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

14
Samples/0_Introduction/matrixMulDrv/matrixMul.h
View File

@@ -30,11 +30,11 @@
// Matrix dimensions
// (chosen as multiples of the thread block size for simplicity)
#define WA (4 * block_size) // Matrix A width
#define HA (6 * block_size) // Matrix A height
#define WB (4 * block_size) // Matrix B width
#define HB WA // Matrix B height
#define WC WB // Matrix C width
#define HC HA // Matrix C height
#define WA (4 * block_size) // Matrix A width
#define HA (6 * block_size) // Matrix A height
#define WB (4 * block_size) // Matrix B width
#define HB WA // Matrix B height
#define WC WB // Matrix C width
#define HC HA // Matrix C height
#endif // _MATRIXMUL_H_
#endif // _MATRIXMUL_H_

429
Samples/0_Introduction/matrixMulDrv/matrixMulDrv.cpp
View File

@@ -46,23 +46,23 @@
// includes, system
#include <builtin_types.h>
#include <math.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <iostream>
#include <cstring>
#include <iostream>
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
// includes, project, CUDA
#include <cstring>
#include <cuda.h>
#include <helper_cuda_drvapi.h>
#include <helper_image.h>
#include <helper_string.h>
#include <helper_timer.h>
#include <cstring>
#include <iostream>
#include <string>
#include "matrixMul.h"
@@ -71,11 +71,9 @@
void runTest(int argc, char **argv);
void randomInit(float *, int);
extern "C" void computeGold(float *, const float *, const float *, unsigned int,
unsigned int, unsigned int);
extern "C" void computeGold(float *, const float *, const float *, unsigned int, unsigned int, unsigned int);
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul,
int *blk_size);
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul, int *blk_size);
#ifndef FATBIN_FILE
#define FATBIN_FILE "matrixMul_kernel64.fatbin"
@@ -84,237 +82,252 @@ static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul,
////////////////////////////////////////////////////////////////////////////////
// Globals
////////////////////////////////////////////////////////////////////////////////
CUdevice cuDevice;
CUdevice cuDevice;
CUcontext cuContext;
CUmodule cuModule;
size_t totalGlobalMem;
CUmodule cuModule;
size_t totalGlobalMem;
const char *sSDKsample = "matrixMulDrv (Driver API)";
void constantInit(float *data, int size, float val) {
for (int i = 0; i < size; ++i) {
data[i] = val;
}
void constantInit(float *data, int size, float val)
{
for (int i = 0; i < size; ++i) {
data[i] = val;
}
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv) {
printf("[ %s ]\n", sSDKsample);
int main(int argc, char **argv)
{
printf("[ %s ]\n", sSDKsample);
runTest(argc, argv);
runTest(argc, argv);
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test for CUDA
////////////////////////////////////////////////////////////////////////////////
void runTest(int argc, char **argv) {
// initialize CUDA
CUfunction matrixMul = NULL;
int block_size = 0;
void runTest(int argc, char **argv)
{
// initialize CUDA
CUfunction matrixMul = NULL;
int block_size = 0;
initCUDA(argc, argv, &matrixMul, &block_size);
initCUDA(argc, argv, &matrixMul, &block_size);
// set seed for rand()
srand(2006);
// set seed for rand()
srand(2006);
// allocate host memory for matrices A and B
unsigned int size_A = WA * HA;
unsigned int mem_size_A = sizeof(float) * size_A;
float *h_A = reinterpret_cast<float *>(malloc(mem_size_A));
unsigned int size_B = WB * HB;
unsigned int mem_size_B = sizeof(float) * size_B;
float *h_B = reinterpret_cast<float *>(malloc(mem_size_B));
// allocate host memory for matrices A and B
unsigned int size_A = WA * HA;
unsigned int mem_size_A = sizeof(float) * size_A;
float *h_A = reinterpret_cast<float *>(malloc(mem_size_A));
unsigned int size_B = WB * HB;
unsigned int mem_size_B = sizeof(float) * size_B;
float *h_B = reinterpret_cast<float *>(malloc(mem_size_B));
// initialize host memory
const float valB = 0.01f;
constantInit(h_A, size_A, 1.0f);
constantInit(h_B, size_B, valB);
// initialize host memory
const float valB = 0.01f;
constantInit(h_A, size_A, 1.0f);
constantInit(h_B, size_B, valB);
// allocate device memory
CUdeviceptr d_A;
checkCudaErrors(cuMemAlloc(&d_A, mem_size_A));
CUdeviceptr d_B;
checkCudaErrors(cuMemAlloc(&d_B, mem_size_B));
// allocate device memory
CUdeviceptr d_A;
checkCudaErrors(cuMemAlloc(&d_A, mem_size_A));
CUdeviceptr d_B;
checkCudaErrors(cuMemAlloc(&d_B, mem_size_B));
// copy host memory to device
checkCudaErrors(cuMemcpyHtoD(d_A, h_A, mem_size_A));
checkCudaErrors(cuMemcpyHtoD(d_B, h_B, mem_size_B));
// copy host memory to device
checkCudaErrors(cuMemcpyHtoD(d_A, h_A, mem_size_A));
checkCudaErrors(cuMemcpyHtoD(d_B, h_B, mem_size_B));
// allocate device memory for result
size_t size_C = WC * HC;
size_t mem_size_C = sizeof(float) * size_C;
// allocate device memory for result
size_t size_C = WC * HC;
size_t mem_size_C = sizeof(float) * size_C;
CUdeviceptr d_C;
checkCudaErrors(cuMemAlloc(&d_C, mem_size_C));
CUdeviceptr d_C;
checkCudaErrors(cuMemAlloc(&d_C, mem_size_C));
// allocate mem for the result on host side
float *h_C = reinterpret_cast<float *>(malloc(mem_size_C));
// allocate mem for the result on host side
float *h_C = reinterpret_cast<float *>(malloc(mem_size_C));
// create and start timer
StopWatchInterface *timer = NULL;
sdkCreateTimer(&timer);
// create and start timer
StopWatchInterface *timer = NULL;
sdkCreateTimer(&timer);
// start the timer
sdkStartTimer(&timer);
// start the timer
sdkStartTimer(&timer);
// There are two ways to launch CUDA kernels via the Driver API.
// In this CUDA Sample, we illustrate both ways to pass parameters
// and specify parameters. By default we use the simpler method.
dim3 block(block_size, block_size, 1);
dim3 grid(WC / block_size, HC / block_size, 1);
// There are two ways to launch CUDA kernels via the Driver API.
// In this CUDA Sample, we illustrate both ways to pass parameters
// and specify parameters. By default we use the simpler method.
dim3 block(block_size, block_size, 1);
dim3 grid(WC / block_size, HC / block_size, 1);
if (1) {
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
// Launching (simplier method)
size_t Matrix_Width_A = (size_t)WA;
size_t Matrix_Width_B = (size_t)WB;
void *args[5] = {&d_C, &d_A, &d_B, &Matrix_Width_A, &Matrix_Width_B};
// new CUDA 4.0 Driver API Kernel launch call
checkCudaErrors(cuLaunchKernel(
matrixMul, grid.x, grid.y, grid.z, block.x, block.y, block.z,
2 * block_size * block_size * sizeof(float), NULL, args, NULL));
} else {
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
// Launching (advanced method)
int offset = 0;
char argBuffer[256];
// pass in launch parameters (not actually de-referencing CUdeviceptr).
// CUdeviceptr is storing the value of the parameters
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_C;
offset += sizeof(d_C);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_A;
offset += sizeof(d_A);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_B;
offset += sizeof(d_B);
size_t Matrix_Width_A = (size_t)WA;
size_t Matrix_Width_B = (size_t)WB;
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_A;
offset += sizeof(Matrix_Width_A);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_B;
offset += sizeof(Matrix_Width_B);
void *kernel_launch_config[5] = {CU_LAUNCH_PARAM_BUFFER_POINTER, argBuffer,
CU_LAUNCH_PARAM_BUFFER_SIZE, &offset,
CU_LAUNCH_PARAM_END};
// new CUDA 4.0 Driver API Kernel launch call
checkCudaErrors(cuLaunchKernel(
matrixMul, grid.x, grid.y, grid.z, block.x, block.y, block.z,
2 * block_size * block_size * sizeof(float), NULL, NULL,
reinterpret_cast<void **>(&kernel_launch_config)));
}
// copy result from device to host
checkCudaErrors(cuMemcpyDtoH(reinterpret_cast<void *>(h_C), d_C, mem_size_C));
// stop and destroy timer
sdkStopTimer(&timer);
printf("Processing time: %f (ms)\n", sdkGetTimerValue(&timer));
sdkDeleteTimer(&timer);
printf("Checking computed result for correctness: ");
bool correct = true;
for (int i = 0; i < static_cast<int>(WC * HC); i++) {
if (fabs(h_C[i] - (WA * valB)) > 1e-5) {
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > 1e-5\n", i,
h_C[i], WA * valB);
correct = false;
if (1) {
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
// Launching (simplier method)
size_t Matrix_Width_A = (size_t)WA;
size_t Matrix_Width_B = (size_t)WB;
void *args[5] = {&d_C, &d_A, &d_B, &Matrix_Width_A, &Matrix_Width_B};
// new CUDA 4.0 Driver API Kernel launch call
checkCudaErrors(cuLaunchKernel(matrixMul,
grid.x,
grid.y,
grid.z,
block.x,
block.y,
block.z,
2 * block_size * block_size * sizeof(float),
NULL,
args,
NULL));
}
}
else {
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
// Launching (advanced method)
int offset = 0;
char argBuffer[256];
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
// pass in launch parameters (not actually de-referencing CUdeviceptr).
// CUdeviceptr is storing the value of the parameters
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_C;
offset += sizeof(d_C);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_A;
offset += sizeof(d_A);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_B;
offset += sizeof(d_B);
printf("\nNOTE: The CUDA Samples are not meant for performance measurements. "
"Results may vary when GPU Boost is enabled.\n");
size_t Matrix_Width_A = (size_t)WA;
size_t Matrix_Width_B = (size_t)WB;
// clean up memory
free(h_A);
free(h_B);
free(h_C);
checkCudaErrors(cuMemFree(d_A));
checkCudaErrors(cuMemFree(d_B));
checkCudaErrors(cuMemFree(d_C));
checkCudaErrors(cuCtxDestroy(cuContext));
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_A;
offset += sizeof(Matrix_Width_A);
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_B;
offset += sizeof(Matrix_Width_B);
void *kernel_launch_config[5] = {
CU_LAUNCH_PARAM_BUFFER_POINTER, argBuffer, CU_LAUNCH_PARAM_BUFFER_SIZE, &offset, CU_LAUNCH_PARAM_END};
// new CUDA 4.0 Driver API Kernel launch call
checkCudaErrors(cuLaunchKernel(matrixMul,
grid.x,
grid.y,
grid.z,
block.x,
block.y,
block.z,
2 * block_size * block_size * sizeof(float),
NULL,
NULL,
reinterpret_cast<void **>(&kernel_launch_config)));
}
// copy result from device to host
checkCudaErrors(cuMemcpyDtoH(reinterpret_cast<void *>(h_C), d_C, mem_size_C));
// stop and destroy timer
sdkStopTimer(&timer);
printf("Processing time: %f (ms)\n", sdkGetTimerValue(&timer));
sdkDeleteTimer(&timer);
printf("Checking computed result for correctness: ");
bool correct = true;
for (int i = 0; i < static_cast<int>(WC * HC); i++) {
if (fabs(h_C[i] - (WA * valB)) > 1e-5) {
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > 1e-5\n", i, h_C[i], WA * valB);
correct = false;
}
}
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
printf("\nNOTE: The CUDA Samples are not meant for performance measurements. "
"Results may vary when GPU Boost is enabled.\n");
// clean up memory
free(h_A);
free(h_B);
free(h_C);
checkCudaErrors(cuMemFree(d_A));
checkCudaErrors(cuMemFree(d_B));
checkCudaErrors(cuMemFree(d_C));
checkCudaErrors(cuCtxDestroy(cuContext));
}
// Allocates a matrix with random float entries.
void randomInit(float *data, int size) {
for (int i = 0; i < size; ++i) {
data[i] = rand() / static_cast<float>(RAND_MAX);
}
}
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul,
int *blk_size) {
CUfunction cuFunction = 0;
int major = 0, minor = 0;
char deviceName[100];
cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
// get compute capabilities and the devicename
checkCudaErrors(cuDeviceGetAttribute(
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice));
checkCudaErrors(cuDeviceGetAttribute(
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice));
checkCudaErrors(cuDeviceGetName(deviceName, sizeof(deviceName), cuDevice));
printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
checkCudaErrors(cuDeviceTotalMem(&totalGlobalMem, cuDevice));
printf(" Total amount of global memory: %llu bytes\n",
(long long unsigned int)totalGlobalMem);
checkCudaErrors(cuCtxCreate(&cuContext, 0, cuDevice));
// first search for the module path before we load the results
std::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()));
// select the suitable kernel function
const char *kernels[] = {"matrixMul_bs32_64bit", "matrixMul_bs16_64bit",
"matrixMul_bs8_64bit"};
int idx = 0;
int block_size = 32;
while (idx < 3) {
int threadsPerBlock = 0;
int blocksPerGrid = 0;
checkCudaErrors(cuModuleGetFunction(&cuFunction, cuModule, kernels[idx]));
checkCudaErrors(cuOccupancyMaxPotentialBlockSize(
&blocksPerGrid, &threadsPerBlock, cuFunction, 0,
2 * block_size * block_size * sizeof(float), 0));
if (block_size * block_size <= threadsPerBlock) {
printf("> %d block size selected\n", block_size);
break;
} else {
block_size /= 2;
void randomInit(float *data, int size)
{
for (int i = 0; i < size; ++i) {
data[i] = rand() / static_cast<float>(RAND_MAX);
}
idx++;
}
*pMatrixMul = cuFunction;
*blk_size = block_size;
return 0;
}
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul, int *blk_size)
{
CUfunction cuFunction = 0;
int major = 0, minor = 0;
char deviceName[100];
cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
// get compute capabilities and the devicename
checkCudaErrors(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice));
checkCudaErrors(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice));
checkCudaErrors(cuDeviceGetName(deviceName, sizeof(deviceName), cuDevice));
printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
checkCudaErrors(cuDeviceTotalMem(&totalGlobalMem, cuDevice));
printf(" Total amount of global memory: %llu bytes\n", (long long unsigned int)totalGlobalMem);
checkCudaErrors(cuCtxCreate(&cuContext, 0, cuDevice));
// first search for the module path before we load the results
std::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()));
// select the suitable kernel function
const char *kernels[] = {"matrixMul_bs32_64bit", "matrixMul_bs16_64bit", "matrixMul_bs8_64bit"};
int idx = 0;
int block_size = 32;
while (idx < 3) {
int threadsPerBlock = 0;
int blocksPerGrid = 0;
checkCudaErrors(cuModuleGetFunction(&cuFunction, cuModule, kernels[idx]));
checkCudaErrors(cuOccupancyMaxPotentialBlockSize(
&blocksPerGrid, &threadsPerBlock, cuFunction, 0, 2 * block_size * block_size * sizeof(float), 0));
if (block_size * block_size <= threadsPerBlock) {
printf("> %d block size selected\n", block_size);
break;
}
else {
block_size /= 2;
}
idx++;
}
*pMatrixMul = cuFunction;
*blk_size = block_size;
return 0;
}

121
Samples/0_Introduction/matrixMulDrv/matrixMul_kernel.cu
View File

@@ -42,86 +42,87 @@
//! wA is A's width and wB is B's width
////////////////////////////////////////////////////////////////////////////////
template <int block_size, typename size_type>
__device__ void matrixMul(float *C, float *A, float *B, size_type wA,
size_type wB) {
// Block index
size_type bx = blockIdx.x;
size_type by = blockIdx.y;
__device__ void matrixMul(float *C, float *A, float *B, size_type wA, size_type wB)
{
// Block index
size_type bx = blockIdx.x;
size_type by = blockIdx.y;
// Thread index
size_type tx = threadIdx.x;
size_type ty = threadIdx.y;
// Thread index
size_type tx = threadIdx.x;
size_type ty = threadIdx.y;
// Index of the first sub-matrix of A processed by the block
size_type aBegin = wA * block_size * by;
// Index of the first sub-matrix of A processed by the block
size_type aBegin = wA * block_size * by;
// Index of the last sub-matrix of A processed by the block
size_type aEnd = aBegin + wA - 1;
// Index of the last sub-matrix of A processed by the block
size_type aEnd = aBegin + wA - 1;
// Step size used to iterate through the sub-matrices of A
size_type aStep = block_size;
// Step size used to iterate through the sub-matrices of A
size_type aStep = block_size;
// Index of the first sub-matrix of B processed by the block
size_type bBegin = block_size * bx;
// Index of the first sub-matrix of B processed by the block
size_type bBegin = block_size * bx;
// Step size used to iterate through the sub-matrices of B
size_type bStep = block_size * wB;
// Step size used to iterate through the sub-matrices of B
size_type bStep = block_size * wB;
// Csub is used to store the element of the block sub-matrix
// that is computed by the thread
float Csub = 0;
// Csub is used to store the element of the block sub-matrix
// that is computed by the thread
float Csub = 0;
// Loop over all the sub-matrices of A and B
// required to compute the block sub-matrix
for (size_type a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
// Declaration of the shared memory array As used to
// store the sub-matrix of A
__shared__ float As[block_size][block_size];
// Loop over all the sub-matrices of A and B
// required to compute the block sub-matrix
for (size_type a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
// Declaration of the shared memory array As used to
// store the sub-matrix of A
__shared__ float As[block_size][block_size];
// Declaration of the shared memory array Bs used to
// store the sub-matrix of B
__shared__ float Bs[block_size][block_size];
// Declaration of the shared memory array Bs used to
// store the sub-matrix of B
__shared__ float Bs[block_size][block_size];
// Load the matrices from device memory
// to shared memory; each thread loads
// one element of each matrix
AS(ty, tx) = A[a + wA * ty + tx];
BS(ty, tx) = B[b + wB * ty + tx];
// Load the matrices from device memory
// to shared memory; each thread loads
// one element of each matrix
AS(ty, tx) = A[a + wA * ty + tx];
BS(ty, tx) = B[b + wB * ty + tx];
// Synchronize to make sure the matrices are loaded
__syncthreads();
// Synchronize to make sure the matrices are loaded
__syncthreads();
// Multiply the two matrices together;
// each thread computes one element
// of the block sub-matrix
// Multiply the two matrices together;
// each thread computes one element
// of the block sub-matrix
#pragma unroll
for (size_type k = 0; k < block_size; ++k) Csub += AS(ty, k) * BS(k, tx);
for (size_type k = 0; k < block_size; ++k)
Csub += AS(ty, k) * BS(k, tx);
// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
__syncthreads();
}
// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
__syncthreads();
}
// Write the block sub-matrix to device memory;
// each thread writes one element
size_type c = wB * block_size * by + block_size * bx;
C[c + wB * ty + tx] = Csub;
// Write the block sub-matrix to device memory;
// each thread writes one element
size_type c = wB * block_size * by + block_size * bx;
C[c + wB * ty + tx] = Csub;
}
// C wrappers around our template kernel
extern "C" __global__ void matrixMul_bs8_64bit(float *C, float *A, float *B,
size_t wA, size_t wB) {
matrixMul<8, size_t>(C, A, B, wA, wB);
extern "C" __global__ void matrixMul_bs8_64bit(float *C, float *A, float *B, size_t wA, size_t wB)
{
matrixMul<8, size_t>(C, A, B, wA, wB);
}
extern "C" __global__ void matrixMul_bs16_64bit(float *C, float *A, float *B,
size_t wA, size_t wB) {
matrixMul<16, size_t>(C, A, B, wA, wB);
extern "C" __global__ void matrixMul_bs16_64bit(float *C, float *A, float *B, size_t wA, size_t wB)
{
matrixMul<16, size_t>(C, A, B, wA, wB);
}
extern "C" __global__ void matrixMul_bs32_64bit(float *C, float *A, float *B,
size_t wA, size_t wB) {
matrixMul<32, size_t>(C, A, B, wA, wB);
extern "C" __global__ void matrixMul_bs32_64bit(float *C, float *A, float *B, size_t wA, size_t wB)
{
matrixMul<32, size_t>(C, A, B, wA, wB);
}
#endif // #ifndef _MATRIXMUL_KERNEL_H_
#endif // #ifndef _MATRIXMUL_KERNEL_H_