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

clean up

Browse Source
This commit is contained in:
Chao Liu
2019-08-13 00:37:23 -05:00
parent fab2f10a55
commit 8bdaba51f8
20 changed files with 505 additions and 651 deletions
Download Patch File Download Diff File Expand all files Collapse all files

140
driver/include/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp
View File

@@ -107,11 +107,11 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<4, 4, 2, 4>;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
// for 3x3, 34x34, v1r2, Pascal, in-block-copy1
constexpr index_t BlockSize = 128;
@@ -137,12 +137,12 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<0, 0, 0, 0>; // not used
constexpr index_t InBlockCopyDataPerRead_N = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<0, 0, 0, 0>; // not used
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 1
// for 3x3, 34x34, v1r3, Pascal
// for 3x3, 28x28, v1r3, Pascal
@@ -170,12 +170,15 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 2, 4>;
constexpr index_t InBlockCopyDataPerRead_N = 4;
using InBlockCopySubLengths_CHWN = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 2, 4>;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
using WeiBlockCopySubLengths_CK = Sequence<2, 4>;
using WeiBlockCopyClusterLengths_CK = Sequence<4, 32>;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
// for 3x3, 34x34, v1r3, Pascal, bad
constexpr index_t BlockSize = 128;
@@ -201,12 +204,12 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<2, 2, 32, 1>;
constexpr index_t InBlockCopyDataPerRead_N = 1;
using InBlockCopyClusterLengths_CHWN = Sequence<2, 2, 32, 1>;
constexpr index_t InBlockCopyDataPerAccess_N = 1;
constexpr index_t WeiBlockCopyDataPerRead_K = 2;
constexpr index_t WeiBlockCopyDataPerAccess_K = 2;
constexpr index_t OutThreadCopyDataPerWrite_N = 1;
constexpr index_t OutThreadCopyDataPerAccess_N = 1;
#elif 0
// for 3x3, 34x34, v1r1, Vega 20
constexpr index_t BlockSize = 256;
@@ -232,12 +235,12 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<4, 4, 2, 8>;
constexpr index_t InBlockCopyDataPerRead_N = 2;
using InBlockCopyClusterLengths_CHWN = Sequence<4, 4, 2, 8>;
constexpr index_t InBlockCopyDataPerAccess_N = 2;
constexpr index_t WeiBlockCopyDataPerRead_K = 2;
constexpr index_t WeiBlockCopyDataPerAccess_K = 2;
constexpr index_t OutThreadCopyDataPerWrite_N = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 4;
#elif 1
// for 3x3, 34x34, v1r3, Vega 20
constexpr index_t BlockSize = 256;
@@ -263,12 +266,12 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 4, 4>;
constexpr index_t InBlockCopyDataPerRead_N = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 4, 4>;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 4;
#elif 0
// for 3x3, 56x56, v1r1, Pascal
constexpr index_t NPerBlock = 32;
@@ -282,13 +285,13 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 2;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
@@ -298,7 +301,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
constexpr index_t BlockSize = 128;
#elif 0
@@ -324,14 +327,14 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 1;
constexpr index_t GemmDataPerReadB = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 4;
constexpr index_t BlockSize = 128;
#elif 0
@@ -347,13 +350,13 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 2;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
@@ -365,7 +368,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
constexpr index_t BlockSize = 128;
#elif 0
@@ -393,12 +396,12 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<4, 2, 4, 4>;
constexpr index_t InBlockCopyDataPerRead_N = 4;
using InBlockCopyClusterLengths_CHWN = Sequence<4, 2, 4, 4>;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
// for 1x1, 28x28, v1r1, Pascal
constexpr index_t NPerBlock = 16;
@@ -413,13 +416,13 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
@@ -429,7 +432,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
constexpr index_t BlockSize = 128;
#elif 0
@@ -453,14 +456,14 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerRead_N = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
constexpr index_t BlockSize = 128;
#endif
@@ -478,9 +481,9 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
#elif 1
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
#elif 1
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
#endif
<GridSize,
@@ -507,10 +510,13 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_CHWN,
InBlockCopyClusterLengths_CHWN,
InBlockCopyDataPerRead_N,
WeiBlockCopyDataPerRead_K,
OutThreadCopyDataPerWrite_N>{};
InBlockCopyDataPerAccess_N,
WeiBlockCopySubLengths_CK,
WeiBlockCopyClusterLengths_CK,
WeiBlockCopyDataPerAccess_K,
OutThreadCopyDataPerAccess_N>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),