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 | 14 ++++++--------
1 files changed, 6 insertions(+), 8 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 5b49091..a150c20 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -11,22 +11,22 @@
__global__ void bias_output_kernel(float *output, float *biases, int n, int size)
{
int offset = blockIdx.x * blockDim.x + threadIdx.x;
- int filter = blockIdx.y % n;
- int batch = blockIdx.y / n;
+ int filter = blockIdx.y;
+ int batch = blockIdx.z;
if(offset < size) output[(batch*n+filter)*size + offset] = biases[filter];
}
void bias_output_gpu(float *output, float *biases, int batch, int n, int size)
{
- dim3 dimGrid((size-1)/BLOCK + 1, n*batch, 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;
@@ -42,13 +42,13 @@
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];
}
}
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());
}
@@ -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;
--
Gitblit v1.10.0