cuda矩阵相乘,CUDA / CUBLAS矩阵向量乘法
I previously posted a question regarding matrix-vector multiplication in CUDA and about writing my own kernel. After doing this, I decided to implement my problem using CUBLAS as suggested by some users (thanks @Robert Crovella ) on SO in the hopes of achieving higher performance (my project is performance driven).
Just to clarify: I want to multiply a NxN matrix with a 1xN vector.
I've been looking at the code pasted below for a couple of days now and I cant figure out why the multiplication is giving me an incorrect result. I fear that i am causing problems by using < vector > arrays (this is part of a much larger system that uses these data types). I don't mean to use this thread as a debugging tool but I think this will also be helpful to other users trying to achieve this as I have not come across a particularly comprehensive source on the internet for my particular problem (and for the cublas v2 API). Thanks in advance!
#include
#include
#include
#include
#include
#include
#include
#include
#include
//#include "timenow.cu"
// 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)
// random data filler
void fillvector(float *data, int N){
for(int i=0; i
data[i] = float(rand() % 10);
}
}
//printer
void printer(bool printOut, float *data, int N){
if(printOut == true){
for(int i=0; i
printf("%2.1f ", data[i]);
}
printf("\n");
}
}
/
/
int main(){
bool printOut = true;
int N;
std::cout << "Enter N: " ;
std::cin >> N;
std::vector x0;
x0.resize(N);
std::vector p;
p.resize(N*N);
// matrix A
std::vector A[N];
for(int i=0;i
A[i].resize(N);
fillvector(A[i].data(), N);
printer(printOut, A[i].data(), N);
}
printf("\n");
fillvector(x0.data(), N);
printer(printOut, x0.data(), N);
printf("\nStarting CUDA computation...");
///double startTime = timenow();
// device pointers
float *d_A, *d_p, *d_b, *d_x0, *d_v, *d_temp;
cudaMalloc((void**)&d_A, N*N*sizeof(float));
cudaMalloc((void**)&d_temp, N*sizeof(float));
cudaMalloc((void**)&d_x0, N*sizeof(float));
cudaCheckErrors("cuda malloc fail");
// might need to flatten A...
cublasSetVector(N, sizeof(float), &x0, 1, d_x0, 1);
//daMemcpy(d_x0, &x0, N*sizeof(float), cudaMemcpyHostToDevice);
cublasSetMatrix(N, N, sizeof(float), &A, N, d_A, N);
cudaCheckErrors("cuda memcpy of A or x0 fail");
float *temp;
temp = (float *)malloc(N*sizeof(temp));
cublasHandle_t handle;
cublasCheckErrors(cublasCreate(&handle));
float alpha = 1.0f;
float beta = 0.0f;
cublasCheckErrors(cublasSgemv(handle, CUBLAS_OP_N, N, N, &alpha, d_A, N, d_x0, 1, &beta, d_temp, 1));
cublasGetVector(N, sizeof(float), &temp, 1, d_temp, 1);
//cudaMemcpy(temp, d_temp, N*sizeof(float), cudaMemcpyDeviceToHost);
cudaCheckErrors("returning to host failed");
printf("\n");
printer(printOut, temp, N);
/*alpha = -1.0;
cublasSaxpy(handle, N, &alpha, d_temp, 1, d_v, 1);
cublasGetVector(N, sizeof(float) * N, d_v, 1, &v, 1);
printf("\n");
for(int i=0; i
printf("%2.1f ",v[i]);
}*/
printf("\nFinished CUDA computations...");
//double endTime = timenow();
//double timeDiff = endTime - startTime;
//printf("\nRuntime: %2.3f seconds \n", timeDiff);
cudaFree(d_temp);
cudaFree(d_A);
cudaFree(d_p);
cudaFree(d_x0);
return 0;
}
解决方案We don't reference the first element of a vector this way:
cublasSetVector(N, sizeof(float), &x0, 1, d_x0,
Instead you should do this:
cublasSetVector(N, sizeof(float), &(x0[0]), 1, d_x0, 1);
And likewise for your SetMatrix call referencing A:
cublasSetMatrix(N, N, sizeof(float), &(A[0]), N, d_A, N);
Your GetVector call has 2 errors:
cublasGetVector(N, sizeof(float), &temp, 1, d_temp, 1);
You have your temp and d_temp parameters reversed (you are copying from device to host) and you should not take the address of temp: it is already a pointer. So do this:
cublasGetVector(N, sizeof(float), d_temp, 1, temp, 1);
You're not doing proper error checking on all cublas calls, such as your get/set matrix/vector calls. Use the same method you are using on other cublas calls for these also.
You are creating A as an array of vectors. This won't work with cublasSetMatrix. Instead we need to create A as a flat vector, of sufficient size (N*N) to store the entire matrix.
Finally, cublas expects the matrices it uses to be stored in column-major order. If you pass C-style arrays in row-major order, you should use the transpose for that matrix in cublasSgemv:
cublasCheckErrors(cublasSgemv(handle, CUBLAS_OP_T, N, N, &alpha, d_A, N, d_x0, 1, &beta, d_temp, 1));
The following code has these various problems fixed:
$ cat t235.cu
#include
#include
#include
#include
#include
#include
#include
#include
#include
//#include "timenow.cu"
// 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)
// random data filler
void fillvector(float *data, int N){
for(int i=0; i
data[i] = float(rand() % 10);
}
}
//printer
void printer(bool printOut, float *data, int N){
if(printOut == true){
for(int i=0; i
printf("%2.1f ", data[i]);
}
printf("\n");
}
}
/
/
int main(){
bool printOut = true;
int N;
std::cout << "Enter N: " ;
std::cin >> N;
std::vector x0;
x0.resize(N);
std::vector p;
p.resize(N*N);
// matrix A
std::vector A(N*N);
fillvector(A.data(), N*N);
for (int i=0; i< N; i++){
printer(printOut, &(A[(i*N)]), N);
printf("\n");}
fillvector(x0.data(), N);
printer(printOut, x0.data(), N);
printf("\nStarting CUDA computation...");
///double startTime = timenow();
// device pointers
float *d_A, *d_x0, *d_temp;
cudaMalloc((void**)&d_A, N*N*sizeof(float));
cudaMalloc((void**)&d_temp, N*sizeof(float));
cudaMalloc((void**)&d_x0, N*sizeof(float));
cudaCheckErrors("cuda malloc fail");
// might need to flatten A...
cublasCheckErrors(cublasSetVector(N, sizeof(float), &(x0[0]), 1, d_x0, 1));
//daMemcpy(d_x0, &x0, N*sizeof(float), cudaMemcpyHostToDevice);
cublasCheckErrors(cublasSetMatrix(N, N, sizeof(float), &(A[0]), N, d_A, N));
//cudaCheckErrors("cuda memcpy of A or x0 fail");
float *temp;
temp = (float *)malloc(N*sizeof(temp));
cublasHandle_t handle;
cublasCheckErrors(cublasCreate(&handle));
float alpha = 1.0f;
float beta = 0.0f;
cublasCheckErrors(cublasSgemv(handle, CUBLAS_OP_T, N, N, &alpha, d_A, N, d_x0, 1, &beta, d_temp, 1));
cublasCheckErrors(cublasGetVector(N, sizeof(float), d_temp, 1, temp, 1));
//cudaMemcpy(temp, d_temp, N*sizeof(float), cudaMemcpyDeviceToHost);
//cudaCheckErrors("returning to host failed");
printf("\n");
printer(printOut, temp, N);
/*alpha = -1.0;
cublasSaxpy(handle, N, &alpha, d_temp, 1, d_v, 1);
cublasGetVector(N, sizeof(float) * N, d_v, 1, &v, 1);
printf("\n");
for(int i=0; i
printf("%2.1f ",v[i]);
}*/
printf("\nFinished CUDA computations...\n");
//double endTime = timenow();
//double timeDiff = endTime - startTime;
//printf("\nRuntime: %2.3f seconds \n", timeDiff);
cudaFree(d_temp);
cudaFree(d_A);
//cudaFree(d_p);
cudaFree(d_x0);
return 0;
}
$ nvcc -arch=sm_20 -O3 -o t235 t235.cu -lcublas
$ ./t235
Enter N: 5
3.0 6.0 7.0 5.0 3.0
5.0 6.0 2.0 9.0 1.0
2.0 7.0 0.0 9.0 3.0
6.0 0.0 6.0 2.0 6.0
1.0 8.0 7.0 9.0 2.0
0.0 2.0 3.0 7.0 5.0
Starting CUDA computation...
83.0 86.0 92.0 62.0 110.0
Finished CUDA computations...
$