ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

enabling vector load on merged dim

Browse Source
This commit is contained in:
Chao Liu
2019-06-24 11:20:19 -05:00
parent 37b82b7e54
commit df29a7e097
3 changed files with 71 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
View File

@@ -59,7 +59,7 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
#if 1
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
@@ -91,6 +91,40 @@ void device_convolution_implicit_gemm_v4_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 WeiBlockCopyDstDataPerWrite_K = 1;
#elif 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
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<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
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;
#endif