ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 01:36:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "feature:tf32:add initial conv3d fwd kernel support (#2763)" (#2848)

Browse Source 
This reverts commit c51102144f.
This commit is contained in:
Illia Silin
2025-09-15 08:27:04 -07:00
committed by GitHub
parent c51102144f
commit 03b59f8c76
44 changed files with 175 additions and 1085 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View File

@@ -59,7 +59,6 @@ template <ck::index_t NDimSpatial,
ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename ComputeDataType = InDataType,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator
{
@@ -328,10 +327,8 @@ struct ReferenceConvFwd : public device::BaseOperator
z,
y,
x);
v_acc += ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_in)) *
ck::type_convert<float>(
ck::type_convert<ComputeDataType>(v_wei));
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
}
}
}

40
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View File

@@ -25,12 +25,6 @@ template <typename ADataType,
typename ComputeTypeB = ComputeTypeA>
struct ReferenceGemm : public device::BaseOperator
{
using ElementDataTypeA =
ck::conditional_t<is_same_v<ComputeTypeA, ck::tf32_t>, float, ComputeTypeA>;
using ElementDataTypeB =
ck::conditional_t<is_same_v<ComputeTypeB, ck::tf32_t>, float, ComputeTypeB>;
// Argument
struct Argument : public device::BaseArgument
{
@@ -69,8 +63,8 @@ struct ReferenceGemm : public device::BaseOperator
const int K = arg.a_m_k_.mDesc.GetLengths()[1];
AccDataType v_acc{0};
ElementDataTypeA v_a{0};
ElementDataTypeB v_b{0};
ComputeTypeA v_a{0};
ComputeTypeB v_b{0};
for(int k = 0; k < K; ++k)
{
@@ -83,16 +77,16 @@ struct ReferenceGemm : public device::BaseOperator
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_a = type_convert<ElementDataTypeA>(i4);
v_a = type_convert<ComputeTypeA>(i4);
}
else if constexpr(is_same_v<ADataType, f4x2_pk_t>)
{
// TODO: add support for ColMajor layout as well
if(k % 2 == 1)
v_a = type_convert<ElementDataTypeA>(
v_a = type_convert<ComputeTypeA>(
f4_t(arg.a_m_k_(m, k).template unpack<>(Number<1>{})));
else
v_a = type_convert<ElementDataTypeA>(
v_a = type_convert<ComputeTypeA>(
f4_t(arg.a_m_k_(m, k).template unpack<>(Number<0>{})));
}
else if constexpr(is_same_v<ADataType, f6x16_pk_t> ||
@@ -100,7 +94,7 @@ struct ReferenceGemm : public device::BaseOperator
is_same_v<ADataType, f6x32_pk_t> ||
is_same_v<ADataType, bf6x32_pk_t>)
{
v_a = type_convert<ElementDataTypeA>(
v_a = type_convert<ComputeTypeA>(
arg.a_m_k_(m, k).unpack(k % ADataType::packed_size));
}
else
@@ -117,16 +111,16 @@ struct ReferenceGemm : public device::BaseOperator
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_b = type_convert<ElementDataTypeB>(i4);
v_b = type_convert<ComputeTypeB>(i4);
}
else if constexpr(is_same_v<BDataType, f4x2_pk_t>)
{
// TODO: add support for RowMajor layout as well
if(k % 2 == 1)
v_b = type_convert<ElementDataTypeB>(
v_b = type_convert<ComputeTypeB>(
f4_t(arg.b_k_n_(k, n).template unpack<>(Number<1>{})));
else
v_b = type_convert<ElementDataTypeB>(
v_b = type_convert<ComputeTypeB>(
f4_t(arg.b_k_n_(k, n).template unpack<>(Number<0>{})));
}
else if constexpr(is_same_v<BDataType, f6x16_pk_t> ||
@@ -134,7 +128,7 @@ struct ReferenceGemm : public device::BaseOperator
is_same_v<BDataType, f6x32_pk_t> ||
is_same_v<BDataType, bf6x32_pk_t>)
{
v_b = type_convert<ElementDataTypeB>(
v_b = type_convert<ComputeTypeB>(
arg.b_k_n_(k, n).unpack(k % BDataType::packed_size));
}
else
@@ -142,18 +136,8 @@ struct ReferenceGemm : public device::BaseOperator
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
}
if constexpr(is_same_v<ComputeTypeA, ComputeTypeB> &&
is_same_v<ComputeTypeA, ck::tf32_t>)
{ // only for tf32 now
v_acc +=
ck::type_convert<AccDataType>(ck::type_convert<ComputeTypeA>(v_a)) *
ck::type_convert<AccDataType>(ck::type_convert<ComputeTypeB>(v_b));
}
else
{
v_acc +=
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
}
v_acc +=
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
}
CDataType v_c{0};

19
library/include/ck/library/reference_tensor_operation/gpu/reference_gemm.hpp
View File

@@ -38,10 +38,6 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
const CDEElementwiseOperation c_element_op)
{
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ElementDataTypeA =
ck::conditional_t<is_same_v<ComputeTypeA, ck::tf32_t>, float, ComputeTypeA>;
using ElementDataTypeB =
ck::conditional_t<is_same_v<ComputeTypeB, ck::tf32_t>, float, ComputeTypeB>;
const int row_idx = blockIdx.x * blockDim.x + threadIdx.x;
const int col_idx = blockIdx.y * blockDim.y + threadIdx.y;
@@ -50,8 +46,8 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
{
AccDataType v_acc{0};
ElementDataTypeA v_a{0};
ElementDataTypeB v_b{0};
ComputeTypeA v_a{0};
ComputeTypeB v_b{0};
CDataType v_c{0};
for(int k_idx = 0; k_idx < k; ++k_idx)
@@ -80,16 +76,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
// apply b_element_op
b_element_op(v_b, p_b_grid[element_idx_b]);
// multiply and accumulate
if constexpr(is_same_v<ComputeTypeA, ComputeTypeB> &&
is_same_v<ComputeTypeA, ck::tf32_t>)
{ // only for tf32 now
v_acc += ck::type_convert<AccDataType>(ck::type_convert<ComputeTypeA>(v_a)) *
ck::type_convert<AccDataType>(ck::type_convert<ComputeTypeB>(v_b));
}
else
{
v_acc += type_convert<AccDataType>(v_a) * type_convert<AccDataType>(v_b);
}
v_acc += type_convert<AccDataType>(v_a) * type_convert<AccDataType>(v_b);
}
// apply c_element_op
c_element_op(v_c, v_acc);

1
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View File

@@ -16,7 +16,6 @@ namespace instance {
// aliasing, for commonly used data type
using F64 = double;
using F32 = float;
using TF32 = ck::tf32_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;

1
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_dynamic_op_instance.hpp
View File

@@ -16,7 +16,6 @@ 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...>;

43
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
View File

@@ -24,7 +24,6 @@ using BF8 = ck::bf8_t;
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...>;
@@ -200,7 +199,7 @@ using device_grouped_conv_fwd_xdl_f16_nchw_instances = std::tuple<
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 8, 1, 8>, 1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 8, 1, 8>, 1>,
// 32x32 instance
// 32x32 instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
@@ -285,45 +284,7 @@ using device_grouped_conv_fwd_xdl_f32_instances = std::tuple<
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 192, 16, 4, 4, 32, 32, 2, 3, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_xdl_f32_tf32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| AComputeType| BComputeType|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| DATATYPE | DATATYPE |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1, TF32, TF32>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 1, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 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<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, TF32, TF32>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 192, 16, 4, 4, 32, 32, 2, 3, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsDataTypes, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;

6
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View File

@@ -443,12 +443,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_inter_instances(
op_ptrs);
}
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<AComputeType, TF32> &&
is_same_v<BComputeType, TF32>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP8

8
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp.hpp
View File

@@ -215,14 +215,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_bias_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_inter_instances(
op_ptrs);
}
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<AComputeType, TF32> &&
is_same_v<BComputeType, TF32>)
{
add_device_grouped_conv3d_fwd_bias_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
op_ptrs);
}
#endif
}
#endif // CK_USE_XDL

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

@@ -578,22 +578,6 @@ void add_device_grouped_conv3d_fwd_bias_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_insta
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
F32,
F32,
Tuple<F32>,
F32,
PassThrough,
PassThrough,
AddClamp,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,

8
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_clamp.hpp
View File

@@ -210,14 +210,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_inter_instances(
op_ptrs);
}
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float> && is_same_v<AComputeType, TF32> &&
is_same_v<BComputeType, TF32>)
{
add_device_grouped_conv3d_fwd_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
op_ptrs);
}
#endif
}
#endif // CK_USE_XDL

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

@@ -578,22 +578,6 @@ void add_device_grouped_conv3d_fwd_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
F32,
F32,
Tuple<>,
F32,
PassThrough,
PassThrough,
Clamp,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_xdl_ndhwgc_gkzyxc_ndhwgk_f32_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,

14
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dynamic_op.hpp
View File

@@ -132,7 +132,6 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_insta
PassThrough,
PassThrough,
DynamicUnaryOp>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
@@ -160,8 +159,7 @@ template <ck::index_t NumDimSpatial,
typename WeiDataType,
typename DDataTypes,
typename OutDataType,
typename AComputeType,
typename BComputeType = AComputeType>
typename ComputeType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial,
InLayout,
@@ -175,8 +173,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::DynamicUnaryOp,
AComputeType,
BComputeType>>
ComputeType>>
{
using DeviceOp =
DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
@@ -191,8 +188,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::DynamicUnaryOp,
AComputeType,
BComputeType>;
ComputeType>;
static auto GetInstances()
{
@@ -211,7 +207,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t>)
is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
@@ -248,7 +244,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t>)
is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_xdl_dynamic_op_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);

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

@@ -559,22 +559,6 @@ void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_tf32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough,
TF32,
TF32>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,