXIT_WAIVED);
}
bmpFile = fopen(InputFile, "rb");
} else if (nImage == 3) {
if (nWidth != 1024 || nHeight != 683) return -1;
const char *fileName = "PCB2_1024x683_8u.raw";
const char *InputFile = sdkFindFilePath(fileName, ".");
if (InputFile == NULL) {
printf("%s file not found.. exiting\n", fileName);
exit(EXIT_WAIVED);
}
bmpFile = fopen(InputFile, "rb");
} else if (nImage == 4) {
if (nWidth != 1280 || nHeight != 720) return -1;
const char *fileName = "PCB_1280x720_8u.raw";
const char *InputFile = sdkFindFilePath(fileName, ".");
if (InputFile == NULL) {
printf("%s file not found.. exiting\n", fileName);
exit(EXIT_WAIVED);
}
bmpFile = fopen(InputFile, "rb");
} else {
printf("Input file load failed.\n");
return -1;
}
if (bmpFile == NULL) return -1;
nSize = fread(pImage, 1, nWidth * nHeight, bmpFile);
if (nSize < nWidth * nHeight) {
fclose(bmpFile);
return -1;
}
fclose(bmpFile);
printf("Input file load succeeded.\n");
return 0;
}
int main(int argc, char **argv) {
int aGenerateLabelsScratchBufferSize[NUMBER_OF_IMAGES];
int aCompressLabelsScratchBufferSize[NUMBER_OF_IMAGES];
int nCompressedLabelCount = 0;
cudaError_t cudaError;
NppStatus nppStatus;
NppStreamContext nppStreamCtx;
FILE *bmpFile;
for (int j = 0; j < NUMBER_OF_IMAGES; j++) {
pInputImageDev[j] = 0;
pInputImageHost[j] = 0;
pUFGenerateLabelsScratchBufferDev[j] = 0;
pUFCompressedLabelsScratchBufferDev[j] = 0;
pUFLabelDev[j] = 0;
pUFLabelHost[j] = 0;
}
nppStreamCtx.hStream = 0; // The NULL stream by default, set this to whatever
// your stream ID is if not the NULL stream.
cudaError = cudaGetDevice(&nppStreamCtx.nCudaDeviceId);
if (cudaError != cudaSuccess) {
printf("CUDA error: no devices supporting CUDA.\n");
return NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY;
}
const NppLibraryVersion *libVer = nppGetLibVersion();
printf("NPP Library Version %d.%d.%d\n", libVer->major, libVer->minor,
libVer->build);
int driverVersion, runtimeVersion;
cudaDriverGetVersion(&driverVersion);
cudaRuntimeGetVersion(&runtimeVersion);
printf("CUDA Driver Version: %d.%d\n", driverVersion / 1000,
(driverVersion % 100) / 10);
printf("CUDA Runtime Version: %d.%d\n\n", runtimeVersion / 1000,
(runtimeVersion % 100) / 10);
cudaError = cudaDeviceGetAttribute(
&nppStreamCtx.nCudaDevAttrComputeCapabilityMajor,
cudaDevAttrComputeCapabilityMajor, nppStreamCtx.nCudaDeviceId);
if (cudaError != cudaSuccess) return NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY;
cudaError = cudaDeviceGetAttribute(
&nppStreamCtx.nCudaDevAttrComputeCapabilityMinor,
cudaDevAttrComputeCapabilityMinor, nppStreamCtx.nCudaDeviceId);
if (cudaError != cudaSuccess) return NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY;
cudaError =
cudaStreamGetFlags(nppStreamCtx.hStream, &nppStreamCtx.nStreamFlags);
cudaDeviceProp oDeviceProperties;
cudaError =
cudaGetDeviceProperties(&oDeviceProperties, nppStreamCtx.nCudaDeviceId);
nppStreamCtx.nMultiProcessorCount = oDeviceProperties.multiProcessorCount;
nppStreamCtx.nMaxThreadsPerMultiProcessor =
oDeviceProperties.maxThreadsPerMultiProcessor;
nppStreamCtx.nMaxThreadsPerBlock = oDeviceProperties.maxThreadsPerBlock;
nppStreamCtx.nSharedMemPerBlock = oDeviceProperties.sharedMemPerBlock;
NppiSize oSizeROI[NUMBER_OF_IMAGES];
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
if (nImage == 0) {
oSizeROI[nImage].width = 512;
oSizeROI[nImage].height = 512;
} else if (nImage == 1) {
oSizeROI[nImage].width = 512;
oSizeROI[nImage].height = 512;
} else if (nImage == 2) {
oSizeROI[nImage].width = 509;
oSizeROI[nImage].height = 335;
} else if (nImage == 3) {
oSizeROI[nImage].width = 1024;
oSizeROI[nImage].height = 683;
} else if (nImage == 4) {
oSizeROI[nImage].width = 1280;
oSizeROI[nImage].height = 720;
}
// NOTE: While using cudaMallocPitch() to allocate device memory for NPP can
// significantly improve the performance of many NPP functions, for UF
// function label markers generation or compression DO NOT USE
// cudaMallocPitch(). Doing so could result in incorrect output.
cudaError = cudaMalloc(
(void **)&pInputImageDev[nImage],
oSizeROI[nImage].width * sizeof(Npp8u) * oSizeROI[nImage].height);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
// For images processed with UF label markers functions ROI width and height
// for label markers generation output AND marker compression functions MUST
// be the same AND line pitch MUST be equal to ROI.width * sizeof(Npp32u).
// Also the image pointer used for label markers generation output must
// start at the same position in the image as it does in the marker
// compression function. Also note that actual input image size and ROI do
// not necessarily need to be related other than ROI being less than or
// equal to image size and image starting position does not necessarily have
// to be at pixel 0 in the input image.
cudaError = cudaMalloc(
(void **)&pUFLabelDev[nImage],
oSizeROI[nImage].width * sizeof(Npp32u) * oSizeROI[nImage].height);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
pInputImageHost[nImage] = reinterpret_cast<Npp8u *>(malloc(
oSizeROI[nImage].width * sizeof(Npp8u) * oSizeROI[nImage].height));
pUFLabelHost[nImage] = reinterpret_cast<Npp32u *>(malloc(
oSizeROI[nImage].width * sizeof(Npp32u) * oSizeROI[nImage].height));
// Use UF functions throughout this sample.
nppStatus = nppiLabelMarkersUFGetBufferSize_32u_C1R(
oSizeROI[nImage], &aGenerateLabelsScratchBufferSize[nImage]);
// One at a time image processing
cudaError = cudaMalloc((void **)&pUFGenerateLabelsScratchBufferDev[nImage],
aGenerateLabelsScratchBufferSize[nImage]);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
if (loadRaw8BitImage(pInputImageHost[nImage],
oSizeROI[nImage].width * sizeof(Npp8u),
oSizeROI[nImage].height, nImage) == 0) {
cudaError = cudaMemcpy2DAsync(
pInputImageDev[nImage], oSizeROI[nImage].width * sizeof(Npp8u),
pInputImageHost[nImage], oSizeROI[nImage].width * sizeof(Npp8u),
oSizeROI[nImage].width * sizeof(Npp8u), oSizeROI[nImage].height,
cudaMemcpyHostToDevice, nppStreamCtx.hStream);
nppStatus = nppiLabelMarkersUF_8u32u_C1R_Ctx(
pInputImageDev[nImage], oSizeROI[nImage].width * sizeof(Npp8u),
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage], nppiNormInf,
pUFGenerateLabelsScratchBufferDev[nImage], nppStreamCtx);
if (nppStatus != NPP_SUCCESS) {
if (nImage == 0)
printf("Lena_LabelMarkersUF_8Way_512x512_32u failed.\n");
else if (nImage == 1)
printf("CT_skull_LabelMarkersUF_8Way_512x512_32u failed.\n");
else if (nImage == 2)
printf("PCB_METAL_LabelMarkersUF_8Way_509x335_32u failed.\n");
else if (nImage == 3)
printf("PCB2_LabelMarkersUF_8Way_1024x683_32u failed.\n");
else if (nImage == 4)
printf("PCB_LabelMarkersUF_8Way_1280x720_32u failed.\n");
tearDown();
return -1;
}
cudaError = cudaMemcpy2DAsync(
pUFLabelHost[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage].width * sizeof(Npp32u), oSizeROI[nImage].height,
cudaMemcpyDeviceToHost, nppStreamCtx.hStream);
// Wait host image read backs to complete, not necessary if no need to
// synchronize
if ((cudaError = cudaStreamSynchronize(nppStreamCtx.hStream)) !=
cudaSuccess) {
printf("Post label generation cudaStreamSynchronize failed\n");
tearDown();
return -1;
}
if (nImage == 0)
bmpFile = fopen(LabelMarkersOutputFile0.c_str(), "wb");
else if (nImage == 1)
bmpFile = fopen(LabelMarkersOutputFile1.c_str(), "wb");
else if (nImage == 2)
bmpFile = fopen(LabelMarkersOutputFile2.c_str(), "wb");
else if (nImage == 3)
bmpFile = fopen(LabelMarkersOutputFile3.c_str(), "wb");
else if (nImage == 4)
bmpFile = fopen(LabelMarkersOutputFile4.c_str(), "wb");
if (bmpFile == NULL) return -1;
size_t nSize = 0;
for (int j = 0; j < oSizeROI[nImage].height; j++) {
nSize += fwrite(&pUFLabelHost[nImage][j * oSizeROI[nImage].width],
sizeof(Npp32u), oSizeROI[nImage].width, bmpFile);
}
fclose(bmpFile);
nppStatus = nppiCompressMarkerLabelsGetBufferSize_32u_C1R(
oSizeROI[nImage].width * oSizeROI[nImage].height,
&aCompressLabelsScratchBufferSize[nImage]);
if (nppStatus != NPP_NO_ERROR) return nppStatus;
cudaError =
cudaMalloc((void **)&pUFCompressedLabelsScratchBufferDev[nImage],
aCompressLabelsScratchBufferSize[nImage]);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
nCompressedLabelCount = 0;
nppStatus = nppiCompressMarkerLabelsUF_32u_C1IR(
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage], oSizeROI[nImage].width * oSizeROI[nImage].height,
&nCompressedLabelCount, pUFCompressedLabelsScratchBufferDev[nImage]);
if (nppStatus != NPP_SUCCESS) {
if (nImage == 0)
printf("Lena_CompressedLabelMarkersUF_8Way_512x512_32u failed.\n");
else if (nImage == 1)
printf(
"CT_Skull_CompressedLabelMarkersUF_8Way_512x512_32u failed.\n");
else if (nImage == 2)
printf(
"PCB_METAL_CompressedLabelMarkersUF_8Way_509x335_32u failed.\n");
else if (nImage == 3)
printf("PCB2_CompressedLabelMarkersUF_8Way_1024x683_32u failed.\n");
else if (nImage == 4)
printf("PCB_CompressedLabelMarkersUF_8Way_1280x720_32u failed.\n");
tearDown();
return -1;
}
cudaError = cudaMemcpy2DAsync(
pUFLabelHost[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage].width * sizeof(Npp32u), oSizeROI[nImage].height,
cudaMemcpyDeviceToHost, nppStreamCtx.hStream);
// Wait for host image read backs to finish, not necessary if no need to
// synchronize
if ((cudaError = cudaStreamSynchronize(nppStreamCtx.hStream)) !=
cudaSuccess ||
nCompressedLabelCount == 0) {
printf("Post label compression cudaStreamSynchronize failed\n");
tearDown();
return -1;
}
if (nImage == 0)
bmpFile = fopen(CompressedMarkerLabelsOutputFile0.c_str(), "wb");
else if (nImage == 1)
bmpFile = fopen(CompressedMarkerLabelsOutputFile1.c_str(), "wb");
else if (nImage == 2)
bmpFile = fopen(CompressedMarkerLabelsOutputFile2.c_str(), "wb");
else if (nImage == 3)
bmpFile = fopen(CompressedMarkerLabelsOutputFile3.c_str(), "wb");
else if (nImage == 4)
bmpFile = fopen(CompressedMarkerLabelsOutputFile4.c_str(), "wb");
if (bmpFile == NULL) return -1;
nSize = 0;
for (int j = 0; j < oSizeROI[nImage].height; j++) {
nSize += fwrite(&pUFLabelHost[nImage][j * oSizeROI[nImage].width],
sizeof(Npp32u), oSizeROI[nImage].width, bmpFile);
}
fclose(bmpFile);
if (nImage == 0)
printf(
"Lena_CompressedMarkerLabelsUF_8Way_512x512_32u succeeded, "
"compressed label count is %d.\n",
nCompressedLabelCount);
else if (nImage == 1)
printf(
"CT_Skull_CompressedMarkerLabelsUF_8Way_512x512_32u succeeded, "
"compressed label count is %d.\n",
nCompressedLabelCount);
else if (nImage == 2)
printf(
"PCB_METAL_CompressedMarkerLabelsUF_8Way_509x335_32u succeeded, "
"compressed label count is %d.\n",
nCompressedLabelCount);
else if (nImage == 3)
printf(
"PCB2_CompressedMarkerLabelsUF_8Way_1024x683_32u succeeded, "
"compressed label count is %d.\n",
nCompressedLabelCount);
else if (nImage == 4)
printf(
"PCB_CompressedMarkerLabelsUF_8Way_1280x720_32u succeeded, "
"compressed label count is %d.\n",
nCompressedLabelCount);
}
}
// Batch image processing
// We want to allocate scratch buffers more efficiently for batch processing
// so first we free up the scratch buffers for image 0 and reallocate them.
// This is not required but helps cudaMalloc to work more efficiently.
cudaFree(pUFCompressedLabelsScratchBufferDev[0]);
int nTotalBatchedUFCompressLabelsScratchBufferDevSize = 0;
for (int k = 0; k < NUMBER_OF_IMAGES; k++)
nTotalBatchedUFCompressLabelsScratchBufferDevSize +=
aCompressLabelsScratchBufferSize[k];
cudaError = cudaMalloc((void **)&pUFCompressedLabelsScratchBufferDev[0],
nTotalBatchedUFCompressLabelsScratchBufferDevSize);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
// Now allocate batch lists
int nBatchImageListBytes = NUMBER_OF_IMAGES * sizeof(NppiImageDescriptor);
cudaError =
cudaMalloc((void **)&pUFBatchSrcImageListDev, nBatchImageListBytes);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
cudaError =
cudaMalloc((void **)&pUFBatchSrcDstImageListDev, nBatchImageListBytes);
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
pUFBatchSrcImageListHost =
reinterpret_cast<NppiImageDescriptor *>(malloc(nBatchImageListBytes));
pUFBatchSrcDstImageListHost =
reinterpret_cast<NppiImageDescriptor *>(malloc(nBatchImageListBytes));
NppiSize oMaxROISize = {0, 0};
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
pUFBatchSrcImageListHost[nImage].pData = pInputImageDev[nImage];
pUFBatchSrcImageListHost[nImage].nStep =
oSizeROI[nImage].width * sizeof(Npp8u);
// src image oSize parameter is ignored in these NPP functions
pUFBatchSrcDstImageListHost[nImage].pData = pUFLabelDev[nImage];
pUFBatchSrcDstImageListHost[nImage].nStep =
oSizeROI[nImage].width * sizeof(Npp32u);
pUFBatchSrcDstImageListHost[nImage].oSize = oSizeROI[nImage];
if (oSizeROI[nImage].width > oMaxROISize.width)
oMaxROISize.width = oSizeROI[nImage].width;
if (oSizeROI[nImage].height > oMaxROISize.height)
oMaxROISize.height = oSizeROI[nImage].height;
}
// Copy label generation batch lists from CPU to GPU
cudaError = cudaMemcpyAsync(pUFBatchSrcImageListDev, pUFBatchSrcImageListHost,
nBatchImageListBytes, cudaMemcpyHostToDevice,
nppStreamCtx.hStream);
if (cudaError != cudaSuccess) return NPP_MEMCPY_ERROR;
cudaError = cudaMemcpyAsync(pUFBatchSrcDstImageListDev,
pUFBatchSrcDstImageListHost, nBatchImageListBytes,
cudaMemcpyHostToDevice, nppStreamCtx.hStream);
if (cudaError != cudaSuccess) return NPP_MEMCPY_ERROR;
// We use 8-way neighbor search throughout this example
nppStatus = nppiLabelMarkersUFBatch_8u32u_C1R_Advanced_Ctx(
pUFBatchSrcImageListDev, pUFBatchSrcDstImageListDev, NUMBER_OF_IMAGES,
oMaxROISize, nppiNormInf, nppStreamCtx);
if (nppStatus != NPP_SUCCESS) {
printf("LabelMarkersUFBatch_8Way_8u32u failed.\n");
tearDown();
return -1;
}
// Now read back generated device images to the host
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
cudaError = cudaMemcpy2DAsync(
pUFLabelHost[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage].width * sizeof(Npp32u), oSizeROI[nImage].height,
cudaMemcpyDeviceToHost, nppStreamCtx.hStream);
}
// Wait for host image read backs to complete, not necessary if no need to
// synchronize
if ((cudaError = cudaStreamSynchronize(nppStreamCtx.hStream)) !=
cudaSuccess) {
printf("Post label generation cudaStreamSynchronize failed\n");
tearDown();
return -1;
}
// Save output to files
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
if (nImage == 0)
bmpFile = fopen(LabelMarkersBatchOutputFile0.c_str(), "wb");
else if (nImage == 1)
bmpFile = fopen(LabelMarkersBatchOutputFile1.c_str(), "wb");
else if (nImage == 2)
bmpFile = fopen(LabelMarkersBatchOutputFile2.c_str(), "wb");
else if (nImage == 3)
bmpFile = fopen(LabelMarkersBatchOutputFile3.c_str(), "wb");
else if (nImage == 4)
bmpFile = fopen(LabelMarkersBatchOutputFile4.c_str(), "wb");
if (bmpFile == NULL) return -1;
size_t nSize = 0;
for (int j = 0; j < oSizeROI[nImage].height; j++) {
nSize += fwrite(&pUFLabelHost[nImage][j * oSizeROI[nImage].width],
sizeof(Npp32u), oSizeROI[nImage].width, bmpFile);
}
fclose(bmpFile);
}
#ifdef USE_BATCHED_LABEL_COMPRESSION
// Now allocate scratch buffer memory for batched label compression
cudaError = cudaMalloc((void **)&pUFBatchSrcDstScratchBufferListDev,
NUMBER_OF_IMAGES * sizeof(NppiBufferDescriptor));
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
cudaError = cudaMalloc((void **)&pUFBatchPerImageCompressedCountListDev,
NUMBER_OF_IMAGES * sizeof(Npp32u));
if (cudaError != cudaSuccess) return NPP_MEMORY_ALLOCATION_ERR;
// Allocate host side scratch buffer point and size list and initialize with
// device scratch buffer pointers
pUFBatchSrcDstScratchBufferListHost =
reinterpret_cast<NppiBufferDescriptor *>(
malloc(NUMBER_OF_IMAGES * sizeof(NppiBufferDescriptor)));
pUFBatchPerImageCompressedCountListHost =
reinterpret_cast<Npp32u *>(malloc(NUMBER_OF_IMAGES * sizeof(Npp32u)));
// Start buffer pointer at beginning of full per image buffer list sized
// pUFCompressedLabelsScratchBufferDev[0]
Npp32u *pCurUFCompressedLabelsScratchBufferDev =
reinterpret_cast<Npp32u *>(pUFCompressedLabelsScratchBufferDev[0]);
int nMaxUFCompressedLabelsScratchBufferSize = 0;
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
// This particular function works on in-place data and SrcDst image batch
// list has already been initialized in batched label generation function
// setup
// Initialize each per image buffer descriptor
pUFBatchSrcDstScratchBufferListHost[nImage].pData =
reinterpret_cast<void *>(pCurUFCompressedLabelsScratchBufferDev);
pUFBatchSrcDstScratchBufferListHost[nImage].nBufferSize =
aCompressLabelsScratchBufferSize[nImage];
if (aCompressLabelsScratchBufferSize[nImage] >
nMaxUFCompressedLabelsScratchBufferSize)
nMaxUFCompressedLabelsScratchBufferSize =
aCompressLabelsScratchBufferSize[nImage];
// Offset buffer pointer to next per image buffer
Npp8u *pTempBuffer =
reinterpret_cast<Npp8u *>(pCurUFCompressedLabelsScratchBufferDev);
pTempBuffer += aCompressLabelsScratchBufferSize[nImage];
pCurUFCompressedLabelsScratchBufferDev =
reinterpret_cast<Npp32u *>((void *)(pTempBuffer));
}
// Copy compression batch scratch buffer list from CPU to GPU
cudaError = cudaMemcpyAsync(pUFBatchSrcDstScratchBufferListDev,
pUFBatchSrcDstScratchBufferListHost,
NUMBER_OF_IMAGES * sizeof(NppiBufferDescriptor),
cudaMemcpyHostToDevice, nppStreamCtx.hStream);
if (cudaError != cudaSuccess) return NPP_MEMCPY_ERROR;
nppStatus = nppiCompressMarkerLabelsUFBatch_32u_C1IR_Advanced_Ctx(
pUFBatchSrcDstImageListDev, pUFBatchSrcDstScratchBufferListDev,
pUFBatchPerImageCompressedCountListDev, NUMBER_OF_IMAGES, oMaxROISize,
nMaxUFCompressedLabelsScratchBufferSize, nppStreamCtx);
if (nppStatus != NPP_SUCCESS) {
printf("BatchCompressedLabelMarkersUF_8Way_32u failed.\n");
tearDown();
return -1;
}
// Copy output compressed label images back to host
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
cudaError = cudaMemcpy2DAsync(
pUFLabelHost[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
pUFLabelDev[nImage], oSizeROI[nImage].width * sizeof(Npp32u),
oSizeROI[nImage].width * sizeof(Npp32u), oSizeROI[nImage].height,
cudaMemcpyDeviceToHost, nppStreamCtx.hStream);
}
// Wait for host image read backs to complete, not necessary if no need to
// synchronize
if ((cudaError = cudaStreamSynchronize(nppStreamCtx.hStream)) !=
cudaSuccess) {
printf("Post label compression cudaStreamSynchronize failed\n");
tearDown();
return -1;
}
// Save compressed label images into files
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
if (nImage == 0)
bmpFile = fopen(CompressedMarkerLabelsBatchOutputFile0.c_str(), "wb");
else if (nImage == 1)
bmpFile = fopen(CompressedMarkerLabelsBatchOutputFile1.c_str(), "wb");
else if (nImage == 2)
bmpFile = fopen(CompressedMarkerLabelsBatchOutputFile2.c_str(), "wb");
else if (nImage == 3)
bmpFile = fopen(CompressedMarkerLabelsBatchOutputFile3.c_str(), "wb");
else if (nImage == 4)
bmpFile = fopen(CompressedMarkerLabelsBatchOutputFile4.c_str(), "wb");
if (bmpFile == NULL) return -1;
size_t nSize = 0;
for (int j = 0; j < oSizeROI[nImage].height; j++) {
nSize += fwrite(&pUFLabelHost[nImage][j * oSizeROI[nImage].width],
sizeof(Npp32u), oSizeROI[nImage].width, bmpFile);
}
fclose(bmpFile);
}
// Read back per image compressed label count.
cudaError = cudaMemcpyAsync(pUFBatchPerImageCompressedCountListHost,
pUFBatchPerImageCompressedCountListDev,
NUMBER_OF_IMAGES * sizeof(Npp32u),
cudaMemcpyDeviceToHost, nppStreamCtx.hStream);
if (cudaError != cudaSuccess) {
tearDown();
return NPP_MEMCPY_ERROR;
}
// Wait for host read back to complete
cudaError = cudaStreamSynchronize(nppStreamCtx.hStream);
printf("\n\n");
for (int nImage = 0; nImage < NUMBER_OF_IMAGES; nImage++) {
if (nImage == 0)
printf(
"Lena_CompressedMarkerLabelsUFBatch_8Way_512x512_32u succeeded, "
"compressed label count is %d.\n",
pUFBatchPerImageCompressedCountListHost[nImage]);
else if (nImage == 1)
printf(
"CT_Skull_CompressedMarkerLabelsUFBatch_8Way_512x512_32u succeeded, "
"compressed label count is %d.\n",
pUFBatchPerImageCompressedCountListHost[nImage]);
else if (nImage == 2)
printf(
"PCB_METAL_CompressedMarkerLabelsUFBatch_8Way_509x335_32u succeeded, "
"compressed label count is %d.\n",
pUFBatchPerImageCompressedCountListHost[nImage]);
else if (nImage == 3)
printf(
"PCB2_CompressedMarkerLabelsUFBatch_8Way_1024x683_32u succeeded, "
"compressed label count is %d.\n",
pUFBatchPerImageCompressedCountListHost[nImage]);
else if (nImage == 4)
printf(
"PCB_CompressedMarkerLabelsUFBatch_8Way_1280x720_32u succeeded, "
"compressed label count is %d.\n",
pUFBatchPerImageCompressedCountListHost[nImage]);
}
#endif // USE_BATCHED_LABEL_COMPRESSION
tearDown();
return 0;
}

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 13.00 Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2013 # Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "matrixMul", "matrixMul_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "batchedLabelMarkersAndLabelCompressionNPP", "batchedLabelMarkersAndLabelCompressionNPP_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,15 +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>nvJPEG_encoder_vs2013</RootNamespace> <RootNamespace>batchedLabelMarkersAndLabelCompressionNPP_vs2015</RootNamespace>
<ProjectName>nvJPEG_encoder</ProjectName> <ProjectName>batchedLabelMarkersAndLabelCompressionNPP</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> <PlatformToolset>v140</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -57,12 +57,12 @@
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>nvjpeg.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>nppisu.lib;nppif.lib;nppc.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>$(OutDir)/nvJPEG_encoder.exe</OutputFile> <OutputFile>$(OutDir)/batchedLabelMarkersAndLabelCompressionNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration></CodeGeneration> <CodeGeneration>compute_35,compute_35;</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 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<ClCompile Include="nvJPEG_encoder.cpp" /> <ClCompile Include="batchedLabelMarkersAndLabelCompressionNPP.cpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 12.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012 # Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "reduction", "reduction_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "batchedLabelMarkersAndLabelCompressionNPP", "batchedLabelMarkersAndLabelCompressionNPP_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,15 +15,20 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>boxFilterNPP_vs2013</RootNamespace> <RootNamespace>batchedLabelMarkersAndLabelCompressionNPP_vs2017</RootNamespace>
<ProjectName>boxFilterNPP</ProjectName> <ProjectName>batchedLabelMarkersAndLabelCompressionNPP</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
<LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
<WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
<TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset> <PlatformToolset>v141</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -33,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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.$(Platform).user.props" />
@@ -53,16 +58,16 @@
<ClCompile> <ClCompile>
<WarningLevel>Level3</WarningLevel> <WarningLevel>Level3</WarningLevel>
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions> <PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);../../Common/UtilNPP;../../Common/FreeImage/include;</AdditionalIncludeDirectories> <AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);</AdditionalIncludeDirectories>
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>nppisu.lib;nppif.lib;nppc.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>nppisu.lib;nppif.lib;nppc.lib;cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);../../Common/FreeImage/lib/x64;../../Common/FreeImage/lib/$(TARGET_OS);../../Common/FreeImage/lib/$(TARGET_OS)/$(TARGET_ARCH);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/boxFilterNPP.exe</OutputFile> <OutputFile>$(OutDir)/batchedLabelMarkersAndLabelCompressionNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
@@ -96,22 +101,12 @@
<TargetMachinePlatform>64</TargetMachinePlatform> <TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='Win32'">
<Link>
</Link>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='x64'">
<Link>
<AdditionalDependencies>freeimage.lib;%(AdditionalDependencies)</AdditionalDependencies>
</Link>
</ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<ClCompile Include="boxFilterNPP.cpp" /> <ClCompile Include="batchedLabelMarkersAndLabelCompressionNPP.cpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 12.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012 # Visual Studio 2019
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "shfl_scan", "shfl_scan_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "batchedLabelMarkersAndLabelCompressionNPP", "batchedLabelMarkersAndLabelCompressionNPP_vs2019.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,15 +15,16 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>conjugateGradientCudaGraphs_vs2012</RootNamespace> <RootNamespace>batchedLabelMarkersAndLabelCompressionNPP_vs2019</RootNamespace>
<ProjectName>conjugateGradientCudaGraphs</ProjectName> <ProjectName>batchedLabelMarkersAndLabelCompressionNPP</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> <PlatformToolset>v142</PlatformToolset>
<WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -33,7 +34,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -57,12 +58,12 @@
</ClCompile> </ClCompile>
<Link> <Link>
<SubSystem>Console</SubSystem> <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> <AdditionalDependencies>nppisu.lib;nppif.lib;nppc.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>$(OutDir)/conjugateGradientCudaGraphs.exe</OutputFile> <OutputFile>$(OutDir)/batchedLabelMarkersAndLabelCompressionNPP.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;compute_75,sm_75;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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="conjugateGradientCudaGraphs.cu" /> <ClCompile Include="batchedLabelMarkersAndLabelCompressionNPP.cpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -0,0 +1,362 @@
################################################################################
# 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.
#
################################################################################
#
# 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 $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa 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 ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(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)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(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 -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(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)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(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)
$(info >>> WARNING - bf16TensorCoreGemm is not supported on Mac OSX - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
# This sample is not supported on ARMv7
ifeq ($(TARGET_ARCH),armv7l)
$(info >>> WARNING - bf16TensorCoreGemm 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 :=
################################################################################
#Detect if installed version of GCC supports required C++11
ifeq ($(TARGET_OS),linux)
empty :=
space := $(empty) $(empty)
GCCVERSIONSTRING := $(shell expr `$(HOST_COMPILER) -dumpversion`)
#Create version number without "."
GCCVERSION := $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f1 -d.)
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f2 -d.)
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f3 -d.)
# Make sure the version number has at least 3 decimals
GCCVERSION += 00
# Remove spaces from the version number
GCCVERSION := $(subst $(space),$(empty),$(GCCVERSION))
# Crop the version number to 3 decimals.
GCCVERSION := $(shell expr `echo $(GCCVERSION)` | cut -b1-3)
#$(warning $(GCCVERSION))
IS_MIN_VERSION := $(shell expr `echo $(GCCVERSION)` \>= 500)
ifeq ($(IS_MIN_VERSION), 1)
$(info >>> GCC Version is greater or equal to 5.0.0 <<<)
else
$(info >>> Waiving build. Minimum GCC version required is 5.0.0<<<)
SAMPLE_ENABLED := 0
endif
endif
# Gencode arguments
SMS ?= 80 86
ifeq ($(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 += --std=c++11
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: bf16TensorCoreGemm
check.deps:
ifeq ($(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
bf16TensorCoreGemm.o:bf16TensorCoreGemm.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
bf16TensorCoreGemm: bf16TensorCoreGemm.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
$(EXEC) ./bf16TensorCoreGemm
clean:
rm -f bf16TensorCoreGemm bf16TensorCoreGemm.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/bf16TensorCoreGemm
clobber: clean

View File

@@ -0,0 +1,68 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>bf16TensorCoreGemm</name>
<cflags>
<flag>--std=c++11</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[A CUDA sample demonstrating __nv_bfloat16 (e8m7) GEMM computation using the Warp Matrix Multiply and Accumulate (WMMA) API introduced with CUDA 11 in Ampere chip family tensor cores for faster matrix operations. This sample also uses async copy provided by cuda pipeline interface for gmem to shmem async loads which improves kernel performance and reduces register presssure.]]></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>
<keyword>matrix multiply</keyword>
<keyword>Async copy</keyword>
<keyword>CPP11</keyword>
<keyword>GCC 5.0.0</keyword>
</keywords>
<libraries>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>bf16TensorCoreGemm.cu</primary_file>
<scopes>
<scope>1:CUDA Basic Topics</scope>
</scopes>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</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>8.0</from>
</supported_sm_architectures>
<title>bfloat16 Tensor Core GEMM</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,70 @@
# bf16TensorCoreGemm - bfloat16 Tensor Core GEMM
## Description
A CUDA sample demonstrating __nv_bfloat16 (e8m7) GEMM computation using the Warp Matrix Multiply and Accumulate (WMMA) API introduced with CUDA 11 in Ampere chip family tensor cores for faster matrix operations. This sample also uses async copy provided by cuda pipeline interface for gmem to shmem async loads which improves kernel performance and reduces register presssure.
## Key Concepts
Matrix Multiply, WMMA, Tensor Cores
## Supported SM Architectures
[SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](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 11.1](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 +1,838 @@
/* Copyright (c) 2020, 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 __nv_bfloat16 (E8M7) GEMM computation using the Warp Matrix Multiply
// and Accumulate API introduced in CUDA 11.0.
// 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_BF16 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 <stdio.h>
#include <cuda.h>
#include <cuda_bf16.h>
#include <mma.h>
#include <cuda_pipeline.h>
// helper functions and utilities to work with CUDA
#include <helper_functions.h>
#include <helper_cuda.h>
// Externally configurable parameters.
// Switch for choosing cpp interface for cuda pipeline
// vs primitives interface.
#define USE_CPP_API 0
#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 0
#endif
// GPU configuration.
#define WARP_SIZE 32
// MMA matrix tile dimensions.
#define M 16
#define N 16
#define K 16
// GEMM configuration.
#define M_TILES 512
#define N_TILES 512
#define K_TILES 512
#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 is (M = 16) * (K = 16) * 8 * (CHUNK_K = 8)
// * sizeof(__nv_bfloat16) = 32 Kb each.
// (i.e. two 8x8 arrays of tiles of 16x16 __nv_bfloat16-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
#endif
#define CHUNK_LINE_BYTES (CHUNK_K * K * sizeof(__nv_bfloat16))
#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 two-byte "__nv_bfloat16" elements is chosen as the minimum possible shift because
// we must keep each row and column 256-bit aligned, as required by nvcuda::wmma::load_matrix_sync.
#define SKEW_BF16 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)
enum kernels
{
bf16mma_shmem_gemm_async_copy = 0, // __nv_bfloat16 MMA shmem using kernel with async_copy
bf16mma_shmem_gemm = 1, // __nv_bfloat16 MMA shmem using kernel normal copy (without async_copy).
simple_bf16mma_gemm = 2 // __nv_bfloat16 MMA non-shmem using simple kernel.
};
const char* kernelNames[] = {"compute_bf16gemm_async_copy", "compute_bf16gemm",
"simple_wmma_bf16gemm"};
using namespace nvcuda;
namespace nvcuda_namespace = nvcuda::experimental;
__host__ void init_host_matrices(__nv_bfloat16 *a, __nv_bfloat16 *b, float *c)
{
for (int i = 0; i < M_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) {
a[i*K_GLOBAL+j] = (__nv_bfloat16)(rand() % 3);
}
}
for (int i = 0; i < N_GLOBAL; i++) {
for (int j = 0; j < K_GLOBAL; j++) {
b[i*K_GLOBAL+j] = (__nv_bfloat16)(rand() % 3);
}
}
for (int t = 0; t < M_GLOBAL * N_GLOBAL; t++) {
c[t] = (float)(rand() % 3);
}
}
__global__ void compute_bf16gemm(const __nv_bfloat16 *A, const __nv_bfloat16 *B, const float *C, float *D, float alpha, float beta)
{
#if __CUDA_ARCH__ >= 800
extern __shared__ __nv_bfloat16 shmem[][CHUNK_K * K + SKEW_BF16];
// 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.
float *shmem_warp_tile_ptr = (float*)&shmem[0][0] + (warpId / BLOCK_ROW_WARPS) * SHMEM_STRIDE * N * BLOCK_ROW_WARPS + (warpId % BLOCK_ROW_WARPS) * SHMEM_OFFSET;
// This pointer is used to stream the C and D matrices block-wide tile to and from shared memory.
float *shmem_warp_stream_ptr = (float*)&shmem[0][0] + warpId * SHMEM_STRIDE * N;
// 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 float *src_gmem_warp_stream_ptr = &C[gmem_idx];
// Stream multiple C tiles to shared memory.
#pragma unroll
for (int i = 0; i < N; i++) {
*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId) =
*((int4*)(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, float> 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 float *tile_ptr = shmem_warp_tile_ptr + i * SHMEM_STRIDE * N + 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 __nv_bfloat16 *warp_ptr = (warpId < (WARPS_PER_BLOCK/2)) ? (&A[block_tile_i * M * K_GLOBAL] + M * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 2) :
(&B[block_tile_j * N * K_GLOBAL] + N * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 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.
const __nv_bfloat16 *lane_ptr = (warp_ptr + tile_k * K + (laneId / CHUNK_COPY_LINE_LANES) * K_GLOBAL);
// 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)) = *((int4*)lane_ptr + (laneId % CHUNK_COPY_LINE_LANES));
// Advance the global memory pointer and the shared memory index.
lane_ptr = 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, __nv_bfloat16, wmma::row_major> a[WARP_COL_TILES];
wmma::fragment<wmma::matrix_b, M, N, K, __nv_bfloat16, wmma::col_major> b[WARP_ROW_TILES];
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
size_t shmem_idx_a = (warpId/BLOCK_ROW_WARPS) * M * BLOCK_ROW_WARPS + (i * M);
const __nv_bfloat16 *tile_ptr = &shmem[shmem_idx_a][k_step * K];
wmma::load_matrix_sync(a[i], tile_ptr, K * CHUNK_K + SKEW_BF16);
#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 __nv_bfloat16 *tile_ptr = &shmem[shmem_idx_b][k_step * K];
wmma::load_matrix_sync(b[j], tile_ptr, K * CHUNK_K + SKEW_BF16);
}
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;
float *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.
float *dst_gmem_warp_stream_ptr = &D[gmem_idx];
#pragma unroll
for (int i = 0; i < N; i++) {
*((int4*)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
}
__syncthreads();
}
#endif
}
__global__ void compute_bf16gemm_async_copy(const __nv_bfloat16 *A, const __nv_bfloat16 *B, const float *C, float *D, float alpha, float beta)
{
#if __CUDA_ARCH__ >= 800
extern __shared__ __nv_bfloat16 shmem[][CHUNK_K * K + SKEW_BF16];
// 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.
float *shmem_warp_tile_ptr = (float*)&shmem[0][0] + (warpId / BLOCK_ROW_WARPS) * SHMEM_STRIDE * N * BLOCK_ROW_WARPS + (warpId % BLOCK_ROW_WARPS) * SHMEM_OFFSET;
// This pointer is used to stream the C and D matrices block-wide tile to and from shared memory.
float *shmem_warp_stream_ptr = (float*)&shmem[0][0] + warpId * SHMEM_STRIDE * N;
// 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;
#if USE_CPP_API
nvcuda_namespace::pipeline pipe;
#endif
// 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 float *src_gmem_warp_stream_ptr = &C[gmem_idx];
// Stream multiple C tiles to shared memory.
#pragma unroll
for (int i = 0; i < N; i++) {
#if USE_CPP_API
nvcuda_namespace::memcpy_async(*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId),
*((int4*)(src_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId),
pipe);
pipe.commit();
#else
__pipeline_memcpy_async((reinterpret_cast<int4*>(&shmem_warp_stream_ptr[(SHMEM_STRIDE * i)])) + laneId,
(reinterpret_cast<const int4*>(&src_gmem_warp_stream_ptr[(GLOBAL_MEM_STRIDE * i)])) + laneId,
sizeof(int4));
__pipeline_commit();
#endif
}
#if USE_CPP_API
pipe.wait_prior<0>();
#else
__pipeline_wait_prior(0);
#endif
__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, float> 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 float *tile_ptr = shmem_warp_tile_ptr + i * SHMEM_STRIDE * N + 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 __nv_bfloat16 *warp_ptr = (warpId < (WARPS_PER_BLOCK/2)) ? (&A[block_tile_i * M * K_GLOBAL] + M * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 2) :
(&B[block_tile_j * N * K_GLOBAL] + N * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 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.
const __nv_bfloat16 *lane_ptr = (warp_ptr + tile_k * K + (laneId / CHUNK_COPY_LINE_LANES) * K_GLOBAL);
// 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.
#if USE_CPP_API
nvcuda_namespace::memcpy_async(*((int4*)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES)),
*((int4*)lane_ptr + (laneId % CHUNK_COPY_LINE_LANES)), pipe);
pipe.commit();
#else
__pipeline_memcpy_async((int4*)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES),
(int4*)lane_ptr + (laneId % CHUNK_COPY_LINE_LANES), sizeof(int4));
__pipeline_commit();
#endif
// Advance the global memory pointer and the shared memory index.
lane_ptr = lane_ptr + K_GLOBAL * CHUNK_COPY_LINES_PER_WARP;
shmem_idx += CHUNK_COPY_LINES_PER_WARP;
}
#if USE_CPP_API
pipe.wait_prior<0>();
#else
__pipeline_wait_prior(0);
#endif
__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, __nv_bfloat16, wmma::row_major> a[WARP_COL_TILES];
wmma::fragment<wmma::matrix_b, M, N, K, __nv_bfloat16, wmma::col_major> b[WARP_ROW_TILES];
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
size_t shmem_idx_a = (warpId / BLOCK_ROW_WARPS) * M * BLOCK_ROW_WARPS + (i * M);
const __nv_bfloat16 *tile_ptr = &shmem[shmem_idx_a][k_step * K];
wmma::load_matrix_sync(a[i], tile_ptr, K * CHUNK_K + SKEW_BF16);
#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 __nv_bfloat16 *tile_ptr = &shmem[shmem_idx_b][k_step * K];
wmma::load_matrix_sync(b[j], tile_ptr, K * CHUNK_K + SKEW_BF16);
}
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;
float *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.
float *dst_gmem_warp_stream_ptr = &D[gmem_idx];
#pragma unroll
for (int i = 0; i < N; i++) {
*((int4*)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
}
__syncthreads();
}
#endif
}
// Performs an MxNxK bf16 GEMM (C=alpha*A*B + beta*C) assuming:
// 1) Matrices are packed in memory.
// 2) M, N and K are multiples of 16, 16 and 16 respectively.
// 3) A is row major, B is column major matrix.
// Note: This is a less performant version of the compute_bf16gemm 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_bf16gemm(__nv_bfloat16 *a, __nv_bfloat16 *b, float *c, float *d, int m_ld, int n_ld, int k_ld, float alpha, float beta)
{
#if __CUDA_ARCH__ >= 800
// Leading dimensions. Packed with no transpositions.
int lda = k_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, M, N, K, __nv_bfloat16, wmma::row_major> a_frag;
wmma::fragment<wmma::matrix_b, M, N, K, __nv_bfloat16, wmma::col_major> b_frag;
wmma::fragment<wmma::accumulator, M, N, K, float> acc_frag;
wmma::fragment<wmma::accumulator, M, N, K, float> c_frag;
wmma::fill_fragment(acc_frag, 0.0f);
// Loop over k
for (int i = 0; i < k_ld; i += K) {
int aCol = i;
int aRow = warpM * M;
int bCol = i;
int bRow = warpN * 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 + bRow + bCol * 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 * N;
int cRow = warpM * 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);
}
#endif
}
__host__ void matMultiplyOnHost(__nv_bfloat16 *A, __nv_bfloat16 *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)
{
printf("Initializing...\n");
int dev = findCudaDevice(argc, (const char **)argv);
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
// Tensor cores require a GPU of Volta (SM8X) architecture or higher.
if (deviceProp.major < 8) {
printf("bf16TensorCoreGemm requires requires SM 8.0 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);
__nv_bfloat16 *A_h = NULL;
__nv_bfloat16 *B_h = NULL;
float *C_h = NULL;
#if CPU_DEBUG
float *result_hD = NULL;
float *result_host = NULL;
#endif
A_h = (__nv_bfloat16*) malloc(sizeof(__nv_bfloat16) * M_GLOBAL * K_GLOBAL);
B_h = (__nv_bfloat16*) malloc(sizeof(__nv_bfloat16) * K_GLOBAL * N_GLOBAL);
C_h = (float*) malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
#if CPU_DEBUG
result_hD = (float*) malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
result_host = (float*) malloc(sizeof(float) * M_GLOBAL * N_GLOBAL);
#endif
__nv_bfloat16 *A = NULL;
__nv_bfloat16 *B = NULL;
float *C = NULL;
float *D = NULL;
checkCudaErrors(cudaMalloc((void**)&A, sizeof(__nv_bfloat16) * M_GLOBAL * K_GLOBAL));
checkCudaErrors(cudaMalloc((void**)&B, sizeof(__nv_bfloat16) * N_GLOBAL * K_GLOBAL));
checkCudaErrors(cudaMalloc((void**)&C, sizeof(float) * M_GLOBAL * N_GLOBAL));
checkCudaErrors(cudaMalloc((void**)&D, sizeof(float) * 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);
printf("Preparing data for GPU...\n");
checkCudaErrors(cudaMemcpy(A, A_h, sizeof(__nv_bfloat16) * M_GLOBAL * K_GLOBAL, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(B, B_h, sizeof(__nv_bfloat16) * N_GLOBAL * K_GLOBAL, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(C, C_h, sizeof(float) * M_GLOBAL * N_GLOBAL, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemset(D, 0, sizeof(float) * M_GLOBAL * N_GLOBAL));
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(__nv_bfloat16) * (BLOCK_COL_TILES * M) * (CHUNK_K * K + SKEW_BF16) * 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);
const float alpha = 1.1f;
const float beta = 1.2f;
cudaEvent_t start, stop;
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
checkCudaErrors(cudaEventRecord(start));
// kernel to run - default (b16mma_shmem_gemm_async_copy == 0)
kernels selected_kernel = bf16mma_shmem_gemm_async_copy;
if (checkCmdLineFlag(argc, (const char **)argv, "kernel")) {
int kernel_number = getCmdLineArgumentInt(argc, (const char **)argv, "kernel");
if (kernel_number < 3) {
selected_kernel = (kernels)kernel_number;
}
else {
printf("Error: kernel number should be between 0 to 2, you have entered %d\n", kernel_number);
exit(EXIT_FAILURE);
}
}
// If enough shared memory available on the GPU use high performant kernel
if ((deviceProp.sharedMemPerMultiprocessor >= SHMEM_SZ) && (selected_kernel != simple_bf16mma_gemm)) {
printf("Computing using high performance kernel = %d - %s\n", selected_kernel, kernelNames[selected_kernel]);
switch (selected_kernel)
{
case bf16mma_shmem_gemm_async_copy :
default:
checkCudaErrors(cudaFuncSetAttribute(compute_bf16gemm_async_copy, cudaFuncAttributeMaxDynamicSharedMemorySize, SHMEM_SZ));
checkKernelErrors((compute_bf16gemm_async_copy<<<deviceProp.multiProcessorCount*2, THREADS_PER_BLOCK, SHMEM_SZ>>>(A, B, C, D, alpha, beta)));
break;
case bf16mma_shmem_gemm :
checkCudaErrors(cudaFuncSetAttribute(compute_bf16gemm, cudaFuncAttributeMaxDynamicSharedMemorySize, SHMEM_SZ));
checkKernelErrors((compute_bf16gemm<<<deviceProp.multiProcessorCount*2, THREADS_PER_BLOCK, SHMEM_SZ>>>(A, B, C, D, alpha, beta)));
break;
}
#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 + (M * blockDim.x / 32 - 1)) / (M * blockDim.x / 32);
gridDim.y = (N_GLOBAL + N * blockDim.y - 1) / (N * blockDim.y);
printf("Computing... using simple_wmma_gemm kernel\n");
simple_wmma_bf16gemm<<<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(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;
checkCudaErrors(cudaEventElapsedTime(&milliseconds, start, stop));
printf("Time: %f ms\n", milliseconds);
printf("TFLOPS: %.2f\n", (((double)M_GLOBAL * N_GLOBAL * K_GLOBAL * 2)/(milliseconds/1000.)) / 1e12);
free(A_h);
free(B_h);
free(C_h);
checkCudaErrors(cudaFree((void*)A));
checkCudaErrors(cudaFree((void*)B));
checkCudaErrors(cudaFree((void*)C));
checkCudaErrors(cudaFree((void*)D));
return 0;
}

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 13.00 Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2013 # Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "NV12toBGRandResize", "NV12toBGRandResize_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bf16TensorCoreGemm", "bf16TensorCoreGemm_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,15 +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_vs2012</RootNamespace> <RootNamespace>bf16TensorCoreGemm_vs2015</RootNamespace>
<ProjectName>cudaTensorCoreGemm</ProjectName> <ProjectName>bf16TensorCoreGemm</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> <PlatformToolset>v140</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -59,10 +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)/bf16TensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_70,sm_70;compute_75,sm_75;</CodeGeneration> <CodeGeneration>compute_80,sm_80;compute_86,sm_86;</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 +97,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="cudaTensorCoreGemm.cu" /> <CudaCompile Include="bf16TensorCoreGemm.cu" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 12.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012 # Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "NV12toBGRandResize", "NV12toBGRandResize_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bf16TensorCoreGemm", "bf16TensorCoreGemm_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -15,15 +15,20 @@
</ItemGroup> </ItemGroup>
<PropertyGroup Label="Globals"> <PropertyGroup Label="Globals">
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid> <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
<RootNamespace>UnifiedMemoryPerf_vs2013</RootNamespace> <RootNamespace>bf16TensorCoreGemm_vs2017</RootNamespace>
<ProjectName>UnifiedMemoryPerf</ProjectName> <ProjectName>bf16TensorCoreGemm</ProjectName>
<CudaToolkitCustomDir /> <CudaToolkitCustomDir />
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
<LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
<WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
<TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup> <PropertyGroup>
<ConfigurationType>Application</ConfigurationType> <ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet> <CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v120</PlatformToolset> <PlatformToolset>v141</PlatformToolset>
</PropertyGroup> </PropertyGroup>
<PropertyGroup Condition="'$(Configuration)'=='Debug'"> <PropertyGroup Condition="'$(Configuration)'=='Debug'">
<UseDebugLibraries>true</UseDebugLibraries> <UseDebugLibraries>true</UseDebugLibraries>
@@ -33,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup> </ImportGroup>
<ImportGroup Label="PropertySheets"> <ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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.$(Platform).user.props" />
@@ -59,10 +64,10 @@
<SubSystem>Console</SubSystem> <SubSystem>Console</SubSystem>
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies> <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories> <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/UnifiedMemoryPerf.exe</OutputFile> <OutputFile>$(OutDir)/bf16TensorCoreGemm.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration> <CodeGeneration>compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions> <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include> <Include>./;../../Common</Include>
<Defines>WIN32</Defines> <Defines>WIN32</Defines>
@@ -97,14 +102,11 @@
</CudaCompile> </CudaCompile>
</ItemDefinitionGroup> </ItemDefinitionGroup>
<ItemGroup> <ItemGroup>
<CudaCompile Include="commonKernels.cu" /> <CudaCompile Include="bf16TensorCoreGemm.cu" />
<ClCompile Include="helperFunctions.cpp" />
<CudaCompile Include="matrixMultiplyPerf.cu" />
<ClInclude Include="commonDefs.hpp" />
<ClInclude Include="commonKernels.hpp" />
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 12.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012 # Visual Studio 2019
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "cudaTensorCoreGemm", "cudaTensorCoreGemm_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "bf16TensorCoreGemm", "bf16TensorCoreGemm_vs2019.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

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>bf16TensorCoreGemm_vs2019</RootNamespace>
<ProjectName>bf16TensorCoreGemm</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v142</PlatformToolset>
<WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(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="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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>$(OutDir)/bf16TensorCoreGemm.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="bf16TensorCoreGemm.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

View File

@@ -0,0 +1,360 @@
################################################################################
# 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.
#
################################################################################
#
# 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 $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32
endif
else
TARGET_SIZE := $(shell getconf LONG_BIT)
endif
else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa 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 ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(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)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(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 -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif
endif
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif
# Install directory of different arch
CUDA_INSTALL_TARGET_DIR :=
ifeq ($(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)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(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 - binaryPartitionCG is not supported on Mac OSX - 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 :=
################################################################################
#Detect if installed version of GCC supports required C++11
ifeq ($(TARGET_OS),linux)
empty :=
space := $(empty) $(empty)
GCCVERSIONSTRING := $(shell expr `$(HOST_COMPILER) -dumpversion`)
#Create version number without "."
GCCVERSION := $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f1 -d.)
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f2 -d.)
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f3 -d.)
# Make sure the version number has at least 3 decimals
GCCVERSION += 00
# Remove spaces from the version number
GCCVERSION := $(subst $(space),$(empty),$(GCCVERSION))
# Crop the version number to 3 decimals.
GCCVERSION := $(shell expr `echo $(GCCVERSION)` | cut -b1-3)
#$(warning $(GCCVERSION))
IS_MIN_VERSION := $(shell expr `echo $(GCCVERSION)` \>= 470)
ifeq ($(IS_MIN_VERSION), 1)
$(info >>> GCC Version is greater or equal to 4.7.0 <<<)
else
$(info >>> Waiving build. Minimum GCC version required is 4.7.0<<<)
SAMPLE_ENABLED := 0
endif
endif
# Gencode arguments
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
SMS ?= 35 37 50 52 60 61 70 72 75 80 86
else
SMS ?= 35 37 50 52 60 61 70 75 80 86
endif
ifeq ($(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 += --std=c++11
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"
endif
################################################################################
# Target rules
all: build
build: binaryPartitionCG
check.deps:
ifeq ($(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
binaryPartitionCG.o:binaryPartitionCG.cu
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
binaryPartitionCG: binaryPartitionCG.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
$(EXEC) ./binaryPartitionCG
clean:
rm -f binaryPartitionCG binaryPartitionCG.o
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/binaryPartitionCG
clobber: clean

View File

@@ -0,0 +1,65 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>binaryPartitionCG</name>
<cflags>
<flag>--std=c++11</flag>
</cflags>
<description><![CDATA[This sample is a simple code that illustrates binary partition cooperative groups and reduce within the thread block.]]></description>
<devicecompilation>whole</devicecompilation>
<includepaths>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">Cooperative Groups</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>Parallel Reduction</keyword>
<keyword>Cooperative Groups</keyword>
<keyword>CPP11</keyword>
</keywords>
<libraries>
</libraries>
<librarypaths>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>binaryPartitionCG.cu</primary_file>
<scopes>
<scope>1:CUDA Basic Topics</scope>
</scopes>
<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>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</sm-arch>
<supported_envs>
<env>
<arch>x86_64</arch>
<platform>linux</platform>
</env>
<env>
<platform>windows7</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>Binary Partition Cooperative Groups</title>
<type>exe</type>
</entry>

View File

@@ -0,0 +1,67 @@
# binaryPartitionCG - Binary Partition Cooperative Groups
## Description
This sample is a simple code that illustrates binary partition cooperative groups and reduce within the thread block.
## Key Concepts
Cooperative Groups
## Supported SM Architectures
[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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, ppc64le, armv7l
## CUDA APIs involved
## Prerequisites
Download and install the [CUDA Toolkit 11.1](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++
```
## References (for more details)

View File

@@ -0,0 +1,155 @@
/* Copyright (c) 2020, 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 illustrates basic usage of binary partition cooperative groups
* within the thread block tile when divergent path exists.
* 1.) Each thread loads a value from random array.
* 2.) then checks if it is odd or even.
* 3.) create binary partition group based on the above predicate
* 4.) we count the number of odd/even in the group based on size of the binary groups
* 5.) write it global counter of odd.
* 6.) sum the values loaded by individual threads(using reduce) and write it to global
* even & odd elements sum.
*
* **NOTE** : binary_partition results in splitting warp into divergent thread groups
this is not good from performance perspective, but in cases where warp
divergence is inevitable one can use binary_partition group.
*/
#include <stdio.h>
#include <cooperative_groups.h>
#include <cooperative_groups/reduce.h>
#include <helper_cuda.h>
namespace cg = cooperative_groups;
void initOddEvenArr(int *inputArr, unsigned int size)
{
for (int i=0; i < size; i++)
{
inputArr[i] = rand() % 50;
}
}
/**
* CUDA kernel device code
*
* Creates cooperative groups and performs odd/even counting & summation.
*/
__global__ void oddEvenCountAndSumCG(int *inputArr, int *numOfOdds, int *sumOfOddAndEvens, unsigned int size)
{
cg::thread_block cta = cg::this_thread_block();
cg::grid_group grid = cg::this_grid();
cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
for (int i = grid.thread_rank(); i < size; i += grid.size())
{
int elem = inputArr[i];
auto subTile = cg::binary_partition(tile32, elem & 1);
if (elem & 1) // Odd numbers group
{
int oddGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
if (subTile.thread_rank() == 0)
{
// Add number of odds present in this group of Odds.
atomicAdd(numOfOdds, subTile.size());
// Add local reduction of odds present in this group of Odds.
atomicAdd(&sumOfOddAndEvens[0], oddGroupSum);
}
}
else // Even numbers group
{
int evenGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
if (subTile.thread_rank() == 0)
{
// Add local reduction of even present in this group of evens.
atomicAdd(&sumOfOddAndEvens[1], evenGroupSum);
}
}
// reconverge warp so for next loop iteration we ensure convergence of
// above diverged threads to perform coalesced loads of inputArr.
cg::sync(tile32);
}
}
/**
* Host main routine
*/
int main(int argc, const char **argv)
{
int deviceId = findCudaDevice(argc, argv);
int *h_inputArr, *d_inputArr;
int *h_numOfOdds, *d_numOfOdds;
int *h_sumOfOddEvenElems, *d_sumOfOddEvenElems;
unsigned int arrSize = 1024 * 100;
h_inputArr = new int[arrSize];
h_numOfOdds = new int[1];
h_sumOfOddEvenElems = new int[2];
initOddEvenArr(h_inputArr, arrSize);
cudaStream_t stream;
checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
checkCudaErrors(cudaMalloc(&d_inputArr, sizeof(int)*arrSize));
checkCudaErrors(cudaMalloc(&d_numOfOdds, sizeof(int)));
checkCudaErrors(cudaMalloc(&d_sumOfOddEvenElems, sizeof(int)*2));
checkCudaErrors(cudaMemcpyAsync(d_inputArr, h_inputArr, sizeof(int)*arrSize, cudaMemcpyHostToDevice, stream));
checkCudaErrors(cudaMemsetAsync(d_numOfOdds, 0, sizeof(int), stream));
checkCudaErrors(cudaMemsetAsync(d_sumOfOddEvenElems, 0, 2*sizeof(int), stream));
//Launch the kernel
int threadsPerBlock=1024;
int blocksPerGrid = arrSize / threadsPerBlock;
printf("\nLaunching %d blocks with %d threads...\n\n",blocksPerGrid, threadsPerBlock);
oddEvenCountAndSumCG<<<blocksPerGrid, threadsPerBlock, 0, stream>>>(d_inputArr, d_numOfOdds, d_sumOfOddEvenElems, arrSize);
checkCudaErrors(cudaMemcpyAsync(h_numOfOdds, d_numOfOdds, sizeof(int), cudaMemcpyDeviceToHost, stream));
checkCudaErrors(cudaMemcpyAsync(h_sumOfOddEvenElems, d_sumOfOddEvenElems, 2*sizeof(int), cudaMemcpyDeviceToHost, stream));
printf("Array size = %d Num of Odds = %d Sum of Odds = %d Sum of Evens %d\n", arrSize, h_numOfOdds[0], h_sumOfOddEvenElems[0], h_sumOfOddEvenElems[1]);
printf("\n...Done.\n\n");
delete[] h_inputArr;
delete[] h_numOfOdds;
delete[] h_sumOfOddEvenElems;
checkCudaErrors(cudaFree(d_inputArr));
checkCudaErrors(cudaFree(d_numOfOdds));
checkCudaErrors(cudaFree(d_sumOfOddEvenElems));
return EXIT_SUCCESS;
}

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 13.00 Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2013 # Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2013.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "binaryPartitionCG", "binaryPartitionCG_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

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>binaryPartitionCG_vs2015</RootNamespace>
<ProjectName>binaryPartitionCG</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="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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>$(OutDir)/binaryPartitionCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="binaryPartitionCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

View File

@@ -1,7 +1,7 @@
 
Microsoft Visual Studio Solution File, Format Version 12.00 Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012 # Visual Studio 2017
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "UnifiedMemoryPerf", "UnifiedMemoryPerf_vs2012.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}" Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "binaryPartitionCG", "binaryPartitionCG_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
EndProject EndProject
Global Global
GlobalSection(SolutionConfigurationPlatforms) = preSolution GlobalSection(SolutionConfigurationPlatforms) = preSolution

View File

@@ -0,0 +1,112 @@
<?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>binaryPartitionCG_vs2017</RootNamespace>
<ProjectName>binaryPartitionCG</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
<LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
<WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
<TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v141</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 11.1.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)/binaryPartitionCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>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;compute_80,sm_80;compute_86,sm_86;</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="binaryPartitionCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

View File

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

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2019
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "binaryPartitionCG", "binaryPartitionCG_vs2019.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>binaryPartitionCG_vs2019</RootNamespace>
<ProjectName>binaryPartitionCG</ProjectName>
<CudaToolkitCustomDir />
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
<PropertyGroup>
<ConfigurationType>Application</ConfigurationType>
<CharacterSet>MultiByte</CharacterSet>
<PlatformToolset>v142</PlatformToolset>
<WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
</PropertyGroup>
<PropertyGroup Condition="'$(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="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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>$(OutDir)/binaryPartitionCG.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="binaryPartitionCG.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -259,8 +280,8 @@ ifeq ($(GENCODE_FLAGS),)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm))) $(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
ifeq ($(SMS),) ifeq ($(SMS),)
# Generate PTX code from SM 30 # Generate PTX code from SM 35
GENCODE_FLAGS += -gencode arch=compute_30,code=compute_30 GENCODE_FLAGS += -gencode arch=compute_35,code=compute_35
endif endif
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility # Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility

View File

@@ -0,0 +1,88 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>boxFilterNPP</name>
<description><![CDATA[A NPP CUDA Sample that demonstrates how to use NPP FilterBox function to perform a Box Filter.]]></description>
<devicecompilation>whole</devicecompilation>
<fallback_min_ptx>true</fallback_min_ptx>
<files>
<file>./Lena.pgm</file>
</files>
<includepaths>
<path>../../Common/UtilNPP</path>
<path>../../Common/FreeImage/include</path>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">Performance Strategies</concept>
<concept level="basic">Image Processing</concept>
<concept level="basic">NPP Library</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>NPP</keyword>
<keyword>Image Processing</keyword>
<keyword>box filter</keyword>
</keywords>
<libraries>
<library>nppisu_static</library>
<library>nppif_static</library>
<library>nppc_static</library>
<library>culibos</library>
<library>freeimage</library>
</libraries>
<librarypaths>
<path>../../Common/FreeImage/lib/$(TARGET_OS)</path>
<path>../../Common/FreeImage/lib/$(TARGET_OS)/$(TARGET_ARCH)</path>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>boxFilterNPP.cpp</primary_file>
<required_dependencies>
<dependency>FreeImage</dependency>
<dependency>NPP</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Basic Topics</scope>
<scope>1:Performance Strategies</scope>
<scope>2:Image Processing</scope>
<scope>2:Computer Vision</scope>
</scopes>
<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>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</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>Box Filter with NPP</title>
<type>exe</type>
</entry>

View File

@@ -10,11 +10,11 @@ Performance Strategies, Image Processing, NPP Library
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux, Windows, MacOSX Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -27,7 +27,7 @@ x86_64, ppc64le, armv7l
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](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
@@ -67,29 +67,5 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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) ## References (for more details)

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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/boxFilterNPP.exe</OutputFile> <OutputFile>$(OutDir)/boxFilterNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
@@ -112,6 +112,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -38,7 +38,7 @@
</PropertyGroup> </PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings"> <ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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" />
@@ -67,7 +67,7 @@
<OutputFile>$(OutDir)/boxFilterNPP.exe</OutputFile> <OutputFile>$(OutDir)/boxFilterNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
@@ -117,6 +117,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/boxFilterNPP.exe</OutputFile> <OutputFile>$(OutDir)/boxFilterNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
@@ -113,6 +113,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

View File

@@ -72,9 +72,9 @@ endif
# architecture # architecture
HOST_ARCH := $(shell uname -m) HOST_ARCH := $(shell uname -m)
TARGET_ARCH ?= $(HOST_ARCH) TARGET_ARCH ?= $(HOST_ARCH)
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le)) ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
TARGET_SIZE := 64 TARGET_SIZE := 64
else ifneq (,$(filter $(TARGET_ARCH),armv7l)) else ifneq (,$(filter $(TARGET_ARCH),armv7l))
TARGET_SIZE := 32 TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
else else
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!) $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
endif endif
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
ifeq ($(HOST_ARCH),aarch64)
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
HOST_ARCH := sbsa
TARGET_ARCH := sbsa
endif
endif
ifneq ($(TARGET_ARCH),$(HOST_ARCH)) ifneq ($(TARGET_ARCH),$(HOST_ARCH))
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-ppc64le)) ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!) $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
endif endif
endif endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
else ifeq ($(TARGET_OS), android) else ifeq ($(TARGET_OS), android)
HOST_COMPILER ?= aarch64-linux-android-clang++ HOST_COMPILER ?= aarch64-linux-android-clang++
endif endif
else ifeq ($(TARGET_ARCH),sbsa)
HOST_COMPILER ?= aarch64-linux-gnu-g++
else ifeq ($(TARGET_ARCH),ppc64le) else ifeq ($(TARGET_ARCH),ppc64le)
HOST_COMPILER ?= powerpc64le-linux-gnu-g++ HOST_COMPILER ?= powerpc64le-linux-gnu-g++
endif endif
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
CCFLAGS += --sysroot=$(TARGET_FS) CCFLAGS += --sysroot=$(TARGET_FS)
endif endif
LDFLAGS += --sysroot=$(TARGET_FS) LDFLAGS += --sysroot=$(TARGET_FS)
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
CCFLAGS += -isystem=$(TARGET_FS)/usr/include CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
endif endif
endif endif
endif ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
ifeq ($(TARGET_OS),qnx)
CCFLAGS += -DWIN_INTERFACE_CUSTOM
LDFLAGS += -lsocket LDFLAGS += -lsocket
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
ifneq ($(TARGET_FS),)
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
endif
ifdef TARGET_OVERRIDE # cuda toolkit targets override
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
endif
endif
endif endif
# Install directory of different arch # Install directory of different arch
@@ -206,6 +225,8 @@ ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/ CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/ CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android) else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
@@ -259,8 +280,8 @@ ifeq ($(GENCODE_FLAGS),)
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm))) $(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
ifeq ($(SMS),) ifeq ($(SMS),)
# Generate PTX code from SM 30 # Generate PTX code from SM 35
GENCODE_FLAGS += -gencode arch=compute_30,code=compute_30 GENCODE_FLAGS += -gencode arch=compute_35,code=compute_35
endif endif
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility # Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility

View File

@@ -0,0 +1,84 @@
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
<entry>
<name>cannyEdgeDetectorNPP</name>
<description><![CDATA[An NPP CUDA Sample that demonstrates the recommended parameters to use with the nppiFilterCannyBorder_8u_C1R Canny Edge Detection image filter function. This function expects a single channel 8-bit grayscale input image. You can generate a grayscale image from a color image by first calling nppiColorToGray() or nppiRGBToGray(). The Canny Edge Detection function combines and improves on the techniques required to produce an edge detection image using multiple steps.]]></description>
<devicecompilation>whole</devicecompilation>
<fallback_min_ptx>true</fallback_min_ptx>
<includepaths>
<path>../../Common/UtilNPP</path>
<path>../../Common/FreeImage/include</path>
<path>./</path>
<path>../</path>
<path>../../common/inc</path>
</includepaths>
<keyconcepts>
<concept level="basic">Performance Strategies</concept>
<concept level="basic">Image Processing</concept>
<concept level="basic">NPP Library</concept>
</keyconcepts>
<keywords>
<keyword>CUDA</keyword>
<keyword>NPP</keyword>
<keyword>Image Processing</keyword>
</keywords>
<libraries>
<library>nppisu_static</library>
<library>nppif_static</library>
<library>nppc_static</library>
<library>culibos</library>
<library>freeimage</library>
</libraries>
<librarypaths>
<path>../../Common/FreeImage/lib/$(TARGET_OS)</path>
<path>../../Common/FreeImage/lib/$(TARGET_OS)/$(TARGET_ARCH)</path>
</librarypaths>
<nsight_eclipse>true</nsight_eclipse>
<primary_file>cannyEdgeDetectorNPP.cpp</primary_file>
<required_dependencies>
<dependency>FreeImage</dependency>
<dependency>NPP</dependency>
</required_dependencies>
<scopes>
<scope>1:CUDA Basic Topics</scope>
<scope>1:Performance Strategies</scope>
<scope>2:Image Processing</scope>
<scope>2:Computer Vision</scope>
</scopes>
<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>
<sm-arch>sm80</sm-arch>
<sm-arch>sm86</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>Canny Edge Detector NPP</title>
<type>exe</type>
</entry>

View File

@@ -10,11 +10,11 @@ Performance Strategies, Image Processing, NPP Library
## 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 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](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) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes ## Supported OSes
Linux, Windows, MacOSX Linux, Windows
## Supported CPU Architecture ## Supported CPU Architecture
@@ -27,7 +27,7 @@ x86_64, ppc64le, armv7l
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit 10.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit 11.1](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
@@ -67,29 +67,5 @@ The samples makefiles can take advantage of certain options:
$ make HOST_COMPILER=g++ $ 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) ## References (for more details)

View File

@@ -47,6 +47,27 @@
#include <helper_cuda.h> #include <helper_cuda.h>
#include <helper_string.h> #include <helper_string.h>
inline int cudaDeviceInit(int argc, const char **argv) {
int deviceCount;
checkCudaErrors(cudaGetDeviceCount(&deviceCount));
if (deviceCount == 0) {
std::cerr << "CUDA error: no devices supporting CUDA." << std::endl;
exit(EXIT_FAILURE);
}
int dev = findCudaDevice(argc, argv);
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, dev);
std::cerr << "cudaSetDevice GPU" << dev << " = " << deviceProp.name
<< std::endl;
checkCudaErrors(cudaSetDevice(dev));
return dev;
}
bool printfNPPinfo(int argc, char *argv[]) { bool printfNPPinfo(int argc, char *argv[]) {
const NppLibraryVersion *libVer = nppGetLibVersion(); const NppLibraryVersion *libVer = nppGetLibVersion();
@@ -74,7 +95,7 @@ int main(int argc, char *argv[]) {
std::string sFilename; std::string sFilename;
char *filePath; char *filePath;
int dev = findCudaDevice(argc, (const char **)argv); cudaDeviceInit(argc, (const char **)argv);
if (printfNPPinfo(argc, argv) == false) { if (printfNPPinfo(argc, argv) == false) {
exit(EXIT_SUCCESS); exit(EXIT_SUCCESS);

View File

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

Microsoft Visual Studio Solution File, Format Version 12.00
# Visual Studio 2012
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "cannyEdgeDetectorNPP", "cannyEdgeDetectorNPP_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

@@ -1,117 +0,0 @@
<?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>cannyEdgeDetectorNPP_vs2012</RootNamespace>
<ProjectName>cannyEdgeDetectorNPP</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>./;$(CudaToolkitDir)/include;../../Common;$(CudaToolkitIncludeDir);../../Common/UtilNPP;../../Common/FreeImage/include;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>nppisu.lib;nppif.lib;nppc.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>$(CudaToolkitLibDir);../../Common/FreeImage/lib/x64;../../Common/FreeImage/lib/$(TARGET_OS);../../Common/FreeImage/lib/$(TARGET_OS)/$(TARGET_ARCH);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/cannyEdgeDetectorNPP.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='Win32'">
<Link>
</Link>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='x64'">
<Link>
<AdditionalDependencies>freeimage.lib;%(AdditionalDependencies)</AdditionalDependencies>
</Link>
</ItemDefinitionGroup>
<ItemGroup>
<ClCompile Include="cannyEdgeDetectorNPP.cpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</Project>

View File

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

Microsoft Visual Studio Solution File, Format Version 13.00
# Visual Studio 2013
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "cannyEdgeDetectorNPP", "cannyEdgeDetectorNPP_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

@@ -1,117 +0,0 @@
<?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>cannyEdgeDetectorNPP_vs2013</RootNamespace>
<ProjectName>cannyEdgeDetectorNPP</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);../../Common/UtilNPP;../../Common/FreeImage/include;</AdditionalIncludeDirectories>
</ClCompile>
<Link>
<SubSystem>Console</SubSystem>
<AdditionalDependencies>nppisu.lib;nppif.lib;nppc.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>$(CudaToolkitLibDir);../../Common/FreeImage/lib/x64;../../Common/FreeImage/lib/$(TARGET_OS);../../Common/FreeImage/lib/$(TARGET_OS)/$(TARGET_ARCH);</AdditionalLibraryDirectories>
<OutputFile>$(OutDir)/cannyEdgeDetectorNPP.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='Win32'">
<Link>
</Link>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Platform)'=='x64'">
<Link>
<AdditionalDependencies>freeimage.lib;%(AdditionalDependencies)</AdditionalDependencies>
</Link>
</ItemDefinitionGroup>
<ItemGroup>
<ClCompile Include="cannyEdgeDetectorNPP.cpp" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" />
</ImportGroup>
</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="$(CUDAPropsPath)\CUDA 10.1.props" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.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>$(OutDir)/cannyEdgeDetectorNPP.exe</OutputFile> <OutputFile>$(OutDir)/cannyEdgeDetectorNPP.exe</OutputFile>
</Link> </Link>
<CudaCompile> <CudaCompile>
<CodeGeneration>compute_30,compute_30;</CodeGeneration> <CodeGeneration>compute_35,compute_35;</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>
@@ -112,6 +112,6 @@
</ItemGroup> </ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" /> <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets"> <ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 10.1.targets" /> <Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup> </ImportGroup>
</Project> </Project>

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