NASP/cuda-samples
85
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-25 22:28:17 +08:00
Code Issues Packages Projects Releases Wiki Activity

Apply consistent code formatting across the repo. Add clang-format and pre-commit hooks.

Browse Source
This commit is contained in:
Rob Armstrong
2025-03-27 10:30:07 -07:00
parent 2cd58fbc9a
commit ceab6e8bcc
782 changed files with 107230 additions and 106548 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
Samples/0_Introduction/vectorAddMMAP/README.md
View File

@@ -30,4 +30,3 @@ cuMemcpyDtoH, cuDeviceCanAccessPeer, cuModuleGetFunction, cuMemSetAccess, cuMemR
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## References (for more details)

298
Samples/0_Introduction/vectorAddMMAP/multidevicealloc_memmap.cpp
View File

@@ -29,172 +29,172 @@
static size_t round_up(size_t x, size_t y) { return ((x + y - 1) / y) * y; }
CUresult simpleMallocMultiDeviceMmap(
CUdeviceptr *dptr, size_t *allocationSize, size_t size,
const std::vector<CUdevice> &residentDevices,
const std::vector<CUdevice> &mappingDevices, size_t align) {
CUresult status = CUDA_SUCCESS;
size_t min_granularity = 0;
size_t stripeSize;
CUresult simpleMallocMultiDeviceMmap(CUdeviceptr *dptr,
size_t *allocationSize,
size_t size,
const std::vector<CUdevice> &residentDevices,
const std::vector<CUdevice> &mappingDevices,
size_t align)
{
CUresult status = CUDA_SUCCESS;
size_t min_granularity = 0;
size_t stripeSize;
// Setup the properties common for all the chunks
// The allocations will be device pinned memory.
// This property structure describes the physical location where the memory
// will be allocated via cuMemCreate allong with additional properties In this
// case, the allocation will be pinnded device memory local to a given device.
CUmemAllocationProp prop = {};
prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
// Setup the properties common for all the chunks
// The allocations will be device pinned memory.
// This property structure describes the physical location where the memory
// will be allocated via cuMemCreate allong with additional properties In this
// case, the allocation will be pinnded device memory local to a given device.
CUmemAllocationProp prop = {};
prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
// Get the minimum granularity needed for the resident devices
// (the max of the minimum granularity of each participating device)
for (int idx = 0; idx < residentDevices.size(); idx++) {
size_t granularity = 0;
// Get the minimum granularity needed for the resident devices
// (the max of the minimum granularity of each participating device)
for (int idx = 0; idx < residentDevices.size(); idx++) {
size_t granularity = 0;
// get the minnimum granularity for residentDevices[idx]
prop.location.id = residentDevices[idx];
status = cuMemGetAllocationGranularity(&granularity, &prop,
CU_MEM_ALLOC_GRANULARITY_MINIMUM);
if (status != CUDA_SUCCESS) {
goto done;
}
if (min_granularity < granularity) {
min_granularity = granularity;
}
}
// Get the minimum granularity needed for the accessing devices
// (the max of the minimum granularity of each participating device)
for (size_t idx = 0; idx < mappingDevices.size(); idx++) {
size_t granularity = 0;
// get the minnimum granularity for mappingDevices[idx]
prop.location.id = mappingDevices[idx];
status = cuMemGetAllocationGranularity(&granularity, &prop,
CU_MEM_ALLOC_GRANULARITY_MINIMUM);
if (status != CUDA_SUCCESS) {
goto done;
}
if (min_granularity < granularity) {
min_granularity = granularity;
}
}
// Round up the size such that we can evenly split it into a stripe size tha
// meets the granularity requirements Essentially size = N *
// residentDevices.size() * min_granularity is the requirement, since each
// piece of the allocation will be stripeSize = N * min_granularity and the
// min_granularity requirement applies to each stripeSize piece of the
// allocation.
size = round_up(size, residentDevices.size() * min_granularity);
stripeSize = size / residentDevices.size();
// Return the rounded up size to the caller for use in the free
if (allocationSize) {
*allocationSize = size;
}
// Reserve the required contiguous VA space for the allocations
status = cuMemAddressReserve(dptr, size, align, 0, 0);
if (status != CUDA_SUCCESS) {
goto done;
}
// Create and map the backings on each gpu
// note: reusing CUmemAllocationProp prop from earlier with prop.type &
// prop.location.type already specified.
for (size_t idx = 0; idx < residentDevices.size(); idx++) {
CUresult status2 = CUDA_SUCCESS;
// Set the location for this chunk to this device
prop.location.id = residentDevices[idx];
// Create the allocation as a pinned allocation on this device
CUmemGenericAllocationHandle allocationHandle;
status = cuMemCreate(&allocationHandle, stripeSize, &prop, 0);
if (status != CUDA_SUCCESS) {
goto done;
// get the minnimum granularity for residentDevices[idx]
prop.location.id = residentDevices[idx];
status = cuMemGetAllocationGranularity(&granularity, &prop, CU_MEM_ALLOC_GRANULARITY_MINIMUM);
if (status != CUDA_SUCCESS) {
goto done;
}
if (min_granularity < granularity) {
min_granularity = granularity;
}
}
// Assign the chunk to the appropriate VA range and release the handle.
// After mapping the memory, it can be referenced by virtual address.
// Since we do not need to make any other mappings of this memory or export
// it, we no longer need and can release the allocationHandle. The
// allocation will be kept live until it is unmapped.
status = cuMemMap(*dptr + (stripeSize * idx), stripeSize, 0,
allocationHandle, 0);
// the handle needs to be released even if the mapping failed.
status2 = cuMemRelease(allocationHandle);
if (status == CUDA_SUCCESS) {
// cuMemRelease should not have failed here
// as the handle was just allocated successfully
// however return an error if it does.
status = status2;
}
// Cleanup in case of any mapping failures.
if (status != CUDA_SUCCESS) {
goto done;
}
}
{
// Each accessDescriptor will describe the mapping requirement for a single
// device
std::vector<CUmemAccessDesc> accessDescriptors;
accessDescriptors.resize(mappingDevices.size());
// Prepare the access descriptor array indicating where and how the backings
// should be visible.
// Get the minimum granularity needed for the accessing devices
// (the max of the minimum granularity of each participating device)
for (size_t idx = 0; idx < mappingDevices.size(); idx++) {
// Specify which device we are adding mappings for.
accessDescriptors[idx].location.type = CU_MEM_LOCATION_TYPE_DEVICE;
accessDescriptors[idx].location.id = mappingDevices[idx];
size_t granularity = 0;
// Specify both read and write access.
accessDescriptors[idx].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
// get the minnimum granularity for mappingDevices[idx]
prop.location.id = mappingDevices[idx];
status = cuMemGetAllocationGranularity(&granularity, &prop, CU_MEM_ALLOC_GRANULARITY_MINIMUM);
if (status != CUDA_SUCCESS) {
goto done;
}
if (min_granularity < granularity) {
min_granularity = granularity;
}
}
// Apply the access descriptors to the whole VA range.
status = cuMemSetAccess(*dptr, size, &accessDescriptors[0],
accessDescriptors.size());
// Round up the size such that we can evenly split it into a stripe size tha
// meets the granularity requirements Essentially size = N *
// residentDevices.size() * min_granularity is the requirement, since each
// piece of the allocation will be stripeSize = N * min_granularity and the
// min_granularity requirement applies to each stripeSize piece of the
// allocation.
size = round_up(size, residentDevices.size() * min_granularity);
stripeSize = size / residentDevices.size();
// Return the rounded up size to the caller for use in the free
if (allocationSize) {
*allocationSize = size;
}
// Reserve the required contiguous VA space for the allocations
status = cuMemAddressReserve(dptr, size, align, 0, 0);
if (status != CUDA_SUCCESS) {
goto done;
goto done;
}
// Create and map the backings on each gpu
// note: reusing CUmemAllocationProp prop from earlier with prop.type &
// prop.location.type already specified.
for (size_t idx = 0; idx < residentDevices.size(); idx++) {
CUresult status2 = CUDA_SUCCESS;
// Set the location for this chunk to this device
prop.location.id = residentDevices[idx];
// Create the allocation as a pinned allocation on this device
CUmemGenericAllocationHandle allocationHandle;
status = cuMemCreate(&allocationHandle, stripeSize, &prop, 0);
if (status != CUDA_SUCCESS) {
goto done;
}
// Assign the chunk to the appropriate VA range and release the handle.
// After mapping the memory, it can be referenced by virtual address.
// Since we do not need to make any other mappings of this memory or export
// it, we no longer need and can release the allocationHandle. The
// allocation will be kept live until it is unmapped.
status = cuMemMap(*dptr + (stripeSize * idx), stripeSize, 0, allocationHandle, 0);
// the handle needs to be released even if the mapping failed.
status2 = cuMemRelease(allocationHandle);
if (status == CUDA_SUCCESS) {
// cuMemRelease should not have failed here
// as the handle was just allocated successfully
// however return an error if it does.
status = status2;
}
// Cleanup in case of any mapping failures.
if (status != CUDA_SUCCESS) {
goto done;
}
}
{
// Each accessDescriptor will describe the mapping requirement for a single
// device
std::vector<CUmemAccessDesc> accessDescriptors;
accessDescriptors.resize(mappingDevices.size());
// Prepare the access descriptor array indicating where and how the backings
// should be visible.
for (size_t idx = 0; idx < mappingDevices.size(); idx++) {
// Specify which device we are adding mappings for.
accessDescriptors[idx].location.type = CU_MEM_LOCATION_TYPE_DEVICE;
accessDescriptors[idx].location.id = mappingDevices[idx];
// Specify both read and write access.
accessDescriptors[idx].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
}
// Apply the access descriptors to the whole VA range.
status = cuMemSetAccess(*dptr, size, &accessDescriptors[0], accessDescriptors.size());
if (status != CUDA_SUCCESS) {
goto done;
}
}
}
done:
if (status != CUDA_SUCCESS) {
if (*dptr) {
simpleFreeMultiDeviceMmap(*dptr, size);
if (status != CUDA_SUCCESS) {
if (*dptr) {
simpleFreeMultiDeviceMmap(*dptr, size);
}
}
}
return status;
return status;
}
CUresult simpleFreeMultiDeviceMmap(CUdeviceptr dptr, size_t size) {
CUresult status = CUDA_SUCCESS;
CUresult simpleFreeMultiDeviceMmap(CUdeviceptr dptr, size_t size)
{
CUresult status = CUDA_SUCCESS;
// Unmap the mapped virtual memory region
// Since the handles to the mapped backing stores have already been released
// by cuMemRelease, and these are the only/last mappings referencing them,
// The backing stores will be freed.
// Since the memory has been unmapped after this call, accessing the specified
// va range will result in a fault (unitll it is remapped).
status = cuMemUnmap(dptr, size);
if (status != CUDA_SUCCESS) {
return status;
}
// Free the virtual address region. This allows the virtual address region
// to be reused by future cuMemAddressReserve calls. This also allows the
// virtual address region to be used by other allocation made through
// opperating system calls like malloc & mmap.
status = cuMemAddressFree(dptr, size);
if (status != CUDA_SUCCESS) {
return status;
}
// Unmap the mapped virtual memory region
// Since the handles to the mapped backing stores have already been released
// by cuMemRelease, and these are the only/last mappings referencing them,
// The backing stores will be freed.
// Since the memory has been unmapped after this call, accessing the specified
// va range will result in a fault (unitll it is remapped).
status = cuMemUnmap(dptr, size);
if (status != CUDA_SUCCESS) {
return status;
}
// Free the virtual address region. This allows the virtual address region
// to be reused by future cuMemAddressReserve calls. This also allows the
// virtual address region to be used by other allocation made through
// opperating system calls like malloc & mmap.
status = cuMemAddressFree(dptr, size);
if (status != CUDA_SUCCESS) {
return status;
}
return status;
return status;
}

10
Samples/0_Introduction/vectorAddMMAP/multidevicealloc_memmap.hpp
View File

@@ -63,10 +63,12 @@
//! handle
//! is not needed after its mappings are set up.
////////////////////////////////////////////////////////////////////////////
CUresult simpleMallocMultiDeviceMmap(
CUdeviceptr *dptr, size_t *allocationSize, size_t size,
const std::vector<CUdevice> &residentDevices,
const std::vector<CUdevice> &mappingDevices, size_t align = 0);
CUresult simpleMallocMultiDeviceMmap(CUdeviceptr *dptr,
size_t *allocationSize,
size_t size,
const std::vector<CUdevice> &residentDevices,
const std::vector<CUdevice> &mappingDevices,
size_t align = 0);
////////////////////////////////////////////////////////////////////////////
//! Frees resources allocated by simpleMallocMultiDeviceMmap

175
Samples/0_Introduction/vectorAddMMAP/vectorAddMMAP.cpp
View File

@@ -36,11 +36,11 @@
*/
// Includes
#include <cstring>
#include <cuda.h>
#include <iostream>
#include <stdio.h>
#include <string.h>
#include <cstring>
#include <iostream>
// includes, project
#include <helper_cuda_drvapi.h>
@@ -54,115 +54,111 @@
using namespace std;
// Variables
CUdevice cuDevice;
CUcontext cuContext;
CUmodule cuModule;
CUfunction vecAdd_kernel;
float *h_A;
float *h_B;
float *h_C;
CUdevice cuDevice;
CUcontext cuContext;
CUmodule cuModule;
CUfunction vecAdd_kernel;
float *h_A;
float *h_B;
float *h_C;
CUdeviceptr d_A;
CUdeviceptr d_B;
CUdeviceptr d_C;
size_t allocationSize = 0;
size_t allocationSize = 0;
// Functions
int CleanupNoFailure();
int CleanupNoFailure();
void RandomInit(float *, int);
//define input fatbin file
// define input fatbin file
#ifndef FATBIN_FILE
#define FATBIN_FILE "vectorAdd_kernel64.fatbin"
#endif
// collect all of the devices whose memory can be mapped from cuDevice.
vector<CUdevice> getBackingDevices(CUdevice cuDevice) {
int num_devices;
vector<CUdevice> getBackingDevices(CUdevice cuDevice)
{
int num_devices;
checkCudaErrors(cuDeviceGetCount(&num_devices));
checkCudaErrors(cuDeviceGetCount(&num_devices));
vector<CUdevice> backingDevices;
backingDevices.push_back(cuDevice);
for (int dev = 0; dev < num_devices; dev++) {
int capable = 0;
int attributeVal = 0;
vector<CUdevice> backingDevices;
backingDevices.push_back(cuDevice);
for (int dev = 0; dev < num_devices; dev++) {
int capable = 0;
int attributeVal = 0;
// The mapping device is already in the backingDevices vector
if (dev == cuDevice) {
continue;
// The mapping device is already in the backingDevices vector
if (dev == cuDevice) {
continue;
}
// Only peer capable devices can map each others memory
checkCudaErrors(cuDeviceCanAccessPeer(&capable, cuDevice, dev));
if (!capable) {
continue;
}
// The device needs to support virtual address management for the required
// apis to work
checkCudaErrors(
cuDeviceGetAttribute(&attributeVal, CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED, cuDevice));
if (attributeVal == 0) {
continue;
}
backingDevices.push_back(dev);
}
// Only peer capable devices can map each others memory
checkCudaErrors(cuDeviceCanAccessPeer(&capable, cuDevice, dev));
if (!capable) {
continue;
}
// The device needs to support virtual address management for the required
// apis to work
checkCudaErrors(cuDeviceGetAttribute(
&attributeVal, CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED,
cuDevice));
if (attributeVal == 0) {
continue;
}
backingDevices.push_back(dev);
}
return backingDevices;
return backingDevices;
}
// Host code
int main(int argc, char **argv) {
printf("Vector Addition (Driver API)\n");
int N = 50000;
size_t size = N * sizeof(float);
int attributeVal = 0;
int main(int argc, char **argv)
{
printf("Vector Addition (Driver API)\n");
int N = 50000;
size_t size = N * sizeof(float);
int attributeVal = 0;
// Initialize
checkCudaErrors(cuInit(0));
// Initialize
checkCudaErrors(cuInit(0));
cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
// Check that the selected device supports virtual address management
checkCudaErrors(cuDeviceGetAttribute(
&attributeVal, CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED,
cuDevice));
printf("Device %d VIRTUAL ADDRESS MANAGEMENT SUPPORTED = %d.\n", cuDevice,
attributeVal);
if (attributeVal == 0) {
printf("Device %d doesn't support VIRTUAL ADDRESS MANAGEMENT.\n", cuDevice);
exit(EXIT_WAIVED);
}
// Check that the selected device supports virtual address management
checkCudaErrors(
cuDeviceGetAttribute(&attributeVal, CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED, cuDevice));
printf("Device %d VIRTUAL ADDRESS MANAGEMENT SUPPORTED = %d.\n", cuDevice, attributeVal);
if (attributeVal == 0) {
printf("Device %d doesn't support VIRTUAL ADDRESS MANAGEMENT.\n", cuDevice);
exit(EXIT_WAIVED);
}
// The vector addition happens on cuDevice, so the allocations need to be
// mapped there.
vector<CUdevice> mappingDevices;
mappingDevices.push_back(cuDevice);
// The vector addition happens on cuDevice, so the allocations need to be
// mapped there.
vector<CUdevice> mappingDevices;
mappingDevices.push_back(cuDevice);
// Collect devices accessible by the mapping device (cuDevice) into the
// backingDevices vector.
vector<CUdevice> backingDevices = getBackingDevices(cuDevice);
// Collect devices accessible by the mapping device (cuDevice) into the
// backingDevices vector.
vector<CUdevice> backingDevices = getBackingDevices(cuDevice);
// Create context
checkCudaErrors(cuCtxCreate(&cuContext, 0, cuDevice));
// Create context
checkCudaErrors(cuCtxCreate(&cuContext, 0, cuDevice));
// first search for the module path before we load the results
string module_path;
std::ostringstream fatbin;
if (!findFatbinPath(FATBIN_FILE, module_path, argv, fatbin))
{
if (!findFatbinPath(FATBIN_FILE, module_path, argv, fatbin)) {
exit(EXIT_FAILURE);
}
else
{
else {
printf("> initCUDA loading module: <%s>\n", module_path.c_str());
}
if (!fatbin.str().size())
{
if (!fatbin.str().size()) {
printf("fatbin file empty. exiting..\n");
exit(EXIT_FAILURE);
}
@@ -204,13 +200,10 @@ int main(int argc, char **argv) {
int threadsPerBlock = 256;
int blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
void *args[] = { &d_A, &d_B, &d_C, &N };
void *args[] = {&d_A, &d_B, &d_C, &N};
// Launch the CUDA kernel
checkCudaErrors(cuLaunchKernel(vecAdd_kernel, blocksPerGrid, 1, 1,
threadsPerBlock, 1, 1,
0,
NULL, args, NULL));
checkCudaErrors(cuLaunchKernel(vecAdd_kernel, blocksPerGrid, 1, 1, threadsPerBlock, 1, 1, 0, NULL, args, NULL));
// Copy result from device memory to host memory
// h_C contains the result in host memory
@@ -219,20 +212,18 @@ int main(int argc, char **argv) {
// Verify result
int i;
for (i = 0; i < N; ++i)
{
for (i = 0; i < N; ++i) {
float sum = h_A[i] + h_B[i];
if (fabs(h_C[i] - sum) > 1e-7f)
{
if (fabs(h_C[i] - sum) > 1e-7f) {
break;
}
}
CleanupNoFailure();
printf("%s\n", (i==N) ? "Result = PASS" : "Result = FAIL");
printf("%s\n", (i == N) ? "Result = PASS" : "Result = FAIL");
exit((i==N) ? EXIT_SUCCESS : EXIT_FAILURE);
exit((i == N) ? EXIT_SUCCESS : EXIT_FAILURE);
}
int CleanupNoFailure()
@@ -243,18 +234,15 @@ int CleanupNoFailure()
checkCudaErrors(simpleFreeMultiDeviceMmap(d_C, allocationSize));
// Free host memory
if (h_A)
{
if (h_A) {
free(h_A);
}
if (h_B)
{
if (h_B) {
free(h_B);
}
if (h_C)
{
if (h_C) {
free(h_C);
}
@@ -265,8 +253,7 @@ int CleanupNoFailure()
// Allocates an array with random float entries.
void RandomInit(float *data, int n)
{
for (int i = 0; i < n; ++i)
{
for (int i = 0; i < n; ++i) {
data[i] = rand() / (float)RAND_MAX;
}
}

9
Samples/0_Introduction/vectorAddMMAP/vectorAdd_kernel.cu
View File

@@ -34,9 +34,10 @@
*/
// Device code
extern "C" __global__ void VecAdd_kernel(const float *A, const float *B,
float *C, int N) {
int i = blockDim.x * blockIdx.x + threadIdx.x;
extern "C" __global__ void VecAdd_kernel(const float *A, const float *B, float *C, int N)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i < N) C[i] = A[i] + B[i];
if (i < N)
C[i] = A[i] + B[i];
}