OrgLion-ML
/
darknet
mirror of https://github.com/AlexeyAB/darknet.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Increased speed on Tensor Cores for csresnext50-panet-spp.cfg model

This commit is contained in:
AlexeyAB 2019-12-20 16:07:34 +03:00
parent 3004ee851c
commit dc7277f152
2 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/convolutional_kernels.cu
View File

@ -420,7 +420,7 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
//if (state.use_mixed_precision) {
int iteration_num = (*state.net.seen) / (state.net.batch*state.net.subdivisions);
if (state.index != 0 && state.net.cudnn_half && !l.xnor && (!state.train || iteration_num > 3*state.net.burn_in) &&
(l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train)
(l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups == 1)
{
//printf("\n CUDNN_HALF!!! state.index = %d \n", state.index);
@ -671,7 +671,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
//#ifdef CUDNN_HALF
int iteration_num = (*state.net.seen) / (state.net.batch*state.net.subdivisions);
if (state.index != 0 && state.net.cudnn_half && !l.xnor && (!state.train || iteration_num > 3*state.net.burn_in) &&
(l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train)
(l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups == 1)
{
const size_t input16_size = l.batch*l.c*l.w*l.h;
const size_t delta16_size = l.batch*l.n*l.out_w*l.out_h;

1
src/dark_cuda.c
View File

@ -116,6 +116,7 @@ static int streamInit[16] = { 0 };
cudaStream_t get_cuda_stream() {
int i = cuda_get_device();
if (!streamInit[i]) {
//printf("Create CUDA-stream \n");
cudaError_t status = cudaStreamCreate(&streamsArray[i]);
//cudaError_t status = cudaStreamCreateWithFlags(&streamsArray[i], cudaStreamNonBlocking);
if (status != cudaSuccess) {