int main()
{
int N, N2;
printf(" \n Input matrix size N x N, N = ");
scanf("%d", &N);
printf(" N = %d \n \n", N);
N2 = N*N;
double *A, *B, *C_cpu, *C_gpu, *D_cpu, *D_gpu, t1, t2, cpu_time, gpu_time;
double r_cpu, *r_gpu, nrmC_cpu, *nrmC_gpu;
A = (double *) malloc(N2*sizeof(double));
B = (double *) malloc(N2*sizeof(double));
C_cpu = (double *) malloc(N2*sizeof(double));
C_gpu = (double *) malloc(N2*sizeof(double));
D_cpu = (double *) malloc(N2*sizeof(double));
D_gpu = (double *) malloc(N2*sizeof(double));
r_gpu = (double *) malloc(1*sizeof(double));
nrmC_gpu = (double *) malloc(1*sizeof(double));
initial(A, B, N);
t1 = clock();
#pragma acc data copyin(A[0:N2], B[0:N2]) copyout(C_cpu[0:N2])
{
cublas_gemm(A, B, C_cpu, N);
}
r_cpu = dot_cpu(C_cpu, B, N2);
axpy_cpu(-1.0*r_cpu, B, C_cpu, N2);
nrmC_cpu = norm_cpu(C_cpu, N2);
copy_cpu(C_cpu, D_cpu, N2);
scal_cpu(1.0/nrmC_cpu, D_cpu, N2);
t2 = clock();
cpu_time = 1.0*(t2 - t1)/CLOCKS_PER_SEC;
t1 = clock();
#pragma acc enter data copyin(A[0:N2], B[0:N2]) create(C_gpu[0:N2], r_gpu[0], nrmC_gpu[0], D_gpu[0:N2])
{
gpu_cublas1(A, B, C_gpu, D_gpu, r_gpu, nrmC_gpu, N, N2);
}
#pragma acc update host(D_gpu[0:N2])
t2 = clock();
gpu_time = 1.0*(t2 - t1)/CLOCKS_PER_SEC;
printf(" gpu part success \n");
printf(" \n error = %f \n", error(D_cpu, D_gpu, N2));
printf(" gpu time = %f, cpu times = %f \n", gpu_time, cpu_time);
return 0;
}
void slerp(float *start, float *end, float s, int n, float *out)
{
float omega = acos(dot_cpu(n, start, 1, end, 1));
float so = sin(omega);
fill_cpu(n, 0, out, 1);
axpy_cpu(n, sin((1-s)*omega)/so, start, 1, out, 1);
axpy_cpu(n, sin(s*omega)/so, end, 1, out, 1);
float mag = mag_array(out, n);
scale_array(out, n, 1./mag);
}
float cuda_compare(float *x_gpu, float *x, size_t n, char *s) {
float *tmp = calloc(n, sizeof(float));
cuda_pull_array(x_gpu, tmp, n);
//int i;
//for(i = 0; i < n; ++i) printf("%f %f\n", tmp[i], x[i]);
axpy_cpu(n, -1, x, 1, tmp, 1);
float err = dot_cpu(n, tmp, 1, tmp, 1);
printf("Error %s: %f\n", s, sqrt(err / n));
free(tmp);
return err;
}
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
if (!state.truth) return;
if (l.cost_type == MASKED) {
mask_ongpu(l.batch*l.inputs, state.input, state.truth);
}
copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1);
axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1);
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
}
void forward_cost_layer(cost_layer l, network_state state)
{
if (!state.truth) return;
if(l.cost_type == MASKED){
int i;
for(i = 0; i < l.batch*l.inputs; ++i){
if(state.truth[i] == 0) state.input[i] = 0;
}
}
copy_cpu(l.batch*l.inputs, state.truth, 1, l.delta, 1);
axpy_cpu(l.batch*l.inputs, -1, state.input, 1, l.delta, 1);
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
//printf("cost: %f\n", *l.output);
}
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
if (!state.truth) return;
if (l.cost_type == MASKED) {
mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
}
if(l.cost_type == SMOOTH){
smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu);
} else {
copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1);
axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1);
}
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
}
void forward_cost_layer(cost_layer l, network_state state)
{
if (!state.truth) return;
if(l.cost_type == MASKED){
int i;
for(i = 0; i < l.batch*l.inputs; ++i){
if(state.truth[i] == SECRET_NUM) state.input[i] = SECRET_NUM;
}
}
if(l.cost_type == SMOOTH){
smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta);
} else {
copy_cpu(l.batch*l.inputs, state.truth, 1, l.delta, 1);
axpy_cpu(l.batch*l.inputs, -1, state.input, 1, l.delta, 1);
}
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
//printf("cost: %f\n", *l.output);
}