Skip to content

GitLab

Commit e7746f82 authored by ltqin's avatar ltqin
Browse files

fix tuning prometer for fp16

parent e790467d
Show whitespace changes
Inline Side-by-side
Showing with 10 additions and 10 deletions
host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
View file @ e7746f82
...@@ -66,13 +66,13 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nh ...@@ -66,13 +66,13 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nh
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4; constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4; constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>; using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>; using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4; constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
...@@ -120,17 +120,17 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nh ...@@ -120,17 +120,17 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nh
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8; constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8; constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>; using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>; using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8; constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif #endif
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment