void backward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
constrain_ongpu(l.outputs*l.batch, 5, l.delta_gpu, 1);
gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
for(i = 0; i < l.batch; ++i){
axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1);
}
if(l.batch_normalize){
backward_batchnorm_layer_gpu(l, state);
}
int m = l.outputs;
int k = l.batch;
int n = l.inputs;
float * a = l.delta_gpu;
float * b = state.input;
float * c = l.weight_updates_gpu;
gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);
m = l.batch;
k = l.outputs;
n = l.inputs;
a = l.delta_gpu;
b = l.weights_gpu;
c = state.delta;
if(c) gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
}
void forward_local_layer_gpu(const local_layer l, network_state state)
{
int out_h = local_out_height(l);
int out_w = local_out_width(l);
int i, j;
int locations = out_h * out_w;
for(i = 0; i < l.batch; ++i) {
copy_ongpu(l.outputs, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1, state.st_handle.stream);
}
for(i = 0; i < l.batch; ++i) {
float *input = state.input + i*l.w*l.h*l.c;
im2col_ongpu(input, l.c, l.h, l.w,
l.size, l.stride, l.pad, l.col_image_gpu, state.st_handle.stream);
float *output = l.output_gpu + i*l.outputs;
for(j = 0; j < locations; ++j) {
float *a = l.weights_gpu + j*l.size*l.size*l.c*l.n;
float *b = l.col_image_gpu + j;
float *c = output + j;
int m = l.n;
int n = 1;
int k = l.size*l.size*l.c;
//printf("passou no local layer forward_local_layer_gpu\n");
gemm_ongpu(0,0,m,n,k,1,a,k,b,locations,1,c,locations, state.st_handle);
}
}
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, state.st_handle.stream);
}
void time_ongpu(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_ongpu(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 backward_local_layer_gpu(local_layer l, network_state state)
{
int i, j;
int locations = l.out_w*l.out_h;
gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu, state.st_handle.stream);
for(i = 0; i < l.batch; ++i) {
axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1, state.st_handle.stream);
}
for(i = 0; i < l.batch; ++i) {
float *input = state.input + i*l.w*l.h*l.c;
im2col_ongpu(input, l.c, l.h, l.w,
l.size, l.stride, l.pad, l.col_image_gpu, state.st_handle.stream);
for(j = 0; j < locations; ++j) {
float *a = l.delta_gpu + i*l.outputs + j;
float *b = l.col_image_gpu + j;
float *c = l.weight_updates_gpu + j*l.size*l.size*l.c*l.n;
int m = l.n;
int n = l.size*l.size*l.c;
int k = 1;
//printf("passou no backward_local_layer_gpu first call\n");
gemm_ongpu(0,1,m,n,k,1,a,locations,b,locations,1,c,n, state.st_handle);
}
if(state.delta) {
for(j = 0; j < locations; ++j) {
float *a = l.weights_gpu + j*l.size*l.size*l.c*l.n;
float *b = l.delta_gpu + i*l.outputs + j;
float *c = l.col_image_gpu + j;
int m = l.size*l.size*l.c;
int n = 1;
int k = l.n;
//printf("passou no backward_local_layer_gpu second call\n");
gemm_ongpu(1,0,m,n,k,1,a,m,b,locations,0,c,locations, state.st_handle);
}
col2im_ongpu(l.col_image_gpu, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w, state.st_handle.stream);
}
}
}
void backward_local_layer_gpu(local_layer l, network_state state)
{
int i, j;
int locations = l.out_w*l.out_h;
gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
for(i = 0; i < l.batch; ++i){
axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1);
}
for(i = 0; i < l.batch; ++i){
float *input = state.input + i*l.w*l.h*l.c;
im2col_ongpu(input, l.c, l.h, l.w,
l.size, l.stride, l.pad, l.col_image_gpu);
for(j = 0; j < locations; ++j){
float *a = l.delta_gpu + i*l.outputs + j;
float *b = l.col_image_gpu + j;
float *c = l.filter_updates_gpu + j*l.size*l.size*l.c*l.n;
int m = l.n;
int n = l.size*l.size*l.c;
int k = 1;
gemm_ongpu(0,1,m,n,k,1,a,locations,b,locations,1,c,n);
}
if(state.delta){
for(j = 0; j < locations; ++j){
float *a = l.filters_gpu + j*l.size*l.size*l.c*l.n;
float *b = l.delta_gpu + i*l.outputs + j;
float *c = l.col_image_gpu + j;
int m = l.size*l.size*l.c;
int n = 1;
int k = l.n;
gemm_ongpu(1,0,m,n,k,1,a,m,b,locations,0,c,locations);
}
col2im_ongpu(l.col_image_gpu, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
}
}
}
void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
float *A_gpu = cuda_make_array(A, (TA ? lda*K:lda*M));
float *B_gpu = cuda_make_array(B, (TB ? ldb*N : ldb*K));
float *C_gpu = cuda_make_array(C, ldc*M);
gemm_ongpu(TA, TB, M, N, K, ALPHA, A_gpu, lda, B_gpu, ldb, BETA, C_gpu, ldc);
cuda_pull_array(C_gpu, C, ldc*M);
cuda_free(A_gpu);
cuda_free(B_gpu);
cuda_free(C_gpu);
}
void forward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
int m = l.batch;
int k = l.inputs;
int n = l.outputs;
float * a = state.input;
float * b = l.weights_gpu;
float * c = l.output_gpu;
gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
if(l.batch_normalize){
forward_batchnorm_layer_gpu(l, state);
}
for(i = 0; i < l.batch; ++i){
axpy_ongpu(l.outputs, 1, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1);
}
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
}
void forward_connected_layer_gpu(connected_layer l, network_state state)
{
int i;
for(i = 0; i < l.batch; ++i){
copy_ongpu_offset(l.outputs, l.biases_gpu, 0, 1, l.output_gpu, i*l.outputs, 1);
}
int m = l.batch;
int k = l.inputs;
int n = l.outputs;
float * a = state.input;
float * b = l.weights_gpu;
float * c = l.output_gpu;
gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
/*
cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
float avg = mean_array(l.output, l.outputs*l.batch);
printf("%f\n", avg);
*/
}