ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 17:26:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

experimenting global and buffer load/store

Browse Source
This commit is contained in:
Chao Liu
2019-09-18 00:15:57 -05:00
parent f58bf38445
commit 9f46cdf5fa
12 changed files with 503 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -48,7 +48,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 1
// each thread hold 64 data
// BlockSize = 256, blockwise-GEMM 128x128, each thread hold 64 data
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
@@ -82,10 +82,47 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopySrcDataPerRead_E = 1;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#elif 0
// each thread hold 32 data
// BlockSize = 64, blockwise-GEMM 64x64, each thread hold 64 data
constexpr index_t BlockSize = 64;
constexpr index_t BPerBlock = 8;
constexpr index_t KPerBlock = 64;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmNRepeat = 2;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 2, 1, 4>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 1, 8, 1>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 2, 1, 3>; // [E, B, N1, N2]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 1;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
using WeiBlockCopySubLengths_E_K = Sequence<4, 2>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 32>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#elif 1
// BlockSize = 256, blockwise-GEMM 64x128, each thread hold 32 data
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
@@ -107,7 +144,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 16, 1>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 2, 1, 3>; // [E, B, N1, N2]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 1;
@@ -119,7 +156,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 2;
constexpr index_t WeiBlockCopySrcDataPerRead_E = 1;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#endif
@@ -133,50 +170,50 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
constexpr auto gridwise_conv =
#if 0
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
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
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
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),