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

use ford/for instead of static_ford/static_for in threadwise copy, somehow register spill is greatly reduced on AMD

Browse Source
This commit is contained in:
Chao Liu
2019-08-07 19:09:13 -05:00
parent 5636576f9b
commit bc9ea646f8
7 changed files with 122 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -112,14 +112,14 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 4, 1>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 4, 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 = 4;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 1;
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>;