void time_gpu(int TA, int TB, int m, int k, int n)
{
int iter = 10;
float *a = random_matrix(m,k);
float *b = random_matrix(k,n);
int lda = (!TA)?k:m;
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
float *a_cl = cuda_make_array(a, m*k);
float *b_cl = cuda_make_array(b, k*n);
float *c_cl = cuda_make_array(c, m*n);
int i;
clock_t start = clock(), end;
for(i = 0; i<iter; ++i){
gemm_gpu(TA,TB,m,n,k,1,a_cl,lda,b_cl,ldb,1,c_cl,n);
cudaThreadSynchronize();
}
double flop = ((double)m)*n*(2.*k + 2.)*iter;
double gflop = flop/pow(10., 9);
end = clock();
double seconds = sec(end-start);
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf s, %lf GFLOPS\n",m,k,k,n, TA, TB, seconds, gflop/seconds);
cuda_free(a_cl);
cuda_free(b_cl);
cuda_free(c_cl);
free(a);
free(b);
free(c);
}
void time_gpu_random_matrix(int TA, int TB, int m, int k, int n)
{
float *a;
if(!TA) a = random_matrix(m,k);
else a = random_matrix(k,m);
int lda = (!TA)?k:m;
float *b;
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
int i;
clock_t start = clock(), end;
for(i = 0; i<32; ++i){
gemm_gpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
}
end = clock();
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf s\n",m,k,k,n, TA, TB, (float)(end-start)/CLOCKS_PER_SEC);
free(a);
free(b);
free(c);
}
void test_gpu_accuracy(int TA, int TB, int m, int k, int n)
{
srand(0);
float *a;
if(!TA) a = random_matrix(m,k);
else a = random_matrix(k,m);
int lda = (!TA)?k:m;
float *b;
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
float *c_gpu = random_matrix(m,n);
memset(c, 0, m*n*sizeof(float));
memset(c_gpu, 0, m*n*sizeof(float));
int i;
//pm(m,k,b);
gemm_gpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c_gpu,n);
//printf("GPU\n");
//pm(m, n, c_gpu);
gemm_cpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
//printf("\n\nCPU\n");
//pm(m, n, c);
double sse = 0;
for(i = 0; i < m*n; ++i) {
//printf("%f %f\n", c[i], c_gpu[i]);
sse += pow(c[i]-c_gpu[i], 2);
}
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %g SSE\n",m,k,k,n, TA, TB, sse/(m*n));
free(a);
free(b);
free(c);
free(c_gpu);
}
