Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-11-06 01:34:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix softmax sum kernel

Browse Source
This commit is contained in:
Matúš Námešný
2024-04-08 22:09:18 +02:00
parent 7bc329a043
commit e419a93408
5 changed files with 49 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/kernels/activation_functions.cu
View File

@@ -44,15 +44,14 @@ __global__ void Kernels::softmax_exp(
__global__ void Kernels::softmax_sum(
const float* __restrict__ d_vector,
float* __restrict__ d_output,
const unsigned int w
float* __restrict__ d_output
) {
__shared__ float partial_sum[BLOCK_SIZE];
int i = blockIdx.x * blockDim.x * 2 + threadIdx.x;
partial_sum[threadIdx.x] = d_vector[i] + d_vector[i + blockDim.x];
int i = blockIdx.x * blockDim.x + threadIdx.x;
partial_sum[threadIdx.x] = d_vector[i];
__syncthreads();
for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {
for (int s = blockDim.x / 2; s > 0; s >>= 1) {
if (threadIdx.x < s) {
partial_sum[threadIdx.x] += partial_sum[threadIdx.x + s];
}

6
src/layers/activation.cu
View File

@@ -42,12 +42,12 @@ void Activation::activate(float* __restrict__ d_input) {
);
Kernels::softmax_sum<<<gridSize, BLOCK_SIZE>>>(
d_input, d_softmax_sum, length
d_input, d_softmax_sum
);
Kernels::softmax_sum<<<1, BLOCK_SIZE>>>(
d_softmax_sum, d_softmax_sum, length
);
d_softmax_sum, d_softmax_sum
);
Kernels::softmax_div<<<gridSize, BLOCK_SIZE>>>(
d_input, d_input, d_softmax_sum, length