From d0b9326a352ed2fbc3ae66fdef40b4533a2f211d Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 11 Aug 2015 06:22:27 +0000
Subject: [PATCH] Hacks to get nightmare to not break gridsizing
---
src/convolutional_kernels.cu | 24 +++++++++++-------------
1 files changed, 11 insertions(+), 13 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 9f0a2f8..a150c20 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -17,16 +17,16 @@
if(offset < size) output[(batch*n+filter)*size + offset] = biases[filter];
}
-extern "C" void bias_output_gpu(float *output, float *biases, int batch, int n, int size)
+void bias_output_gpu(float *output, float *biases, int batch, int n, int size)
{
- dim3 dimBlock(BLOCK, 1, 1);
dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
+ dim3 dimBlock(BLOCK, 1, 1);
bias_output_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
check_error(cudaPeekAtLastError());
}
-__global__ void backward_bias_kernel(float *bias_updates, float *delta, int batch, int n, int size, float scale)
+__global__ void backward_bias_kernel(float *bias_updates, float *delta, int batch, int n, int size)
{
__shared__ float part[BLOCK];
int i,b;
@@ -42,17 +42,17 @@
part[p] = sum;
__syncthreads();
if(p == 0){
- for(i = 0; i < BLOCK; ++i) bias_updates[filter] += scale * part[i];
+ for(i = 0; i < BLOCK; ++i) bias_updates[filter] += part[i];
}
}
-extern "C" void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
+void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
{
- backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size, 1);
+ backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size);
check_error(cudaPeekAtLastError());
}
-extern "C" void forward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
+void forward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
{
int i;
int m = layer.n;
@@ -71,7 +71,7 @@
activate_array_ongpu(layer.output_gpu, m*n*layer.batch, layer.activation);
}
-extern "C" void backward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
+void backward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
{
int i;
int m = layer.n;
@@ -82,8 +82,6 @@
gradient_array_ongpu(layer.output_gpu, m*k*layer.batch, layer.activation, layer.delta_gpu);
backward_bias_gpu(layer.bias_updates_gpu, layer.delta_gpu, layer.batch, layer.n, k);
- if(state.delta) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, state.delta, 1);
-
for(i = 0; i < layer.batch; ++i){
float * a = layer.delta_gpu;
float * b = layer.col_image_gpu;
@@ -105,7 +103,7 @@
}
}
-extern "C" void pull_convolutional_layer(convolutional_layer layer)
+void pull_convolutional_layer(convolutional_layer layer)
{
cuda_pull_array(layer.filters_gpu, layer.filters, layer.c*layer.n*layer.size*layer.size);
cuda_pull_array(layer.biases_gpu, layer.biases, layer.n);
@@ -113,7 +111,7 @@
cuda_pull_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
}
-extern "C" void push_convolutional_layer(convolutional_layer layer)
+void push_convolutional_layer(convolutional_layer layer)
{
cuda_push_array(layer.filters_gpu, layer.filters, layer.c*layer.n*layer.size*layer.size);
cuda_push_array(layer.biases_gpu, layer.biases, layer.n);
@@ -121,7 +119,7 @@
cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
}
-extern "C" void update_convolutional_layer_gpu(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
+void update_convolutional_layer_gpu(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
{
int size = layer.size*layer.size*layer.c*layer.n;
--
Gitblit v1.10.0