| | |
| | | __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; |
| | |
| | | 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()); |
| | | } |
| | | |