ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 04:49:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

removing dependency on old tensor descriptor

Browse Source 
[ROCm/composable_kernel commit: 51a9fa1dbd]
This commit is contained in:
Chao Liu
2019-09-26 11:49:05 -05:00
parent eb0a575ad1
commit bd8f263b70
10 changed files with 162 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

88
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp
View File

@@ -177,52 +177,52 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded(InDesc,
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
constexpr auto gridwise_conv =
#if 0
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
BPerBlock,
KPerBlock,
EPerBlock,
GemmNRepeat,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>{};
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
#if 0
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
BPerBlock,
KPerBlock,
EPerBlock,
GemmNRepeat,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),
dim3(BlockSize),

2
driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded.hpp
View File

@@ -51,7 +51,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded(InDesc,
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 1
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 128;