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

Conv:TF32: add more instances - 2 (#2879)

Browse Source 
* add instances of device_grouped_conv_fwd_xdl_f32_comp_instances
* add instances of device_grouped_conv_fwd_xdl_f32_tf32_mem_instances
* add instances of device_grouped_conv_fwd_xdl_large_tensor_f32_tf32_instances
* tf32:conv:add instances for base class DeviceConvFwd
* tf32:conv:add instances for base class DeviceGroupedConvBwdDataMultipleD
* tf32:conv:add instances for base class DeviceGroupedConvBwdWeight
* add tf32 in profiler
* remove gnhwc/ngchw/ngcdhw instances
* remove non-ndhwgc/nhwgc/nhwc instances
* add check in IsSupportedArgument()
This commit is contained in:
yinglu
2025-10-10 15:28:17 +08:00
committed by GitHub
parent ad7a215aba
commit fada1a3cae
56 changed files with 2119 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View File

@@ -28,6 +28,7 @@ template <ck::index_t NDimSpatial,
ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename ComputeDataType = OutDataType,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvBwdData : public device::BaseOperator
{
@@ -142,8 +143,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
c,
x);
v_acc += ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
v_acc += ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_out)) *
ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_wei));
}
}
}
@@ -235,8 +238,11 @@ struct ReferenceConvBwdData : public device::BaseOperator
y,
x);
v_acc += ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
v_acc +=
ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_out)) *
ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_wei));
}
}
}
@@ -354,8 +360,12 @@ struct ReferenceConvBwdData : public device::BaseOperator
x);
v_acc +=
ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
ck::type_convert<float>(
ck::type_convert<
ComputeDataType>(v_out)) *
ck::type_convert<float>(
ck::type_convert<
ComputeDataType>(v_wei));
}
}
}

7
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_transpose_xdl_instance.hpp
View File

@@ -18,6 +18,7 @@ namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
using BF8 = ck::bf8_t;
using F8 = ck::f8_t;
@@ -84,17 +85,17 @@ using device_grouped_conv_bwd_data_transpose_xdl_bf16_instances =
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, make_default_loop_scheduler(), BF16, BF16, 4, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 4, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 4, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 4, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, make_default_loop_scheduler(), BF16, BF16, 1, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 1, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 1, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 1, 2>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 1>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 1>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, make_default_loop_scheduler(), BF16, BF16, 2, 1>,

38
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp
View File

@@ -18,6 +18,7 @@ namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
@@ -143,6 +144,43 @@ using device_grouped_conv_bwd_data_xdl_bilinear_f32_instances =
// clang-format on
>;
// f32_f32_f32_f32 tf32
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_bilinear_f32_tf32_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, make_default_loop_scheduler(), TF32, TF32>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 32, 1, 4>, 1, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4> , 4, make_default_loop_scheduler(), TF32, TF32>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<F32>, F32, PassThrough, PassThrough, Bilinear, ConvSpec, true, true, 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, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4> , 8, make_default_loop_scheduler(), TF32, TF32>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation

113
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
View File

@@ -18,6 +18,7 @@ namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
using BF8 = ck::bf8_t;
using F8 = ck::f8_t;
@@ -375,6 +376,41 @@ using device_grouped_conv_bwd_data_xdl_f32_optimized_loads_instances =
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// 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, 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, 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, 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, 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, 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>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
@@ -411,6 +447,83 @@ using device_grouped_conv_bwd_data_xdl_f32_instances =
// clang-format on
>;
// f32_f32_f32_f32 tf32
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_f32_tf32_generic_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, make_default_loop_scheduler(), TF32, TF32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_f32_tf32_16_16_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 16, 64, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 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, 64, 16, 64, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 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, 64, 16, 64, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 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, 64, 64, 16, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 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, 64, 64, 16, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 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, 64, 64, 16, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, make_default_loop_scheduler(), TF32, TF32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_f32_tf32_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, make_default_loop_scheduler(), TF32, TF32>,
// instances for small conv.K and conv.C
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, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 32, 1, 4>, 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, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 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, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 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, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 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, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 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, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 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, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 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, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 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, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 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, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 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, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, 4, make_default_loop_scheduler(), TF32, TF32>
// clang-format on
>;
// f16_f16_f16_comp_f8
template <index_t NDimSpatial,
typename ALayout,

38
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_scale_instance.hpp
View File

@@ -18,6 +18,7 @@ namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
@@ -143,6 +144,43 @@ using device_grouped_conv_bwd_data_xdl_scale_f32_instances =
// clang-format on
>;
// f32_f32_f32_f32 tf32
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_scale_f32_tf32_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 32, 1, 4>, 1>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Tuple<>, F32, PassThrough, PassThrough, Scale, ConvSpec, true, true, 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, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation

19
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp
View File

@@ -18,6 +18,7 @@ using namespace ck::tensor_layout::convolution;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
@@ -58,6 +59,24 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_instances = std::tuple
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion, TF32, TF32>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,

34
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp
View File

@@ -18,6 +18,7 @@ using namespace ck::tensor_layout::convolution;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
@@ -74,6 +75,39 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_bilinear_instances = std:
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_bilinear_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| DsData| InData| WeiData| OutData| AccData| DsData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer|
//#########################################| Dim| | | | Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector|
//#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| |
// generic instance
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple<BLayout>, F32, F32, F32, F32, Tuple<F32>, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,

57
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
View File

@@ -18,6 +18,7 @@ using namespace ck::tensor_layout::convolution;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
@@ -96,6 +97,62 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances = std::tuple<
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_generic_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| |
// generic instance
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
index_t MaxTransposeTransferSrcScalarPerVector = 1,
index_t MaxTransposeTransferDstScalarPerVector = 1>
using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| Compute| Compute| MaxTranspose| MaxTranspose|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| TypeA| TypeB| TransferSrc| TransferDst|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | |
// generic instance
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,

34
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp
View File

@@ -18,6 +18,7 @@ using namespace ck::tensor_layout::convolution;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using TF32 = ck::tf32_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
@@ -74,6 +75,39 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_scale_instances = std::tu
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_scale_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| DsData| InData| WeiData| OutData| AccData| DsData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer|
//#########################################| Dim| | | | Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector|
//#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| |
// generic instance
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>,
DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,

54
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
View File

@@ -117,14 +117,28 @@ struct DeviceOperationInstanceFactory<
#endif
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_16_16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_optimized_loads_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: this operator requires the same compute type");
if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_16_16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_optimized_loads_instances(
op_ptrs);
}
else
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_16_16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_optimized_loads_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_BF16
@@ -275,12 +289,26 @@ struct DeviceOperationInstanceFactory<
is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_16_16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_optimized_loads_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: this operator requires the same compute type");
if constexpr(is_same_v<ComputeTypeA, F32>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_16_16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_optimized_loads_instances(
op_ptrs);
}
else if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_16_16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_optimized_loads_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_BF16

96
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_xdl.inc
View File

@@ -127,6 +127,38 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_16_16_instance
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_16_16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
@@ -140,6 +172,22 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_optimized_load
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
@@ -479,6 +527,38 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_16_16_insta
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_16_16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
@@ -492,6 +572,22 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_optimized_l
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(

83
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View File

@@ -349,20 +349,37 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: ComputeTypeA and ComputeTypeB should be the same");
if constexpr(is_same_v<ComputeTypeA, float>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_instances(
op_ptrs);
}
else if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev5_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_FP16
@@ -595,20 +612,36 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: ComputeTypeA and ComputeTypeB should be the same");
if constexpr(is_same_v<ComputeTypeA, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev5_instances(
op_ptrs);
}
else if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev5_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev5_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_FP16

32
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_bilinear.hpp
View File

@@ -62,6 +62,21 @@ void add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_
PassThrough,
Bilinear,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeightMultipleD<3,
NDHWGC,
GKZYXC,
NDHWGK,
Tuple<GKZYXC>,
F32,
F32,
F32,
Tuple<F32>,
PassThrough,
Bilinear,
PassThrough,
TF32,
TF32>>>& instances);
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
void add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_f8_instances(
@@ -138,11 +153,20 @@ struct DeviceOperationInstanceFactory<
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: this operator requires the same compute type");
if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
op_ptrs);
}
else
{
add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_FP16

33
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp
View File

@@ -62,6 +62,22 @@ void add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_ins
PassThrough,
Scale,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeightMultipleD<3,
NDHWGC,
GKZYXC,
NDHWGK,
Tuple<>,
F32,
F32,
F32,
Tuple<>,
PassThrough,
Scale,
PassThrough,
TF32,
TF32>>>& instances);
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
void add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_f8_instances(
@@ -138,11 +154,20 @@ struct DeviceOperationInstanceFactory<
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
static_assert(is_same_v<ComputeTypeA, ComputeTypeB>,
"Error: this operator requires the same compute type");
if constexpr(is_same_v<ComputeTypeA, TF32>)
{
add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
op_ptrs);
}
else
{
add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_FP16

142
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_xdl.inc
View File

@@ -570,6 +570,20 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_ngchw_gkyxc_ngkhw_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW,
@@ -606,6 +620,20 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipe
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -618,6 +646,20 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipe
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -630,6 +672,20 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -641,6 +697,20 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
#endif
// conv3d backward weight
#ifdef CK_ENABLE_BF16
@@ -1177,7 +1247,7 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipe
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
@@ -1188,6 +1258,76 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipe
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_f8_instances(

16
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp_xdl.inc
View File

@@ -570,6 +570,22 @@ void add_device_grouped_conv2d_fwd_bias_clamp_xdl_merged_groups_nhwgc_gkyxc_nhwg
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
F32,
F32,
Tuple<F32>,
F32,
PassThrough,
PassThrough,
AddClamp,
TF32,
TF32>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,

3
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
View File

@@ -7,12 +7,15 @@ add_instance_library(
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_16_16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_16_16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_16_16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_16_16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_optimized_loads_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_optimized_loads_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_optimized_loads_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_optimized_loads_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkyxc_ngkhw_f16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkyxc_ngkhw_bf16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkyxc_ngkhw_f32_instance.cpp

51
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_16_16_instance.cpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_16_16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_16_16_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_16_16_instances<
2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

51
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

52
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_optimized_loads_instance.cpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_tf32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances<
2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

5
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
View File

@@ -10,6 +10,7 @@ set(GROUPED_CONV2D_BWD_WEIGHT
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_default_pipev2_instance.cpp
@@ -21,9 +22,13 @@ set(GROUPED_CONV2D_BWD_WEIGHT
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_pad0_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_pad0_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instance.cpp
xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instance.cpp

42
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev2_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev5_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_default_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v5>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

48
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances<2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev2_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev5_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_tf32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v5>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

3
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
View File

@@ -6,12 +6,15 @@ set(GROUPED_CONV3D_BWD_DATA
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_16_16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_16_16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_16_16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_16_16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_optimized_loads_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_optimized_loads_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_optimized_loads_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_optimized_loads_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ngcdhw_gkzyxc_ngkdhw_f16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_ngcdhw_gkzyxc_ngkdhw_f32_instance.cpp

51
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_16_16_instance.cpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_16_16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_16_16_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_16_16_instances<
3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

51
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

52
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_optimized_loads_instance.cpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_tf32_optimized_loads_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_f32_tf32_optimized_loads_instances<
3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

5
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
View File

@@ -6,6 +6,7 @@ set(GROUPED_CONV3D_BWD_WEIGHT
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev2_instance.cpp
@@ -17,9 +18,13 @@ set(GROUPED_CONV3D_BWD_WEIGHT
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instance.cpp
xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instance.cpp

42
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev2_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev5_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_default_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v5>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

48
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances<3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev2_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

42
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev5_instance.cpp Normal file
View File

@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_pad0_pipev5_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v5>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

1
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_bilinear/CMakeLists.txt
View File

@@ -2,6 +2,7 @@
set(GROUPED_CONV3D_BWD_WEIGHT_BILINEAR
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)

52
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_bilinear/xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeightMultipleD<3,
NDHWGC,
GKZYXC,
NDHWGK,
Tuple<GKZYXC>,
F32,
F32,
F32,
Tuple<F32>,
PassThrough,
Bilinear,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_bilinear_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_bilinear_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

1
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt
View File

@@ -2,6 +2,7 @@
set(GROUPED_CONV3D_BWD_WEIGHT_SCALE
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)

52
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instance.cpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeightMultipleD<3,
NDHWGC,
GKZYXC,
NDHWGK,
Tuple<>,
F32,
F32,
F32,
Tuple<>,
PassThrough,
Scale,
PassThrough,
TF32,
TF32>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_scale_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_scale_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck