ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

debugging ds_read asm

Browse Source
This commit is contained in:
Chao Liu
2019-04-26 15:34:55 -05:00
parent b93d2e1b57
commit 3ce77700b6
6 changed files with 118 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
driver/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp
View File

@@ -475,7 +475,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
#elif 1
#elif 0
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
#elif 1
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn

109
driver/device_convolution_implicit_gemm_v1_nchw_cyxk_khwn.hpp
View File

@@ -65,6 +65,76 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 0
// for 3x3, 34x34, v1r3, Pascal
constexpr index_t BlockSize = 128;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 16;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockReorderSrcSubLengths_NCHW = Sequence<2, 1, 2, 1>;
using InBlockReorderSrcClusterLengths_NCHW = Sequence<1, 8, 1, 16>;
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
constexpr index_t InBlockReorderDataPerWrite_N = 2;
using WeiBlockCopyClusterLengths = Sequence<0, 0>; // not used
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
#elif 1
// for 3x3, 34x34, v1r3, Vega 20
constexpr index_t BlockSize = 256;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 16;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockReorderSrcSubLengths_NCHW = Sequence<2, 1, 2, 1>;
using InBlockReorderSrcClusterLengths_NCHW = Sequence<1, 8, 2, 16>;
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
constexpr index_t InBlockReorderDataPerWrite_N = 2;
using WeiBlockCopyClusterLengths = Sequence<0, 0>; // not used
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
#elif 0
// for 3x3, 28x28, v1r2, Pascal
constexpr index_t BlockSize = 128;
@@ -133,41 +203,6 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
using WeiBlockCopyClusterLengths = Sequence<0, 0>; // not used
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
#elif 1
// for 3x3, 34x34, v1r3, Pascal
constexpr index_t BlockSize = 128;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 16;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockReorderSrcSubLengths_NCHW = Sequence<2, 1, 2, 1>;
using InBlockReorderSrcClusterLengths_NCHW = Sequence<1, 8, 1, 16>;
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
constexpr index_t InBlockReorderDataPerWrite_N = 2;
using WeiBlockCopyClusterLengths = Sequence<0, 0>; // not used
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
#endif
@@ -182,9 +217,9 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
constexpr auto gridwise_conv =
#if 0
GridwiseConvolutionImplicitGemm_v1r2_nchw_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_khwn
#elif 1
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
#endif
<GridSize,

2
driver/driver.hip.cpp
View File

@@ -606,7 +606,7 @@ int main(int argc, char* argv[])
device_direct_convolution_2_nchw_kcyx_nkhw
#elif 0
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
#elif 1
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn
#elif 1
device_convolution_implicit_gemm_v1_nchw_cyxk_khwn