From ab75d5c5783db4792e400f933d584984f3aa7bf0 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 20 Jun 2016 05:39:37 +0000
Subject: [PATCH] t1
---
src/convolutional_kernels.cu | 104 +++++++++++++++++++++++++++++++++++++++------------
1 files changed, 79 insertions(+), 25 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 62d6079..2376835 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -72,10 +72,6 @@
void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
int i;
- int m = l.n;
- int k = l.size*l.size*l.c;
- int n = convolutional_out_height(l)*
- convolutional_out_width(l);
fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
if(l.binary){
@@ -85,69 +81,127 @@
if(l.xnor){
binarize_filters_gpu(l.filters_gpu, l.n, l.c*l.size*l.size, l.binary_filters_gpu);
- //binarize_gpu(l.filters_gpu, l.n*l.c*l.size*l.size, l.binary_filters_gpu);
swap_binary(&l);
- for(i = 0; i < l.batch; ++i){
- binarize_input_gpu(state.input + i*l.inputs, l.c, l.h*l.w, l.binary_input_gpu + i*l.inputs);
- }
+ binarize_gpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input_gpu);
state.input = l.binary_input_gpu;
}
+#ifdef CUDNN
+ float one = 1;
+ cudnnConvolutionForward(cudnn_handle(),
+ &one,
+ l.srcTensorDesc,
+ state.input,
+ l.filterDesc,
+ l.filters_gpu,
+ l.convDesc,
+ l.fw_algo,
+ state.workspace,
+ l.workspace_size,
+ &one,
+ l.dstTensorDesc,
+ l.output_gpu);
+
+#else
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = l.out_w*l.out_h;
for(i = 0; i < l.batch; ++i){
- im2col_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, l.col_image_gpu);
+ im2col_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, state.workspace);
float * a = l.filters_gpu;
- float * b = l.col_image_gpu;
+ float * b = state.workspace;
float * c = l.output_gpu;
gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
}
+#endif
if (l.batch_normalize) {
forward_batchnorm_layer_gpu(l, state);
}
- add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, n);
+ 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, m*n*l.batch, l.activation);
+ activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
//if(l.dot > 0) dot_error_gpu(l);
if(l.binary || l.xnor) swap_binary(&l);
}
void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
{
- int i;
- int m = l.n;
- int n = l.size*l.size*l.c;
- int k = convolutional_out_height(l)*
- convolutional_out_width(l);
+ gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
- gradient_array_ongpu(l.output_gpu, m*k*l.batch, l.activation, l.delta_gpu);
-
- backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.n, k);
+ backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.n, l.out_w*l.out_h);
if(l.batch_normalize){
backward_batchnorm_layer_gpu(l, state);
}
+ float *original_input = state.input;
if(l.xnor) state.input = l.binary_input_gpu;
+#ifdef CUDNN
+ float one = 1;
+ cudnnConvolutionBackwardFilter(cudnn_handle(),
+ &one,
+ l.srcTensorDesc,
+ state.input,
+ l.ddstTensorDesc,
+ l.delta_gpu,
+ l.convDesc,
+ l.bf_algo,
+ state.workspace,
+ l.workspace_size,
+ &one,
+ l.dfilterDesc,
+ l.filter_updates_gpu);
+
+ if(state.delta){
+ if(l.binary || l.xnor) swap_binary(&l);
+ cudnnConvolutionBackwardData(cudnn_handle(),
+ &one,
+ l.filterDesc,
+ l.filters_gpu,
+ l.ddstTensorDesc,
+ l.delta_gpu,
+ l.convDesc,
+ l.bd_algo,
+ state.workspace,
+ l.workspace_size,
+ &one,
+ l.dsrcTensorDesc,
+ state.delta);
+ if(l.binary || l.xnor) swap_binary(&l);
+ if(l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
+ }
+
+#else
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = l.out_w*l.out_h;
+
+ int i;
for(i = 0; i < l.batch; ++i){
float * a = l.delta_gpu;
- float * b = l.col_image_gpu;
+ float * b = state.workspace;
float * c = l.filter_updates_gpu;
- im2col_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, l.col_image_gpu);
+ im2col_ongpu(state.input + i*l.c*l.h*l.w, l.c, l.h, l.w, l.size, l.stride, l.pad, state.workspace);
gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
if(l.binary || l.xnor) swap_binary(&l);
float * a = l.filters_gpu;
float * b = l.delta_gpu;
- float * c = l.col_image_gpu;
+ float * c = state.workspace;
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
- 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);
- if(l.binary || l.xnor) swap_binary(&l);
+ col2im_ongpu(state.workspace, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta + i*l.c*l.h*l.w);
+ if(l.binary || l.xnor) {
+ swap_binary(&l);
+ }
+ if(l.xnor) gradient_array_ongpu(original_input + i*l.c*l.h*l.w, l.c*l.h*l.w, HARDTAN, state.delta + i*l.c*l.h*l.w);
}
}
+#endif
}
void pull_convolutional_layer(convolutional_layer layer)
--
Gitblit v1.10.0