mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 01:10:17 +00:00
Add ScaleAddScaleAddRelu post op for conv fwd (#1006)
* Add ScaleAddScaleAddRelu post op for conv fwd * Fixes * Fix instance file name * Minor fix
This commit is contained in:
Bartłomiej Kocot
@@ -42,6 +42,7 @@ template <ck::index_t NDimSpatial,
|
||||
typename InElementwiseOperation,
|
||||
typename WeiElementwiseOperation,
|
||||
typename OutElementwiseOperation,
|
||||
ck::index_t NumDTensor = 0,
|
||||
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
|
||||
struct ReferenceConvFwd : public device::BaseOperator
|
||||
{
|
||||
@@ -57,10 +58,12 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
std::vector<ck::index_t> input_right_pads,
|
||||
InElementwiseOperation in_element_op,
|
||||
WeiElementwiseOperation wei_element_op,
|
||||
OutElementwiseOperation out_element_op)
|
||||
OutElementwiseOperation out_element_op,
|
||||
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors)
|
||||
: input_{input},
|
||||
weight_{weight},
|
||||
output_{output},
|
||||
d_tensors_{d_tensors},
|
||||
conv_strides_{conv_filter_strides},
|
||||
conv_dilations_{conv_filter_dilations},
|
||||
in_left_pads_{input_left_pads},
|
||||
@@ -75,6 +78,8 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
const Tensor<WeiDataType>& weight_;
|
||||
Tensor<OutDataType>& output_;
|
||||
|
||||
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors_;
|
||||
|
||||
std::vector<index_t> conv_strides_;
|
||||
std::vector<index_t> conv_dilations_;
|
||||
std::vector<index_t> in_left_pads_;
|
||||
@@ -129,7 +134,26 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
}
|
||||
|
||||
OutDataType v_out;
|
||||
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
|
||||
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
|
||||
if constexpr(NumDTensor == 0)
|
||||
{
|
||||
arg.out_element_op_(v_out, v_acc_converted);
|
||||
}
|
||||
else if constexpr(NumDTensor == 1)
|
||||
{
|
||||
arg.out_element_op_(v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, wo));
|
||||
}
|
||||
else if constexpr(NumDTensor == 2)
|
||||
{
|
||||
arg.out_element_op_(v_out,
|
||||
v_acc_converted,
|
||||
arg.d_tensors_[0](g, n, k, wo),
|
||||
arg.d_tensors_[1](g, n, k, wo));
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Output ElementOp not supported in reference.");
|
||||
}
|
||||
arg.output_(g, n, k, wo) = v_out;
|
||||
};
|
||||
|
||||
@@ -183,7 +207,27 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
}
|
||||
|
||||
OutDataType v_out;
|
||||
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
|
||||
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
|
||||
if constexpr(NumDTensor == 0)
|
||||
{
|
||||
arg.out_element_op_(v_out, v_acc_converted);
|
||||
}
|
||||
else if constexpr(NumDTensor == 1)
|
||||
{
|
||||
arg.out_element_op_(
|
||||
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, ho, wo));
|
||||
}
|
||||
else if constexpr(NumDTensor == 2)
|
||||
{
|
||||
arg.out_element_op_(v_out,
|
||||
v_acc_converted,
|
||||
arg.d_tensors_[0](g, n, k, ho, wo),
|
||||
arg.d_tensors_[1](g, n, k, ho, wo));
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Output ElementOp not supported in reference.");
|
||||
}
|
||||
arg.output_(g, n, k, ho, wo) = v_out;
|
||||
};
|
||||
|
||||
@@ -250,7 +294,27 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
}
|
||||
|
||||
OutDataType v_out;
|
||||
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
|
||||
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
|
||||
if constexpr(NumDTensor == 0)
|
||||
{
|
||||
arg.out_element_op_(v_out, v_acc_converted);
|
||||
}
|
||||
else if constexpr(NumDTensor == 1)
|
||||
{
|
||||
arg.out_element_op_(
|
||||
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, d_o, ho, wo));
|
||||
}
|
||||
else if constexpr(NumDTensor == 2)
|
||||
{
|
||||
arg.out_element_op_(v_out,
|
||||
v_acc_converted,
|
||||
arg.d_tensors_[0](g, n, k, d_o, ho, wo),
|
||||
arg.d_tensors_[1](g, n, k, d_o, ho, wo));
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Output ElementOp not supported in reference.");
|
||||
}
|
||||
arg.output_(g, n, k, d_o, ho, wo) = v_out;
|
||||
};
|
||||
|
||||
@@ -294,7 +358,8 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
std::vector<ck::index_t> input_right_pads,
|
||||
InElementwiseOperation in_element_op,
|
||||
WeiElementwiseOperation wei_element_op,
|
||||
OutElementwiseOperation out_element_op)
|
||||
OutElementwiseOperation out_element_op,
|
||||
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors = {})
|
||||
{
|
||||
return Argument{input,
|
||||
weight,
|
||||
@@ -305,7 +370,8 @@ struct ReferenceConvFwd : public device::BaseOperator
|
||||
input_right_pads,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op};
|
||||
out_element_op,
|
||||
d_tensors};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
@@ -0,0 +1,131 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using namespace ck::tensor_layout::convolution;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ScaleAddScaleAddRelu = ck::tensor_operation::element_wise::ScaleAddScaleAddRelu;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 = ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto ConvFwdOddC =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
|
||||
|
||||
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
|
||||
|
||||
template <index_t NDimSpatial,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_bf16_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|
|
||||
//########################################| 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|
|
||||
//########################################| | | | | | | | | | | | 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
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, ck::Tuple<BF16, BF16>, BF16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
// instances for small conv.K and conv.C
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, ck::Tuple<BF16, BF16>, BF16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, ck::Tuple<BF16, BF16>, BF16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, ck::Tuple<BF16, BF16>, BF16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <index_t NDimSpatial,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_f16_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|
|
||||
//########################################| 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|
|
||||
//########################################| | | | | | | | | | | | 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
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, ck::Tuple<F16, F16>, F16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
// instances for small conv.K and conv.C
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, ck::Tuple<F16, F16>, F16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, ck::Tuple<F16, F16>, F16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, ck::Tuple<F16, F16>, F16, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <index_t NDimSpatial,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_f32_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|
|
||||
//########################################| 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|
|
||||
//########################################| | | | | | | | | | | | 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
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, ck::Tuple<F32, F32>, F32, PassThrough, PassThrough, ScaleAddScaleAddRelu, 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>,
|
||||
// instances for small conv.K and conv.C
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, ck::Tuple<F32, F32>, F32, PassThrough, PassThrough, ScaleAddScaleAddRelu, 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>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, ck::Tuple<F32, F32>, F32, PassThrough, PassThrough, ScaleAddScaleAddRelu, 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>,
|
||||
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, ck::Tuple<F32, F32>, F32, PassThrough, PassThrough, ScaleAddScaleAddRelu, 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>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <index_t NDimSpatial,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_int8_instances = std::tuple<
|
||||
// clang-format off
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | 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
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, ck::Tuple<F32, F32>, int8_t, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
// instances for small conv.K and conv.C
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, ck::Tuple<F32, F32>, int8_t, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, ck::Tuple<F32, F32>, int8_t, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, ck::Tuple<F32, F32>, int8_t, PassThrough, PassThrough, ScaleAddScaleAddRelu, ConvSpec, GemmMNKPadding, 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, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,176 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
#include <memory>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ScaleAddScaleAddRelu = ck::tensor_operation::element_wise::ScaleAddScaleAddRelu;
|
||||
|
||||
#ifdef CK_ENABLE_BF16
|
||||
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
|
||||
void add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
|
||||
NDHWGC,
|
||||
GKZYXC,
|
||||
ck::Tuple<NDHWGK, NDHWGK>,
|
||||
NDHWGK,
|
||||
BF16,
|
||||
BF16,
|
||||
ck::Tuple<BF16, BF16>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
ScaleAddScaleAddRelu>>>& instances);
|
||||
#endif
|
||||
|
||||
#ifdef CK_ENABLE_FP16
|
||||
void add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
|
||||
NDHWGC,
|
||||
GKZYXC,
|
||||
ck::Tuple<NDHWGK, NDHWGK>,
|
||||
NDHWGK,
|
||||
F16,
|
||||
F16,
|
||||
ck::Tuple<F16, F16>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
ScaleAddScaleAddRelu>>>& instances);
|
||||
#endif
|
||||
|
||||
#ifdef CK_ENABLE_FP32
|
||||
void add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
|
||||
NDHWGC,
|
||||
GKZYXC,
|
||||
ck::Tuple<NDHWGK, NDHWGK>,
|
||||
NDHWGK,
|
||||
F32,
|
||||
F32,
|
||||
ck::Tuple<F32, F32>,
|
||||
F32,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
ScaleAddScaleAddRelu>>>& instances);
|
||||
#endif
|
||||
|
||||
#ifdef CK_ENABLE_INT8
|
||||
void add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
|
||||
NDHWGC,
|
||||
GKZYXC,
|
||||
ck::Tuple<NDHWGK, NDHWGK>,
|
||||
NDHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
ck::Tuple<F32, F32>,
|
||||
int8_t,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
ScaleAddScaleAddRelu>>>& instances);
|
||||
#endif
|
||||
|
||||
template <ck::index_t NumDimSpatial,
|
||||
typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename DLayouts,
|
||||
typename OutLayout,
|
||||
typename InDataType,
|
||||
typename WeiDataType,
|
||||
typename DDataTypes,
|
||||
typename OutDataType,
|
||||
typename ComputeType>
|
||||
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
|
||||
NumDimSpatial,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
DLayouts,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
DDataTypes,
|
||||
OutDataType,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::ScaleAddScaleAddRelu,
|
||||
ComputeType>>
|
||||
{
|
||||
using DeviceOp =
|
||||
DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
DLayouts,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
DDataTypes,
|
||||
OutDataType,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::ScaleAddScaleAddRelu,
|
||||
ComputeType>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
|
||||
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
|
||||
{
|
||||
#ifdef CK_ENABLE_FP32
|
||||
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
|
||||
is_same_v<OutDataType, float>)
|
||||
{
|
||||
add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
#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<ComputeType, half_t>)
|
||||
{
|
||||
add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
#endif
|
||||
#ifdef CK_ENABLE_BF16
|
||||
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
|
||||
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
|
||||
{
|
||||
add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
#endif
|
||||
#ifdef CK_ENABLE_INT8
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<OutDataType, int8_t>)
|
||||
{
|
||||
add_device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user