added implicit_gemm_v1 lds double_buffer

This commit is contained in:
Chao Liu
2019-04-08 14:11:55 -05:00
parent c075d3f7d9
commit cc0fa73acd
8 changed files with 594 additions and 14 deletions

View File

@@ -3,6 +3,7 @@
#include "device.hpp"
#include "gridwise_convolution_wrapper.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
@@ -289,7 +290,12 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn<GridSize,
#if 0
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn
#else
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_chwn_desc),

View File

@@ -189,7 +189,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t BlockSize = 256;
#elif 0
#elif 1
// 1x1, 14x14, Pascal, enable lds_double_buffer, disable register double buffer
constexpr index_t BPerBlock = 64;
constexpr index_t KPerBlock = 128;
@@ -217,7 +217,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
#elif 1
#elif 0
// 1x1, 14x14, Vega 20, enable lds_double_buffer, disable register_double_buffer
constexpr index_t BPerBlock = 128;
constexpr index_t KPerBlock = 128;

View File

@@ -663,7 +663,7 @@ int main(int argc, char* argv[])
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
#elif 1
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
#elif 0
#elif 1
device_implicit_gemm_convolution_2_chwn_cyxk_khwn
#endif
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);