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 forward_compact_layer_gpu(const layer l, network_state state)
{
int i, b;
for (b=0;b<l.batch;b++)
{
if (l.method==0) // add
{
// copy first section
copy_ongpu(l.outputs, state.input+b*l.inputs, 1, l.output_gpu+b*l.outputs, 1);
// add other splits
for (i=1;i<l.index;i++)
{
axpy_ongpu(l.outputs, 1, state.input+b*l.inputs+i*l.outputs, 1, l.output_gpu+b*l.outputs, 1);
}
} else if (l.method==1) // sub
{
// copy first section
copy_ongpu(l.outputs, state.input+b*l.inputs, 1, l.output_gpu+b*l.outputs, 1);
// sub other splits
for (i=1;i<l.index;i++)
{
axpy_ongpu(l.outputs, -1, state.input+b*l.inputs+i*l.outputs, 1, l.output_gpu+b*l.outputs, 1);
}
} else if (l.method==2) // max
{
compact_forward_max_gpu(l.w, l.h, l.c, l.index,
state.input+b*l.inputs,
l.output_gpu+b*l.outputs,
l.indexes_gpu);
} else if (l.method==10)
{
compact_forward_padd_gpu(l.w, l.h, l.c,
state.input+b*l.inputs,
l.output_gpu+b*l.outputs);
} else if (l.method==12)
{
compact_forward_pmax_gpu(l.w, l.h, l.c,
state.input+b*l.inputs,
l.output_gpu+b*l.outputs,
l.indexes_gpu);
}
}
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;
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);
*/
}
void forward_shortcut_layer_gpu(const layer l, network_state state)
{
copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
shortcut_gpu(l.batch, l.w, l.h, l.c, state.net.layers[l.index].output_gpu, l.out_w, l.out_h, l.out_c, l.output_gpu);
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
}
