Skip to content

GitLab

Commit b725e3fc authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into miopen_downstream-dynamic_reduction_pr

parents df0d6810 f3acd251
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2002 additions and 829 deletions
composable_kernel/include/utility/tuple_helper.hpp
View file @ b725e3fc
...@@ -14,9 +14,7 @@ struct is_known_at_compile_time<Tuple<Ts...>> ...@@ -14,9 +14,7 @@ struct is_known_at_compile_time<Tuple<Ts...>>
return container_reduce( return container_reduce(
Tuple<Ts...>{}, Tuple<Ts...>{},
[](auto x, bool r) { [](auto x, bool r) {
return is_known_at_compile_time< return is_known_at_compile_time<remove_cvref_t<decltype(x)>>::value & r;
remove_cv_t<remove_reference_t<decltype(x)>>>::value &
r;
}, },
true); true);
} }
......
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw.cpp
View file @ b725e3fc
...@@ -374,13 +374,8 @@ extern "C" __global__ void ...@@ -374,13 +374,8 @@ extern "C" __global__ void
CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1{}, CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1{},
CGridBlockCluster_BlockId_To_GM10_GN10{})); CGridBlockCluster_BlockId_To_GM10_GN10{}));
const auto desc_tuple = *reinterpret_cast<const DescTuple*>( const auto desc_tuple =
#pragma clang diagnostic push *reinterpret_cast<const DescTuple*>(cast_pointer_to_generic_address_space(p_desc_tuple));
#pragma clang diagnostic ignored "-Wold-style-cast"
// TODO: how to cast?
(const void*)p_desc_tuple
#pragma clang diagnostic pop
);
const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = desc_tuple[I0]; const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = desc_tuple[I0];
const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = desc_tuple[I1]; const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = desc_tuple[I1];
......
host/driver_offline/CMakeLists.txt
View file @ b725e3fc
...@@ -13,9 +13,15 @@ include_directories(BEFORE ...@@ -13,9 +13,15 @@ include_directories(BEFORE
set(CONV_FWD_DRIVER_OFFLINE_SOURCE src/conv_fwd_driver_offline.cpp) set(CONV_FWD_DRIVER_OFFLINE_SOURCE src/conv_fwd_driver_offline.cpp)
set(CONV_BWD_DRIVER_OFFLINE_SOURCE src/conv_bwd_driver_offline.cpp) set(CONV_BWD_DRIVER_OFFLINE_SOURCE src/conv_bwd_driver_offline.cpp)
set(CONV_WRW_DRIVER_OFFLINE_SOURCE src/conv_wrw_driver_offline.cpp)
set(GEMM_DRIVER_OFFLINE_SOURCE src/gemm_driver_offline.cpp)
add_executable(conv_fwd_driver_offline ${CONV_FWD_DRIVER_OFFLINE_SOURCE}) add_executable(conv_fwd_driver_offline ${CONV_FWD_DRIVER_OFFLINE_SOURCE})
add_executable(conv_bwd_driver_offline ${CONV_BWD_DRIVER_OFFLINE_SOURCE}) add_executable(conv_bwd_driver_offline ${CONV_BWD_DRIVER_OFFLINE_SOURCE})
add_executable(conv_wrw_driver_offline ${CONV_WRW_DRIVER_OFFLINE_SOURCE})
add_executable(gemm_driver_offline ${GEMM_DRIVER_OFFLINE_SOURCE})
target_link_libraries(conv_fwd_driver_offline PRIVATE host_tensor) target_link_libraries(conv_fwd_driver_offline PRIVATE host_tensor)
target_link_libraries(conv_bwd_driver_offline PRIVATE host_tensor) target_link_libraries(conv_bwd_driver_offline PRIVATE host_tensor)
target_link_libraries(conv_wrw_driver_offline PRIVATE host_tensor)
target_link_libraries(gemm_driver_offline PRIVATE host_tensor)
host/driver_offline/include/device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk.hpp
View file @ b725e3fc
...@@ -56,9 +56,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -56,9 +56,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4; constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -84,9 +84,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -84,9 +84,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -112,9 +112,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -112,9 +112,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -140,9 +140,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -140,9 +140,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 256; constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4; constexpr index_t NRepeat = 4;
...@@ -168,9 +168,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -168,9 +168,9 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4; constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -208,40 +208,42 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -208,40 +208,42 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
// HACK: hacks that control index calculation when iterating over A, B, C matrix // HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks = constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: gemmk0 make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: gemmm Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: gemmk1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: Gemmk0 make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: Gemmm Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: Gemmk1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple( constexpr auto out_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: gemmk0 make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0>{}, // 1+: gemmn Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: gemmk1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: gemmk0 make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0>{}, // 1-: gemmn Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: gemmk1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_m0_m1_m2_n_grid_step_hacks = make_tuple( // clang-format off
constexpr auto in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks = make_tuple(
make_tuple( make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: MRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 1+: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 3+: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}), // 7+: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple( make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: MRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 1-: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 3-: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})); // 7-: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
//clang-format on
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{};
...@@ -263,8 +265,8 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -263,8 +265,8 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
GemmMPerBlock, GemmMPerBlock,
GemmNPerBlock, GemmNPerBlock,
GemmKPerBlock, GemmKPerBlock,
GemmMPerWave, GemmMPerXDL,
GemmNPerWave, GemmNPerXDL,
GemmK1, GemmK1,
MRepeat, MRepeat,
NRepeat, NRepeat,
...@@ -289,7 +291,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -289,7 +291,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_step_hacks), decltype(out_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(in_m0_m1_m2_n_grid_step_hacks), decltype(in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks), decltype(out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false // CAccessOrderMRepeatNRepeat false // CAccessOrderMRepeatNRepeat
...@@ -301,7 +303,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk( ...@@ -301,7 +303,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
in_gemmm_gemmn_grid_desc, in_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks, wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_step_hacks, out_gemmk0_gemmn_gemmk1_grid_step_hacks,
in_m0_m1_m2_n_grid_step_hacks, in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks, wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks, out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat); nrepeat);
......
host/driver_offline/include/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk.hpp
View file @ b725e3fc
...@@ -195,25 +195,27 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk ...@@ -195,25 +195,27 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: Gemmn Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: Gemmn
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: Gemmk1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: Gemmk1
constexpr auto in_m0_m1_m2_n_grid_step_hacks = make_tuple( // clang-format off
constexpr auto in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks = make_tuple(
make_tuple( make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 0+: MRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 2+: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 4+: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 5+: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 6+: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple( make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 0-: MRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 2-: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 4-: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 5-: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 6-: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
//clang-format on
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0>{};
...@@ -265,7 +267,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk ...@@ -265,7 +267,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks), decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks), decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(in_m0_m1_m2_n_grid_step_hacks), decltype(in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks), decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks), decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
true // CAccessOrderMRepeatNRepeat true // CAccessOrderMRepeatNRepeat
...@@ -277,7 +279,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk ...@@ -277,7 +279,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
in_gemmm_gemmn_grid_desc, in_gemmm_gemmn_grid_desc,
out_gemmk0_gemmm_gemmk1_grid_step_hacks, out_gemmk0_gemmm_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_step_hacks, wei_gemmk0_gemmn_gemmk1_grid_step_hacks,
in_m0_m1_m2_n_grid_step_hacks, in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks, out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks, wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat); nrepeat);
......
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp → host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
View file @ b725e3fc
#include <unistd.h> #include <unistd.h>
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp" #include "transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
#include "driver_gemm_xdlops_v2r2.hpp" #include "driver_gemm_xdlops_v2r3.hpp"
template <typename TInWei, template <typename TInWei,
typename TAcc, typename TAcc,
...@@ -14,17 +14,17 @@ template <typename TInWei, ...@@ -14,17 +14,17 @@ template <typename TInWei,
typename ConvDilations, typename ConvDilations,
typename InLeftPads, typename InLeftPads,
typename InRightPads> typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk( void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const InLengths& in_n_hi_wi_c_lengths, const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_y_x_c_lengths, const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_ho_wo_k_lengths, const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides, const ConvStrides& conv_strides,
const ConvDilations& conv_dilations, const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads, const InLeftPads& in_left_pads,
const InRightPads& in_right_pads, const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_hi_wi_c, const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_y_x_c, Tensor<TInWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_ho_wo_k, const Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat) ck::index_t nrepeat)
{ {
using namespace ck; using namespace ck;
...@@ -34,138 +34,149 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk( ...@@ -34,138 +34,149 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk(
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{}; constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{}; constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace()); DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace()); DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace()); DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data()); in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data()); wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data()); out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths); const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths); const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths); const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1 #if 1
// [M, N, K0, K1] = [256, 128, 4, 4] for fp32 // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256; constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64; constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 64; constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1; constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
// using vector load 4, so config's wo*ho must be a multiple of 4
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4; constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
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, 8>;
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 = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1 #elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16 // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256; constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64; constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 64; constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 1; constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
// using vector load 4, so config's wo*ho must be a multiple of 4
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8; constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 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, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>; using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8; constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif #endif
const auto descs = const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(wei_k_y_x_c_desc, wei_k_c_y_x_desc,
in_n_hi_wi_c_desc, in_n_c_hi_wi_desc,
out_n_ho_wo_k_desc, out_n_k_ho_wo_desc,
conv_strides, conv_strides,
conv_dilations, conv_dilations,
in_left_pads, in_left_pads,
in_right_pads, in_right_pads,
Number<GemmK1>{}); Number<GemmK1>{});
const auto wei_gemmk0_gemmm_gemmk1_grid_desc = descs[I0]; const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1]; const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2]; const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix // HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple( constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}), make_tuple(make_tuple(Sequence<0, 0, 1, 0, 0>{}, // 0+: GemmK0
make_tuple( Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})); Sequence<0, 0, 1, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 2, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks = constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}), Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})); // 2-: GemmK1
constexpr auto out_m0_m1_m2_n_grid_step_hacks = constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 1, 0, 0>{}), Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 2, 0, 0>{})); Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 1, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks = constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0>{};
for(index_t i = 0; i < 5; ++i) for(index_t i = 0; i < 5; ++i)
{ {
float ave_time = driver_gemm_xdlops_v2r2< float ave_time = driver_gemm_xdlops_v2r3<
BlockSize, BlockSize,
TInWei, TInWei,
TAcc, TAcc,
TOut, TOut,
InMemoryDataOperationEnum_t::Set, InMemoryDataOperationEnum_t::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc), decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc), decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc), decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock, GemmMPerBlock,
GemmNPerBlock, GemmNPerBlock,
GemmKPerBlock, GemmKPerBlock,
GemmMPerWave, GemmMPerWave,
GemmNPerWave, GemmNPerWave,
GemmK1,
MRepeat, MRepeat,
NRepeat, NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1, GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
...@@ -181,49 +192,37 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk( ...@@ -181,49 +192,37 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk(
Sequence<1, 0, 2>, Sequence<1, 0, 2>,
Sequence<1, 0, 2>, Sequence<1, 0, 2>,
2, 2,
GemmBBlockTransferSrcScalarPerVector_GemmK1, GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1, GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1>, Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
2, 7,
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks), decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks), decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_m1_m2_n_grid_step_hacks), decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks), decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks)>( decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()), false>(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()), static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()), static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc, out_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc, in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc, wei_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks, out_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks, in_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_m1_m2_n_grid_step_hacks, wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks, out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks, in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat); nrepeat);
{ float perf = static_cast<float>(calculate_convolution_flops(
const auto N = out_n_ho_wo_k_lengths[I0]; in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
const auto K = out_n_ho_wo_k_lengths[I3]; (std::size_t(1000) * 1000 * 1000) / ave_time;
const auto C = wei_k_y_x_c_lengths[I3];
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = (float)(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
} }
// copy result back to host // copy result back to host
out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data()); wei_k_c_y_x_device_buf.FromDevice(wei_k_c_y_x.mData.data());
} }
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ df0d6810
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 1;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 4;
constexpr index_t MRepeat = 1;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#endif
const auto descs =
#if 1
transform_forward_convolution_into_gemm_v4r4_xdlops_nchw_kcyx_nkhw_pad
#else
transform_forward_convolution_into_gemm_v4r4_xdlops_nchw_kcyx_nkhw_1x1
#endif
<TInWei, GemmMPerBlock, GemmNPerBlock, GemmMPerWave, GemmNPerWave, GemmKPack>(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads);
for(index_t i = 0; i < 5; ++i)
{
#if 0
float ave_time = launch_kernel_gemm_xdlops_v1
#else
float ave_time = launch_kernel_gemm_xdlops_v2
#endif
<BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmKPack,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_KPack,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<0, 2, 1>,
Sequence<1, 0, 2>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_KPack,
false, // don't move back src coordinate after threadwise copy, which will be fused
// with MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
3,
GemmCThreadTransferDstScalarPerVector_GemmN1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (float)calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());
}
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
View file @ b725e3fc
...@@ -47,7 +47,35 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -47,7 +47,35 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths); const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths); const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1 #if 0
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16 // [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
...@@ -92,36 +120,39 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -92,36 +120,39 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const auto out_gemmm_gemmn_grid_desc = descs[I2]; const auto out_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix // HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple( constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}), make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
make_tuple( Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{})); Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks = constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}), Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_m0_m1_m2_n_grid_step_hacks = constexpr auto out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 1, 0, 0>{}, Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 1, 0, 0>{}, Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 1, 0, 0>{}), Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0>{}, make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 2, 0, 0>{}, Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 2, 0, 0>{}, Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 2, 0, 0>{})); Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{}; Sequence<0, 0, 0, 0, 0>{};
...@@ -169,7 +200,7 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -169,7 +200,7 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks), decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_m1_m2_n_grid_step_hacks), decltype(out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks), decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false>(static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()), false>(static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
...@@ -180,7 +211,7 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -180,7 +211,7 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
out_gemmm_gemmn_grid_desc, out_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks, wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks, in_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_m1_m2_n_grid_step_hacks, out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks, wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks, in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat); nrepeat);
......
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ df0d6810
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 1
// [M, N, K0, K1] = [256, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 0
// [M, N, K0, K1] = [256, 256, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto wei_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}));
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmK1,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
6,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_m1_m2_n_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false // CAccessOrderMRepeatNRepeat
>(static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_m1_m2_n_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Hi = in_n_hi_wi_c_lengths[I1];
const auto Wi = in_n_hi_wi_c_lengths[I2];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = (float)(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data());
}
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
View file @ b725e3fc
...@@ -56,8 +56,8 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -56,8 +56,8 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4; constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 4;
...@@ -84,9 +84,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -84,9 +84,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4; constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -112,9 +112,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -112,9 +112,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 256; constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 4; constexpr index_t NRepeat = 4;
...@@ -140,9 +140,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -140,9 +140,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -168,9 +168,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -168,9 +168,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 256; constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4; constexpr index_t NRepeat = 4;
...@@ -196,9 +196,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -196,9 +196,9 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32; constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
...@@ -249,23 +249,23 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -249,23 +249,23 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1 Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_m0_m1_m2_n_grid_step_hacks = constexpr auto out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: MRepeat make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0>{}, // 1+: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0>{}, // 2+: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0>{}, // 3+: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0>{}, // 4+: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0>{}, // 5+: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0>{}, // 6+: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0>{}), // 7+: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: MRepeat make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0>{}, // 1-: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0>{}, // 2-: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0>{}, // 3-: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0>{}, // 4-: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0>{}, // 5-: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0>{}, // 6-: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0>{})); // 7-: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
...@@ -287,8 +287,8 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -287,8 +287,8 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
GemmMPerBlock, GemmMPerBlock,
GemmNPerBlock, GemmNPerBlock,
GemmKPerBlock, GemmKPerBlock,
GemmMPerWave, GemmMPerXDL,
GemmNPerWave, GemmNPerXDL,
GemmK1, GemmK1,
MRepeat, MRepeat,
NRepeat, NRepeat,
...@@ -313,7 +313,7 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -313,7 +313,7 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(in_gemmk0_gemmm_gemmk1_grid_step_hacks), decltype(in_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks), decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_m1_m2_n_grid_step_hacks), decltype(out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks), decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks), decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false // CAccessOrderMRepeatNRepeat false // CAccessOrderMRepeatNRepeat
...@@ -325,7 +325,7 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -325,7 +325,7 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
out_gemmm_gemmn_grid_desc, out_gemmm_gemmn_grid_desc,
in_gemmk0_gemmm_gemmk1_grid_step_hacks, in_gemmk0_gemmm_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_step_hacks, wei_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_m1_m2_n_grid_step_hacks, out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks, in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks, wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat); nrepeat);
......
host/driver_offline/include/device_gemm_xdlops_km_kn_mn.hpp 0 → 100644
View file @ b725e3fc
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename ABType,
typename AccType,
typename CType,
typename ADesc,
typename BDesc,
typename CDesc>
void device_gemm_xdlops_km_kn_mn(const ADesc& a_k_m_grid_desc,
const BDesc& b_k_n_grid_desc,
const CDesc& c_m_n_grid_desc,
const Tensor<ABType>& a_k_m,
const Tensor<ABType>& b_k_n,
Tensor<CType>& c_m_n,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
DeviceMem a_k_m_device_buf(sizeof(ABType) * a_k_m.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(ABType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CType) * c_m_n.mDesc.GetElementSpace());
a_k_m_device_buf.ToDevice(a_k_m.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n.mData.data());
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t MPerBlock = 256;
constexpr index_t NPerBlock = 128;
constexpr index_t KPerBlock = 4;
constexpr index_t MPerXDL = 32;
constexpr index_t NPerXDL = 32;
constexpr index_t K1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using ABlockTransferThreadSliceLengths_K0_M_K1 = Sequence<1, 4, 4>;
using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<4, 64, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_M = 4;
constexpr index_t ABlockTransferDstScalarPerVector_K1 = 4;
using BBlockTransferThreadSliceLengths_K0_N_K1 = Sequence<1, 2, 4>;
using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<4, 64, 1>;
constexpr index_t BBlockTransferSrcScalarPerVector_N = 2;
constexpr index_t BBlockTransferDstScalarPerVector_K1 = 4;
constexpr index_t CThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t MPerBlock = 256;
constexpr index_t NPerBlock = 128;
constexpr index_t KPerBlock = 4;
constexpr index_t MPerXDL = 32;
constexpr index_t NPerXDL = 32;
constexpr index_t K1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using ABlockTransferThreadSliceLengths_K0_M_K1 = Sequence<1, 4, 8>;
using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<4, 64, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_M = 4;
constexpr index_t ABlockTransferDstScalarPerVector_K1 = 8;
using BBlockTransferThreadSliceLengths_K0_N_K1 = Sequence<1, 2, 8>;
using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<4, 64, 1>;
constexpr index_t BBlockTransferSrcScalarPerVector_N = 2;
constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
constexpr index_t CThreadTransferDstScalarPerVector = 1;
#endif
const auto K = a_k_m_grid_desc.GetLength(I0);
const auto M = a_k_m_grid_desc.GetLength(I1);
const auto N = b_k_n_grid_desc.GetLength(I1);
constexpr auto K1Number = Number<K1>{};
const auto K0 = K / K1Number;
const auto a_k0_m_k1_grid_desc =
transform_tensor_descriptor(a_k_m_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
const auto b_k0_n_k1_grid_desc =
transform_tensor_descriptor(b_k_n_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto a_k0_m_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0>{})); // 2-: K1
constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0>{})); // 2-: K1
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time =
driver_gemm_xdlops_v2r3<BlockSize,
ABType,
AccType,
CType,
InMemoryDataOperationEnum_t::Set,
decltype(a_k0_m_k1_grid_desc),
decltype(b_k0_n_k1_grid_desc),
decltype(c_m_n_grid_desc),
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
Sequence<0, 2, 1>,
Sequence<0, 2, 1>,
1,
ABlockTransferSrcScalarPerVector_M,
ABlockTransferDstScalarPerVector_K1,
false, // don't move back src coordinate after threadwise copy
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
Sequence<0, 2, 1>,
Sequence<0, 2, 1>,
1,
BBlockTransferSrcScalarPerVector_N,
BBlockTransferDstScalarPerVector_K1,
false, // don't move back src coordinate after threadwise copy
Sequence<0, 2, 4, 5, 6, 1, 3, 7>,
7,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks),
decltype(b_k0_n_k1_grid_step_hacks),
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks),
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks),
false // CAccessOrderMRepeatNRepeat
>(static_cast<ABType*>(a_k_m_device_buf.GetDeviceBuffer()),
static_cast<ABType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CType*>(c_m_n_device_buf.GetDeviceBuffer()),
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m_n_grid_desc,
a_k0_m_k1_grid_step_hacks,
b_k0_n_k1_grid_step_hacks,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
a_k0_m_k1_grid_move_slice_window_step_hacks,
b_k0_n_k1_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>((std::size_t(2) * M * N * K)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
c_m_n_device_buf.FromDevice(c_m_n.mData.data());
}
host/driver_offline/include/device_gemm_xdlops_km_nk_mn.hpp 0 → 100644
View file @ b725e3fc
This diff is collapsed.
host/driver_offline/include/device_gemm_xdlops_mk_kn_mn.hpp 0 → 100644
View file @ b725e3fc
This diff is collapsed.
host/driver_offline/include/device_gemm_xdlops_mk_nk_mn.hpp 0 → 100644
View file @ b725e3fc
This diff is collapsed.
host/driver_offline/src/conv_wrw_driver_offline.cpp 0 → 100644
View file @ b725e3fc
This diff is collapsed.
host/driver_offline/src/gemm_driver_offline.cpp 0 → 100644
View file @ b725e3fc
This diff is collapsed.
host/host_tensor/include/gemm_common.hpp 0 → 100644
View file @ b725e3fc
#ifndef GEMM_COMMON_HPP
#define GEMM_COMMON_HPP
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
#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