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

131
Samples/0_Introduction/simpleMPI/simpleMPI.cpp
View File

@@ -26,15 +26,15 @@
*/
/* Simple example demonstrating how to use MPI with CUDA
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.cpp: main program, compiled with mpicxx on linux/Mac platforms
* on Windows, please download the Microsoft HPC Pack SDK 2008
*/
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.cpp: main program, compiled with mpicxx on linux/Mac platforms
* on Windows, please download the Microsoft HPC Pack SDK 2008
*/
// MPI include
#include <mpi.h>
@@ -42,87 +42,88 @@
// System includes
#include <iostream>
using std::cout;
using std::cerr;
using std::cout;
using std::endl;
// User include
#include "simpleMPI.h"
// Error handling macros
#define MPI_CHECK(call) \
if ((call) != MPI_SUCCESS) { \
cerr << "MPI error calling \"" #call "\"\n"; \
my_abort(-1); \
}
#define MPI_CHECK(call) \
if ((call) != MPI_SUCCESS) { \
cerr << "MPI error calling \"" #call "\"\n"; \
my_abort(-1); \
}
// Host code
// No CUDA here, only MPI
int main(int argc, char *argv[]) {
// Dimensions of the dataset
int blockSize = 256;
int gridSize = 10000;
int dataSizePerNode = gridSize * blockSize;
int main(int argc, char *argv[])
{
// Dimensions of the dataset
int blockSize = 256;
int gridSize = 10000;
int dataSizePerNode = gridSize * blockSize;
// Initialize MPI state
MPI_CHECK(MPI_Init(&argc, &argv));
// Initialize MPI state
MPI_CHECK(MPI_Init(&argc, &argv));
// Get our MPI node number and node count
int commSize, commRank;
MPI_CHECK(MPI_Comm_size(MPI_COMM_WORLD, &commSize));
MPI_CHECK(MPI_Comm_rank(MPI_COMM_WORLD, &commRank));
// Get our MPI node number and node count
int commSize, commRank;
MPI_CHECK(MPI_Comm_size(MPI_COMM_WORLD, &commSize));
MPI_CHECK(MPI_Comm_rank(MPI_COMM_WORLD, &commRank));
// Generate some random numbers on the root node (node 0)
int dataSizeTotal = dataSizePerNode * commSize;
float *dataRoot = NULL;
// Generate some random numbers on the root node (node 0)
int dataSizeTotal = dataSizePerNode * commSize;
float *dataRoot = NULL;
// Are we the root node?
if (commRank == 0) {
cout << "Running on " << commSize << " nodes" << endl;
dataRoot = new float[dataSizeTotal];
initData(dataRoot, dataSizeTotal);
}
// Are we the root node?
if (commRank == 0) {
cout << "Running on " << commSize << " nodes" << endl;
dataRoot = new float[dataSizeTotal];
initData(dataRoot, dataSizeTotal);
}
// Allocate a buffer on each node
float *dataNode = new float[dataSizePerNode];
// Allocate a buffer on each node
float *dataNode = new float[dataSizePerNode];
// Dispatch a portion of the input data to each node
MPI_CHECK(MPI_Scatter(dataRoot, dataSizePerNode, MPI_FLOAT, dataNode,
dataSizePerNode, MPI_FLOAT, 0, MPI_COMM_WORLD));
// Dispatch a portion of the input data to each node
MPI_CHECK(
MPI_Scatter(dataRoot, dataSizePerNode, MPI_FLOAT, dataNode, dataSizePerNode, MPI_FLOAT, 0, MPI_COMM_WORLD));
if (commRank == 0) {
// No need for root data any more
delete[] dataRoot;
}
if (commRank == 0) {
// No need for root data any more
delete[] dataRoot;
}
// On each node, run computation on GPU
computeGPU(dataNode, blockSize, gridSize);
// On each node, run computation on GPU
computeGPU(dataNode, blockSize, gridSize);
// Reduction to the root node, computing the sum of output elements
float sumNode = sum(dataNode, dataSizePerNode);
float sumRoot;
// Reduction to the root node, computing the sum of output elements
float sumNode = sum(dataNode, dataSizePerNode);
float sumRoot;
MPI_CHECK(
MPI_Reduce(&sumNode, &sumRoot, 1, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD));
MPI_CHECK(MPI_Reduce(&sumNode, &sumRoot, 1, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD));
if (commRank == 0) {
float average = sumRoot / dataSizeTotal;
cout << "Average of square roots is: " << average << endl;
}
if (commRank == 0) {
float average = sumRoot / dataSizeTotal;
cout << "Average of square roots is: " << average << endl;
}
// Cleanup
delete[] dataNode;
MPI_CHECK(MPI_Finalize());
// Cleanup
delete[] dataNode;
MPI_CHECK(MPI_Finalize());
if (commRank == 0) {
cout << "PASSED\n";
}
if (commRank == 0) {
cout << "PASSED\n";
}
return 0;
return 0;
}
// Shut down MPI cleanly if something goes wrong
void my_abort(int err) {
cout << "Test FAILED\n";
MPI_Abort(MPI_COMM_WORLD, err);
void my_abort(int err)
{
cout << "Test FAILED\n";
MPI_Abort(MPI_COMM_WORLD, err);
}

92
Samples/0_Introduction/simpleMPI/simpleMPI.cu
View File

@@ -26,14 +26,14 @@
*/
/* Simple example demonstrating how to use MPI with CUDA
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.cu: GPU part, compiled with nvcc
*/
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.cu: GPU part, compiled with nvcc
*/
#include <iostream>
using std::cerr;
@@ -42,61 +42,63 @@ using std::endl;
#include "simpleMPI.h"
// Error handling macro
#define CUDA_CHECK(call) \
if ((call) != cudaSuccess) { \
cudaError_t err = cudaGetLastError(); \
cerr << "CUDA error calling \"" #call "\", code is " << err << endl; \
my_abort(err); \
}
#define CUDA_CHECK(call) \
if ((call) != cudaSuccess) { \
cudaError_t err = cudaGetLastError(); \
cerr << "CUDA error calling \"" #call "\", code is " << err << endl; \
my_abort(err); \
}
// Device code
// Very simple GPU Kernel that computes square roots of input numbers
__global__ void simpleMPIKernel(float *input, float *output) {
int tid = blockIdx.x * blockDim.x + threadIdx.x;
output[tid] = sqrt(input[tid]);
__global__ void simpleMPIKernel(float *input, float *output)
{
int tid = blockIdx.x * blockDim.x + threadIdx.x;
output[tid] = sqrt(input[tid]);
}
// Initialize an array with random data (between 0 and 1)
void initData(float *data, int dataSize) {
for (int i = 0; i < dataSize; i++) {
data[i] = (float)rand() / RAND_MAX;
}
void initData(float *data, int dataSize)
{
for (int i = 0; i < dataSize; i++) {
data[i] = (float)rand() / RAND_MAX;
}
}
// CUDA computation on each node
// No MPI here, only CUDA
void computeGPU(float *hostData, int blockSize, int gridSize) {
int dataSize = blockSize * gridSize;
void computeGPU(float *hostData, int blockSize, int gridSize)
{
int dataSize = blockSize * gridSize;
// Allocate data on GPU memory
float *deviceInputData = NULL;
CUDA_CHECK(cudaMalloc((void **)&deviceInputData, dataSize * sizeof(float)));
// Allocate data on GPU memory
float *deviceInputData = NULL;
CUDA_CHECK(cudaMalloc((void **)&deviceInputData, dataSize * sizeof(float)));
float *deviceOutputData = NULL;
CUDA_CHECK(cudaMalloc((void **)&deviceOutputData, dataSize * sizeof(float)));
float *deviceOutputData = NULL;
CUDA_CHECK(cudaMalloc((void **)&deviceOutputData, dataSize * sizeof(float)));
// Copy to GPU memory
CUDA_CHECK(cudaMemcpy(deviceInputData, hostData, dataSize * sizeof(float),
cudaMemcpyHostToDevice));
// Copy to GPU memory
CUDA_CHECK(cudaMemcpy(deviceInputData, hostData, dataSize * sizeof(float), cudaMemcpyHostToDevice));
// Run kernel
simpleMPIKernel<<<gridSize, blockSize>>>(deviceInputData, deviceOutputData);
// Run kernel
simpleMPIKernel<<<gridSize, blockSize>>>(deviceInputData, deviceOutputData);
// Copy data back to CPU memory
CUDA_CHECK(cudaMemcpy(hostData, deviceOutputData, dataSize * sizeof(float),
cudaMemcpyDeviceToHost));
// Copy data back to CPU memory
CUDA_CHECK(cudaMemcpy(hostData, deviceOutputData, dataSize * sizeof(float), cudaMemcpyDeviceToHost));
// Free GPU memory
CUDA_CHECK(cudaFree(deviceInputData));
CUDA_CHECK(cudaFree(deviceOutputData));
// Free GPU memory
CUDA_CHECK(cudaFree(deviceInputData));
CUDA_CHECK(cudaFree(deviceOutputData));
}
float sum(float *data, int size) {
float accum = 0.f;
float sum(float *data, int size)
{
float accum = 0.f;
for (int i = 0; i < size; i++) {
accum += data[i];
}
for (int i = 0; i < size; i++) {
accum += data[i];
}
return accum;
return accum;
}

27
Samples/0_Introduction/simpleMPI/simpleMPI.h
View File

@@ -26,19 +26,20 @@
*/
/* Simple example demonstrating how to use MPI with CUDA
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.h: common header file
*/
*
* Generate some random numbers on one node.
* Dispatch them to all nodes.
* Compute their square root on each node's GPU.
* Compute the average of the results using MPI.
*
* simpleMPI.h: common header file
*/
// Forward declarations
extern "C" {
void initData(float *data, int dataSize);
void computeGPU(float *hostData, int blockSize, int gridSize);
float sum(float *data, int size);
void my_abort(int err);
extern "C"
{
void initData(float *data, int dataSize);
void computeGPU(float *hostData, int blockSize, int gridSize);
float sum(float *data, int size);
void my_abort(int err);
}