18 Commits
v13.1 ... v13.2

Author SHA1 Message Date
Rob Armstrong
a4526d5229 Merge branch 'shawnz_bug_fix' into 'master'
Fixing Bug 3570274, 3570274, 5952280, 5970525 and 5970583

See merge request cuda-samples/cuda-samples!150
2026-03-30 08:09:12 -07:00
Shawn Zeng
f69622b6b6 Bug 5970583: Update the incorrect links of dependencies for Samples 2026-03-12 16:51:36 +08:00
Shawn Zeng
722c5dd069 Bug 5970525: Update a minor issue of Samples/0_Introduction/fp16ScalarProduct 2026-03-11 16:22:10 +08:00
Shawn Zeng
e32d17e67e Bug 5952280: Add cuModuleUnload(Module) to release resource for 6 samples 2026-03-10 16:05:27 +08:00
Shawn Zeng
5cf83b5a98 Bug 3570274 and 3570274: Check Unified Memory is supported or not for executing Unified Memory samples 2026-03-03 12:20:07 +08:00
Rob Armstrong
191900a7b4 Update CUDA Programming Guide URL 2026-02-04 09:21:29 -08:00
Rob Armstrong
338df69d0f Merge branch 'Bug_5809031' into 'master'
Bug 5809031: Explicitly adding P2H2P fallback as cudaMemcpyPeerAsync...

See merge request cuda-samples/cuda-samples!148
2026-01-29 12:50:46 -08:00
Shawn Zeng
9b7a6f907a Bug 5809031: Explicitly ensure P2P is disabled before fallback detecting 2026-01-26 11:37:52 +08:00
Shawn Zeng
5bbae96fd3 Bug 5809031: Explicitly adding P2H2P fallback as cudaMemcpyPeerAsync automatically fallback not supported with P2P off and CC mode on 2026-01-23 18:56:38 +08:00
Rob Armstrong
ce9407c240 Merge branch 'master_bug_fix' into 'master'
Bug fix for CUDA 13.2 samples

See merge request cuda-samples/cuda-samples!146
2026-01-21 09:07:07 -08:00
Rob Armstrong
cdedff42d2 Merge branch 'make_install' into 'master'
Make_install: enable setting CUDA_SAMPLES_INSTALL_DIR

See merge request cuda-samples/cuda-samples!145
2026-01-21 09:06:50 -08:00
Shawn Zeng
6c4d183ba3 Bug 5812203: Replace thrust::make_tuple,tuple,get with the cuda::std ones 2026-01-21 15:48:47 +08:00
Shawn Zeng
af49216f0c Modify CUDA Toolkit version to 13.2 2026-01-21 15:48:47 +08:00
Shawn Zeng
17113ea87e Bug 5701423: Add MSVC-specific flag /Zc:preprocessor as CCCL change the warning of this flag to error 2026-01-21 15:48:47 +08:00
Shawn Zeng
70fcdd353f Bug 5742096: Using uint32_t instead of int to fix the overflow undefine dehavior issue of the sample 0_Introduction/simpleOccupancy 2026-01-21 15:48:47 +08:00
Shawn Zeng
5b2dd19a21 Bug 5056041: Fix the VUID-vkAcquireNextImageKHR-semaphore-01779 error of simpleVulkan 2026-01-21 15:48:47 +08:00
Shawn Zeng
37dcb9b61d Bug 5671906: Enable vulkan validaitonLayers for Vulkan samples with Debug build only 2026-01-21 15:48:47 +08:00
Shawn Zeng
1bc6774795 Make_install: enable setting CUDA_SAMPLES_INSTALL_DIR 2026-01-21 14:51:36 +08:00
80 changed files with 375 additions and 165 deletions

View File

@@ -1,5 +1,8 @@
## Changelog ## Changelog
### CUDA 13.2
* Added the MSVC compile flag `-Xcompiler=/Zc:preprocessor` in CMakeLists.txt to comply with CUDA13.2 CCCL. Previously, using the traditional preprocessor triggered the warning “MSVC/cl.exe with traditional preprocessor is used…”, which now leads to a build error.
### CUDA 13.1 ### CUDA 13.1
* Minor bug fixes and enhancements, no structural or functional changes * Minor bug fixes and enhancements, no structural or functional changes

View File

@@ -22,6 +22,11 @@ endif()
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --extended-lambda") set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --extended-lambda")
# Add MSVC-specific flags for standard-conforming preprocessor (required for CCCL)
if(MSVC)
add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/Zc:preprocessor>)
endif()
# Include installation configuration before processing samples # Include installation configuration before processing samples
include(cmake/InstallSamples.cmake) include(cmake/InstallSamples.cmake)

View File

@@ -1,6 +1,6 @@
# CUDA Samples # CUDA Samples
Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 13.1](https://developer.nvidia.com/cuda-downloads). Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 13.2](https://developer.nvidia.com/cuda-downloads).
## Release Notes ## Release Notes
@@ -604,6 +604,11 @@ NVJPEG library provides high-performance, GPU accelerated JPEG decoding function
NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++. NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++.
#### NVJITLINK
NVJITLINK is a CUDA runtime library that links multiple GPU device code objects at runtime and enables just‑in‑time link‑time optimization (JIT LTO).
It can take device code from tools such as NVCC and NVRTC, link and optimize it on the fly, and produce the final GPU binary.For more details, see: https://docs.nvidia.com/cuda/nvjitlink/index.html
#### Stream Priorities #### Stream Priorities
Stream Priorities allows the creation of streams with specified priorities. Stream Priorities is only available on GPUs with SM architecture of 3.5 or above. Stream Priorities allows the creation of streams with specified priorities. Stream Priorities is only available on GPUs with SM architecture of 3.5 or above.
@@ -636,7 +641,7 @@ Answers to frequently asked questions about CUDA can be found at http://develope
## References ## References
* [CUDA Programming Guide](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html) * [CUDA Programming Guide](https://docs.nvidia.com/cuda/cuda-programming-guide/index.html)
* [Accelerated Computing Blog](https://developer.nvidia.com/blog/?tags=accelerated-computing) * [Accelerated Computing Blog](https://developer.nvidia.com/blog/?tags=accelerated-computing)
## Attributions ## Attributions

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaStreamDestroy, cudaFree, cudaMallocManaged, cudaStreamAttachMemAsync, cudaSetDevice, cudaDeviceSynchronize, cudaStreamSynchronize, cudaStreamCreate, cudaGetDeviceProperties cudaStreamDestroy, cudaFree, cudaMallocManaged, cudaStreamAttachMemAsync, cudaSetDevice, cudaDeviceSynchronize, cudaStreamSynchronize, cudaStreamCreate, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[OpenMP](../../../README.md#openmp), [UVM](../../../README.md#uvm), [CUBLAS](../../../README.md#cublas) [OpenMP](../../../README.md#openmp), [UVM](../../../README.md#unified-virtual-memory), [CUBLAS](../../../README.md#cublas)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaMemcpy, cudaFree, cudaMallocHost, cudaFreeHost, cudaMalloc, cudaGetDeviceProperties cudaMemcpy, cudaFree, cudaMallocHost, cudaFreeHost, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[FP16](../../../README.md#fp16) [FP16](../../../README.md#16-bit-floating-point)
## Prerequisites ## Prerequisites

View File

@@ -94,7 +94,7 @@ scalarProductKernel_intrinsics(half2 const *const a, half2 const *const b, float
shArray[threadIdx.x] = __float2half2_rn(0.f); shArray[threadIdx.x] = __float2half2_rn(0.f);
half2 value = __float2half2_rn(0.f); half2 value = __float2half2_rn(0.f);
for (int i = threadIdx.x + blockDim.x + blockIdx.x; i < size; i += stride) { for (int i = threadIdx.x + blockDim.x * blockIdx.x; i < size; i += stride) {
value = __hfma2(a[i], b[i], value); value = __hfma2(a[i], b[i], value);
} }
@@ -118,7 +118,7 @@ scalarProductKernel_native(half2 const *const a, half2 const *const b, float *co
half2 value(0.f, 0.f); half2 value(0.f, 0.f);
shArray[threadIdx.x] = value; shArray[threadIdx.x] = value;
for (int i = threadIdx.x + blockDim.x + blockIdx.x; i < size; i += stride) { for (int i = threadIdx.x + blockDim.x * blockIdx.x; i < size; i += stride) {
value = a[i] * b[i] + value; value = a[i] * b[i] + value;
} }

View File

@@ -255,6 +255,7 @@ void runTest(int argc, char **argv)
checkCudaErrors(cuMemFree(d_A)); checkCudaErrors(cuMemFree(d_A));
checkCudaErrors(cuMemFree(d_B)); checkCudaErrors(cuMemFree(d_B));
checkCudaErrors(cuMemFree(d_C)); checkCudaErrors(cuMemFree(d_C));
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(cuContext)); checkCudaErrors(cuCtxDestroy(cuContext));
} }

View File

@@ -68,6 +68,7 @@ extern "C" void computeGold(float *, const float *, const float *, unsigned int,
// Globals // Globals
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
CUcontext g_cuContext; CUcontext g_cuContext;
CUmodule g_cuModule;
bool noprompt = false; bool noprompt = false;
static const char *sSDKsample = "matrixMulDynlinkJIT (CUDA dynamic linking)"; static const char *sSDKsample = "matrixMulDynlinkJIT (CUDA dynamic linking)";
@@ -90,7 +91,6 @@ CUresult initCUDA(int argc, char **argv, CUfunction *pMatrixMul, int *block_size
{ {
CUresult status; CUresult status;
CUdevice cuDevice; CUdevice cuDevice;
CUmodule cuModule;
CUfunction cuFunction; CUfunction cuFunction;
int major, minor, block_size, devID = 0; int major, minor, block_size, devID = 0;
char deviceName[256]; char deviceName[256];
@@ -184,11 +184,11 @@ CUresult initCUDA(int argc, char **argv, CUfunction *pMatrixMul, int *block_size
printf("> Compiling CUDA module\n"); printf("> Compiling CUDA module\n");
#if defined(_WIN64) || defined(__LP64__) #if defined(_WIN64) || defined(__LP64__)
status = status = cuModuleLoadDataEx(
cuModuleLoadDataEx(&cuModule, matrixMul_kernel_64_ptxdump, jitNumOptions, jitOptions, (void **)jitOptVals); &g_cuModule, matrixMul_kernel_64_ptxdump, jitNumOptions, jitOptions, (void **)jitOptVals);
#else #else
status = status = cuModuleLoadDataEx(
cuModuleLoadDataEx(&cuModule, matrixMul_kernel_32_ptxdump, jitNumOptions, jitOptions, (void **)jitOptVals); &g_cuModule, matrixMul_kernel_32_ptxdump, jitNumOptions, jitOptions, (void **)jitOptVals);
#endif #endif
printf("> PTX JIT log:\n%s\n", jitLogBuffer); printf("> PTX JIT log:\n%s\n", jitLogBuffer);
@@ -206,9 +206,10 @@ CUresult initCUDA(int argc, char **argv, CUfunction *pMatrixMul, int *block_size
// retrieve CUDA function from the compiled module // retrieve CUDA function from the compiled module
status = cuModuleGetFunction( status = cuModuleGetFunction(
&cuFunction, cuModule, (block_size == 16) ? "matrixMul_bs16_32bit" : "matrixMul_bs32_32bit"); &cuFunction, g_cuModule, (block_size == 16) ? "matrixMul_bs16_32bit" : "matrixMul_bs32_32bit");
if (CUDA_SUCCESS != status) { if (CUDA_SUCCESS != status) {
cuModuleUnload(g_cuModule);
cuCtxDestroy(g_cuContext); cuCtxDestroy(g_cuContext);
exit(EXIT_FAILURE); exit(EXIT_FAILURE);
} }
@@ -337,6 +338,7 @@ int main(int argc, char **argv)
checkCudaErrors(cuMemFree(d_A)); checkCudaErrors(cuMemFree(d_A));
checkCudaErrors(cuMemFree(d_B)); checkCudaErrors(cuMemFree(d_B));
checkCudaErrors(cuMemFree(d_C)); checkCudaErrors(cuMemFree(d_C));
checkCudaErrors(cuModuleUnload(g_cuModule));
checkCudaErrors(cuCtxDestroy(g_cuContext)); checkCudaErrors(cuCtxDestroy(g_cuContext));
printf("Test run %s\n", (1 == res) ? "success!" : "failed!"); printf("Test run %s\n", (1 == res) ? "success!" : "failed!");

View File

@@ -26,7 +26,7 @@ x86_64, armv7l, aarch64
cudaStreamCreateWithFlags, cudaFree, cudaDeviceGetAttribute, cudaMallocHost, cudaFreeHost, cudaStreamSynchronize, cudaLaunchCooperativeKernel, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaMemcpyAsync, cudaOccupancyMaxPotentialBlockSize cudaStreamCreateWithFlags, cudaFree, cudaDeviceGetAttribute, cudaMallocHost, cudaFreeHost, cudaStreamSynchronize, cudaLaunchCooperativeKernel, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaMemcpyAsync, cudaOccupancyMaxPotentialBlockSize
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11), [MBCG](../../../README.md#mbcg) [C++11 CUDA](../../../README.md#c11-cuda), [MBCG](../../../README.md#multi-block-cooperative-groups)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64
cudaHostAlloc, cudaGraphicsUnmapResources, cudaMalloc, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaMemcpyToArray, cudaDeviceSynchronize, cudaProcess, cudaGraphicsUnregisterResource, cudaGraphicsSubResourceGetMappedArray, cudaGraphicsGLRegisterBuffer, cudaGraphicsGLRegisterImage cudaHostAlloc, cudaGraphicsUnmapResources, cudaMalloc, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaMemcpyToArray, cudaDeviceSynchronize, cudaProcess, cudaGraphicsUnregisterResource, cudaGraphicsSubResourceGetMappedArray, cudaGraphicsGLRegisterBuffer, cudaGraphicsGLRegisterImage
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, ppc64le
cudaSetDevice, cudaIpcCloseMemHandle, cudaEventDestroy, cudaGetDeviceCount, cudaMemcpyAsync, cudaDeviceCanAccessPeer, cudaStreamCreateWithFlags, cudaStreamDestroy, cudaGetLastError, cudaIpcOpenEventHandle, cudaIpcOpenMemHandle, cudaIpcGetEventHandle, cudaStreamWaitEvent, cudaEventCreate, cudaFree, cudaEventSynchronize, cudaEventRecord, cudaIpcGetMemHandle, cudaStreamSynchronize, cudaDeviceEnablePeerAccess, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties cudaSetDevice, cudaIpcCloseMemHandle, cudaEventDestroy, cudaGetDeviceCount, cudaMemcpyAsync, cudaDeviceCanAccessPeer, cudaStreamCreateWithFlags, cudaStreamDestroy, cudaGetLastError, cudaIpcOpenEventHandle, cudaIpcOpenMemHandle, cudaIpcGetEventHandle, cudaStreamWaitEvent, cudaEventCreate, cudaFree, cudaEventSynchronize, cudaEventRecord, cudaIpcGetMemHandle, cudaStreamSynchronize, cudaDeviceEnablePeerAccess, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[IPC](../../../README.md#ipc) [IPC](../../../README.md#cuda-interprocess-communication)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMalloc, cudaGetLastError, cudaMemcpy, cudaFree cudaMalloc, cudaGetLastError, cudaMemcpy, cudaFree
## Dependencies needed to build/run ## Dependencies needed to build/run
[MPI](../../../README.md#mpi) [MPI](../../../README.md#message-passing-interface)
## Prerequisites ## Prerequisites

View File

@@ -25,6 +25,7 @@
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/ */
#include <cstdint>
#include <helper_cuda.h> // helper functions for CUDA error check #include <helper_cuda.h> // helper functions for CUDA error check
#include <iostream> #include <iostream>
@@ -38,7 +39,7 @@ const int manualBlockSize = 32;
// execution configuration, including anything the launch configurator // execution configuration, including anything the launch configurator
// API suggests. // API suggests.
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
__global__ void square(int *array, int arrayCount) __global__ void square(uint32_t *array, int arrayCount)
{ {
extern __shared__ int dynamicSmem[]; extern __shared__ int dynamicSmem[];
int idx = threadIdx.x + blockIdx.x * blockDim.x; int idx = threadIdx.x + blockIdx.x * blockDim.x;
@@ -99,7 +100,7 @@ static double reportPotentialOccupancy(void *kernel, int blockSize, size_t dynam
// This function configures the launch based on the "automatic" // This function configures the launch based on the "automatic"
// argument, records the runtime, and reports occupancy and runtime. // argument, records the runtime, and reports occupancy and runtime.
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
static int launchConfig(int *array, int arrayCount, bool automatic) static int launchConfig(uint32_t *array, int arrayCount, bool automatic)
{ {
int blockSize; int blockSize;
int minGridSize; int minGridSize;
@@ -166,20 +167,20 @@ static int launchConfig(int *array, int arrayCount, bool automatic)
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
static int test(bool automaticLaunchConfig, const int count = 1000000) static int test(bool automaticLaunchConfig, const int count = 1000000)
{ {
int *array; uint32_t *array;
int *dArray; uint32_t *dArray;
int size = count * sizeof(int); int size = count * sizeof(uint32_t);
array = new int[count]; array = new uint32_t[count];
for (int i = 0; i < count; i += 1) { for (uint32_t i = 0; i < count; i += 1) {
array[i] = i; array[i] = i;
} }
checkCudaErrors(cudaMalloc(&dArray, size)); checkCudaErrors(cudaMalloc(&dArray, size));
checkCudaErrors(cudaMemcpy(dArray, array, size, cudaMemcpyHostToDevice)); checkCudaErrors(cudaMemcpy(dArray, array, size, cudaMemcpyHostToDevice));
for (int i = 0; i < count; i += 1) { for (uint32_t i = 0; i < count; i += 1) {
array[i] = 0; array[i] = 0;
} }
@@ -189,8 +190,9 @@ static int test(bool automaticLaunchConfig, const int count = 1000000)
checkCudaErrors(cudaFree(dArray)); checkCudaErrors(cudaFree(dArray));
// Verify the return data // Verify the return data
// Both GPU and CPU use uint32_t * uint32_t, which has well-defined overflow behavior (modulo 2^32)
// //
for (int i = 0; i < count; i += 1) { for (uint32_t i = 0; i < count; i += 1) {
if (array[i] != i * i) { if (array[i] != i * i) {
std::cout << "element " << i << " expected " << i * i << " actual " << array[i] << std::endl; std::cout << "element " << i << " expected " << i * i << " actual " << array[i] << std::endl;
return 1; return 1;

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFreeArray, cudaFree, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaExtent, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaFreeArray, cudaFree, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaExtent, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -297,6 +297,7 @@ void runTest(int argc, char **argv)
free(image_path); free(image_path);
free(ref_path); free(ref_path);
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(cuContext)); checkCudaErrors(cuCtxDestroy(cuContext));
exit(bTestResults ? EXIT_SUCCESS : EXIT_FAILURE); exit(bTestResults ? EXIT_SUCCESS : EXIT_FAILURE);

View File

@@ -26,7 +26,7 @@ x86_64, ppc64le
cudaDeviceSynchronize, cudaMallocManaged, cudaGetDeviceProperties, cudaFree cudaDeviceSynchronize, cudaMallocManaged, cudaGetDeviceProperties, cudaFree
## Dependencies needed to build/run ## Dependencies needed to build/run
[UVM](../../../README.md#uvm) [UVM](../../../README.md#unified-virtual-memory)
## Prerequisites ## Prerequisites

View File

@@ -211,6 +211,7 @@ int CleanupNoFailure()
free(h_C); free(h_C);
} }
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(cuContext)); checkCudaErrors(cuCtxDestroy(cuContext));
return EXIT_SUCCESS; return EXIT_SUCCESS;

View File

@@ -247,6 +247,7 @@ int CleanupNoFailure()
free(h_C); free(h_C);
} }
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(cuContext)); checkCudaErrors(cuCtxDestroy(cuContext));
return EXIT_SUCCESS; return EXIT_SUCCESS;

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMemcpyFromSymbol, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMemcpyFromSymbol, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMallocArray, cudaFreeArray, cudaFree, cudaGetErrorString, cudaMemcpy, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMallocArray, cudaFreeArray, cudaFree, cudaGetErrorString, cudaMemcpy, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,8 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites
Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform. Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.

View File

@@ -39,10 +39,7 @@ if(${OpenGL_FOUND})
# Add target for particles # Add target for particles
add_executable(particles particleSystem.cpp particleSystem_cuda.cu particles.cpp render_particles.cpp shaders.cpp) add_executable(particles particleSystem.cpp particleSystem_cuda.cu particles.cpp render_particles.cpp shaders.cpp)
target_compile_options(particles PRIVATE target_compile_options(particles PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>
$<$<AND:$<COMPILE_LANGUAGE:CUDA>,$<BOOL:${MSVC}>>:-Xcompiler="/Zc:preprocessor">
)
target_compile_features(particles PRIVATE cxx_std_17 cuda_std_17) target_compile_features(particles PRIVATE cxx_std_17 cuda_std_17)

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaMemset, cudaMemcpyToSymbol, cudaGraphicsGLRegisterBuffer, cudaGraphicsUnregisterResource, cudaMalloc, cudaInit, cudaGLInit cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaMemset, cudaMemcpyToSymbol, cudaGraphicsGLRegisterBuffer, cudaGraphicsUnregisterResource, cudaMalloc, cudaInit, cudaGLInit
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -135,8 +135,8 @@ extern "C"
thrust::device_ptr<float4> d_pos4((float4 *)pos); thrust::device_ptr<float4> d_pos4((float4 *)pos);
thrust::device_ptr<float4> d_vel4((float4 *)vel); thrust::device_ptr<float4> d_vel4((float4 *)vel);
thrust::for_each(thrust::make_zip_iterator(thrust::make_tuple(d_pos4, d_vel4)), thrust::for_each(thrust::make_zip_iterator(d_pos4, d_vel4),
thrust::make_zip_iterator(thrust::make_tuple(d_pos4 + numParticles, d_vel4 + numParticles)), thrust::make_zip_iterator(d_pos4 + numParticles, d_vel4 + numParticles),
integrate_functor(deltaTime)); integrate_functor(deltaTime));
} }

View File

@@ -41,6 +41,9 @@
#include "thrust/iterator/zip_iterator.h" #include "thrust/iterator/zip_iterator.h"
#include "thrust/sort.h" #include "thrust/sort.h"
// for cuda::std::get
#include <cuda/std/utility>
namespace cg = cooperative_groups; namespace cg = cooperative_groups;
#include "helper_math.h" #include "helper_math.h"
#include "math_constants.h" #include "math_constants.h"
@@ -60,8 +63,8 @@ struct integrate_functor
template <typename Tuple> __device__ void operator()(Tuple t) template <typename Tuple> __device__ void operator()(Tuple t)
{ {
volatile float4 posData = thrust::get<0>(t); volatile float4 posData = cuda::std::get<0>(t);
volatile float4 velData = thrust::get<1>(t); volatile float4 velData = cuda::std::get<1>(t);
float3 pos = make_float3(posData.x, posData.y, posData.z); float3 pos = make_float3(posData.x, posData.y, posData.z);
float3 vel = make_float3(velData.x, velData.y, velData.z); float3 vel = make_float3(velData.x, velData.y, velData.z);
@@ -107,8 +110,8 @@ struct integrate_functor
} }
// store new position and velocity // store new position and velocity
thrust::get<0>(t) = make_float4(pos, posData.w); cuda::std::get<0>(t) = make_float4(pos, posData.w);
thrust::get<1>(t) = make_float4(vel, velData.w); cuda::std::get<1>(t) = make_float4(vel, velData.w);
} }
}; };

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaMemcpy, cudaFree, cudaSetDevice, cudaDeviceSynchronize, cudaGetDevice, cudaMalloc, cudaGetDeviceProperties cudaMemcpy, cudaFree, cudaSetDevice, cudaDeviceSynchronize, cudaGetDevice, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMemcpy, cudaFree, cudaSetDevice, cudaDeviceSynchronize, cudaLaunchCooperativeKernel, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties, cudaOccupancyMaxPotentialBlockSize cudaMemcpy, cudaFree, cudaSetDevice, cudaDeviceSynchronize, cudaLaunchCooperativeKernel, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties, cudaOccupancyMaxPotentialBlockSize
## Dependencies needed to build/run ## Dependencies needed to build/run
[MBCG](../../../README.md#mbcg), [CPP11](../../../README.md#cpp11) [MBCG](../../../README.md#multi-block-cooperative-groups), [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -505,13 +505,12 @@ private:
thrust::device_ptr<uint> dMinScannedEdges = pools.uintEdges.get(); thrust::device_ptr<uint> dMinScannedEdges = pools.uintEdges.get();
thrust::device_ptr<float> dMinScannedWeights = pools.floatEdges.get(); thrust::device_ptr<float> dMinScannedWeights = pools.floatEdges.get();
thrust::inclusive_scan_by_key( thrust::inclusive_scan_by_key(dEdgesFlags,
dEdgesFlags,
dEdgesFlags + edgesCount_, dEdgesFlags + edgesCount_,
thrust::make_zip_iterator(thrust::make_tuple(dWeights_, dEdges_)), thrust::make_zip_iterator(dWeights_, dEdges_),
thrust::make_zip_iterator(thrust::make_tuple(dMinScannedWeights, dMinScannedEdges)), thrust::make_zip_iterator(dMinScannedWeights, dMinScannedEdges),
thrust::greater_equal<uint>(), thrust::greater_equal<uint>(),
thrust::minimum<thrust::tuple<float, uint>>()); thrust::minimum<cuda::std::tuple<float, uint>>());
// To make things clear. // To make things clear.
// Let "dEdgesFlags" denote groups of edges that // Let "dEdgesFlags" denote groups of edges that
@@ -562,11 +561,10 @@ private:
thrust::sequence(dClusteredVerticesIDs, dClusteredVerticesIDs + verticesCount_); thrust::sequence(dClusteredVerticesIDs, dClusteredVerticesIDs + verticesCount_);
thrust::sort(thrust::make_zip_iterator(thrust::make_tuple(thrust::device_ptr<uint>(dSuccessors), thrust::sort(thrust::make_zip_iterator(thrust::device_ptr<uint>(dSuccessors),
thrust::device_ptr<uint>(dClusteredVerticesIDs))), thrust::device_ptr<uint>(dClusteredVerticesIDs)),
thrust::make_zip_iterator( thrust::make_zip_iterator(thrust::device_ptr<uint>(dSuccessors + verticesCount_),
thrust::make_tuple(thrust::device_ptr<uint>(dSuccessors + verticesCount_), thrust::device_ptr<uint>(dClusteredVerticesIDs + verticesCount_)));
thrust::device_ptr<uint>(dClusteredVerticesIDs + verticesCount_))));
// Mark those groups. // Mark those groups.
thrust::device_ptr<uint> dVerticesFlags_ = pools.uintVertices.get(); thrust::device_ptr<uint> dVerticesFlags_ = pools.uintVertices.get();
@@ -653,9 +651,8 @@ private:
// the step of the algorithm to that one. Hence we need to // the step of the algorithm to that one. Hence we need to
// preserve the structure of the original graph: neighbours and // preserve the structure of the original graph: neighbours and
// weights should be grouped by vertex. // weights should be grouped by vertex.
thrust::sort(thrust::make_zip_iterator(thrust::make_tuple(dNewStartpoints, dSurvivedEdgesIDs)), thrust::sort(thrust::make_zip_iterator(dNewStartpoints, dSurvivedEdgesIDs),
thrust::make_zip_iterator( thrust::make_zip_iterator(dNewStartpoints + edgesCount_, dSurvivedEdgesIDs + edgesCount_));
thrust::make_tuple(dNewStartpoints + edgesCount_, dSurvivedEdgesIDs + edgesCount_)));
// Find the group of contracted edges. // Find the group of contracted edges.
uint *invalidEdgesPtr = uint *invalidEdgesPtr =

View File

@@ -24,7 +24,7 @@ x86_64, armv7l, aarch64
cudaMemcpy, cudaFree, cudaMallocHost, cudaEventSynchronize, cudaEventRecord, cudaFreeHost, cudaGetDevice, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate cudaMemcpy, cudaFree, cudaMallocHost, cudaEventSynchronize, cudaEventRecord, cudaFreeHost, cudaGetDevice, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaMemcpy, cudaGetMipmappedArrayLevel, cudaGraphicsMapResources, cudaDestroySurfaceObject, cudaExtent, cudaDeviceSynchronize, cudaCreateSurfaceObject, cudaMallocMipmappedArray, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaCreateTextureObject, cudaGraphicsUnmapResources, cudaMallocArray, cudaFreeArray, cudaArrayGetInfo, cudaGetLastError, cudaDestroyTextureObject, cudaGraphicsGLRegisterBuffer, cudaFreeMipmappedArray, cudaFree, cudaGraphicsUnregisterResource, cudaMalloc cudaMemcpy, cudaGetMipmappedArrayLevel, cudaGraphicsMapResources, cudaDestroySurfaceObject, cudaExtent, cudaDeviceSynchronize, cudaCreateSurfaceObject, cudaMallocMipmappedArray, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaCreateTextureObject, cudaGraphicsUnmapResources, cudaMallocArray, cudaFreeArray, cudaArrayGetInfo, cudaGetLastError, cudaDestroyTextureObject, cudaGraphicsGLRegisterBuffer, cudaFreeMipmappedArray, cudaFree, cudaGraphicsUnregisterResource, cudaMalloc
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaStreamCreateWithFlags, cudaMemcpy, cudaMemcpyAsync, cudaFree, cudaGetErrorString, cudaGetLastError, cudaPeekAtLastError, cudaDeviceSynchronize, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate cudaStreamCreateWithFlags, cudaMemcpy, cudaMemcpyAsync, cudaFree, cudaGetErrorString, cudaGetLastError, cudaPeekAtLastError, cudaDeviceSynchronize, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CDP](../../../README.md#cdp) [CDP](../../../README.md#cuda-dynamic-parallellism)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaMemcpy, cudaFree, cudaGetDeviceCount, cudaMalloc, cudaGetDeviceProperties cudaMemcpy, cudaFree, cudaGetDeviceCount, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[CDP](../../../README.md#cdp) [CDP](../../../README.md#cuda-dynamic-parallellism)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaMemcpy, cudaFree, cudaGetLastError, cudaDeviceSetLimit, cudaMalloc, cudaGetDeviceProperties cudaMemcpy, cudaFree, cudaGetLastError, cudaDeviceSetLimit, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[CDP](../../../README.md#cdp) [CDP](../../../README.md#cuda-dynamic-parallellism)
## Prerequisites ## Prerequisites

View File

@@ -30,6 +30,9 @@
#include <thrust/host_vector.h> #include <thrust/host_vector.h>
#include <thrust/random.h> #include <thrust/random.h>
// for cuda::std::tuple
#include <cuda/std/tuple>
namespace cg = cooperative_groups; namespace cg = cooperative_groups;
#include <helper_cuda.h> #include <helper_cuda.h>
@@ -610,7 +613,7 @@ struct Random_generator
return a; return a;
} }
__host__ __device__ __forceinline__ thrust::tuple<float, float> operator()() __host__ __device__ __forceinline__ cuda::std::tuple<float, float> operator()()
{ {
#ifdef __CUDA_ARCH__ #ifdef __CUDA_ARCH__
unsigned seed = hash(blockIdx.x * blockDim.x + threadIdx.x + count); unsigned seed = hash(blockIdx.x * blockDim.x + threadIdx.x + count);
@@ -622,7 +625,7 @@ struct Random_generator
#endif #endif
thrust::default_random_engine rng(seed); thrust::default_random_engine rng(seed);
thrust::random::uniform_real_distribution<float> distrib; thrust::random::uniform_real_distribution<float> distrib;
return thrust::make_tuple(distrib(rng), distrib(rng)); return cuda::std::make_tuple(distrib(rng), distrib(rng));
} }
}; };
@@ -644,9 +647,8 @@ bool cdpQuadtree(int warp_size)
// Generate random points. // Generate random points.
Random_generator rnd; Random_generator rnd;
thrust::generate(thrust::make_zip_iterator(thrust::make_tuple(x_d0.begin(), y_d0.begin())), thrust::generate(
thrust::make_zip_iterator(thrust::make_tuple(x_d0.end(), y_d0.end())), thrust::make_zip_iterator(x_d0.begin(), y_d0.begin()), thrust::make_zip_iterator(x_d0.end(), y_d0.end()), rnd);
rnd);
// Host structures to analyze the device ones. // Host structures to analyze the device ones.
Points points_init[2]; Points points_init[2];

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaDeviceSynchronize, cudaGetLastError, cudaGetDeviceProperties, cudaDeviceSetLimit cudaDeviceSynchronize, cudaGetLastError, cudaGetDeviceProperties, cudaDeviceSetLimit
## Dependencies needed to build/run ## Dependencies needed to build/run
[CDP](../../../README.md#cdp) [CDP](../../../README.md#cuda-dynamic-parallellism)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaStreamCreateWithFlags, cudaMemcpy, cudaStreamDestroy, cudaFree, cudaDeviceSynchronize, cudaDeviceSetLimit, cudaMalloc, cudaGetDeviceProperties cudaStreamCreateWithFlags, cudaMemcpy, cudaStreamDestroy, cudaFree, cudaDeviceSynchronize, cudaDeviceSetLimit, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[CDP](../../../README.md#cdp) [CDP](../../../README.md#cuda-dynamic-parallellism)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l, aarch64
cudaStreamCreateWithFlags, cudaMalloc, cudaDeviceGetAttribute, cudaFree, cudaMallocHost, cudaEventSynchronize, cudaEventRecord, cudaFreeHost, cudaStreamSynchronize, cudaEventDestroy, cudaEventElapsedTime, cudaMemsetAsync, cudaMemcpyAsync, cudaEventCreate cudaStreamCreateWithFlags, cudaMalloc, cudaDeviceGetAttribute, cudaFree, cudaMallocHost, cudaEventSynchronize, cudaEventRecord, cudaFreeHost, cudaStreamSynchronize, cudaEventDestroy, cudaEventElapsedTime, cudaMemsetAsync, cudaMemcpyAsync, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l, aarch64
cuDeviceCanAccessPeer, cuMemImportFromShareableHandle, cuModuleLoadDataEx, cuModuleGetFunction, cuMemSetAccess, cuModuleLoad, cuStreamCreate, cuMemRelease, cuInit, cuLaunchKernel, cuMemcpyDtoHAsync, cuMemCreate, cuDeviceGet, cuCtxDestroy, cuDeviceGetCount, cuMemMap, cuMemExportToShareableHandle, cuStreamSynchronize, cuCtxEnablePeerAccess, cuDeviceGetAttribute, cuOccupancyMaxActiveBlocksPerMultiprocessor, cuCtxSetCurrent, cuMemGetAllocationGranularity, cuMemAddressFree, cuMemUnmap, cuCtxCreate, cuStreamDestroy, cuMemAddressReserve cuDeviceCanAccessPeer, cuMemImportFromShareableHandle, cuModuleLoadDataEx, cuModuleGetFunction, cuMemSetAccess, cuModuleLoad, cuStreamCreate, cuMemRelease, cuInit, cuLaunchKernel, cuMemcpyDtoHAsync, cuMemCreate, cuDeviceGet, cuCtxDestroy, cuDeviceGetCount, cuMemMap, cuMemExportToShareableHandle, cuStreamSynchronize, cuCtxEnablePeerAccess, cuDeviceGetAttribute, cuOccupancyMaxActiveBlocksPerMultiprocessor, cuCtxSetCurrent, cuMemGetAllocationGranularity, cuMemAddressFree, cuMemUnmap, cuCtxCreate, cuStreamDestroy, cuMemAddressReserve
## Dependencies needed to build/run ## Dependencies needed to build/run
[IPC](../../../README.md#ipc) [IPC](../../../README.md#cuda-interprocess-communication)
## Prerequisites ## Prerequisites

View File

@@ -419,6 +419,7 @@ static void childProcess(int devId, int id, char **argv)
// Clean up! // Clean up!
checkCudaErrors(cuStreamDestroy(stream)); checkCudaErrors(cuStreamDestroy(stream));
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(ctx)); checkCudaErrors(cuCtxDestroy(ctx));
// Unmap the allocations from our address space. Unmapping will also free the // Unmap the allocations from our address space. Unmapping will also free the

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate cudaMemcpy, cudaFree, cudaGetErrorString, cudaGetLastError, cudaEventSynchronize, cudaFuncSetAttribute, cudaEventRecord, cudaMemset, cudaMalloc, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[CPP11](../../../README.md#cpp11) [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaFree, cudaMallocManaged, cudaDeviceSynchronize, cudaEventRecord, cudaLaunchCooperativeKernel, cudaEventDestroy, cudaEventElapsedTime, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties, cudaEventCreate cudaFree, cudaMallocManaged, cudaDeviceSynchronize, cudaEventRecord, cudaLaunchCooperativeKernel, cudaEventDestroy, cudaEventElapsedTime, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[UVM](../../../README.md#uvm), [MBCG](../../../README.md#mbcg) [UVM](../../../README.md#unified-virtual-memory), [MBCG](../../../README.md#multi-block-cooperative-groups)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaHostAlloc, cudaMemPrefetchAsync, cudaFree, cudaLaunchCooperativeKernel, cudaMallocManaged, cudaSetDevice, cudaGetDeviceCount, cudaGetDeviceProperties, cudaFreeHost, cudaMemset, cudaStreamCreate, cudaStreamSynchronize, cudaDeviceEnablePeerAccess, cudaMemAdvise, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaDeviceCanAccessPeer cudaHostAlloc, cudaMemPrefetchAsync, cudaFree, cudaLaunchCooperativeKernel, cudaMallocManaged, cudaSetDevice, cudaGetDeviceCount, cudaGetDeviceProperties, cudaFreeHost, cudaMemset, cudaStreamCreate, cudaStreamSynchronize, cudaDeviceEnablePeerAccess, cudaMemAdvise, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaDeviceCanAccessPeer
## Dependencies needed to build/run ## Dependencies needed to build/run
[UVM](../../../README.md#uvm), [MDCG](../../../README.md#mdcg), [CPP11](../../../README.md#cpp11) [UVM](../../../README.md#unified-virtual-memory), [MDCG](../../../README.md#multi-device-cooperative-groups), [C++11 CUDA](../../../README.md#c11-cuda)
## Prerequisites ## Prerequisites

View File

@@ -24,7 +24,7 @@ x86_64, armv7l
cudaFree, cudaMallocManaged, cudaDeviceSynchronize, cudaMalloc, cudaGetDeviceProperties cudaFree, cudaMallocManaged, cudaDeviceSynchronize, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[UVM](../../../README.md#uvm), [CUBLAS](../../../README.md#cublas), [CUSPARSE](../../../README.md#cusparse) [UVM](../../../README.md#unified-virtual-memory), [CUBLAS](../../../README.md#cublas), [CUSPARSE](../../../README.md#cusparse)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMalloc, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaCalculateSlopeKernel, cudaGraphicsMapResources, cudaUpdateHeightmapKernel, cudaGraphicsUnregisterResource, cudaGenerateSpectrumKernel, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGraphicsUnmapResources, cudaMemcpy, cudaMalloc, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaCalculateSlopeKernel, cudaGraphicsMapResources, cudaUpdateHeightmapKernel, cudaGraphicsUnregisterResource, cudaGenerateSpectrumKernel, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl), [CUFFT](../../../README.md#cufft) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew), [CUFFT](../../../README.md#cufft)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaMalloc, cudaGetErrorString, cudaMemcpy, cudaFree cudaMalloc, cudaGetErrorString, cudaMemcpy, cudaFree
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl), [CURAND](../../../README.md#curand) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew), [CURAND](../../../README.md#curand)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaMemcpy, cudaFree, cudaMemcpyFromSymbol, cudaGetDevice, cudaMalloc, cudaGetDeviceProperties cudaMemcpy, cudaFree, cudaMemcpyFromSymbol, cudaGetDevice, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[callback](../../../README.md#callback), [CUFFT](../../../README.md#cufft) [CUFFT Callback Routines](../../../README.md#cufft-callback-routines), [CUFFT](../../../README.md#cufft)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGLUnmapBufferObject, cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGLMapBufferObject, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaGLUnregisterBufferObject, cudaDeviceSynchronize, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGLUnmapBufferObject, cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGLMapBufferObject, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaGLUnregisterBufferObject, cudaDeviceSynchronize, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaGetErrorString, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaGetErrorString, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMallocArray, cudaFreeArray, cudaFree, cudaMemcpy, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMallocArray, cudaFreeArray, cudaFree, cudaMemcpy, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaRuntimeGetVersion, cudaGraphicsUnmapResources, cudaMallocPitch, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaRuntimeGetVersion, cudaGraphicsUnmapResources, cudaMallocPitch, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMallocPitch, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMallocPitch, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl), [CUFFT](../../../README.md#cufft) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew), [CUFFT](../../../README.md#cufft)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGLUnmapBufferObject, cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMemcpy, cudaFree, cudaGLMapBufferObject, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaGLUnregisterBufferObject, cudaCreateTextureObject, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGLUnmapBufferObject, cudaGraphicsUnmapResources, cudaCreateChannelDesc, cudaMemcpy, cudaFree, cudaGLMapBufferObject, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaGLUnregisterBufferObject, cudaCreateTextureObject, cudaGLRegisterBufferObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceCount, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaStreamQuery, cudaEventDestroy, cudaEventElapsedTime, cudaDeviceCanAccessPeer, cudaEventCreate cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceCount, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaStreamQuery, cudaEventDestroy, cudaEventElapsedTime, cudaDeviceCanAccessPeer, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -101,6 +101,64 @@ void printHelp(void)
printf("--numElems=<NUM_OF_INT_ELEMS> Number of integer elements to be used in p2p copy.\n"); printf("--numElems=<NUM_OF_INT_ELEMS> Number of integer elements to be used in p2p copy.\n");
} }
///////////////////////////////////////////////////////////////////////////
// Detect if cudaMemcpyPeerAsync will automatically fall back to
// host-staged copies when P2P is disabled.
//
// We probe a single representative pair (device 0 -> device 1).
// On a given system, confidential-computing (CC) and security
// policies are uniform across GPUs, so if this pair is blocked
// with cudaErrorNotSupported in P2P-off mode, it is reasonable
// to assume all cross-GPU pairs behave the same.
//
// For a production application that must handle heterogeneous
// environments, users may want to probe all device pairs.
///////////////////////////////////////////////////////////////////////////
bool detectFallback(int numGPUs)
{
if (numGPUs <= 1)
return false;
cudaSetDevice(0);
int *tmp0 = nullptr, *tmp1 = nullptr;
cudaStream_t s;
cudaStreamCreateWithFlags(&s, cudaStreamNonBlocking);
size_t testElems = 1;
cudaMalloc(&tmp0, testElems * sizeof(int));
cudaSetDevice(1);
cudaMalloc(&tmp1, testElems * sizeof(int));
cudaCheckError();
// Explicitly ensure P2P is disabled for this test
// (Clear any pre-existing P2P access if it happens to be enabled)
cudaSetDevice(0);
cudaDeviceDisablePeerAccess(1);
cudaGetLastError(); // Clear error if peer access was not enabled
cudaSetDevice(1);
cudaDeviceDisablePeerAccess(0);
cudaGetLastError(); // Clear error if peer access was not enabled
bool needsFallback = false;
cudaError_t testErr = cudaMemcpyPeerAsync(tmp1, 1, tmp0, 0, testElems * sizeof(int), s);
if (testErr == cudaErrorPeerAccessNotEnabled || testErr == cudaErrorNotSupported) {
needsFallback = true;
printf("Note: cudaMemcpyPeerAsync reported '%s' - will use host-mediated copy when P2P is disabled\n",
cudaGetErrorString(testErr));
cudaGetLastError();
}
cudaStreamSynchronize(s);
cudaFree(tmp0);
cudaFree(tmp1);
cudaStreamDestroy(s);
cudaCheckError();
return needsFallback;
}
void checkP2Paccess(int numGPUs) void checkP2Paccess(int numGPUs)
{ {
for (int i = 0; i < numGPUs; i++) { for (int i = 0; i < numGPUs; i++) {
@@ -128,7 +186,9 @@ void performP2PCopy(int *dest,
int num_elems, int num_elems,
int repeat, int repeat,
bool p2paccess, bool p2paccess,
cudaStream_t streamToRun) cudaStream_t streamToRun,
bool useFallback,
int *hostBuffer)
{ {
int blockSize = 0; int blockSize = 0;
int numBlocks = 0; int numBlocks = 0;
@@ -140,15 +200,33 @@ void performP2PCopy(int *dest,
for (int r = 0; r < repeat; r++) { for (int r = 0; r < repeat; r++) {
copyp2p<<<numBlocks, blockSize, 0, streamToRun>>>((int4 *)dest, (int4 *)src, num_elems / 4); copyp2p<<<numBlocks, blockSize, 0, streamToRun>>>((int4 *)dest, (int4 *)src, num_elems / 4);
} }
cudaCheckError();
}
else if (useFallback && srcDevice != destDevice) {
// Use host-mediated copy for cross-GPU transfers when cudaMemcpyPeerAsync is not supported
for (int r = 0; r < repeat; r++) {
cudaMemcpyAsync(hostBuffer, src, sizeof(int) * num_elems, cudaMemcpyDeviceToHost, streamToRun);
cudaMemcpyAsync(dest, hostBuffer, sizeof(int) * num_elems, cudaMemcpyHostToDevice, streamToRun);
}
cudaCheckError();
}
else if (useFallback && srcDevice == destDevice) {
// Same device copy
for (int r = 0; r < repeat; r++) {
cudaMemcpyAsync(dest, src, sizeof(int) * num_elems, cudaMemcpyDeviceToDevice, streamToRun);
}
cudaCheckError();
} }
else { else {
// Use cudaMemcpyPeerAsync
for (int r = 0; r < repeat; r++) { for (int r = 0; r < repeat; r++) {
cudaMemcpyPeerAsync(dest, destDevice, src, srcDevice, sizeof(int) * num_elems, streamToRun); cudaMemcpyPeerAsync(dest, destDevice, src, srcDevice, sizeof(int) * num_elems, streamToRun);
} }
cudaCheckError();
} }
} }
void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer p2p_method) void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer p2p_method, bool needsFallback)
{ {
int repeat = 5; int repeat = 5;
volatile int *flag = NULL; volatile int *flag = NULL;
@@ -178,6 +256,16 @@ void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer
cudaCheckError(); cudaCheckError();
} }
// Decide if fallback is needed based on global detection result and current test scenario
bool useFallback = (!p2p && needsFallback);
int *hostBuffer = nullptr;
if (useFallback) {
// Allocate hostBuffer for this function's numElems
cudaHostAlloc((void **)&hostBuffer, sizeof(int) * numElems, cudaHostAllocDefault);
cudaCheckError();
}
vector<double> bandwidthMatrix(numGPUs * numGPUs); vector<double> bandwidthMatrix(numGPUs * numGPUs);
for (int i = 0; i < numGPUs; i++) { for (int i = 0; i < numGPUs; i++) {
@@ -215,16 +303,20 @@ void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer
if (i == j) { if (i == j) {
// Perform intra-GPU, D2D copies // Perform intra-GPU, D2D copies
performP2PCopy(buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
else { else {
if (p2p_method == P2P_WRITE) { if (p2p_method == P2P_WRITE) {
performP2PCopy(buffers[j], j, buffers[i], i, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[j], j, buffers[i], i, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
else { else {
performP2PCopy(buffers[i], i, buffers[j], j, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[i], i, buffers[j], j, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
} }
cudaCheckError();
cudaEventRecord(stop[i], stream[i]); cudaEventRecord(stop[i], stream[i]);
cudaCheckError(); cudaCheckError();
@@ -284,11 +376,16 @@ void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer
cudaCheckError(); cudaCheckError();
} }
if (hostBuffer) {
cudaFreeHost(hostBuffer);
cudaCheckError();
}
cudaFreeHost((void *)flag); cudaFreeHost((void *)flag);
cudaCheckError(); cudaCheckError();
} }
void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p) void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p, bool needsFallback)
{ {
int repeat = 5; int repeat = 5;
volatile int *flag = NULL; volatile int *flag = NULL;
@@ -319,6 +416,16 @@ void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p)
cudaCheckError(); cudaCheckError();
} }
// Decide if fallback is needed based on global detection result and current test scenario
bool useFallback = (!p2p && needsFallback);
int *hostBuffer = nullptr;
if (useFallback) {
// Allocate hostBuffer for this function's numElems
cudaHostAlloc((void **)&hostBuffer, sizeof(int) * numElems, cudaHostAllocDefault);
cudaCheckError();
}
vector<double> bandwidthMatrix(numGPUs * numGPUs); vector<double> bandwidthMatrix(numGPUs * numGPUs);
for (int i = 0; i < numGPUs; i++) { for (int i = 0; i < numGPUs; i++) {
@@ -362,18 +469,22 @@ void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p)
if (i == j) { if (i == j) {
// For intra-GPU perform 2 memcopies buffersD2D <-> buffers // For intra-GPU perform 2 memcopies buffersD2D <-> buffers
performP2PCopy(buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream0[i]); performP2PCopy(
performP2PCopy(buffersD2D[i], i, buffers[i], i, numElems, repeat, access, stream1[i]); buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream0[i], useFallback, hostBuffer);
performP2PCopy(
buffersD2D[i], i, buffers[i], i, numElems, repeat, access, stream1[i], useFallback, hostBuffer);
} }
else { else {
if (access && p2p_mechanism == SM) { if (access && p2p_mechanism == SM) {
cudaSetDevice(j); cudaSetDevice(j);
} }
performP2PCopy(buffers[i], i, buffers[j], j, numElems, repeat, access, stream1[j]); performP2PCopy(
buffers[i], i, buffers[j], j, numElems, repeat, access, stream1[j], useFallback, hostBuffer);
if (access && p2p_mechanism == SM) { if (access && p2p_mechanism == SM) {
cudaSetDevice(i); cudaSetDevice(i);
} }
performP2PCopy(buffers[j], j, buffers[i], i, numElems, repeat, access, stream0[i]); performP2PCopy(
buffers[j], j, buffers[i], i, numElems, repeat, access, stream0[i], useFallback, hostBuffer);
} }
// Notify stream0 that stream1 is complete and record the time of // Notify stream0 that stream1 is complete and record the time of
@@ -439,11 +550,16 @@ void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p)
cudaCheckError(); cudaCheckError();
} }
if (hostBuffer) {
cudaFreeHost(hostBuffer);
cudaCheckError();
}
cudaFreeHost((void *)flag); cudaFreeHost((void *)flag);
cudaCheckError(); cudaCheckError();
} }
void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method) void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method, bool needsFallback)
{ {
int repeat = 100; int repeat = 100;
int numElems = 4; // perform 1-int4 transfer. int numElems = 4; // perform 1-int4 transfer.
@@ -478,6 +594,16 @@ void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method)
cudaCheckError(); cudaCheckError();
} }
// Decide if fallback is needed based on global detection result and current test scenario
bool useFallback = (!p2p && needsFallback);
int *hostBuffer = nullptr;
if (useFallback) {
// Allocate hostBuffer for this function's numElems
cudaHostAlloc((void **)&hostBuffer, sizeof(int) * numElems, cudaHostAllocDefault);
cudaCheckError();
}
vector<double> gpuLatencyMatrix(numGPUs * numGPUs); vector<double> gpuLatencyMatrix(numGPUs * numGPUs);
vector<double> cpuLatencyMatrix(numGPUs * numGPUs); vector<double> cpuLatencyMatrix(numGPUs * numGPUs);
@@ -513,14 +639,17 @@ void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method)
sdkResetTimer(&stopWatch); sdkResetTimer(&stopWatch);
if (i == j) { if (i == j) {
// Perform intra-GPU, D2D copies // Perform intra-GPU, D2D copies
performP2PCopy(buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[i], i, buffersD2D[i], i, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
else { else {
if (p2p_method == P2P_WRITE) { if (p2p_method == P2P_WRITE) {
performP2PCopy(buffers[j], j, buffers[i], i, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[j], j, buffers[i], i, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
else { else {
performP2PCopy(buffers[i], i, buffers[j], j, numElems, repeat, access, stream[i]); performP2PCopy(
buffers[i], i, buffers[j], j, numElems, repeat, access, stream[i], useFallback, hostBuffer);
} }
} }
float cpu_time_ms = sdkGetTimerValue(&stopWatch); float cpu_time_ms = sdkGetTimerValue(&stopWatch);
@@ -597,6 +726,11 @@ void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method)
sdkDeleteTimer(&stopWatch); sdkDeleteTimer(&stopWatch);
if (hostBuffer) {
cudaFreeHost(hostBuffer);
cudaCheckError();
}
cudaFreeHost((void *)flag); cudaFreeHost((void *)flag);
cudaCheckError(); cudaCheckError();
} }
@@ -645,6 +779,9 @@ int main(int argc, char **argv)
checkP2Paccess(numGPUs); checkP2Paccess(numGPUs);
// Environment detection: One-time check if cudaMemcpyPeerAsync is supported when P2P is disabled
bool needsFallback = detectFallback(numGPUs);
// Check peer-to-peer connectivity // Check peer-to-peer connectivity
printf("P2P Connectivity Matrix\n"); printf("P2P Connectivity Matrix\n");
printf(" D\\D"); printf(" D\\D");
@@ -671,25 +808,25 @@ int main(int argc, char **argv)
} }
printf("Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n"); printf("Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n");
outputBandwidthMatrix(numElems, numGPUs, false, P2P_WRITE); outputBandwidthMatrix(numElems, numGPUs, false, P2P_WRITE, needsFallback);
printf("Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\n"); printf("Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\n");
outputBandwidthMatrix(numElems, numGPUs, true, P2P_WRITE); outputBandwidthMatrix(numElems, numGPUs, true, P2P_WRITE, needsFallback);
if (p2p_method == P2P_READ) { if (p2p_method == P2P_READ) {
printf("Unidirectional P2P=Enabled Bandwidth (P2P Reads) Matrix (GB/s)\n"); printf("Unidirectional P2P=Enabled Bandwidth (P2P Reads) Matrix (GB/s)\n");
outputBandwidthMatrix(numElems, numGPUs, true, p2p_method); outputBandwidthMatrix(numElems, numGPUs, true, p2p_method, needsFallback);
} }
printf("Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n"); printf("Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n");
outputBidirectionalBandwidthMatrix(numElems, numGPUs, false); outputBidirectionalBandwidthMatrix(numElems, numGPUs, false, needsFallback);
printf("Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)\n"); printf("Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)\n");
outputBidirectionalBandwidthMatrix(numElems, numGPUs, true); outputBidirectionalBandwidthMatrix(numElems, numGPUs, true, needsFallback);
printf("P2P=Disabled Latency Matrix (us)\n"); printf("P2P=Disabled Latency Matrix (us)\n");
outputLatencyMatrix(numGPUs, false, P2P_WRITE); outputLatencyMatrix(numGPUs, false, P2P_WRITE, needsFallback);
printf("P2P=Enabled Latency (P2P Writes) Matrix (us)\n"); printf("P2P=Enabled Latency (P2P Writes) Matrix (us)\n");
outputLatencyMatrix(numGPUs, true, P2P_WRITE); outputLatencyMatrix(numGPUs, true, P2P_WRITE, needsFallback);
if (p2p_method == P2P_READ) { if (p2p_method == P2P_READ) {
printf("P2P=Enabled Latency (P2P Reads) Matrix (us)\n"); printf("P2P=Enabled Latency (P2P Reads) Matrix (us)\n");
outputLatencyMatrix(numGPUs, true, p2p_method); outputLatencyMatrix(numGPUs, true, p2p_method, needsFallback);
} }
printf("\nNOTE: The CUDA Samples are not meant for performance measurements. " printf("\nNOTE: The CUDA Samples are not meant for performance measurements. "

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaHostAlloc, cudaGraphicsUnmapResources, cudaMalloc, cudaFree, cudaGetChannelDesc, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaMemcpyToArray, cudaDeviceSynchronize, cudaCreateTextureObject, cudaProcess, cudaGraphicsUnregisterResource, cudaGraphicsSubResourceGetMappedArray, cudaGraphicsGLRegisterBuffer, cudaGraphicsGLRegisterImage cudaHostAlloc, cudaGraphicsUnmapResources, cudaMalloc, cudaFree, cudaGetChannelDesc, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaMemcpyToArray, cudaDeviceSynchronize, cudaCreateTextureObject, cudaProcess, cudaGraphicsUnregisterResource, cudaGraphicsSubResourceGetMappedArray, cudaGraphicsGLRegisterBuffer, cudaGraphicsGLRegisterImage
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGetDevice, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGetDevice, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaStreamCreateWithFlags, cudaExternalMemoryGetMappedBuffer, cudaSignalSemaphore, cudaWaitExternalSemaphoresAsync, cudaVertMem, cudaImportExternalSemaphore, cudaWaitSemaphore, cudaHeightMap, cudaSetDevice, cudaGetDeviceCount, cudaSignalExternalSemaphoresAsync, cudaTimelineSemaphore, cudaStreamSynchronize, cudaDestroyExternalMemory, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaImportExternalMemory, cudaGetDeviceProperties, cudaDestroyExternalSemaphore cudaStreamCreateWithFlags, cudaExternalMemoryGetMappedBuffer, cudaSignalSemaphore, cudaWaitExternalSemaphoresAsync, cudaVertMem, cudaImportExternalSemaphore, cudaWaitSemaphore, cudaHeightMap, cudaSetDevice, cudaGetDeviceCount, cudaSignalExternalSemaphoresAsync, cudaTimelineSemaphore, cudaStreamSynchronize, cudaDestroyExternalMemory, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaImportExternalMemory, cudaGetDeviceProperties, cudaDestroyExternalSemaphore
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan) [X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan), [GLFW](../../../README.md#glfw)
## Prerequisites ## Prerequisites

View File

@@ -144,8 +144,9 @@ VulkanBaseApp::~VulkanBaseApp()
} }
#ifdef _VK_TIMELINE_SEMAPHORE #ifdef _VK_TIMELINE_SEMAPHORE
if (m_vkPresentationSemaphore != VK_NULL_HANDLE) { for (size_t i = 0; i < m_vkImageAcquiredSemaphores.size(); i++) {
vkDestroySemaphore(m_device, m_vkPresentationSemaphore, nullptr); vkDestroySemaphore(m_device, m_vkImageAcquiredSemaphores[i], nullptr);
vkDestroySemaphore(m_device, m_vkRenderCompleteSemaphores[i], nullptr);
} }
#endif /* _VK_TIMELINE_SEMAPHORE */ #endif /* _VK_TIMELINE_SEMAPHORE */
@@ -1351,8 +1352,15 @@ void VulkanBaseApp::createSyncObjects()
} }
#ifdef _VK_TIMELINE_SEMAPHORE #ifdef _VK_TIMELINE_SEMAPHORE
if (vkCreateSemaphore(m_device, &semaphoreInfo, nullptr, &m_vkPresentationSemaphore) != VK_SUCCESS) { m_vkImageAcquiredSemaphores.resize(MAX_FRAMES_IN_FLIGHT);
throw std::runtime_error("Failed to create binary semaphore!"); m_vkRenderCompleteSemaphores.resize(MAX_FRAMES_IN_FLIGHT);
for (size_t i = 0; i < MAX_FRAMES_IN_FLIGHT; i++) {
if (vkCreateSemaphore(m_device, &semaphoreInfo, nullptr, &m_vkImageAcquiredSemaphores[i]) != VK_SUCCESS) {
throw std::runtime_error("Failed to create image acquired semaphore!");
}
if (vkCreateSemaphore(m_device, &semaphoreInfo, nullptr, &m_vkRenderCompleteSemaphores[i]) != VK_SUCCESS) {
throw std::runtime_error("Failed to create render complete semaphore!");
}
} }
#endif /* _VK_TIMELINE_SEMAPHORE */ #endif /* _VK_TIMELINE_SEMAPHORE */
} }
@@ -1634,6 +1642,12 @@ void VulkanBaseApp::drawFrame()
static uint64_t waitValue = 0; static uint64_t waitValue = 0;
static uint64_t signalValue = 1; static uint64_t signalValue = 1;
size_t currentFrameIdx = m_currentFrame % MAX_FRAMES_IN_FLIGHT;
// Wait for this frame's fence to avoid VUID-vkAcquireNextImageKHR-semaphore-01779
// Ensures previous frame's GPU work completes before reusing per-frame semaphores
vkWaitForFences(m_device, 1, &m_inFlightFences[currentFrameIdx], VK_TRUE, std::numeric_limits<uint64_t>::max());
VkSemaphoreWaitInfo semaphoreWaitInfo = {}; VkSemaphoreWaitInfo semaphoreWaitInfo = {};
semaphoreWaitInfo.sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO; semaphoreWaitInfo.sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO;
semaphoreWaitInfo.pSemaphores = &m_vkTimelineSemaphore; semaphoreWaitInfo.pSemaphores = &m_vkTimelineSemaphore;
@@ -1641,20 +1655,26 @@ void VulkanBaseApp::drawFrame()
semaphoreWaitInfo.pValues = &waitValue; semaphoreWaitInfo.pValues = &waitValue;
vkWaitSemaphores(m_device, &semaphoreWaitInfo, std::numeric_limits<uint64_t>::max()); vkWaitSemaphores(m_device, &semaphoreWaitInfo, std::numeric_limits<uint64_t>::max());
// Use separate binary semaphores for swapchain sync to avoid VUID-01779
// Timeline semaphore (m_vkTimelineSemaphore) is dedicated to Vulkan-CUDA interop
uint32_t imageIndex; uint32_t imageIndex;
VkResult result = vkAcquireNextImageKHR(m_device, VkResult result = vkAcquireNextImageKHR(m_device,
m_swapChain, m_swapChain,
std::numeric_limits<uint64_t>::max(), std::numeric_limits<uint64_t>::max(),
m_vkPresentationSemaphore, m_vkImageAcquiredSemaphores[currentFrameIdx],
VK_NULL_HANDLE, VK_NULL_HANDLE,
&imageIndex); &imageIndex);
if (result == VK_ERROR_OUT_OF_DATE_KHR) { if (result == VK_ERROR_OUT_OF_DATE_KHR) {
recreateSwapChain(); recreateSwapChain();
return;
} }
else if (result != VK_SUCCESS && result != VK_SUBOPTIMAL_KHR) { else if (result != VK_SUCCESS && result != VK_SUBOPTIMAL_KHR) {
throw std::runtime_error("Failed to acquire swap chain image!"); throw std::runtime_error("Failed to acquire swap chain image!");
} }
// Reset the fence for this frame
vkResetFences(m_device, 1, &m_inFlightFences[currentFrameIdx]);
updateUniformBuffer(imageIndex); updateUniformBuffer(imageIndex);
VkSubmitInfo submitInfo = {}; VkSubmitInfo submitInfo = {};
@@ -1663,6 +1683,8 @@ void VulkanBaseApp::drawFrame()
std::vector<VkSemaphore> waitSemaphores; std::vector<VkSemaphore> waitSemaphores;
std::vector<VkPipelineStageFlags> waitStages; std::vector<VkPipelineStageFlags> waitStages;
waitSemaphores.push_back(m_vkImageAcquiredSemaphores[currentFrameIdx]);
waitStages.push_back(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT);
waitSemaphores.push_back(m_vkTimelineSemaphore); waitSemaphores.push_back(m_vkTimelineSemaphore);
waitStages.push_back(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT); waitStages.push_back(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT);
@@ -1675,26 +1697,30 @@ void VulkanBaseApp::drawFrame()
std::vector<VkSemaphore> signalSemaphores; std::vector<VkSemaphore> signalSemaphores;
signalSemaphores.push_back(m_vkTimelineSemaphore); signalSemaphores.push_back(m_vkTimelineSemaphore);
signalSemaphores.push_back(m_vkRenderCompleteSemaphores[currentFrameIdx]);
submitInfo.signalSemaphoreCount = (uint32_t)signalSemaphores.size(); submitInfo.signalSemaphoreCount = (uint32_t)signalSemaphores.size();
submitInfo.pSignalSemaphores = signalSemaphores.data(); submitInfo.pSignalSemaphores = signalSemaphores.data();
VkTimelineSemaphoreSubmitInfo timelineInfo = {}; VkTimelineSemaphoreSubmitInfo timelineInfo = {};
timelineInfo.sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO; timelineInfo.sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO;
timelineInfo.waitSemaphoreValueCount = 1; timelineInfo.waitSemaphoreValueCount = 2;
timelineInfo.pWaitSemaphoreValues = &waitValue; uint64_t waitValues[] = {0, waitValue};
timelineInfo.signalSemaphoreValueCount = 1; timelineInfo.pWaitSemaphoreValues = waitValues;
timelineInfo.pSignalSemaphoreValues = &signalValue; timelineInfo.signalSemaphoreValueCount = 2;
uint64_t signalValues[] = {signalValue, 0};
timelineInfo.pSignalSemaphoreValues = signalValues;
submitInfo.pNext = &timelineInfo; submitInfo.pNext = &timelineInfo;
if (vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, VK_NULL_HANDLE) != VK_SUCCESS) { // Submit with fence to track when this frame completes
if (vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, m_inFlightFences[currentFrameIdx]) != VK_SUCCESS) {
throw std::runtime_error("failed to submit draw command buffer!"); throw std::runtime_error("failed to submit draw command buffer!");
} }
VkPresentInfoKHR presentInfo = {}; VkPresentInfoKHR presentInfo = {};
presentInfo.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR; presentInfo.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR;
presentInfo.waitSemaphoreCount = 1; presentInfo.waitSemaphoreCount = 1;
presentInfo.pWaitSemaphores = &m_vkPresentationSemaphore; presentInfo.pWaitSemaphores = &m_vkRenderCompleteSemaphores[currentFrameIdx];
VkSwapchainKHR swapChains[] = {m_swapChain}; VkSwapchainKHR swapChains[] = {m_swapChain};
presentInfo.swapchainCount = 1; presentInfo.swapchainCount = 1;

View File

@@ -107,7 +107,8 @@ protected:
std::vector<VkFence> m_inFlightFences; std::vector<VkFence> m_inFlightFences;
std::vector<VkBuffer> m_uniformBuffers; std::vector<VkBuffer> m_uniformBuffers;
std::vector<VkDeviceMemory> m_uniformMemory; std::vector<VkDeviceMemory> m_uniformMemory;
VkSemaphore m_vkPresentationSemaphore; std::vector<VkSemaphore> m_vkImageAcquiredSemaphores;
std::vector<VkSemaphore> m_vkRenderCompleteSemaphores;
VkSemaphore m_vkTimelineSemaphore; VkSemaphore m_vkTimelineSemaphore;
VkDescriptorSetLayout m_descriptorSetLayout; VkDescriptorSetLayout m_descriptorSetLayout;
VkDescriptorPool m_descriptorPool; VkDescriptorPool m_descriptorPool;

View File

@@ -38,7 +38,7 @@
typedef float vec2[2]; typedef float vec2[2];
std::string execution_path; std::string execution_path;
#ifndef NDEBUG #ifdef NDEBUG
#define ENABLE_VALIDATION (false) #define ENABLE_VALIDATION (false)
#else #else
#define ENABLE_VALIDATION (true) #define ENABLE_VALIDATION (true)

View File

@@ -29,7 +29,7 @@ cuMemCreate, cuMemAddressReserve, cuMemGetAllocationGranularity, cuMemAddressFre
cudaWaitExternalSemaphoresAsync, cudaImportExternalSemaphore, cudaDeviceGetAttribute, cudaSetDevice, cudaLaunchHostFunc, cudaMallocHost, cudaSignalExternalSemaphoresAsync, cudaFreeHost, cudaMemsetAsync, cudaMemcpyAsync, cudaGetDeviceCount, cudaStreamCreateWithFlags, cudaStreamDestroy, cudaDestroyExternalSemaphore, cudaSignalSemaphore, cudaWaitSemaphore, cudaFree, cudaStreamSynchronize, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties cudaWaitExternalSemaphoresAsync, cudaImportExternalSemaphore, cudaDeviceGetAttribute, cudaSetDevice, cudaLaunchHostFunc, cudaMallocHost, cudaSignalExternalSemaphoresAsync, cudaFreeHost, cudaMemsetAsync, cudaMemcpyAsync, cudaGetDeviceCount, cudaStreamCreateWithFlags, cudaStreamDestroy, cudaDestroyExternalSemaphore, cudaSignalSemaphore, cudaWaitSemaphore, cudaFree, cudaStreamSynchronize, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan) [X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan), [GLFW](../../../README.md#glfw)
## Prerequisites ## Prerequisites

View File

@@ -44,7 +44,7 @@
#include "helper_multiprocess.h" #include "helper_multiprocess.h"
// #define DEBUG // #define DEBUG
#ifndef DEBUG #ifdef NDEBUG
#define ENABLE_VALIDATION (false) #define ENABLE_VALIDATION (false)
#else #else
#define ENABLE_VALIDATION (true) #define ENABLE_VALIDATION (true)

View File

@@ -127,9 +127,9 @@ extern "C"
thrust::device_ptr<float4> d_oldVel(oldVel); thrust::device_ptr<float4> d_oldVel(oldVel);
thrust::for_each( thrust::for_each(
thrust::make_zip_iterator(thrust::make_tuple(d_newPos, d_newVel, d_oldPos, d_oldVel)), thrust::make_zip_iterator(d_newPos, d_newVel, d_oldPos, d_oldVel),
thrust::make_zip_iterator(thrust::make_tuple( thrust::make_zip_iterator(
d_newPos + numParticles, d_newVel + numParticles, d_oldPos + numParticles, d_oldVel + numParticles)), d_newPos + numParticles, d_newVel + numParticles, d_oldPos + numParticles, d_oldVel + numParticles),
integrate_functor(deltaTime, noiseTex)); integrate_functor(deltaTime, noiseTex));
} }
@@ -143,8 +143,8 @@ extern "C"
thrust::device_ptr<float> d_keys(keys); thrust::device_ptr<float> d_keys(keys);
thrust::device_ptr<uint> d_indices(indices); thrust::device_ptr<uint> d_indices(indices);
thrust::for_each(thrust::make_zip_iterator(thrust::make_tuple(d_pos, d_keys)), thrust::for_each(thrust::make_zip_iterator(d_pos, d_keys),
thrust::make_zip_iterator(thrust::make_tuple(d_pos + numParticles, d_keys + numParticles)), thrust::make_zip_iterator(d_pos + numParticles, d_keys + numParticles),
calcDepth_functor(sortVector)); calcDepth_functor(sortVector));
thrust::sequence(d_indices, d_indices + numParticles); thrust::sequence(d_indices, d_indices + numParticles);

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaExtent, cudaPitchedPtr, cudaCreateTextureObject, cudaMemcpyToSymbol cudaExtent, cudaPitchedPtr, cudaCreateTextureObject, cudaMemcpyToSymbol
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -40,6 +40,9 @@
#include "thrust/iterator/zip_iterator.h" #include "thrust/iterator/zip_iterator.h"
#include "thrust/sort.h" #include "thrust/sort.h"
// for cuda::std::get
#include <cuda/std/utility>
cudaTextureObject_t noiseTex; cudaTextureObject_t noiseTex;
// simulation parameters // simulation parameters
__constant__ SimParams cudaParams; __constant__ SimParams cudaParams;
@@ -65,8 +68,8 @@ struct integrate_functor
template <typename Tuple> __device__ void operator()(Tuple t) template <typename Tuple> __device__ void operator()(Tuple t)
{ {
volatile float4 posData = thrust::get<2>(t); volatile float4 posData = cuda::std::get<2>(t);
volatile float4 velData = thrust::get<3>(t); volatile float4 velData = cuda::std::get<3>(t);
float3 pos = make_float3(posData.x, posData.y, posData.z); float3 pos = make_float3(posData.x, posData.y, posData.z);
float3 vel = make_float3(velData.x, velData.y, velData.z); float3 vel = make_float3(velData.x, velData.y, velData.z);
@@ -95,8 +98,8 @@ struct integrate_functor
vel *= cudaParams.globalDamping; vel *= cudaParams.globalDamping;
// store new position and velocity // store new position and velocity
thrust::get<0>(t) = make_float4(pos, age); cuda::std::get<0>(t) = make_float4(pos, age);
thrust::get<1>(t) = make_float4(vel, velData.w); cuda::std::get<1>(t) = make_float4(vel, velData.w);
} }
}; };
@@ -111,10 +114,10 @@ struct calcDepth_functor
template <typename Tuple> __host__ __device__ void operator()(Tuple t) template <typename Tuple> __host__ __device__ void operator()(Tuple t)
{ {
volatile float4 p = thrust::get<0>(t); volatile float4 p = cuda::std::get<0>(t);
float key = -dot(make_float3(p.x, p.y, p.z), float key = -dot(make_float3(p.x, p.y, p.z),
sortVector); // project onto sort vector sortVector); // project onto sort vector
thrust::get<1>(t) = key; cuda::std::get<1>(t) = key;
} }
}; };

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaMemcpy, cudaGraphicsMapResources, cudaDestroySurfaceObject, cudaExtent, cudaDeviceSynchronize, cudaCreateSurfaceObject, cudaMemcpyToSymbol, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaCreateTextureObject, cudaGraphicsUnmapResources, cudaMallocArray, cudaFreeArray, cudaDestroyTextureObject, cudaMemset, cudaGraphicsGLRegisterBuffer, cudaFree, cudaGraphicsUnregisterResource, cudaMalloc cudaMemcpy, cudaGraphicsMapResources, cudaDestroySurfaceObject, cudaExtent, cudaDeviceSynchronize, cudaCreateSurfaceObject, cudaMemcpyToSymbol, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaCreateTextureObject, cudaGraphicsUnmapResources, cudaMallocArray, cudaFreeArray, cudaDestroyTextureObject, cudaMemset, cudaGraphicsGLRegisterBuffer, cudaFree, cudaGraphicsUnregisterResource, cudaMalloc
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, armv7l
cudaProfilerStop, cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaExtent, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemset, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaProfilerStop, cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaPitchedPtr, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaExtent, cudaDeviceSynchronize, cudaCreateTextureObject, cudaMemset, cudaMemcpyToSymbol, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GL](../../../README.md#gl) [X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ x86_64, aarch64
cudaVkSemaphoreSignal, cudaWaitExternalSemaphoresAsync, cudaMemcpy, cudaVkImportSemaphore, cudaImportExternalSemaphore, cudaGetMipmappedArrayLevel, cudaSetDevice, cudaDestroySurfaceObject, cudaExtent, cudaMallocMipmappedArray, cudaCreateSurfaceObject, cudaStreamCreate, cudaSignalExternalSemaphoresAsync, cudaUpdateVkImage, cudaCreateTextureObject, cudaGetDeviceCount, cudaExternalMemoryGetMappedMipmappedArray, cudaDestroyTextureObject, cudaVkImportImageMem, cudaDestroyExternalMemory, cudaVkSemaphoreWait, cudaImportExternalMemory, cudaDestroyExternalSemaphore, cudaFreeMipmappedArray, cudaFree, cudaMalloc, cudaGetDeviceProperties cudaVkSemaphoreSignal, cudaWaitExternalSemaphoresAsync, cudaMemcpy, cudaVkImportSemaphore, cudaImportExternalSemaphore, cudaGetMipmappedArrayLevel, cudaSetDevice, cudaDestroySurfaceObject, cudaExtent, cudaMallocMipmappedArray, cudaCreateSurfaceObject, cudaStreamCreate, cudaSignalExternalSemaphoresAsync, cudaUpdateVkImage, cudaCreateTextureObject, cudaGetDeviceCount, cudaExternalMemoryGetMappedMipmappedArray, cudaDestroyTextureObject, cudaVkImportImageMem, cudaDestroyExternalMemory, cudaVkSemaphoreWait, cudaImportExternalMemory, cudaDestroyExternalSemaphore, cudaFreeMipmappedArray, cudaFree, cudaMalloc, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan) [X11](../../../README.md#x11), [VULKAN](../../../README.md#vulkan), [GLFW](../../../README.md#glfw)
## Prerequisites ## Prerequisites

View File

@@ -69,9 +69,9 @@ const int MAX_FRAMES = 4;
const std::vector<const char *> validationLayers = {"VK_LAYER_KHRONOS_validation"}; const std::vector<const char *> validationLayers = {"VK_LAYER_KHRONOS_validation"};
#ifdef NDEBUG #ifdef NDEBUG
const bool enableValidationLayers = true;
#else
const bool enableValidationLayers = false; const bool enableValidationLayers = false;
#else
const bool enableValidationLayers = true;
#endif #endif
std::string execution_path; std::string execution_path;

View File

@@ -24,7 +24,7 @@ x86_64, armv7l, aarch64
cudaMemcpy, cudaStreamDestroy, cudaMemPrefetchAsync, cudaFree, cudaMallocHost, cudaMallocManaged, cudaStreamAttachMemAsync, cudaHostGetDevicePointer, cudaFreeHost, cudaStreamSynchronize, cudaMalloc, cudaMemcpyAsync, cudaStreamCreate, cudaGetDeviceProperties cudaMemcpy, cudaStreamDestroy, cudaMemPrefetchAsync, cudaFree, cudaMallocHost, cudaMallocManaged, cudaStreamAttachMemAsync, cudaHostGetDevicePointer, cudaFreeHost, cudaStreamSynchronize, cudaMalloc, cudaMemcpyAsync, cudaStreamCreate, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[UVM](../../../README.md#uvm) [UVM](../../../README.md#unified-virtual-memory)
## Prerequisites ## Prerequisites

View File

@@ -664,7 +664,17 @@ int main(int argc, char **argv)
verboseResults = 1; verboseResults = 1;
} }
// set device
cudaDeviceProp device_prop;
int device_id = findCudaDevice(argc, (const char **)argv); int device_id = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&device_prop, device_id));
if (!device_prop.managedMemory) {
// This samples requires being run on a device that supports Unified Memory
fprintf(stderr, "Unified Memory not supported on this device\n");
exit(EXIT_WAIVED);
}
matrixMultiplyPerfRunner(reportAsBandwidth, print_launch_transfer_results, print_std_deviation, device_id); matrixMultiplyPerfRunner(reportAsBandwidth, print_launch_transfer_results, print_std_deviation, device_id);

View File

@@ -29,7 +29,7 @@ cuEventRecord, cuDeviceGetAttribute, cuEventCreate, cuCtxSynchronize, cuEventDes
cudaMemcpy, cudaGetErrorString, cudaFree, cudaDeviceSynchronize, cudaGetValueMismatch, cudaMalloc cudaMemcpy, cudaGetErrorString, cudaFree, cudaDeviceSynchronize, cudaGetValueMismatch, cudaMalloc
## Dependencies needed to build/run ## Dependencies needed to build/run
[EGL](../../../README.md#egl), [EGLSync](../../../README.md#eglsync), [X11](../../../README.md#x11), [GLES](../../../README.md#gles) [EGL](../../../README.md#egl), [EGLSync](../../../README.md#eglsync), [X11](../../../README.md#x11), [OpenGL ES](../../../README.md#opengl-es)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMallocPitch, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties cudaGraphicsUnmapResources, cudaMemcpy, cudaMallocArray, cudaFreeArray, cudaFree, cudaMallocPitch, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDestroyTextureObject, cudaCreateTextureObject, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer, cudaGetDeviceProperties
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GLES](../../../README.md#gles), [CUFFT](../../../README.md#cufft) [X11](../../../README.md#x11), [OpenGL ES](../../../README.md#opengl-es), [CUFFT](../../../README.md#cufft)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ armv7l
cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaStreamQuery, cudaEventDestroy, cudaEventElapsedTime, cudaGetDeviceCount, cudaEventCreate cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaStreamQuery, cudaEventDestroy, cudaEventElapsedTime, cudaGetDeviceCount, cudaEventCreate
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GLES](../../../README.md#gles) [X11](../../../README.md#x11), [OpenGL ES](../../../README.md#opengl-es)
## Prerequisites ## Prerequisites

View File

@@ -26,7 +26,7 @@ armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[X11](../../../README.md#x11), [GLES](../../../README.md#gles) [X11](../../../README.md#x11), [OpenGL ES](../../../README.md#opengl-es)
## Prerequisites ## Prerequisites

View File

@@ -31,7 +31,7 @@ armv7l
cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer cudaGraphicsUnmapResources, cudaMemcpy, cudaFree, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaDeviceSynchronize, cudaGraphicsUnregisterResource, cudaMalloc, cudaGraphicsGLRegisterBuffer
## Dependencies needed to build/run ## Dependencies needed to build/run
[EGLOutput](../../../README.md#egloutput), [GLES](../../../README.md#gles) [EGLOutput](../../../README.md#egloutput), [OpenGL ES](../../../README.md#opengl-es)
## Prerequisites ## Prerequisites

View File

@@ -79,12 +79,22 @@ if(NOT CUDA_SAMPLES_INSTALL_CONFIGURED)
if(NOT EXISTS "${ROOT_BINARY_DIR}/CMakeCache.txt") if(NOT EXISTS "${ROOT_BINARY_DIR}/CMakeCache.txt")
set(ROOT_BINARY_DIR "${CMAKE_BINARY_DIR}") set(ROOT_BINARY_DIR "${CMAKE_BINARY_DIR}")
endif() endif()
# If user didn't specify CMAKE_INSTALL_PREFIX, default to build/bin
set(CMAKE_INSTALL_PREFIX "${ROOT_BINARY_DIR}/bin" CACHE PATH "Installation directory" FORCE) set(CMAKE_INSTALL_PREFIX "${ROOT_BINARY_DIR}/bin" CACHE PATH "Installation directory" FORCE)
endif() endif()
# Create the installation path: bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE) # Create the installation path: bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
# For multi-config generators, this will be evaluated at install time # For multi-config generators, this will be evaluated at install time
set(CUDA_SAMPLES_INSTALL_DIR "${CMAKE_INSTALL_PREFIX}/${TARGET_ARCH}/${TARGET_OS}/${BUILD_TYPE_EXPR}" CACHE INTERNAL "Install directory for samples") if(NOT DEFINED CUDA_SAMPLES_INSTALL_DIR)
set(CUDA_SAMPLES_INSTALL_DIR "${CMAKE_INSTALL_PREFIX}/${TARGET_ARCH}/${TARGET_OS}/${BUILD_TYPE_EXPR}" CACHE PATH "Install directory for samples")
endif()
# Check if we should use the default dynamic installation logic
if("${CUDA_SAMPLES_INSTALL_DIR}" STREQUAL "${CMAKE_INSTALL_PREFIX}/${TARGET_ARCH}/${TARGET_OS}/${BUILD_TYPE_EXPR}")
set(CUDA_SAMPLES_INSTALL_USE_DEFAULT TRUE)
else()
set(CUDA_SAMPLES_INSTALL_USE_DEFAULT FALSE)
endif()
# Print installation configuration # Print installation configuration
message(STATUS "CUDA Samples installation configured:") message(STATUS "CUDA Samples installation configured:")
@@ -102,10 +112,10 @@ endif()
# Function to setup installation for regular samples # Function to setup installation for regular samples
# This should be called after all targets are defined # This should be called after all targets are defined
function(setup_samples_install) function(setup_samples_install)
# Create an install script that will copy executables and specific file types # Create an install script that copies executables, data files, and shared libraries
# All files are copied to flat root directory # This script only installs - build must be done separately
# This script runs at install time, after the build is complete
install(CODE " install(CODE "
if(\"${CUDA_SAMPLES_INSTALL_USE_DEFAULT}\" STREQUAL \"TRUE\")
# Determine the actual install directory based on the configuration # Determine the actual install directory based on the configuration
# CMAKE_INSTALL_CONFIG_NAME is available at install time for multi-config generators # CMAKE_INSTALL_CONFIG_NAME is available at install time for multi-config generators
if(DEFINED CMAKE_INSTALL_CONFIG_NAME) if(DEFINED CMAKE_INSTALL_CONFIG_NAME)
@@ -114,6 +124,9 @@ function(setup_samples_install)
set(INSTALL_BUILD_TYPE \"${BUILD_TYPE_EXPR}\") set(INSTALL_BUILD_TYPE \"${BUILD_TYPE_EXPR}\")
endif() endif()
set(INSTALL_DIR \"${CMAKE_INSTALL_PREFIX}/${TARGET_ARCH}/${TARGET_OS}/\${INSTALL_BUILD_TYPE}\") set(INSTALL_DIR \"${CMAKE_INSTALL_PREFIX}/${TARGET_ARCH}/${TARGET_OS}/\${INSTALL_BUILD_TYPE}\")
else()
set(INSTALL_DIR \"${CUDA_SAMPLES_INSTALL_DIR}\")
endif()
# Search in the current project's binary directory for built executables # Search in the current project's binary directory for built executables
file(GLOB_RECURSE BINARY_FILES file(GLOB_RECURSE BINARY_FILES