ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-13 10:37:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add merged groups instances for all 2D vanilla grouped conv fwd types and layouts.

Browse Source
This commit is contained in:
kiefer
2025-08-27 10:18:13 +00:00
parent 73521fe091
commit a06e27650e
10 changed files with 355 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_merged_groups_instance.hpp
View File

@@ -78,6 +78,28 @@ using device_grouped_conv_fwd_wmma_cshufflev3_merged_groups_f16_instances = std:
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_merged_groups_int8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| TODO: Other pipelines
//########################################| | | | | | | | | | | | 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_NWaveNPerWmma| _NWaveNPerWmma|
//########################################| | | | | | | | | | | | | | | | | TODO: ??| | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
// TODO: I had to change A and B srcScalarPerVector from 8 to 1 in order to get these instances to be compatible with the device implementation. I am pretty sure they will not work for XDL either.
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int8_t, DsDataTypes, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 32, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, int8_t, int8_t, 8>
// DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsDataTypes, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 16>,
// DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsDataTypes, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 32>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation

20
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View File

@@ -800,8 +800,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_comp_2x_instances(op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_comp_part2_instances(
@@ -825,8 +825,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances(op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instances(
@@ -846,8 +846,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_int8_instances(
// op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_int8_comp_instances(op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_int8_mem_intra_instances(
// op_ptrs);
@@ -866,8 +866,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_f16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_f16_16x16_instances(
@@ -890,8 +890,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_bf16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_16x16_instances(

200
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_wmma_cshufflev3_merged_groups.inc
View File

@@ -10,112 +10,82 @@ namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
// void
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Empty_Tuple,
// NHWGK,
// BF16,
// BF16,
// Empty_Tuple,
// BF16,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// void
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NGCHW,
// GKCYX,
// Empty_Tuple,
// NGKHW,
// BF16,
// BF16,
// Empty_Tuple,
// BF16,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
// void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Empty_Tuple,
// NHWGK,
// F16,
// F16,
// Empty_Tuple,
// F16,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NGCHW,
// GKCYX,
// Empty_Tuple,
// NGKHW,
// F16,
// F16,
// Empty_Tuple,
// F16,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
// void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Empty_Tuple,
// NHWGK,
// F32,
// F32,
// Empty_Tuple,
// F32,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
// void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_f32_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NGCHW,
// GKCYX,
// Empty_Tuple,
// NGKHW,
// F32,
// F32,
// Empty_Tuple,
// F32,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
// void
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Empty_Tuple,
// NHWGK,
// int8_t,
// int8_t,
// Empty_Tuple,
// int8_t,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// TODO: The XDL version of this function is forward declared but never defined! Oversight in XDL
// implementation?
// void
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_ngchw_gkcyx_ngkhw_int8_instances(
@@ -166,23 +136,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ndhwgc_gkzyxc_n
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
// void
// add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NDHWGC,
// GKZYXC,
// Empty_Tuple,
// NDHWGK,
// F32,
// F32,
// Empty_Tuple,
// F32,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
#endif
// grouped conv3d forward, NGCDHW/GKCZYX/NGKDHW
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ngcdhw_gkczyx_ngkdhw_bf16_instances(
@@ -216,23 +169,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ngcdhw_gkczyx_n
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
// void
// add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ngcdhw_gkczyx_ngkdhw_f32_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NGCDHW,
// GKCZYX,
// Empty_Tuple,
// NGKDHW,
// F32,
// F32,
// Empty_Tuple,
// F32,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation