ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix:bf16x3:enable all instances on gfx950 (#3248)

Browse Source 
* fix:bf16x3:enable all instances on gfx950

* fix clang-format fail

* fix clang-format fail

* fix:modified wrong params previously
This commit is contained in:
yinglu
2025-11-21 09:09:43 +08:00
committed by GitHub
parent 938b8ed3bf
commit 4155eb24f9
3 changed files with 25 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
View File

@@ -376,6 +376,11 @@ using device_grouped_conv_bwd_data_xdl_f32_optimized_loads_instances =
// clang-format on
>;
#if defined(__gfx950__)
constexpr auto _k_per_block = 32;
#else
constexpr auto _k_per_block = 16;
#endif
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
@@ -392,22 +397,22 @@ using device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances =
// A K1 one access for each thread per load
// 32x32
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 32, 8, 8, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 32, 1, 8>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 16, 4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 4, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 32, 1, 8>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32,_k_per_block,4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 4, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 32, 1, 8>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 32, 8, 8, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 16, 1, 16>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 16, 4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 16, 1, 16>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32,_k_per_block,4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 16, 1, 16>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 32, 8, 8, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 32>, 1, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32, 16, 4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 32>, 1, make_default_loop_scheduler(), TF32, TF32>,
// 16x16
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 32,_k_per_block,4, 4, 32, 32, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 32>, 1, make_default_loop_scheduler(), TF32, TF32>,
// 16x16
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 64, 1, 4>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 16, 4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 2, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 64, 1, 4>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16,_k_per_block,4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 2, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, 1, 1, 1, S<1, 64, 1, 4>, 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 32, 1, 8>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 16, 4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 32, 1, 8>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16,_k_per_block,4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, 1, 1, 1, S<1, 32, 1, 8>, 2, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, 1, S<4, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16, 16, 4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, make_default_loop_scheduler(), TF32, TF32>
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 16,_k_per_block,4, 4, 16, 16, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, make_default_loop_scheduler(), TF32, TF32>
// clang-format on
>;

12
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp
View File

@@ -143,6 +143,12 @@ using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple<
// clang-format on
>;
#if defined(__gfx950__)
constexpr auto _k_per_block = 32;
#else
constexpr auto _k_per_block = 16;
#endif
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
@@ -158,9 +164,9 @@ using device_grouped_conv_fwd_xdl_merged_groups_f32_tf32_instances = std::tuple<
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 32>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16,_k_per_block,4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16,_k_per_block,4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 16,_k_per_block,4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, LoopScheduler::Default, 32>
// clang-format on
>;

4
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View File

@@ -94,8 +94,8 @@ function(add_instance_library INSTANCE_NAME)
message(DEBUG "removing gemm_universal_preshuffle_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
# Only build tf32 instances for gfx942
if(NOT INST_TARGETS MATCHES "gfx942" AND source_name MATCHES "_tf32_")
# Only build tf32 instances for gfx942 & gfx950
if(NOT (INST_TARGETS MATCHES "gfx942" OR INST_TARGETS MATCHES "gfx950") AND source_name MATCHES "_tf32_")
message(DEBUG "removing tf32 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()