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

Merge remote-tracking branch 'origin/master' into add_padding

Browse Source
This commit is contained in:
Chao Liu
2019-08-15 13:48:45 -05:00
parent 4fb81e008c 0979fb4af9
commit 86ceded98b
8 changed files with 82 additions and 240 deletions
Download Patch File Download Diff File Expand all files Collapse all files

64
driver/include/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp
View File

@@ -112,38 +112,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
// for 3x3, 34x34, v1r2, Pascal, in-block-copy1
constexpr index_t BlockSize = 128;
constexpr index_t NPerBlock = 4;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 8;
constexpr index_t NPerThread = 4;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 2;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
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 InBlockCopyClusterLengths_CHWN = Sequence<0, 0, 0, 0>; // not used
constexpr index_t InBlockCopyDataPerAccess_N = 4;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
#elif 1
// for 3x3, 34x34, v1r3, Pascal
// for 3x3, 28x28, v1r3, Pascal
// for 3x3, 14x14, v1r3, Pascal
@@ -179,37 +148,6 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
#elif 0
// for 3x3, 34x34, v1r3, Pascal, bad
constexpr index_t BlockSize = 128;
constexpr index_t NPerBlock = 1;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 32;
constexpr index_t NPerThread = 1;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 8;
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 InBlockCopyClusterLengths_CHWN = Sequence<2, 2, 32, 1>;
constexpr index_t InBlockCopyDataPerAccess_N = 1;
constexpr index_t WeiBlockCopyDataPerAccess_K = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 1;
#elif 0
// for 3x3, 34x34, v1r1, Vega 20
constexpr index_t BlockSize = 256;

14
driver/include/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded.hpp
View File

@@ -81,9 +81,9 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(InDesc,
#if 1
// v1r3, 3x3, 32x32, 1x1 pad
constexpr index_t BlockSize = 128;
constexpr index_t BlockSize = 256;
constexpr index_t NPerBlock = 16;
constexpr index_t NPerBlock = 32;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 2;
@@ -97,7 +97,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(InDesc,
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
@@ -105,14 +105,14 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(InDesc,
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_CHWN = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 2, 4>;
using InBlockCopyClusterLengths_CHWN = Sequence<8, 2, 2, 8>;
constexpr index_t InBlockCopyDataPerAccess_N = 4;
using WeiBlockCopySubLengths_CK = Sequence<2, 4>;
using WeiBlockCopyClusterLengths_CK = Sequence<4, 32>;
using WeiBlockCopySubLengths_CK = Sequence<1, 4>;
using WeiBlockCopyClusterLengths_CK = Sequence<8, 32>;
constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
constexpr index_t OutThreadCopyDataPerAccess_N = 2;
constexpr index_t OutThreadCopyDataPerAccess_N = 4;
#endif
constexpr index_t GridSize =

37
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -91,43 +91,6 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#elif 1
// each thread hold 64 data
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
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 = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 2, 2>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 8, 2>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 2;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 2;
using WeiBlockCopySubLengths_E_K = Sequence<2, 2>;
using WeiBlockCopyClusterLengths_E_K = Sequence<4, 64>;
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 = 2;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 2;
#elif 0
// each thread hold 32 data
constexpr index_t BlockSize = 256;

2
driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
View File

@@ -54,7 +54,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 0
#if 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 128;