NASP/cuda-samples
84
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-13 00:08:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

CUDA 13.2 samples update (#432)

Browse Source 
- Added Python samples for CUDA Python 1.0 release
- Renamed top-level `Samples` directory to `cpp` to accommodate Python samples.
This commit is contained in:
Dheemanth
2026-05-13 15:13:18 -07:00
committed by GitHub
parent a4526d5229
commit aeab82ff30
2072 changed files with 15077 additions and 1970 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
cpp/0_Introduction/vectorAddMMAP/.vscode/c_cpp_properties.json vendored Normal file
View File

@@ -0,0 +1,18 @@
{
"configurations": [
{
"name": "Linux",
"includePath": [
"${workspaceFolder}/**",
"${workspaceFolder}/../../../Common"
],
"defines": [],
"compilerPath": "/usr/local/cuda/bin/nvcc",
"cStandard": "gnu17",
"cppStandard": "gnu++14",
"intelliSenseMode": "linux-gcc-x64",
"configurationProvider": "ms-vscode.makefile-tools"
}
],
"version": 4
}

7
cpp/0_Introduction/vectorAddMMAP/.vscode/extensions.json vendored Normal file
View File

@@ -0,0 +1,7 @@
{
"recommendations": [
"nvidia.nsight-vscode-edition",
"ms-vscode.cpptools",
"ms-vscode.makefile-tools"
]
}

67
cpp/0_Introduction/vectorAddMMAP/CMakeLists.txt Normal file
View File

@@ -0,0 +1,67 @@
cmake_minimum_required(VERSION 3.20)
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
project(vectorAddMMAP LANGUAGES C CXX CUDA)
find_package(CUDAToolkit REQUIRED)
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
set(CMAKE_CUDA_ARCHITECTURES 75 80 86 87 89 90 100 110 120)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
if(ENABLE_CUDA_DEBUG)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (may significantly affect performance on some targets)
else()
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -lineinfo") # add line information to all builds for debug tools (exclusive to -G option)
endif()
# Include directories and libraries
include_directories(../../../Common)
# Source file
# Add target for vectorAddMMAP
if(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
message(STATUS "Will not build sample vectorAddMMAP - not supported on aarch64")
else()
add_executable(vectorAddMMAP vectorAddMMAP.cpp multidevicealloc_memmap.cpp)
target_compile_options(vectorAddMMAP PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
target_compile_features(vectorAddMMAP PRIVATE cxx_std_17 cuda_std_17)
set_target_properties(vectorAddMMAP PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
target_include_directories(vectorAddMMAP PRIVATE
${CUDAToolkit_INCLUDE_DIRS}
)
target_link_libraries(vectorAddMMAP PUBLIC
CUDA::cuda_driver
)
set(CUDA_FATBIN_FILE "${CMAKE_CURRENT_BINARY_DIR}/vectorAdd_kernel64.fatbin")
set(CUDA_KERNEL_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/vectorAdd_kernel.cu")
# Construct GENCODE_FLAGS explicitly from CUDA architectures
set(GENCODE_FLAGS "")
foreach(arch ${CMAKE_CUDA_ARCHITECTURES})
list(APPEND GENCODE_FLAGS "-gencode=arch=compute_${arch},code=sm_${arch}")
endforeach()
add_custom_command(
OUTPUT ${CUDA_FATBIN_FILE}
COMMAND ${CMAKE_CUDA_COMPILER} ${INCLUDES} ${ALL_CCFLAGS} -Wno-deprecated-gpu-targets ${GENCODE_FLAGS} -o ${CUDA_FATBIN_FILE} -fatbin ${CUDA_KERNEL_SOURCE}
DEPENDS ${CUDA_KERNEL_SOURCE}
COMMENT "Building CUDA fatbin: ${CUDA_FATBIN_FILE}"
)
# Create a dummy target for fatbin generation
add_custom_target(generate_fatbin_vectorAddMMAP ALL DEPENDS ${CUDA_FATBIN_FILE})
# Ensure matrixMulDrv depends on the fatbin
add_dependencies(vectorAddMMAP generate_fatbin_vectorAddMMAP)
endif()
# Include installation configuration
include(${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/InstallSamples.cmake)
setup_samples_install()

32
cpp/0_Introduction/vectorAddMMAP/README.md Normal file
View File

@@ -0,0 +1,32 @@
# vectorAddMMAP - Vector Addition cuMemMap
## Description
This sample replaces the device allocation in the vectorAddDrv sample with cuMemMap-ed allocations. This sample demonstrates that the cuMemMap api allows the user to specify the physical properties of their memory while retaining the contiguous nature of their access, thus not requiring a change in their program structure.
## Key Concepts
CUDA Driver API, Vector Addition, MMAP
## Supported SM Architectures
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
## Supported OSes
Linux, Windows
## Supported CPU Architecture
x86_64, ppc64le
## CUDA APIs involved
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
cuMemcpyDtoH, cuDeviceCanAccessPeer, cuModuleGetFunction, cuMemSetAccess, cuMemRelease, cuInit, cuMemcpyHtoD, cuLaunchKernel, cuMemCreate, cuModuleLoadData, cuCtxDestroy, cuDeviceGetCount, cuMemMap, cuDeviceGetAttribute, cuMemGetAllocationGranularity, cuMemAddressFree, cuMemUnmap, cuCtxCreate, cuMemAddressReserve
## Prerequisites
Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## References (for more details)

200
cpp/0_Introduction/vectorAddMMAP/multidevicealloc_memmap.cpp Normal file
View File

@@ -0,0 +1,200 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "multidevicealloc_memmap.hpp"
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;
// 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 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;
}
// 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);
}
}
return status;
}
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;
}
return status;
}

79
cpp/0_Introduction/vectorAddMMAP/multidevicealloc_memmap.hpp Normal file
View File

@@ -0,0 +1,79 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once
#include <cuda.h>
#include <vector>
////////////////////////////////////////////////////////////////////////////
//! Allocate virtually contiguous memory backed on separate devices
//! @return CUresult error code on failure.
//! @param[out] dptr Virtual address reserved for allocation
//! @param[out] allocationSize Actual amount of virtual address space reserved.
//! AllocationSize is needed in the free operation.
//! @param[in] size The minimum size to allocate (will be rounded up
//! to accomodate
//! required granularity).
//! @param[in] residentDevices Specifies what devices the allocation should be
//! striped across.
//! @param[in] mappingDevices Specifies what devices need to read/write to the
//! allocation.
//! @align Additional allignment requirement if desired.
//! @note The VA mappings will look like the following:
//!
//! v-stripeSize-v v-rounding -v
//! +-----------------------------------------+
//! | D1 | D2 | D3 |
//! +-----------------------------------------+
//! ^-- dptr ^-- dptr + size
//!
//! Each device in the residentDevices list will get an equal sized stripe.
//! Excess memory allocated will be that meets the minimum
//! granularity requirements of all the devices.
//!
//! @note uses cuMemGetAllocationGranularity cuMemCreate cuMemMap and
//! cuMemSetAccess
//! function calls to organize the va space
//!
//! @note uses cuMemRelease to release the allocationHandle. The allocation
//! 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);
////////////////////////////////////////////////////////////////////////////
//! Frees resources allocated by simpleMallocMultiDeviceMmap
//! @CUresult CUresult error code on failure.
//! @param[in] dptr Virtual address reserved by simpleMallocMultiDeviceMmap
//! @param[in] size allocationSize returned by simpleMallocMultiDeviceMmap
////////////////////////////////////////////////////////////////////////////
CUresult simpleFreeMultiDeviceMmap(CUdeviceptr dptr, size_t size);

261
cpp/0_Introduction/vectorAddMMAP/vectorAddMMAP.cpp Normal file
View File

@@ -0,0 +1,261 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
/* Vector addition: C = A + B.
*
* This sample replaces the device allocation in the vectorAddDrvsample with
* cuMemMap-ed allocations. This sample demonstrates that the cuMemMap api
* allows the user to specify the physical properties of their memory while
* retaining the contiguos nature of their access, thus not requiring a change
* in their program structure.
*
*/
// Includes
#include <cstring>
#include <cuda.h>
#include <iostream>
#include <stdio.h>
#include <string.h>
// includes, project
#include <helper_cuda_drvapi.h>
#include <helper_functions.h>
// includes, CUDA
#include <builtin_types.h>
#include "multidevicealloc_memmap.hpp"
using namespace std;
// Variables
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;
// Functions
int CleanupNoFailure();
void RandomInit(float *, int);
// 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;
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;
// 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);
}
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;
// Initialize
checkCudaErrors(cuInit(0));
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);
}
// 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);
CUctxCreateParams ctxCreateParams = {};
// Create context
checkCudaErrors(cuCtxCreate(&cuContext, &ctxCreateParams, 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)) {
exit(EXIT_FAILURE);
}
else {
printf("> initCUDA loading module: <%s>\n", module_path.c_str());
}
if (!fatbin.str().size()) {
printf("fatbin file empty. exiting..\n");
exit(EXIT_FAILURE);
}
// Create module from binary file (FATBIN)
checkCudaErrors(cuModuleLoadData(&cuModule, fatbin.str().c_str()));
// Get function handle from module
checkCudaErrors(cuModuleGetFunction(&vecAdd_kernel, cuModule, "VecAdd_kernel"));
// Allocate input vectors h_A and h_B in host memory
h_A = (float *)malloc(size);
h_B = (float *)malloc(size);
h_C = (float *)malloc(size);
// Initialize input vectors
RandomInit(h_A, N);
RandomInit(h_B, N);
// Allocate vectors in device memory
// note that a call to cuCtxEnablePeerAccess is not needed even though
// the backing devices and mapping device are not the same.
// This is because the cuMemSetAccess call explicitly specifies
// the cross device mapping.
// cuMemSetAccess is still subject to the constraints of cuDeviceCanAccessPeer
// for cross device mappings (hence why we checked cuDeviceCanAccessPeer earlier).
checkCudaErrors(simpleMallocMultiDeviceMmap(&d_A, &allocationSize, size, backingDevices, mappingDevices));
checkCudaErrors(simpleMallocMultiDeviceMmap(&d_B, NULL, size, backingDevices, mappingDevices));
checkCudaErrors(simpleMallocMultiDeviceMmap(&d_C, NULL, size, backingDevices, mappingDevices));
// Copy vectors from host memory to device memory
checkCudaErrors(cuMemcpyHtoD(d_A, h_A, size));
checkCudaErrors(cuMemcpyHtoD(d_B, h_B, size));
// This is the new CUDA 4.0 API for Kernel Parameter Passing and Kernel Launch (simpler method)
// Grid/Block configuration
int threadsPerBlock = 256;
int blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
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));
// Copy result from device memory to host memory
// h_C contains the result in host memory
checkCudaErrors(cuMemcpyDtoH(h_C, d_C, size));
// Verify result
int i;
for (i = 0; i < N; ++i) {
float sum = h_A[i] + h_B[i];
if (fabs(h_C[i] - sum) > 1e-7f) {
break;
}
}
CleanupNoFailure();
printf("%s\n", (i == N) ? "Result = PASS" : "Result = FAIL");
exit((i == N) ? EXIT_SUCCESS : EXIT_FAILURE);
}
int CleanupNoFailure()
{
// Free device memory
checkCudaErrors(simpleFreeMultiDeviceMmap(d_A, allocationSize));
checkCudaErrors(simpleFreeMultiDeviceMmap(d_B, allocationSize));
checkCudaErrors(simpleFreeMultiDeviceMmap(d_C, allocationSize));
// Free host memory
if (h_A) {
free(h_A);
}
if (h_B) {
free(h_B);
}
if (h_C) {
free(h_C);
}
checkCudaErrors(cuModuleUnload(cuModule));
checkCudaErrors(cuCtxDestroy(cuContext));
return EXIT_SUCCESS;
}
// Allocates an array with random float entries.
void RandomInit(float *data, int n)
{
for (int i = 0; i < n; ++i) {
data[i] = rand() / (float)RAND_MAX;
}
}

43
cpp/0_Introduction/vectorAddMMAP/vectorAdd_kernel.cu Normal file
View File

@@ -0,0 +1,43 @@
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of NVIDIA CORPORATION nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
/* Vector addition: C = A + B.
*
* This sample is a very basic sample that implements element by element
* vector addition. It is the same as the sample illustrating Chapter 3
* of the programming guide with some additions like error checking.
*
*/
// 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;
if (i < N)
C[i] = A[i] + B[i];
}