Related
I receive Segmentation fault (core dumped) when i run this code.
I know the cudaMalloc is the problem, but I have no idea how to solve it. I just started learning some CUDA programming and im not familiar with it. I'm working on wsl if it matters.
#include<stdio.h>
#define SIZE 20
__global__ void VectorAdd(int *a,int *b,int *c,int n){
int i = threadIdx.x;
if(i<n)
c[i]=a[i]+b[i];
}
int main(){
int *a,*b,*c;
cudaMalloc((void**)&a,SIZE *sizeof(int));
cudaMalloc((void**)&b, SIZE *sizeof(int));
cudaMalloc((void**)&c,SIZE *sizeof(int));
for (int i=1;i<SIZE;i++){
a[i]=i;
b[i]=i;
c[i]=0;
}
VectorAdd<<<1, SIZE>>>(a,b,c,SIZE);
cudaDeviceSynchronize();
for(int i=1;i<SIZE;i++){
printf("%d \n",c[i]);
}
cudaFree(a);
cudaFree(b);
cudaFree(c);
return 0;
}
As the comments already suggested, you have to initialize values for arrays a and b on the host, copy them to device array, and once computation is completed you have to copy data from c back to the host.
#include<stdio.h>
#define SIZE 20
__global__ void VectorAdd(int *a,int *b,int *c,int n){
int i = threadIdx.x;
if(i<n)
c[i]=a[i]+b[i];
}
int main(){
int *a,*b,*c;
int *h_a, *h_b, *h_c; /*declare pointers to host arrays*/
cudaMalloc((void**)&a,SIZE *sizeof(int));
cudaMalloc((void**)&b, SIZE *sizeof(int));
cudaMalloc((void**)&c,SIZE *sizeof(int));
/* allocate memory for host arrays */
h_a = new int[SIZE];
h_b = new int[SIZE];
h_c = new int[SIZE];
/* initialize values on host arrays */
for (int i = 0; i < SIZE; i++){
h_a[i]=i;
h_b[i]=i;
}
/*copy data from host to device */
cudaMemcpy(a, h_a, SIZE*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(b, h_b, SIZE*sizeof(int), cudaMemcpyHostToDevice);
VectorAdd<<<1, SIZE>>>(a,b,c,SIZE);
// cudaDeviceSynchronize(); /* this is not needed because cudaMemcpy implies sync. */
/*copy results from device to host*/
cudaMemcpy(h_c, c, SIZE*sizeof(int), cudaMemcpyDeviceToHost);
for(int i = 0; i < SIZE; i++){
printf("%d \n",h_c[i]);
}
cudaFree(a);
cudaFree(b);
cudaFree(c);
/* free host memory */
delete [] h_a;
delete [] h_b;
delete [] h_c;
return 0;
}
Notes
For some reason you start iterating from position 1 instead 0 in you for loops! If this is wrong by accident I fixed it!
cudaMemcpy always performs synchronization between host and device. So, cudaDeviceSynchronize() is not necessary after kernel invocation.
To avoid explicit handling of separated host and device data, you could use cudaMallocManaged instead of cudaMalloc.
I'm running a toy CUDA sample on my GeForce 1080 Ti (Pascal) on windows 10 and CUDA 9.2.
Goal is to test cudaMemPrefetchAsync to the CPU, as it's supposed to work.
However, I get a CUDA error (invalid device ordinal) on this particular line.
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <cstdio>
#include <cstdlib>
void fill(int* a, int val, int N) {
for (int k = 0; k < N; ++k) {
a[k] = val;
}
}
__global__ void add(int* a, int* b, int N)
{
for (int i = threadIdx.x + blockIdx.x * blockDim.x; i < N; i += blockDim.x * gridDim.x) {
a[i] += b[i];
}
}
inline void check(cudaError_t err, const char* file, int line) {
if (err != cudaSuccess) {
::fprintf(stderr, "ERROR at %s[%d] : %s\n", file, line, cudaGetErrorString(err));
abort();
}
}
#define CUDA_CHECK(err) do { check(err, __FILE__, __LINE__); } while(0)
int main()
{
int deviceId;
CUDA_CHECK(cudaGetDevice(&deviceId));
const int N = 1024*1024*32;
int *a, *b;
CUDA_CHECK(cudaMallocManaged(&a, N * sizeof(int)));
CUDA_CHECK(cudaMallocManaged(&b, N * sizeof(int)));
CUDA_CHECK(cudaMemPrefetchAsync(a, N * sizeof(int), cudaCpuDeviceId)); // program breaks here
CUDA_CHECK(cudaMemPrefetchAsync(b, N * sizeof(int), cudaCpuDeviceId));
fill(a, 1, N);
fill(a, 2, N);
CUDA_CHECK(cudaMemPrefetchAsync(a, N * sizeof(int), deviceId));
CUDA_CHECK(cudaMemPrefetchAsync(b, N * sizeof(int), deviceId));
add<<<32, 256>>>(a, b, N);
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());
return 0;
}
Is that a hardware/driver/OS limitation? Can I simply ignore the error?
Is that a hardware/driver/OS limitation?
Yes, the latter. Quoting from the documentation
GPUs with SM architecture 6.x or higher (Pascal class or newer)
provide additional Unified Memory features such as on-demand page
migration and GPU memory oversubscription that are outlined throughout
this document. Note that currently these features are only supported
on Linux operating systems.
So asynchronous page migration is not supported in Windows at the moment and that it why you get an error when you try to enable it.
I am trying to call cudaMemsetAsync from kernel (so called "dynamic parallelism"). But no matter what value I use, it always set memory to 0.
Here is my test code:
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "cuda_device_runtime_api.h"
#include <stdio.h>
const int size = 5;
__global__ void kernel(int *c)
{
cudaMemsetAsync(c, 0x7FFFFFFF, size * 4, NULL);
}
int main()
{
cudaError_t cudaStatus;
int c[size] = { 12, 12, 12, 12, 12 };
int *dev_c = 0;
cudaStatus = cudaSetDevice(0);
cudaStatus = cudaMalloc((void**)&dev_c, size * sizeof(int));
cudaStatus = cudaMemcpy(dev_c, c, size * sizeof(int), cudaMemcpyHostToDevice);
kernel <<< 1, 1 >>>(dev_c);
cudaStatus = cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dev_c);
cudaStatus = cudaDeviceReset();
printf("%d\n", cudaStatus);
printf("{%d,%d,%d,%d,%d}\n", c[0], c[1], c[2], c[3], c[4]);
return 0;
}
And if I run it, I got output like this:
>nvcc -run kernel.cu -gencode=arch=compute_35,code=\"sm_35,compute_35\" -rdc=true -lcudadevrt
kernel.cu
Creating library a.lib and object a.exp
0
{0,0,0,0,0}
When I call memory set, I use value 0x7FFFFFFF. I'm expecting non-zero numbers, but it always shows zero.
Is this a bug? or I did something wrong? I'm using CUDA 8.0
I can confirm this appears not to work in CUDA 8 on the systems I tested it with.
If you want a single thread to perform the operation, you can use memset directly in device code (it, like memcpy, has been supported forever). The kernel will emit a byte sized loop inline within your kernel and the operation will be handled by each running thread.
If you want a dynamic parallelism style memset operation, then the easiest thing is to make your own. A trivial (and very, very lightly tested) implementation in the code you posted might look like this:
#include <cstring>
#include <cstdio>
const int size = 5;
__global__ void myMemset_kernel(void* p, unsigned char val, size_t sz)
{
size_t tid = threadIdx.x + blockDim.x * blockIdx.x;
unsigned char* _p = (unsigned char*)p;
for(; tid < sz; tid += blockDim.x * gridDim.x) {
_p[tid] = val;
}
}
__device__ void myMemset(void* p, unsigned int val, size_t sz, cudaStream_t s=NULL)
{
const dim3 blocksz(256,1,1);
size_t nblocks = (sz + blocksz.x -1) / blocksz.x;
unsigned charval = val & 0xff;
myMemset_kernel<<< dim3(nblocks,1,1), blocksz, 0, s >>>(p, charval, sz);
}
__global__ void kernel(int *c)
{
cudaStream_t s;
cudaStreamCreateWithFlags(&s, cudaStreamNonBlocking);
myMemset(c, 0x7FFFFFFF, size * 4, s);
cudaDeviceSynchronize();
}
int main()
{
int c[size];
int *dev_c;
memset(&c[0], 0xffffff0c, size * sizeof(int));
printf("{%08x,%08x,%08x,%08x,%08x}\n", c[0], c[1], c[2], c[3], c[4]);
cudaMalloc((void**)&dev_c, size * sizeof(int));
cudaMemcpy(dev_c, c, size * sizeof(int), cudaMemcpyHostToDevice);
kernel <<< 1, 1 >>>(dev_c);
cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dev_c);
printf("{%08x,%08x,%08x,%08x,%08x}\n", c[0], c[1], c[2], c[3], c[4]);
return 0;
}
which compiles and does this:
$ nvcc -rdc=true -arch=sm_52 -o memset memset.cu -lcudadevrt
$ ./memset
{0c0c0c0c,0c0c0c0c,0c0c0c0c,0c0c0c0c,0c0c0c0c}
{ffffffff,ffffffff,ffffffff,ffffffff,ffffffff}
A final point -- note the values above and read this question and answer. In your code, it is not possible to use cudaMemset to apply a value of 0x7FFFFFFF. Although the value argument is an unsigned integer, cudaMemset and its relatives work like regular memset and set byte values. Only the least significant byte of the 32 bit argument is used to set values. If your objective is to set 32 bit values, then you will need to make your own version of memset for that purpose anyway.
While playing with CUBLAS matrix multiplication sample I realised that nvprof profiler shows an extra call of cudaMemcpy Host to Device.
While 2 appear in source code, 3 actual calls are issued.
Why would that be? Is it an intrinsic effect of using CUBLAS?
Code from CUDA CUBLAS sample:
compiled with flags: -lcublas -I/usr/local/cuda-7.5/samples/common/inc
//////////////////////////////////////////////////////////////////////////
// Utilities and system includes
#include <assert.h>
#include <helper_string.h> // helper for shared functions common to CUDA Samples
// CUDA runtime
#include <cuda_runtime.h>
#include <cublas_v2.h>
// CUDA and CUBLAS functions
#include <helper_functions.h>
#include <helper_cuda.h>
#ifndef min
#define min(a,b) ((a < b) ? a : b)
#endif
#ifndef max
#define max(a,b) ((a > b) ? a : b)
#endif
typedef struct _matrixSize // Optional Command-line multiplier for matrix sizes
{
unsigned int uiWA, uiHA, uiWB, uiHB, uiWC, uiHC;
} sMatrixSize;
////////////////////////////////////////////////////////////////////////////////
//! Compute reference data set matrix multiply on CPU
//! C = A * B
//! #param C reference data, computed but preallocated
//! #param A matrix A as provided to device
//! #param B matrix B as provided to device
//! #param hA height of matrix A
//! #param wB width of matrix B
////////////////////////////////////////////////////////////////////////////////
void
matrixMulCPU(float *C, const float *A, const float *B, unsigned int hA, unsigned int wA, unsigned int wB)
{
for (unsigned int i = 0; i < hA; ++i)
for (unsigned int j = 0; j < wB; ++j)
{
double sum = 0;
for (unsigned int k = 0; k < wA; ++k)
{
double a = A[i * wA + k];
double b = B[k * wB + j];
sum += a * b;
}
C[i * wB + j] = (float)sum;
}
}
// Allocates a matrix with random float entries.
void randomInit(float *data, int size)
{
for (int i = 0; i < size; ++i)
data[i] = rand() / (float)RAND_MAX;
}
void printDiff(float *data1, float *data2, int width, int height, int iListLength, float fListTol)
{
printf("Listing first %d Differences > %.6f...\n", iListLength, fListTol);
int i,j,k;
int error_count=0;
for (j = 0; j < height; j++)
{
if (error_count < iListLength)
{
printf("\n Row %d:\n", j);
}
for (i = 0; i < width; i++)
{
k = j * width + i;
float fDiff = fabs(data1[k] - data2[k]);
if (fDiff > fListTol)
{
if (error_count < iListLength)
{
printf(" Loc(%d,%d)\tCPU=%.5f\tGPU=%.5f\tDiff=%.6f\n", i, j, data1[k], data2[k], fDiff);
}
error_count++;
}
}
}
printf(" \n Total Errors = %d\n", error_count);
}
void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMatrixSize &matrix_size)
{
// By default, we use device 0, otherwise we override the device ID based on what is provided at the command line
cudaError_t error;
devID = 0;
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
devID = getCmdLineArgumentInt(argc, (const char **)argv, "device");
error = cudaSetDevice(devID);
if (error != cudaSuccess)
{
printf("cudaSetDevice returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
}
// get number of SMs on this GPU
error = cudaGetDevice(&devID);
if (error != cudaSuccess)
{
printf("cudaGetDevice returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
if (checkCmdLineFlag(argc, (const char **)argv, "sizemult"))
{
iSizeMultiple = getCmdLineArgumentInt(argc, (const char **)argv, "sizemult");
}
iSizeMultiple = min(iSizeMultiple, 10);
iSizeMultiple = max(iSizeMultiple, 1);
cudaDeviceProp deviceProp;
error = cudaGetDeviceProperties(&deviceProp, devID);
if (error != cudaSuccess)
{
printf("cudaGetDeviceProperties returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", devID, deviceProp.name, deviceProp.major, deviceProp.minor);
// use a larger block size for Fermi and above
int block_size = (deviceProp.major < 2) ? 16 : 32;
matrix_size.uiWA = 3 * block_size * iSizeMultiple;
matrix_size.uiHA = 4 * block_size * iSizeMultiple;
matrix_size.uiWB = 2 * block_size * iSizeMultiple;
matrix_size.uiHB = 3 * block_size * iSizeMultiple;
matrix_size.uiWC = 2 * block_size * iSizeMultiple;
matrix_size.uiHC = 4 * block_size * iSizeMultiple;
printf("MatrixA(%u,%u), MatrixB(%u,%u), MatrixC(%u,%u)\n",
matrix_size.uiHA, matrix_size.uiWA,
matrix_size.uiHB, matrix_size.uiWB,
matrix_size.uiHC, matrix_size.uiWC);
if( matrix_size.uiWA != matrix_size.uiHB ||
matrix_size.uiHA != matrix_size.uiHC ||
matrix_size.uiWB != matrix_size.uiWC)
{
printf("ERROR: Matrix sizes do not match!\n");
exit(-1);
}
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test matrix multiply using CUBLAS
////////////////////////////////////////////////////////////////////////////////
int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
{
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
// use a larger block size for Fermi and above
int block_size = (deviceProp.major < 2) ? 16 : 32;
// set seed for rand()
srand(2006);
// allocate host memory for matrices A and B
unsigned int size_A = matrix_size.uiWA * matrix_size.uiHA;
unsigned int mem_size_A = sizeof(float) * size_A;
float *h_A = (float *)malloc(mem_size_A);
unsigned int size_B = matrix_size.uiWB * matrix_size.uiHB;
unsigned int mem_size_B = sizeof(float) * size_B;
float *h_B = (float *)malloc(mem_size_B);
// set seed for rand()
srand(2006);
// initialize host memory
randomInit(h_A, size_A);
randomInit(h_B, size_B);
// allocate device memory
float *d_A, *d_B, *d_C;
unsigned int size_C = matrix_size.uiWC * matrix_size.uiHC;
unsigned int mem_size_C = sizeof(float) * size_C;
// allocate host memory for the result
float *h_C = (float *) malloc(mem_size_C);
float *h_CUBLAS = (float *) malloc(mem_size_C);
checkCudaErrors(cudaMalloc((void **) &d_A, mem_size_A));
checkCudaErrors(cudaMalloc((void **) &d_B, mem_size_B));
checkCudaErrors(cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMalloc((void **) &d_C, mem_size_C));
// setup execution parameters
dim3 threads(block_size, block_size);
dim3 grid(matrix_size.uiWC / threads.x, matrix_size.uiHC / threads.y);
// create and start timer
printf("Computing result using CUBLAS...");
// execute the kernel
int nIter = 30;
// CUBLAS version 2.0
{
const float alpha = 1.0f;
const float beta = 0.0f;
cublasHandle_t handle;
cudaEvent_t start, stop;
checkCudaErrors(cublasCreate(&handle));
//Perform warmup operation with cublas
checkCudaErrors(cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB));
// Allocate CUDA events that we'll use for timing
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
// Record the start event
checkCudaErrors(cudaEventRecord(start, NULL));
for (int j = 0; j < nIter; j++)
{
//note cublas is column primary!
//need to transpose the order
checkCudaErrors(cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB));
}
printf("done.\n");
// Record the stop event
checkCudaErrors(cudaEventRecord(stop, NULL));
// Wait for the stop event to complete
checkCudaErrors(cudaEventSynchronize(stop));
float msecTotal = 0.0f;
checkCudaErrors(cudaEventElapsedTime(&msecTotal, start, stop));
// Compute and print the performance
float msecPerMatrixMul = msecTotal / nIter;
double flopsPerMatrixMul = 2.0 * (double)matrix_size.uiHC * (double)matrix_size.uiWC * (double)matrix_size.uiHB;
double gigaFlops = (flopsPerMatrixMul * 1.0e-9f) / (msecPerMatrixMul / 1000.0f);
printf(
"Performance= %.2f GFlop/s, Time= %.3f msec, Size= %.0f Ops\n",
gigaFlops,
msecPerMatrixMul,
flopsPerMatrixMul);
// copy result from device to host
checkCudaErrors(cudaMemcpy(h_CUBLAS, d_C, mem_size_C, cudaMemcpyDeviceToHost));
// Destroy the handle
checkCudaErrors(cublasDestroy(handle));
}
// compute reference solution
printf("Computing result using host CPU...");
float *reference = (float *)malloc(mem_size_C);
matrixMulCPU(reference, h_A, h_B, matrix_size.uiHA, matrix_size.uiWA, matrix_size.uiWB);
printf("done.\n");
// check result (CUBLAS)
bool resCUBLAS = sdkCompareL2fe(reference, h_CUBLAS, size_C, 1.0e-6f);
if (resCUBLAS != true)
{
printDiff(reference, h_CUBLAS, matrix_size.uiWC, matrix_size.uiHC, 100, 1.0e-5f);
}
printf("Comparing CUBLAS Matrix Multiply with CPU results: %s\n", (true == resCUBLAS) ? "PASS" : "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);
free(reference);
checkCudaErrors(cudaFree(d_A));
checkCudaErrors(cudaFree(d_B));
checkCudaErrors(cudaFree(d_C));
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
if (resCUBLAS == true)
{
return EXIT_SUCCESS; // return value = 1
}
else
{
return EXIT_FAILURE; // return value = 0
}
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
printf("[Matrix Multiply CUBLAS] - Starting...\n");
int devID = 0, sizeMult = 5;
sMatrixSize matrix_size;
initializeCUDA(argc, argv, devID, sizeMult, matrix_size);
int matrix_result = matrixMultiply(argc, argv, devID, matrix_size);
return matrix_result;
}
The additional memory transfer seems to be caused by the CUBLAS library and is triggered by a call to cublasInit. You can confirm this by profiling the following code:
#include <cublas_v2.h>
int main()
{
cublasHandle_t handle;
cublasCreate(&handle);
cudaDeviceReset();
return 0;
}
which nvprof reports as calling cudaMemcpy:
$ nvprof ./a.out
==9536== NVPROF is profiling process 9536, command: ./a.out
==9536== Profiling application: ./a.out
==9536== Profiling result:
Time(%) Time Calls Avg Min Max Name
100.00% 1.1190us 1 1.1190us 1.1190us 1.1190us [CUDA memcpy HtoD]
==9536== API calls:
Time(%) Time Calls Avg Min Max Name
76.51% 348.53ms 1 348.53ms 348.53ms 348.53ms cudaFree
23.26% 105.97ms 1 105.97ms 105.97ms 105.97ms cudaDeviceReset
0.09% 420.25us 178 2.3600us 125ns 103.52us cuDeviceGetAttribute
0.08% 349.37us 2 174.69us 110.59us 238.78us cuDeviceTotalMem
0.04% 202.10us 3 67.366us 9.3750us 109.43us cudaMalloc
0.01% 55.217us 2 27.608us 24.529us 30.688us cuDeviceGetName
0.00% 14.365us 1 14.365us 14.365us 14.365us cudaMemcpy
0.00% 10.016us 16 626ns 434ns 2.0440us cudaEventCreateWithFlags
0.00% 4.5000us 11 409ns 271ns 1.2730us cudaDeviceGetAttribute
0.00% 3.4510us 4 862ns 251ns 2.3370us cuDeviceGetCount
0.00% 2.3200us 4 580ns 281ns 1.0350us cuDeviceGet
0.00% 1.3600us 1 1.3600us 1.3600us 1.3600us cudaGetDevice
0.00% 630ns 1 630ns 630ns 630ns cuInit
0.00% 339ns 1 339ns 339ns 339ns cuDriverGetVersion
I doubt that anyone without access to the current CUBLAS source will be able to explain why initialising the CUBLAS library triggers a host to device transfer, but that seems to be the cause of your observation.
This function performs the symmetric matrix-matrix multiplication using CUDA. Although, I succeeded in using the nonsymmetric version "cublas{t}gemm()" I couldn't use the "cublas{t}symm()" function properly.
I know that CUBLAS library uses column-major matrix storage. I am using row-major C/C++ matrix and I know how to solve this issue for "cublas{t}gemm()" by replacing the input matrices and etc. However, I couldn't solve it for the symmetric case. The problem is even if I use column-major matrix storage I find unexpectable results. Matrices contain complex floats (cuComplex). I assume I have row-major matrices. Here is the code and the output:
// Matrix multiplication: C = A * B.
// Host code.
//
// Utilities and system includes
#include <assert.h>
#include <helper_string.h> // helper for shared functions common to CUDA SDK samples
// CUDA runtime
#include <cuda_runtime.h>
#include <cublas_v2.h>
#ifndef min
#define min(a,b) ((a < b) ? a : b)
#endif
#ifndef max
#define max(a,b) ((a > b) ? a : b)
#endif
////////////////////////////////////////////////////////////////////////////////
// These are CUDA Helper functions (in addition to helper_cuda.h)
void inline checkError(cublasStatus_t status, const char *msg)
{
if (status != CUBLAS_STATUS_SUCCESS)
{
printf("%s", msg);
exit(EXIT_FAILURE);
}
}
// end of CUDA Helper Functions
// Allocates a matrix with random float entries.
void randomCmplxInit(cuComplex *data, int size)
{
for (int i = 0; i < size; ++i)
data[i] = make_cuComplex( rand() / (float)RAND_MAX, rand() / (float)RAND_MAX);
}
//void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMatrixSize &matrix_size)
void initializeCUDA(int argc, char **argv, int &devID)
{
// By default, we use device 0, otherwise we override the device ID based on what is provided at the command line
cudaError_t error;
devID = 0;
int m,n,k;
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
devID = getCmdLineArgumentInt(argc, (const char **)argv, "device");
error = cudaSetDevice(devID);
if (error != cudaSuccess)
{
printf("cudaSetDevice returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
}
// get number of SMs on this GPU
error = cudaGetDevice(&devID);
cudaDeviceProp deviceProp;
error = cudaGetDeviceProperties(&deviceProp, devID);
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", devID, deviceProp.name, deviceProp.major, deviceProp.minor);
// use a larger block size for Fermi and above
int block_size = (deviceProp.major < 2) ? 16 : 32;
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test matrix multiply using CUBLAS
////////////////////////////////////////////////////////////////////////////////
int matrixMultiply(int argc, char **argv, int devID)
{
int i,j;
unsigned int m,n,k;
cudaDeviceProp deviceProp;
cudaError_t error;
error = cudaGetDeviceProperties(&deviceProp, devID);
if (error != cudaSuccess)
{
printf("cudaGetDeviceProperties returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
// use a larger block size for Fermi and above
int block_size = (deviceProp.major < 2) ? 16 : 32;
m=3; //number of rows of matrix op(A) and C. A--> (m x k)
n=2; //number of columns of matrix op(B) and C. B--> (k x n)
k=m; //number of columns of op(A) and rows of op(B). C--> (m x n)
// I want to compute C = A*B in row-major format,
//so I must find C(T)=B(T)A(T) = C(T)A in column-major format
// allocate host memory for matrices A and B
unsigned int size_A = m*(m+1)/2; //size of a symmetric matrix
unsigned int mem_size_A = sizeof(cuComplex) * size_A;
cuComplex *h_A = (cuComplex *)malloc(mem_size_A);
unsigned int size_B = m*n;
unsigned int mem_size_B = sizeof(cuComplex) * size_B;
cuComplex *h_B = (cuComplex *)malloc(mem_size_B);
// initialize host memory
for (i = 0; i < size_A; ++i)
h_A[i] = make_cuComplex( (float)(i+1),(float)0);
for (i = 0; i < size_B; ++i)
h_B[i] = make_cuComplex((float)(i+2), (float)0);
// allocate device memory
cuComplex *d_A, *d_B, *d_C;
unsigned int size_C = m*n;
unsigned int mem_size_C = sizeof(cuComplex) * size_C;
// allocate host memory for the result
cuComplex *h_C = (cuComplex *) malloc(mem_size_C);
cuComplex *h_CUBLAS = (cuComplex *) malloc(mem_size_C);
error = cudaMalloc((void **) &d_A, mem_size_A);
error = cudaMalloc((void **) &d_B, mem_size_B);
// copy host memory to device
error = cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);
error = cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);
error = cudaMalloc((void **) &d_C, mem_size_C);
// setup execution parameters
dim3 threads(block_size, block_size);
dim3 grid(n / threads.x, m / threads.y);
// create and start timer
printf("Computing result using CUBLAS...");
// CUBLAS version 2.0
{
cublasHandle_t handle;
cublasStatus_t ret;
ret = cublasCreate(&handle);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCreate returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
const cuComplex alpha = make_cuComplex(1.0f,0.0f);
const cuComplex beta = make_cuComplex(0.0f,0.0f);
//Perform operation with cublas
ret = cublasCsymm(handle, CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, n,m,&alpha,d_A,m,d_B,m,&beta,d_C,m);
// copy result from device to host
error = cudaMemcpy(h_CUBLAS, d_C, mem_size_C, cudaMemcpyDeviceToHost);
checkError(cublasDestroy(handle), "cublasDestroy() error!\n");
}
printf ("\nComputations completed.\n\n");
printf (" symm matrix A: \n");
int s=0;
for (i=0; i<min(m,4); i++) {
for (j=0; j<=i; j++) {
//printf ("%7.5G + j(%7.5G)", h_A[j+i*k].x,h_A[j+i*k].y);
printf ("%7.5G", h_A[s].x);
s++;
}
printf ("\n");
}
printf ("\n matrix B: \n");
for (i=0; i<min(k,4); i++) {
for (j=0; j<min(n,4); j++) {
//printf ("%7.5G + j(%7.5G)", h_B[j+i*n].x,h_B[j+i*n].y);
printf ("%7.5G", h_B[j+i*n].x);
}
printf ("\n");
}
printf ("\n matrix C=A*B: \n");
for (i=0; i<min(m,4); i++) {
for (j=0; j<min(n,4); j++) {
//printf ("%7.5G + j(%7.5G)", h_CUBLAS[j+i*n].x,h_CUBLAS[j+i*n].y);
printf ("%7.5G", h_CUBLAS[j+i*n].x);
}
printf ("\n");
}
// clean up memory
free(h_A);
free(h_B);
free(h_C);
//free(reference);
cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);
cudaDeviceReset();
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
printf("[Matrix Multiply CUBLAS] - Starting...\n");
int devID = 0, sizeMult = 5;
initializeCUDA(argc, argv, devID);
int matrix_result = matrixMultiply(argc, argv, devID);
}
I suppose that I have the following matrices for the multiplication:
A =
1 2 4
2 3 5
4 5 6
B =
2 3
4 5
6 7
and expect to obtain
A*B =
34 41
46 56
64 79
But the obtained OUTPUT is as follows:
symm matrix A:
1
2 3
4 5 6
matrix B:
2 3
4 5
6 7
matrix C=A*B:
78 90
74 97
114 146
What am I missing in this code ? Probably the arguments of "cublasCsymm" function are wrong.
Thanks,
Kagan
EDIT:
Based on questions posed below, I elected to re-work my answer and example code.
You can handle row-major storage without transpose at least for these operations. And this observation is further facilitated by the fact that the symm function does not used the packed storage.
So to answer the additional questions:
the cublasCsymm function does not use a packed storage format (like some other functions such as cublasCspmv for example), because the cublasCsymm function is intended to duplicate the functionality of the corresponding netlib function, which also does not use a packed storage format. Based on my review of the cublas API, I don't see a symmetric-packed-storage matrix-matrix multiply function available.
You can use row-major storage (e.g. C-style) with cublas, without transposing, at least for these operations (matrix-matrix multiply, without packed storage) by following the advice given here.
What follows is a re-worked version of my previous example, that incorporates the information in item 2 above.
// Matrix multiplication: C = A * B.
// Host code.
//
// Utilities and system includes
#include <assert.h>
#include <helper_string.h> // helper for shared functions common to CUDA SDK sa
mples
// CUDA runtime
#include <cuda_runtime.h>
#include <cublas_v2.h>
// error check macros
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
// for CUBLAS V2 API
#define cublasCheckErrors(fn) \
do { \
cublasStatus_t __err = fn; \
if (__err != CUBLAS_STATUS_SUCCESS) { \
fprintf(stderr, "Fatal cublas error: %d (at %s:%d)\n", \
(int)(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
#ifndef min
#define min(a,b) ((a < b) ? a : b)
#endif
#ifndef max
#define max(a,b) ((a > b) ? a : b)
#endif
////////////////////////////////////////////////////////////////////////////////
// These are CUDA Helper functions (in addition to helper_cuda.h)
void inline checkError(cublasStatus_t status, const char *msg)
{
if (status != CUBLAS_STATUS_SUCCESS)
{
printf("%s", msg);
exit(EXIT_FAILURE);
}
}
// end of CUDA Helper Functions
// Allocates a matrix with random float entries.
void randomCmplxInit(cuComplex *data, int size)
{
for (int i = 0; i < size; ++i)
data[i] = make_cuComplex( rand() / (float)RAND_MAX, rand() / (float)RAND
_MAX);
}
//void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMa
trixSize &matrix_size)
void initializeCUDA(int argc, char **argv, int &devID)
{
// By default, we use device 0, otherwise we override the device ID based on
what is provided at the command line
cudaError_t error;
devID = 0;
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
devID = getCmdLineArgumentInt(argc, (const char **)argv, "device");
error = cudaSetDevice(devID);
if (error != cudaSuccess)
{
printf("cudaSetDevice returned error code %d, line(%d)\n", error, __
LINE__);
exit(EXIT_FAILURE);
}
}
// get number of SMs on this GPU
error = cudaGetDevice(&devID);
cudaDeviceProp deviceProp;
error = cudaGetDeviceProperties(&deviceProp, devID);
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", devID, dev
iceProp.name, deviceProp.major, deviceProp.minor);
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test matrix multiply using CUBLAS
////////////////////////////////////////////////////////////////////////////////
int matrixMultiply(int argc, char **argv, int devID)
{
int i,j;
unsigned int m,n,k;
cudaDeviceProp deviceProp;
cudaError_t error;
error = cudaGetDeviceProperties(&deviceProp, devID);
if (error != cudaSuccess)
{
printf("cudaGetDeviceProperties returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
// use a larger block size for Fermi and above
m=3; //number of rows of matrix op(A) and C. A--> (m x k)
n=2; //number of columns of matrix op(B) and C. B--> (k x n)
k=m; //number of columns of op(A) and rows of op(B). C--> (m x n)
// I want to compute C = A*B in row-major format,
//so I must find C(T)=B(T)A(T) = C(T)A in column-major format
// allocate host memory for matrices A and B
unsigned int size_A = m*m; //size of a symmetric matrix
printf("size_A = %d\n", size_A);
unsigned int mem_size_A = sizeof(cuComplex) * size_A;
cuComplex *h_A = (cuComplex *)malloc(mem_size_A);
unsigned int size_B = m*n;
unsigned int mem_size_B = sizeof(cuComplex) * size_B;
cuComplex *h_B = (cuComplex *)malloc(mem_size_B);
// initialize host memory
// for (i = 0; i < size_A; ++i)
// h_A[i] = make_cuComplex( (float)(i+1),(float)0);
h_A[0] = make_cuComplex((float)1, (float)0);
h_A[1] = make_cuComplex((float)2, (float)0);
h_A[2] = make_cuComplex((float)4, (float)0);
h_A[3] = make_cuComplex((float)0, (float)0);
h_A[4] = make_cuComplex((float)3, (float)0);
h_A[5] = make_cuComplex((float)5, (float)0);
h_A[6] = make_cuComplex((float)0, (float)0);
h_A[7] = make_cuComplex((float)0, (float)0);
h_A[8] = make_cuComplex((float)6, (float)0);
// for (i = 0; i < size_B; ++i)
// h_B[i] = make_cuComplex((float)(i+2), (float)0);
h_B[0] = make_cuComplex((float)2, (float)0);
h_B[1] = make_cuComplex((float)3, (float)0);
h_B[2] = make_cuComplex((float)4, (float)0);
h_B[3] = make_cuComplex((float)5, (float)0);
h_B[4] = make_cuComplex((float)6, (float)0);
h_B[5] = make_cuComplex((float)7, (float)0);
// allocate device memory
cuComplex *d_A, *d_B, *d_C;
unsigned int size_C = m*n;
unsigned int mem_size_C = sizeof(cuComplex) * size_C;
// allocate host memory for the result
cuComplex *h_C = (cuComplex *) malloc(mem_size_C);
cuComplex *h_CUBLAS = (cuComplex *) malloc(mem_size_C);
error = cudaMalloc((void **) &d_A, mem_size_A);
error = cudaMalloc((void **) &d_B, mem_size_B);
// copy host memory to device
error = cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);
error = cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);
error = cudaMalloc((void **) &d_C, mem_size_C);
// create and start timer
printf("Computing result using CUBLAS...");
// CUBLAS version 2.0
{
cublasHandle_t handle;
cublasStatus_t ret;
ret = cublasCreate(&handle);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCreate returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
const cuComplex alpha = make_cuComplex(1.0f,0.0f);
const cuComplex beta = make_cuComplex(0.0f,0.0f);
//Perform operation with cublas
ret = cublasCsymm(handle, CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_LOWER, n,m,&alpha,d_A,m,d_B,n,&beta,d_C,n);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCsymm returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
// copy result from device to host
error = cudaMemcpy(h_CUBLAS, d_C, mem_size_C, cudaMemcpyDeviceToHost);
checkError(cublasDestroy(handle), "cublasDestroy() error!\n");
}
printf ("\nComputations completed.\n\n");
printf (" symm matrix A: \n");
// int s=0;
for (i=0; i<min(m,4); i++) {
for (j=0; j<min(m,4); j++) {
//printf ("%7.5G + j(%7.5G)", h_A[j+i*k].x,h_A[j+i*k].y);
// printf ("%7.5G", h_A[s].x);
printf ("%7.5G", h_A[j+(i*m)].x);
// s++;
}
printf ("\n");
}
printf ("\n matrix B: \n");
for (i=0; i<min(k,4); i++) {
for (j=0; j<min(n,4); j++) {
//printf ("%7.5G + j(%7.5G)", h_B[j+i*n].x,h_B[j+i*n].y);
printf ("%7.5G", h_B[j+(i*n)].x);
}
printf ("\n");
}
printf ("\n matrix C=A*B: \n");
for (i=0; i<min(m,4); i++) {
for (j=0; j<min(n,4); j++) {
//printf ("%7.5G + j(%7.5G)", h_CUBLAS[j+i*n].x,h_CUBLAS[j+i*n].y);
printf ("%7.5G", h_CUBLAS[j+(i*n)].x);
}
printf ("\n");
}
// clean up memory
free(h_A);
free(h_B);
free(h_C);
//free(reference);
cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);
cudaDeviceReset();
return 0;
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
printf("[Matrix Multiply CUBLAS] - Starting...\n");
int devID = 0;
initializeCUDA(argc, argv, devID);
int matrix_result = matrixMultiply(argc, argv, devID);
cudaCheckErrors("some error");
return 0;
}
$ ./t213
[Matrix Multiply CUBLAS] - Starting...
GPU Device 0: "Tesla M2070" with compute capability 2.0
size_A = 9
Computing result using CUBLAS...
Computations completed.
symm matrix A:
1 2 4
0 3 5
0 0 6
matrix B:
2 3
4 5
6 7
matrix C=A*B:
34 41
46 56
64 79
$
ORIGINAL RESPONSE:
Several problems:
When I run your code as you have it posted right now, I don't get the
results that you show. Here's what I get:
[Matrix Multiply CUBLAS] - Starting...
GPU Device 0: "Tesla M2070" with compute capability 2.0
Computing result using CUBLAS...
Computations completed.
symm matrix A:
1
2 3
4 5 6
matrix B:
2 3
4 5
6 7
matrix C=A*B:
-131 -128
260 -122
-115 266
The code compiles with a number of warnings and also you're not doing proper error checking (for example you're not checking the return value from cublasCsymm
You are wanting to multiply C = A*B This means A is on the LEFT,
but you are passing CUBLAS_SIDE_RIGHT to cublasCsymm Several other cublasCsymm parameters were wrong as well. I think maybe you thought you could do A*B as (B(T)*A(T)) but that only works for square matrices. Not sure what you were thinking, exactly.
You having row-major storage on your matrices and passing them to cublas which interprets them in column-major order. For the following matrix:
1 2
3 4
row-major storage looks like this:
1 2 3 4
column-major storage looks like this:
1 3 2 4
You can transpose these matrices if you wish, using cublasCgeam or you can manually modify your storage.
You're making some sort of assumption about some kind of compressed
storage format for the symmetric matrix A which is not correct.
Read carefully the defintion of the storage
type.
It doesn't say the portion of the matrix that is "supplied" or
"present" it says the portion of the matrix that is filled.
Here is a complete code that has the above problems fixed:
// Matrix multiplication: C = A * B.
// Host code.
//
// Utilities and system includes
#include <assert.h>
#include <helper_string.h> // helper for shared functions common to CUDA SDK sa
mples
// CUDA runtime
#include <cuda_runtime.h>
#include <cublas_v2.h>
// error check macros
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
// for CUBLAS V2 API
#define cublasCheckErrors(fn) \
do { \
cublasStatus_t __err = fn; \
if (__err != CUBLAS_STATUS_SUCCESS) { \
fprintf(stderr, "Fatal cublas error: %d (at %s:%d)\n", \
(int)(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
#ifndef min
#define min(a,b) ((a < b) ? a : b)
#endif
#ifndef max
#define max(a,b) ((a > b) ? a : b)
#endif
////////////////////////////////////////////////////////////////////////////////
// These are CUDA Helper functions (in addition to helper_cuda.h)
void inline checkError(cublasStatus_t status, const char *msg)
{
if (status != CUBLAS_STATUS_SUCCESS)
{
printf("%s", msg);
exit(EXIT_FAILURE);
}
}
// end of CUDA Helper Functions
// Allocates a matrix with random float entries.
void randomCmplxInit(cuComplex *data, int size)
{
for (int i = 0; i < size; ++i)
data[i] = make_cuComplex( rand() / (float)RAND_MAX, rand() / (float)RAND_MAX);
}
//void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMatrixSize &matrix_size)
void initializeCUDA(int argc, char **argv, int &devID)
{
// By default, we use device 0, otherwise we override the device ID based on what is provided at the command line
cudaError_t error;
devID = 0;
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
devID = getCmdLineArgumentInt(argc, (const char **)argv, "device");
error = cudaSetDevice(devID);
if (error != cudaSuccess)
{
printf("cudaSetDevice returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
}
// get number of SMs on this GPU
error = cudaGetDevice(&devID);
cudaDeviceProp deviceProp;
error = cudaGetDeviceProperties(&deviceProp, devID);
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n", devID, deviceProp.name, deviceProp.major, deviceProp.minor);
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test matrix multiply using CUBLAS
////////////////////////////////////////////////////////////////////////////////
int matrixMultiply(int argc, char **argv, int devID)
{
int i,j;
unsigned int m,n,k;
cudaDeviceProp deviceProp;
cudaError_t error;
error = cudaGetDeviceProperties(&deviceProp, devID);
if (error != cudaSuccess)
{
printf("cudaGetDeviceProperties returned error code %d, line(%d)\n", error, __LINE__);
exit(EXIT_FAILURE);
}
// use a larger block size for Fermi and above
m=3; //number of rows of matrix op(A) and C. A--> (m x k)
n=2; //number of columns of matrix op(B) and C. B--> (k x n)
k=m; //number of columns of op(A) and rows of op(B). C--> (m x n)
// I want to compute C = A*B in row-major format,
//so I must find C(T)=B(T)A(T) = C(T)A in column-major format
// allocate host memory for matrices A and B
unsigned int size_A = m*m; //size of a symmetric matrix
printf("size_A = %d\n", size_A);
unsigned int mem_size_A = sizeof(cuComplex) * size_A;
cuComplex *h_A = (cuComplex *)malloc(mem_size_A);
unsigned int size_B = m*n;
unsigned int mem_size_B = sizeof(cuComplex) * size_B;
cuComplex *h_B = (cuComplex *)malloc(mem_size_B);
// initialize host memory
// for (i = 0; i < size_A; ++i)
// h_A[i] = make_cuComplex( (float)(i+1),(float)0);
h_A[0] = make_cuComplex((float)1, (float)0);
h_A[1] = make_cuComplex((float)2, (float)0);
h_A[2] = make_cuComplex((float)4, (float)0);
h_A[3] = make_cuComplex((float)0, (float)0);
h_A[4] = make_cuComplex((float)3, (float)0);
h_A[5] = make_cuComplex((float)5, (float)0);
h_A[6] = make_cuComplex((float)0, (float)0);
h_A[7] = make_cuComplex((float)0, (float)0);
h_A[8] = make_cuComplex((float)6, (float)0);
// for (i = 0; i < size_B; ++i)
// h_B[i] = make_cuComplex((float)(i+2), (float)0);
h_B[0] = make_cuComplex((float)2, (float)0);
h_B[1] = make_cuComplex((float)4, (float)0);
h_B[2] = make_cuComplex((float)6, (float)0);
h_B[3] = make_cuComplex((float)3, (float)0);
h_B[4] = make_cuComplex((float)5, (float)0);
h_B[5] = make_cuComplex((float)7, (float)0);
// allocate device memory
cuComplex *d_A, *d_B, *d_C;
unsigned int size_C = m*n;
unsigned int mem_size_C = sizeof(cuComplex) * size_C;
// allocate host memory for the result
cuComplex *h_C = (cuComplex *) malloc(mem_size_C);
cuComplex *h_CUBLAS = (cuComplex *) malloc(mem_size_C);
error = cudaMalloc((void **) &d_A, mem_size_A);
error = cudaMalloc((void **) &d_B, mem_size_B);
// copy host memory to device
error = cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);
error = cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);
error = cudaMalloc((void **) &d_C, mem_size_C);
// create and start timer
printf("Computing result using CUBLAS...");
// CUBLAS version 2.0
{
cublasHandle_t handle;
cublasStatus_t ret;
ret = cublasCreate(&handle);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCreate returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
const cuComplex alpha = make_cuComplex(1.0f,0.0f);
const cuComplex beta = make_cuComplex(0.0f,0.0f);
//Perform operation with cublas
ret = cublasCsymm(handle, CUBLAS_SIDE_LEFT, CUBLAS_FILL_MODE_LOWER, m,n,&alpha,d_A,m,d_B,m,&beta,d_C,m);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCsymm returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
Here is the output:
[Matrix Multiply CUBLAS] - Starting...
GPU Device 0: "Tesla M2070" with compute capability 2.0
size_A = 9
Computing result using CUBLAS...
Computations completed.
symm matrix A:
1 0 0
2 3 0
4 5 6
matrix B:
2 3
4 5
6 7
matrix C=A*B:
34 41
46 56
64 79