Skip to content

GitLab

Commit 3bd51021 authored by Chao Liu's avatar Chao Liu
Browse files

tune implicit_gemm

parent 216e3da6
Hide whitespace changes
Inline Side-by-side
Showing with 26 additions and 26 deletions
driver/conv.cu
View file @ 3bd51021
...@@ -404,7 +404,7 @@ int main() ...@@ -404,7 +404,7 @@ int main()
#elif 1 #elif 1
device_implicit_gemm_convolution_nchw_kcsr( device_implicit_gemm_convolution_nchw_kcsr(
in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device); in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device);
#elif 0 #elif 1
device_implicit_gemm_convolution_nchw_srck( device_implicit_gemm_convolution_nchw_srck(
in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device); in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device);
#elif 0 #elif 0
......
driver/device_implicit_gemm_convolution_nchw_kcsr.cuh
View file @ 3bd51021
...@@ -38,20 +38,20 @@ void device_implicit_gemm_convolution_nchw_kcsr( ...@@ -38,20 +38,20 @@ void device_implicit_gemm_convolution_nchw_kcsr(
constexpr unsigned WoPerThread = 2; constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 16; constexpr unsigned BlockSize = 16;
#elif 0 #elif 1
constexpr unsigned NPerBlock = 2; constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 32; constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 4; constexpr unsigned CPerBlock = 2;
constexpr unsigned HoPerBlock = 2; constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32; constexpr unsigned WoPerBlock = 32;
constexpr unsigned KPerThread = 4; constexpr unsigned KPerThread = 16;
constexpr unsigned CPerThread = 2; constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 2; constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2; constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 128; constexpr unsigned BlockSize = 128;
#elif 1 #elif 0
constexpr unsigned NPerBlock = 1; constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 64; constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 4; constexpr unsigned CPerBlock = 4;
......
driver/device_implicit_gemm_convolution_nchw_srck.cuh
View file @ 3bd51021
...@@ -63,32 +63,32 @@ void device_implicit_gemm_convolution_nchw_srck(InDesc, ...@@ -63,32 +63,32 @@ void device_implicit_gemm_convolution_nchw_srck(InDesc,
constexpr unsigned WoPerThread = 2; constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 16; constexpr unsigned BlockSize = 16;
#elif 1 #elif 0
constexpr unsigned NPerBlock = 2; constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 32; constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 4; constexpr unsigned CPerBlock = 2;
constexpr unsigned HoPerBlock = 2; constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32; constexpr unsigned WoPerBlock = 32;
constexpr unsigned KPerThread = 4; constexpr unsigned KPerThread = 8;
constexpr unsigned CPerThread = 2; constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 2; constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2; constexpr unsigned WoPerThread = 4;
constexpr unsigned BlockSize = 128; constexpr unsigned BlockSize = 128;
#elif 0 #elif 1
constexpr unsigned NPerBlock = 2; constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 64; constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 4; constexpr unsigned CPerBlock = 2;
constexpr unsigned HoPerBlock = 2; constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32; constexpr unsigned WoPerBlock = 32;
constexpr unsigned KPerThread = 4; constexpr unsigned KPerThread = 16;
constexpr unsigned CPerThread = 2; constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 2; constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2; constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 256; constexpr unsigned BlockSize = 128;
#endif #endif
constexpr unsigned GridSize = constexpr unsigned GridSize =
......
src/include/gridwise_implicit_gemm_convolution_nchw_kcsr.cuh
View file @ 3bd51021
...@@ -185,7 +185,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc, ...@@ -185,7 +185,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
in_nchw_block_desc.GetLengths()); in_nchw_block_desc.GetLengths());
#endif #endif
#if 0 #if 1
// weight: global mem to LDS, // weight: global mem to LDS,
// convert [K,C,S,R] to [S,R,C,K] // convert [K,C,S,R] to [S,R,C,K]
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>( blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>(
...@@ -238,7 +238,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc, ...@@ -238,7 +238,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
const unsigned k_thread_data_begin = matrix_c_index.row_begin; const unsigned k_thread_data_begin = matrix_c_index.row_begin;
const unsigned wo_thread_data_begin = matrix_c_index.col_begin / NPerThread; const unsigned wo_thread_data_begin = matrix_c_index.col_begin / NPerThread;
#if 0 #if 1
// output: register to global mem, // output: register to global mem,
// convert out_thread[Ho,K,Wo,N] to out_global[N,K,Ho,Wo] // convert out_thread[Ho,K,Wo,N] to out_global[N,K,Ho,Wo]
constexpr auto reorder_nkhw_from_hkwn = Sequence<3, 1, 0, 2>{}; constexpr auto reorder_nkhw_from_hkwn = Sequence<3, 1, 0, 2>{};
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment