mirror/darknet
1
0
Fork 0
mirror of https://github.com/pjreddie/darknet.git synced 2023-08-10 21:13:14 +03:00
Code Issues Projects Releases Wiki Activity

Fixed openmp bugs for XNOR

Browse Source
This commit is contained in:
AlexeyAB
2018-09-12 16:22:54 +03:00
parent c0e01fd63c
commit ca43bbdaae
3 changed files with 219 additions and 161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
src/convolutional_kernels.cu
View File

@ -141,70 +141,39 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
size_t t_intput_size = new_ldb * n;
size_t t_bit_input_size = t_intput_size / 8;// +1;
/*
int i = 0;
im2col_align_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, l.align_workspace_gpu, l.bit_align);
//cudaDeviceSynchronize();
// should be optimized
float_to_bit_gpu(l.align_workspace_gpu, (unsigned char *)state.workspace, l.align_workspace_size);
//cudaDeviceSynchronize();
//im2col_align_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, state.workspace, l.bit_align);
transpose_bin_gpu((unsigned char *)state.workspace, (unsigned char *)l.transposed_align_workspace_gpu, k, n, l.bit_align, new_ldb, 8);
//cudaDeviceSynchronize();
// should be optimized
gemm_nn_custom_bin_mean_transposed_gpu(m, n, k,
(unsigned char *)l.align_bit_weights_gpu, new_ldb, (unsigned char *)l.transposed_align_workspace_gpu, new_ldb, l.output_gpu, n, l.mean_arr_gpu);
//cudaDeviceSynchronize();
//check_error(status);
*/
{
//
int i = 0;
im2col_align_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, l.align_workspace_gpu, l.bit_align);
//cudaDeviceSynchronize();
/*
float *input_cpu = (float *)calloc(input_size, sizeof(float));
status = cudaMemcpy(input_cpu, state.input, input_size* sizeof(float), cudaMemcpyDeviceToHost);
check_error(status);
// should be optimized
float_to_bit_gpu(l.align_workspace_gpu, (unsigned char *)state.workspace, l.align_workspace_size);
//cudaDeviceSynchronize();
// swaped(binary_weights <-> l.weights)
convolve_cpu(input_cpu, l.weights, l.output, l.w, l.h, l.c, l.n, l.size, l.pad); // CPU
status = cudaMemcpy(l.output_gpu, l.output, l.outputs * sizeof(float), cudaMemcpyHostToDevice);
check_error(status);
free(input_cpu);
*/
//im2col_align_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, state.workspace, l.bit_align);
/*
float *input_cpu = (float *)calloc(input_size, sizeof(float));
float *input_bin_cpu = (float *)calloc(input_size, sizeof(char));
//float *weights_bin_cpu = (float *)calloc(l.n*l.c*l.size*l.size, sizeof(char));
status = cudaMemcpy(input_cpu, state.input, input_size * sizeof(float), cudaMemcpyDeviceToHost);
check_error(status);
float_to_bit(input_cpu, (unsigned char *)input_bin_cpu, input_size);
//float_to_bit(l.weights, (unsigned char *)weights_bin_cpu, l.n*l.c*l.size*l.size); // l.align_bit_weights
transpose_bin_gpu((unsigned char *)state.workspace, (unsigned char *)l.transposed_align_workspace_gpu, k, n, l.bit_align, new_ldb, 8);
//cudaDeviceSynchronize();
convolve_bin_cpu(input_bin_cpu, (float *)l.align_bit_weights, l.output, l.w, l.h, l.c, l.n, l.size, l.pad, l.new_lda, l.mean_arr); // CPU
status = cudaMemcpy(l.output_gpu, l.output, l.outputs * sizeof(float), cudaMemcpyHostToDevice);
check_error(status);
//free(weights_bin_cpu);
free(input_bin_cpu);
free(input_cpu);
*/
// should be optimized
gemm_nn_custom_bin_mean_transposed_gpu(m, n, k,
(unsigned char *)l.align_bit_weights_gpu, new_ldb, (unsigned char *)l.transposed_align_workspace_gpu, new_ldb, l.output_gpu, n, l.mean_arr_gpu);
//cudaDeviceSynchronize();
//check_error(status);
}
/*
{
float_to_bit_gpu(state.input, (unsigned char *)l.align_workspace_gpu, input_size);
convolve_bin_gpu(l.align_workspace_gpu, (float *)l.align_bit_weights_gpu, l.output_gpu, l.w, l.h, l.c, l.n, l.size, l.pad, l.new_lda, l.mean_arr_gpu);
//convolve_gpu(state.input, l.weights_gpu, l.output_gpu, l.w, l.h, l.c, l.n, l.size, l.pad);
//cudaDeviceSynchronize();
//check_error(status);
}
*/
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, l.out_w*l.out_h);
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);