Skip to content

GitLab

Commit ad09ebdb authored by carlushuang's avatar carlushuang
Browse files

add kyxck8

parent d6d37ea9
Hide whitespace changes
Inline Side-by-side
Showing with 2335 additions and 927 deletions
include/ck/tensor_operation/cpu/device/device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp
View file @ ad09ebdb
#ifndef DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_HPP
#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_HPP
#include <iostream>
#include <sstream>
#include <numeric>
#include "device.hpp"
#include "device_base_cpu.hpp"
#include "device_conv_fwd_cpu.hpp"
#include "convolution_forward_specialization_cpu.hpp"
#include "common_header.hpp"
#include "../../gpu/device/tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_avx2.hpp"
#include "threadwise_gemm_avx2.hpp"
#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
ck::index_t NumDimSpatial,
ck::index_t MPerBlock, // block means data are designed to fit in cache (L1/L2/L3)
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerThread,
ck::index_t NPerThread,
bool UseALocalBuffer,
bool UseBLocalBuffer,
bool UseCLocalBuffer>
struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
: public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
{
using DeviceOp = DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
using ADataType = InDataType;
using BDataType = WeiDataType;
using CDataType = OutDataType;
using AElementwiseOperation = InElementwiseOperation;
using BElementwiseOperation = WeiElementwiseOperation;
using CElementwiseOperation = OutElementwiseOperation;
// TODO make A/B datatype different
using ABDataType = InDataType;
static constexpr index_t NDimSpatial = NumDimSpatial;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr bool NonTemporalStore = false;
static constexpr auto GetBlockMNKAccessOrder()
{
if constexpr(BlockLoopOverSpecialization == DefaultBlockLoopOver ||
BlockLoopOverSpecialization == LoopOver_MNK)
return ck::Sequence<0, 1, 2>{};
else if constexpr(BlockLoopOverSpecialization == LoopOver_MKN)
return ck::Sequence<0, 2, 1>{};
}
using BlockMNKAccessOrder = decltype(GetBlockMNKAccessOrder());
static constexpr auto GetThreadwiseGemm_Dispatch()
{
if constexpr(MPerThread == 4 && NPerThread == 24)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else if constexpr(MPerThread == 6 && NPerThread == 16)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_6x16_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else
{
// static_assert(false, "invalid Mr/Nr");
}
}
using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
static constexpr auto GetInputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
}
static constexpr auto GetWeightBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(
math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
KPerBlock,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
}
static constexpr auto GetOutputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
}
static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
{
ck::index_t gemm_n_padded =
math::integer_least_multiple(gemm_n, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
const auto wei_gemm_n_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_n, gemm_k));
const auto wei_gemm_padn_k_grid_desc = transform_tensor_descriptor(
wei_gemm_n_k_grid_desc,
make_tuple(make_right_pad_transform(gemm_n, gemm_n_padded - gemm_n),
make_pass_through_transform(gemm_k)),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
const auto wei_gemm_n0_k_n1_grid_desc = transform_tensor_descriptor(
wei_gemm_padn_k_grid_desc,
ck::make_tuple(
ck::make_unmerge_transform(
ck::make_tuple(wei_gemm_padn_k_grid_desc.GetLength(I0) /
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize)),
ck::make_pass_through_transform(wei_gemm_padn_k_grid_desc.GetLength(I1))),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
ck::make_tuple(ck::Sequence<0, 2>{}, ck::Sequence<1>{}));
return wei_gemm_n0_k_n1_grid_desc;
}
static auto GetOutputTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n)
{
const auto out_gemm_m_n_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_n));
return out_gemm_m_n_grid_desc;
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Wi = input_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[0];
const index_t ConvStrideW = conv_filter_strides[0];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t X = filter_spatial_lengths[0];
const index_t ConvDilationW = conv_filter_dilations[0];
const index_t InLeftPadW = input_left_pads[0];
const index_t InRightPadW = input_right_pads[0];
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)),
make_merge_transform(make_tuple(X, C))),
make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t ConvStrideH = conv_filter_strides[0];
const index_t ConvStrideW = conv_filter_strides[1];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Y = filter_spatial_lengths[0];
const index_t X = filter_spatial_lengths[1];
const index_t ConvDilationH = conv_filter_dilations[0];
const index_t ConvDilationW = conv_filter_dilations[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_merge_transform(make_tuple(Y, X, C))),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
ck::index_t gemm_m_pad,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Di = input_spatial_lengths[0];
const index_t Hi = input_spatial_lengths[1];
const index_t Wi = input_spatial_lengths[2];
const index_t Do = output_spatial_lengths[0];
const index_t Ho = output_spatial_lengths[1];
const index_t Wo = output_spatial_lengths[2];
const index_t ConvStrideD = conv_filter_strides[0];
const index_t ConvStrideH = conv_filter_strides[1];
const index_t ConvStrideW = conv_filter_strides[2];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_do_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Z = filter_spatial_lengths[0];
const index_t Y = filter_spatial_lengths[1];
const index_t X = filter_spatial_lengths[2];
const index_t ConvDilationD = conv_filter_dilations[0];
const index_t ConvDilationH = conv_filter_dilations[1];
const index_t ConvDilationW = conv_filter_dilations[2];
const index_t InLeftPadD = input_left_pads[0];
const index_t InLeftPadH = input_left_pads[1];
const index_t InLeftPadW = input_left_pads[2];
const index_t InRightPadD = input_right_pads[0];
const index_t InRightPadH = input_right_pads[1];
const index_t InRightPadW = input_right_pads[2];
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{},
Sequence<1, 2>{},
Sequence<3, 4>{},
Sequence<5, 6>{},
Sequence<7>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_merge_transform(make_tuple(Z, Y, X, C))),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
static index_t GetGemmM(ck::index_t N, const std::vector<ck::index_t>& output_spatial_lengths)
{
return N * std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmK(ck::index_t C, const std::vector<ck::index_t>& filter_spatial_lengths)
{
return C * std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmN(ck::index_t K)
{
// return ck::math::integer_least_multiple(K,
// ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
return K;
}
static auto MakeABCGridDescriptor(ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
using namespace ck;
const index_t GemmM = GetGemmM(N, output_spatial_lengths);
const index_t GemmN = GetGemmN(K);
const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
// A:
const auto in_gemm_m_k_grid_desc =
GetInputTensorDescriptor<NumDimSpatial>(N,
C,
GemmM,
GemmK,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
// B:
const auto wei_gemm_n0_k_n1_grid_desc = GetWeightTensorDescriptor(GemmK, GemmN);
// C:
const auto out_gemm_m_n_grid_desc = GetOutputTensorDescriptor(GemmM, GemmN);
return make_tuple(
in_gemm_m_k_grid_desc, wei_gemm_n0_k_n1_grid_desc, out_gemm_m_n_grid_desc);
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1});
}
using ABCGridDescs = decltype(GetABCGridDesc<NumDimSpatial>());
using AGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
using BGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// static constexpr bool UseCLocalBuffer = false;
using AThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC<
InDataType,
InDataType,
AGridDesc,
decltype(GetInputBlockDescriptor()),
InElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using BThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC<
WeiDataType,
WeiDataType,
BGridDesc,
decltype(GetWeightBlockDescriptor()),
WeiElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using CThreadwiseCopy = ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN<
OutDataType,
OutDataType,
CGridDesc,
decltype(GetOutputBlockDescriptor()),
OutElementwiseOperation,
!UseCLocalBuffer,
ConvForwardSpecialization,
GemmKSpecialization>;
using GridwiseGemm =
ck::cpu::GridwiseGemmAvx2_MxN<InDataType, // InDataType,
WeiDataType, // WeiDataType,
OutDataType, // OutDataType,
AGridDesc, // AGridDesc,
BGridDesc, // BGridDesc,
CGridDesc, // CGridDesc,
AElementwiseOperation, // AElementwiseOperation,
BElementwiseOperation, // BElementwiseOperation,
CElementwiseOperation, // CElementwiseOperation,
MPerBlock, // MPerBlock,
NPerBlock, // NPerBlock,
KPerBlock, // KPerBlock,
ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
AThreadwiseCopy, // AThreadwiseCopy
BThreadwiseCopy, // BThreadwiseCopy
CThreadwiseCopy, // CThreadwiseCopy
BlockMNKAccessOrder, // BlockMNKAccessOrder,
ck::Sequence<0, 1>, // ThreadMNAccessOrder
UseALocalBuffer, // UseALocalBuffer
UseBLocalBuffer, // UseBLocalBuffer
UseCLocalBuffer // UseCLocalBuffer
>;
// Argument
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
: p_a_grid_{p_in_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_out_grid},
a_grid_desc_{},
b_grid_desc_{},
c_grid_desc_{},
a_element_op_{in_element_op},
b_element_op_{wei_element_op},
c_element_op_{out_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
filter_spatial_lengths_{filter_spatial_lengths},
conv_filter_strides_{conv_filter_strides},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
const auto descs = DeviceOp::MakeABCGridDescriptor(N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
a_grid_desc_ = descs[I0];
b_grid_desc_ = descs[I1];
c_grid_desc_ = descs[I2];
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc a_grid_desc_;
BGridDesc b_grid_desc_;
CGridDesc c_grid_desc_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
index_t Conv_C_;
std::vector<index_t> filter_spatial_lengths_;
std::vector<index_t> conv_filter_strides_;
std::vector<index_t> input_left_pads_;
std::vector<index_t> input_right_pads_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1)
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
{
throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
}
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
const auto kernel = ck::cpu::kernel_gemm_avx_mxn<GridwiseGemm,
InDataType,
WeiDataType,
OutDataType,
AGridDesc,
BGridDesc,
CGridDesc,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
float ave_time = 0;
if(nrepeat != 1)
ave_time = launch_and_time_cpu_kernel(kernel,
nrepeat,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
// TODO: this is for benchmark purpose, so last time we clear c buffer and calculate the
// result
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
launch_cpu_kernel(kernel,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
return ave_time;
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config, nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
// check if it's 1x1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
if constexpr(GemmKSpecialization ==
ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
{
if(!(arg.Conv_C_ % KPerBlock == 0))
return false;
}
// Gridwise GEMM size
return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_grid,
const void* p_wei_grid,
void* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
static_cast<OutDataType*>(p_out_grid),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
auto string_local_buffer = [](bool is_local_buffer) {
if(is_local_buffer)
return "L";
else
return "G";
};
// clang-format off
str << "DeviceConv" << std::to_string(NumDimSpatial)
<< "DFwdAvx2_NHWC_KYXC"
<<"_FS"<< static_cast<int>(ConvForwardSpecialization)
<<"_KS"<< static_cast<int>(GemmKSpecialization)
<<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
<< "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
<< "_TT" << MPerThread << "x" << NPerThread
<< "_A" << string_local_buffer(UseALocalBuffer)
<< "_B" << string_local_buffer(UseBLocalBuffer)
<< "_C" << string_local_buffer(UseCLocalBuffer)
;
if constexpr (!std::is_same<OutElementwiseOperation,
ck::tensor_operation::cpu::element_wise::PassThrough>::value)
{
str << "_" << OutElementwiseOperation::Name();
}
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
#endif
#ifndef DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_HPP
#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_HPP
#include <iostream>
#include <sstream>
#include <numeric>
#include "device.hpp"
#include "device_base_cpu.hpp"
#include "device_conv_fwd_cpu.hpp"
#include "convolution_forward_specialization_cpu.hpp"
#include "common_header.hpp"
#include "../../gpu/device/tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_avx2.hpp"
#include "threadwise_gemm_avx2.hpp"
#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
ck::index_t NumDimSpatial,
ck::index_t MPerBlock, // block means data are designed to fit in cache (L1/L2/L3)
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerThread,
ck::index_t NPerThread,
bool UseALocalBuffer,
bool UseBLocalBuffer,
bool UseCLocalBuffer>
struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
: public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
{
using DeviceOp = DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
using ADataType = InDataType;
using BDataType = WeiDataType;
using CDataType = OutDataType;
using AElementwiseOperation = InElementwiseOperation;
using BElementwiseOperation = WeiElementwiseOperation;
using CElementwiseOperation = OutElementwiseOperation;
// TODO make A/B datatype different
using ABDataType = InDataType;
static constexpr index_t NDimSpatial = NumDimSpatial;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr bool NonTemporalStore = false;
static constexpr auto GetBlockMNKAccessOrder()
{
if constexpr(BlockLoopOverSpecialization == DefaultBlockLoopOver ||
BlockLoopOverSpecialization == LoopOver_MNK)
return ck::Sequence<0, 1, 2>{};
else if constexpr(BlockLoopOverSpecialization == LoopOver_MKN)
return ck::Sequence<0, 2, 1>{};
}
using BlockMNKAccessOrder = decltype(GetBlockMNKAccessOrder());
static constexpr auto GetThreadwiseGemm_Dispatch()
{
if constexpr(MPerThread == 4 && NPerThread == 24)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else if constexpr(MPerThread == 6 && NPerThread == 16)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_6x16_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else
{
// static_assert(false, "invalid Mr/Nr");
}
}
using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
static constexpr auto GetInputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
}
static constexpr auto GetWeightBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(
math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
KPerBlock,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
}
static constexpr auto GetOutputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
}
static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
{
ck::index_t gemm_n_padded =
math::integer_least_multiple(gemm_n, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
const auto wei_gemm_n_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_n, gemm_k));
const auto wei_gemm_padn_k_grid_desc = transform_tensor_descriptor(
wei_gemm_n_k_grid_desc,
make_tuple(make_right_pad_transform(gemm_n, gemm_n_padded - gemm_n),
make_pass_through_transform(gemm_k)),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
const auto wei_gemm_n0_k_n1_grid_desc = transform_tensor_descriptor(
wei_gemm_padn_k_grid_desc,
ck::make_tuple(
ck::make_unmerge_transform(
ck::make_tuple(wei_gemm_padn_k_grid_desc.GetLength(I0) /
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize)),
ck::make_pass_through_transform(wei_gemm_padn_k_grid_desc.GetLength(I1))),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
ck::make_tuple(ck::Sequence<0, 2>{}, ck::Sequence<1>{}));
return wei_gemm_n0_k_n1_grid_desc;
}
static auto GetOutputTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n)
{
const auto out_gemm_m_n_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_n));
return out_gemm_m_n_grid_desc;
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Wi = input_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[0];
const index_t ConvStrideW = conv_filter_strides[0];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t X = filter_spatial_lengths[0];
const index_t ConvDilationW = conv_filter_dilations[0];
const index_t InLeftPadW = input_left_pads[0];
const index_t InRightPadW = input_right_pads[0];
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)),
make_merge_transform(make_tuple(X, C))),
make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t ConvStrideH = conv_filter_strides[0];
const index_t ConvStrideW = conv_filter_strides[1];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Y = filter_spatial_lengths[0];
const index_t X = filter_spatial_lengths[1];
const index_t ConvDilationH = conv_filter_dilations[0];
const index_t ConvDilationW = conv_filter_dilations[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_merge_transform(make_tuple(Y, X, C))),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
ck::index_t gemm_m_pad,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Di = input_spatial_lengths[0];
const index_t Hi = input_spatial_lengths[1];
const index_t Wi = input_spatial_lengths[2];
const index_t Do = output_spatial_lengths[0];
const index_t Ho = output_spatial_lengths[1];
const index_t Wo = output_spatial_lengths[2];
const index_t ConvStrideD = conv_filter_strides[0];
const index_t ConvStrideH = conv_filter_strides[1];
const index_t ConvStrideW = conv_filter_strides[2];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_do_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Z = filter_spatial_lengths[0];
const index_t Y = filter_spatial_lengths[1];
const index_t X = filter_spatial_lengths[2];
const index_t ConvDilationD = conv_filter_dilations[0];
const index_t ConvDilationH = conv_filter_dilations[1];
const index_t ConvDilationW = conv_filter_dilations[2];
const index_t InLeftPadD = input_left_pads[0];
const index_t InLeftPadH = input_left_pads[1];
const index_t InLeftPadW = input_left_pads[2];
const index_t InRightPadD = input_right_pads[0];
const index_t InRightPadH = input_right_pads[1];
const index_t InRightPadW = input_right_pads[2];
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{},
Sequence<1, 2>{},
Sequence<3, 4>{},
Sequence<5, 6>{},
Sequence<7>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_merge_transform(make_tuple(Z, Y, X, C))),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
static index_t GetGemmM(ck::index_t N, const std::vector<ck::index_t>& output_spatial_lengths)
{
return N * std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmK(ck::index_t C, const std::vector<ck::index_t>& filter_spatial_lengths)
{
return C * std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmN(ck::index_t K)
{
// return ck::math::integer_least_multiple(K,
// ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
return K;
}
static auto MakeABCGridDescriptor(ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
using namespace ck;
const index_t GemmM = GetGemmM(N, output_spatial_lengths);
const index_t GemmN = GetGemmN(K);
const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
// A:
const auto in_gemm_m_k_grid_desc =
GetInputTensorDescriptor<NumDimSpatial>(N,
C,
GemmM,
GemmK,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
// B:
const auto wei_gemm_n0_k_n1_grid_desc = GetWeightTensorDescriptor(GemmK, GemmN);
// C:
const auto out_gemm_m_n_grid_desc = GetOutputTensorDescriptor(GemmM, GemmN);
return make_tuple(
in_gemm_m_k_grid_desc, wei_gemm_n0_k_n1_grid_desc, out_gemm_m_n_grid_desc);
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1});
}
using ABCGridDescs = decltype(GetABCGridDesc<NumDimSpatial>());
using AGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
using BGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// static constexpr bool UseCLocalBuffer = false;
using AThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC<
InDataType,
InDataType,
AGridDesc,
decltype(GetInputBlockDescriptor()),
InElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using BThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXC<
WeiDataType,
WeiDataType,
BGridDesc,
decltype(GetWeightBlockDescriptor()),
WeiElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using CThreadwiseCopy = ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN<
OutDataType,
OutDataType,
CGridDesc,
decltype(GetOutputBlockDescriptor()),
OutElementwiseOperation,
!UseCLocalBuffer,
ConvForwardSpecialization,
GemmKSpecialization>;
using GridwiseGemm =
ck::cpu::GridwiseGemmAvx2_MxN<InDataType, // InDataType,
WeiDataType, // WeiDataType,
OutDataType, // OutDataType,
AGridDesc, // AGridDesc,
BGridDesc, // BGridDesc,
CGridDesc, // CGridDesc,
AElementwiseOperation, // AElementwiseOperation,
BElementwiseOperation, // BElementwiseOperation,
CElementwiseOperation, // CElementwiseOperation,
MPerBlock, // MPerBlock,
NPerBlock, // NPerBlock,
KPerBlock, // KPerBlock,
ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
AThreadwiseCopy, // AThreadwiseCopy
BThreadwiseCopy, // BThreadwiseCopy
CThreadwiseCopy, // CThreadwiseCopy
BlockMNKAccessOrder, // BlockMNKAccessOrder,
ck::Sequence<0, 1>, // ThreadMNAccessOrder
UseALocalBuffer, // UseALocalBuffer
UseBLocalBuffer, // UseBLocalBuffer
UseCLocalBuffer // UseCLocalBuffer
>;
// Argument
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
: p_a_grid_{p_in_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_out_grid},
a_grid_desc_{},
b_grid_desc_{},
c_grid_desc_{},
a_element_op_{in_element_op},
b_element_op_{wei_element_op},
c_element_op_{out_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
filter_spatial_lengths_{filter_spatial_lengths},
conv_filter_strides_{conv_filter_strides},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
const auto descs = DeviceOp::MakeABCGridDescriptor(N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
a_grid_desc_ = descs[I0];
b_grid_desc_ = descs[I1];
c_grid_desc_ = descs[I2];
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc a_grid_desc_;
BGridDesc b_grid_desc_;
CGridDesc c_grid_desc_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
index_t Conv_C_;
std::vector<index_t> filter_spatial_lengths_;
std::vector<index_t> conv_filter_strides_;
std::vector<index_t> input_left_pads_;
std::vector<index_t> input_right_pads_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1)
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
{
throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
}
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
const auto kernel = ck::cpu::kernel_gemm_avx_mxn<GridwiseGemm,
InDataType,
WeiDataType,
OutDataType,
AGridDesc,
BGridDesc,
CGridDesc,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
float ave_time = 0;
if(nrepeat != 1)
ave_time = launch_and_time_cpu_kernel(kernel,
nrepeat,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
// TODO: this is for benchmark purpose, so last time we clear c buffer and calculate the
// result
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
launch_cpu_kernel(kernel,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
return ave_time;
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config, nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
// check if it's 1x1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
if constexpr(GemmKSpecialization ==
ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
{
if(!(arg.Conv_C_ % KPerBlock == 0))
return false;
}
// Gridwise GEMM size
return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_grid,
const void* p_wei_grid,
void* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
static_cast<OutDataType*>(p_out_grid),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
auto string_local_buffer = [](bool is_local_buffer) {
if(is_local_buffer)
return "L";
else
return "G";
};
// clang-format off
str << "DeviceConv" << std::to_string(NumDimSpatial)
<< "DFwdAvx2_NHWC_KYXC"
<<"_FS"<< static_cast<int>(ConvForwardSpecialization)
<<"_KS"<< static_cast<int>(GemmKSpecialization)
<<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
<< "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
<< "_TT" << MPerThread << "x" << NPerThread
<< "_A" << string_local_buffer(UseALocalBuffer)
<< "_B" << string_local_buffer(UseBLocalBuffer)
<< "_C" << string_local_buffer(UseCLocalBuffer)
;
if constexpr (!std::is_same<OutElementwiseOperation,
ck::tensor_operation::cpu::element_wise::PassThrough>::value)
{
str << "_" << OutElementwiseOperation::Name();
}
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
#endif
include/ck/tensor_operation/cpu/device/device_convnd_fwd_avx2_nhwc_kyxck8_nhwk.hpp 0 → 100644
View file @ ad09ebdb
#ifndef DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_HPP
#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_HPP
#include <iostream>
#include <sstream>
#include <numeric>
#include "device.hpp"
#include "device_base_cpu.hpp"
#include "device_conv_fwd_cpu.hpp"
#include "convolution_forward_specialization_cpu.hpp"
#include "common_header.hpp"
#include "../../gpu/device/tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_avx2.hpp"
#include "threadwise_gemm_avx2.hpp"
#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
ck::index_t NumDimSpatial,
ck::index_t MPerBlock, // block means data are designed to fit in cache (L1/L2/L3)
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerThread,
ck::index_t NPerThread,
bool UseALocalBuffer,
bool UseBLocalBuffer,
bool UseCLocalBuffer>
struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
: public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
{
using DeviceOp = DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K;
using ADataType = InDataType;
using BDataType = WeiDataType;
using CDataType = OutDataType;
using AElementwiseOperation = InElementwiseOperation;
using BElementwiseOperation = WeiElementwiseOperation;
using CElementwiseOperation = OutElementwiseOperation;
// TODO make A/B datatype different
using ABDataType = InDataType;
static constexpr index_t NDimSpatial = NumDimSpatial;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr bool NonTemporalStore = false;
static constexpr auto GetBlockMNKAccessOrder()
{
if constexpr(BlockLoopOverSpecialization == DefaultBlockLoopOver ||
BlockLoopOverSpecialization == LoopOver_MNK)
return ck::Sequence<0, 1, 2>{};
else if constexpr(BlockLoopOverSpecialization == LoopOver_MKN)
return ck::Sequence<0, 2, 1>{};
}
using BlockMNKAccessOrder = decltype(GetBlockMNKAccessOrder());
static constexpr auto GetThreadwiseGemm_Dispatch()
{
if constexpr(MPerThread == 4 && NPerThread == 24)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else if constexpr(MPerThread == 6 && NPerThread == 16)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_6x16_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else
{
// static_assert(false, "invalid Mr/Nr");
}
}
using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
static constexpr auto GetInputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
}
static constexpr auto GetWeightBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(
math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
KPerBlock,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
}
static constexpr auto GetOutputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
}
static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
{
return make_naive_tensor_descriptor_packed(make_tuple(gemm_n / 8, gemm_k, 8));
}
static auto GetOutputTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n)
{
const auto out_gemm_m_n_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_n));
return out_gemm_m_n_grid_desc;
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Wi = input_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[0];
const index_t ConvStrideW = conv_filter_strides[0];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t X = filter_spatial_lengths[0];
const index_t ConvDilationW = conv_filter_dilations[0];
const index_t InLeftPadW = input_left_pads[0];
const index_t InRightPadW = input_right_pads[0];
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)),
make_merge_transform(make_tuple(X, C))),
make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t ConvStrideH = conv_filter_strides[0];
const index_t ConvStrideW = conv_filter_strides[1];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Y = filter_spatial_lengths[0];
const index_t X = filter_spatial_lengths[1];
const index_t ConvDilationH = conv_filter_dilations[0];
const index_t ConvDilationW = conv_filter_dilations[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_merge_transform(make_tuple(Y, X, C))),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
ck::index_t gemm_m_pad,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Di = input_spatial_lengths[0];
const index_t Hi = input_spatial_lengths[1];
const index_t Wi = input_spatial_lengths[2];
const index_t Do = output_spatial_lengths[0];
const index_t Ho = output_spatial_lengths[1];
const index_t Wo = output_spatial_lengths[2];
const index_t ConvStrideD = conv_filter_strides[0];
const index_t ConvStrideH = conv_filter_strides[1];
const index_t ConvStrideW = conv_filter_strides[2];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_do_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Z = filter_spatial_lengths[0];
const index_t Y = filter_spatial_lengths[1];
const index_t X = filter_spatial_lengths[2];
const index_t ConvDilationD = conv_filter_dilations[0];
const index_t ConvDilationH = conv_filter_dilations[1];
const index_t ConvDilationW = conv_filter_dilations[2];
const index_t InLeftPadD = input_left_pads[0];
const index_t InLeftPadH = input_left_pads[1];
const index_t InLeftPadW = input_left_pads[2];
const index_t InRightPadD = input_right_pads[0];
const index_t InRightPadH = input_right_pads[1];
const index_t InRightPadW = input_right_pads[2];
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{},
Sequence<1, 2>{},
Sequence<3, 4>{},
Sequence<5, 6>{},
Sequence<7>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_merge_transform(make_tuple(Z, Y, X, C))),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
static index_t GetGemmM(ck::index_t N, const std::vector<ck::index_t>& output_spatial_lengths)
{
return N * std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmK(ck::index_t C, const std::vector<ck::index_t>& filter_spatial_lengths)
{
return C * std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmN(ck::index_t K)
{
// return ck::math::integer_least_multiple(K,
// ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
return K;
}
static auto MakeABCGridDescriptor(ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
using namespace ck;
const index_t GemmM = GetGemmM(N, output_spatial_lengths);
const index_t GemmN = GetGemmN(K);
const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
// A:
const auto in_gemm_m_k_grid_desc =
GetInputTensorDescriptor<NumDimSpatial>(N,
C,
GemmM,
GemmK,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
// B:
const auto wei_gemm_n0_k_n1_grid_desc = GetWeightTensorDescriptor(GemmK, GemmN);
// C:
const auto out_gemm_m_n_grid_desc = GetOutputTensorDescriptor(GemmM, GemmN);
return make_tuple(
in_gemm_m_k_grid_desc, wei_gemm_n0_k_n1_grid_desc, out_gemm_m_n_grid_desc);
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1});
}
using ABCGridDescs = decltype(GetABCGridDesc<NumDimSpatial>());
using AGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
using BGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// static constexpr bool UseCLocalBuffer = false;
using AThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC<
InDataType,
InDataType,
AGridDesc,
decltype(GetInputBlockDescriptor()),
InElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using BThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXCK8<
WeiDataType,
WeiDataType,
BGridDesc,
decltype(GetWeightBlockDescriptor()),
WeiElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using CThreadwiseCopy = ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN<
OutDataType,
OutDataType,
CGridDesc,
decltype(GetOutputBlockDescriptor()),
OutElementwiseOperation,
!UseCLocalBuffer,
ConvForwardSpecialization,
GemmKSpecialization>;
using GridwiseGemm =
ck::cpu::GridwiseGemmAvx2_MxN<InDataType, // InDataType,
WeiDataType, // WeiDataType,
OutDataType, // OutDataType,
AGridDesc, // AGridDesc,
BGridDesc, // BGridDesc,
CGridDesc, // CGridDesc,
AElementwiseOperation, // AElementwiseOperation,
BElementwiseOperation, // BElementwiseOperation,
CElementwiseOperation, // CElementwiseOperation,
MPerBlock, // MPerBlock,
NPerBlock, // NPerBlock,
KPerBlock, // KPerBlock,
ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
AThreadwiseCopy, // AThreadwiseCopy
BThreadwiseCopy, // BThreadwiseCopy
CThreadwiseCopy, // CThreadwiseCopy
BlockMNKAccessOrder, // BlockMNKAccessOrder,
ck::Sequence<0, 1>, // ThreadMNAccessOrder
UseALocalBuffer, // UseALocalBuffer
UseBLocalBuffer, // UseBLocalBuffer
UseCLocalBuffer // UseCLocalBuffer
>;
// Argument
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
: p_a_grid_{p_in_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_out_grid},
a_grid_desc_{},
b_grid_desc_{},
c_grid_desc_{},
a_element_op_{in_element_op},
b_element_op_{wei_element_op},
c_element_op_{out_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
filter_spatial_lengths_{filter_spatial_lengths},
conv_filter_strides_{conv_filter_strides},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
const auto descs = DeviceOp::MakeABCGridDescriptor(N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
a_grid_desc_ = descs[I0];
b_grid_desc_ = descs[I1];
c_grid_desc_ = descs[I2];
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc a_grid_desc_;
BGridDesc b_grid_desc_;
CGridDesc c_grid_desc_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
index_t Conv_C_;
std::vector<index_t> filter_spatial_lengths_;
std::vector<index_t> conv_filter_strides_;
std::vector<index_t> input_left_pads_;
std::vector<index_t> input_right_pads_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1)
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
{
throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
}
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
const auto kernel = ck::cpu::kernel_gemm_avx_mxn<GridwiseGemm,
InDataType,
WeiDataType,
OutDataType,
AGridDesc,
BGridDesc,
CGridDesc,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
float ave_time = 0;
if(nrepeat != 1)
ave_time = launch_and_time_cpu_kernel(kernel,
nrepeat,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
// TODO: this is for benchmark purpose, so last time we clear c buffer and calculate the
// result
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
launch_cpu_kernel(kernel,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
return ave_time;
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config, nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
// check if it's 1x1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
if constexpr(GemmKSpecialization ==
ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
{
if(!(arg.Conv_C_ % KPerBlock == 0))
return false;
}
if(!(arg.Conv_K_ % 8 == 0))
return false;
// Gridwise GEMM size
return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_grid,
const void* p_wei_grid,
void* p_out_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
static_cast<OutDataType*>(p_out_grid),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
auto string_local_buffer = [](bool is_local_buffer) {
if(is_local_buffer)
return "L";
else
return "G";
};
// clang-format off
str << "DeviceConv" << std::to_string(NumDimSpatial)
<< "DFwdAvx2_NHWC_KYXCK8"
<<"_FS"<< static_cast<int>(ConvForwardSpecialization)
<<"_KS"<< static_cast<int>(GemmKSpecialization)
<<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
<< "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
<< "_TT" << MPerThread << "x" << NPerThread
<< "_A" << string_local_buffer(UseALocalBuffer)
<< "_B" << string_local_buffer(UseBLocalBuffer)
<< "_C" << string_local_buffer(UseCLocalBuffer)
;
if constexpr (!std::is_same<OutElementwiseOperation,
ck::tensor_operation::cpu::element_wise::PassThrough>::value)
{
str << "_" << OutElementwiseOperation::Name();
}
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
#endif
include/ck/tensor_operation/cpu/thread/threadwise_tensor_slice_transfer_avx2_specialization.hpp
View file @ ad09ebdb
......@@ -484,8 +484,10 @@ struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC
while(i_m_itr > 0)
{
if((*reinterpret_cast<uint32_t*>(&i_hi_itr) < Hi) &&
(*reinterpret_cast<uint32_t*>(&i_wi_itr) < Wi))
if((*reinterpret_cast<uint32_t*>(&i_hi_itr) <
*reinterpret_cast<uint32_t*>(&Hi)) &&
(*reinterpret_cast<uint32_t*>(&i_wi_itr) <
*reinterpret_cast<uint32_t*>(&Wi)))
avx2_util::memcpy32_avx2(p_dst, p_src, k_per_block, element_op_);
else
avx2_util::memset32_avx2(p_dst, 0, k_per_block);
......@@ -543,8 +545,10 @@ struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC
// printf("current_k_block_along_c:%d, i_c_itr_k:%d, k_per_block:%d\n",
// current_k_block_along_c, i_c_itr_k,k_per_block); fflush(stdout);
if((*reinterpret_cast<uint32_t*>(&i_hi_itr_k) < Hi) &&
(*reinterpret_cast<uint32_t*>(&i_wi_itr_k) < Wi))
if((*reinterpret_cast<uint32_t*>(&i_hi_itr_k) <
*reinterpret_cast<uint32_t*>(&Hi)) &&
(*reinterpret_cast<uint32_t*>(&i_wi_itr_k) <
*reinterpret_cast<uint32_t*>(&Wi)))
avx2_util::memcpy32_avx2(
p_dst_k, p_src_k, current_k_block_along_c, element_op_);
else
......@@ -715,7 +719,7 @@ template <typename SrcData,
bool BypassTransfer,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization>
struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC
struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXC
{
static constexpr ck::index_t nDim = SrcDesc::GetNumOfDimension();
using Index = MultiIndex<nDim>;
......@@ -723,7 +727,7 @@ struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC
// using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
// using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
constexpr ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC(
constexpr ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXC(
const SrcDesc& src_desc,
const Index& src_slice_origin,
const DstDesc& dst_desc,
......@@ -927,6 +931,190 @@ struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC
intptr_t src_offset;
};
template <typename SrcData,
typename DstData,
typename SrcDesc,
typename DstDesc,
typename ElementwiseOperation,
bool BypassTransfer,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization>
struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXCK8
{
static constexpr ck::index_t nDim = SrcDesc::GetNumOfDimension();
using Index = MultiIndex<nDim>;
constexpr ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXCK8(
const SrcDesc& src_desc,
const Index& src_slice_origin,
const DstDesc& dst_desc,
const Index& dst_slice_origin,
const ElementwiseOperation& element_op)
: element_op_(element_op)
{
GemmN1 =
src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}]; // Need to be 8
GemmN = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
GemmK = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
}
void SetSrcSliceOrigin(const SrcDesc&, const Index& src_slice_origin_idx)
{
ck::index_t idx_n0 = src_slice_origin_idx[Number<0>{}];
ck::index_t idx_k = src_slice_origin_idx[Number<1>{}];
ck::index_t idx_n1 = src_slice_origin_idx[Number<2>{}];
src_offset = idx_n0 * GemmK * GemmN1 + idx_k * GemmN1 + idx_n1;
// printf("xxxx i_gemm_n:%d, i_gemm_k:%d, src_offset:%d\n", i_gemm_n, i_gemm_k,
// src_offset);
}
void SetDstSliceOrigin(const DstDesc&, const Index&) {}
template <typename SrcBuffer, typename DstBuffer, typename SliceLengths>
void RunRead(const SrcDesc&,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const SliceLengths& slice_length)
{
if constexpr(BypassTransfer) {}
else
{
const ck::index_t n0_per_block = slice_length[Number<0>{}];
const ck::index_t k_n1_per_block =
slice_length[Number<1>{}] * slice_length[Number<2>{}];
const ck::index_t SrcStride_K_N1 = GemmK * slice_length[Number<2>{}];
// printf(" >>>> %d, %d, %d -> %d(%dx%d), %d\n", GemmN, GemmK, GemmN1, n_per_block,
// dst_desc.GetTransforms()[Number<0>{}]
// .GetUpperLengths()[Number<0>{}],
// dst_desc.GetTransforms()[Number<0>{}]
// .GetUpperLengths()[Number<2>{}],
// k_per_block);
const float* p_src = reinterpret_cast<const float*>(src_buf.p_data_) + src_offset;
float* p_dst = reinterpret_cast<float*>(dst_buf.p_data_);
// n0 * k * n1
index_t i_n0_itr = n0_per_block;
while(i_n0_itr >= 8)
{
avx2_util::memcpy32_avx2(p_dst + 0 * k_n1_per_block,
p_src + 0 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 1 * k_n1_per_block,
p_src + 1 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 2 * k_n1_per_block,
p_src + 2 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 3 * k_n1_per_block,
p_src + 3 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 4 * k_n1_per_block,
p_src + 4 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 5 * k_n1_per_block,
p_src + 5 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 6 * k_n1_per_block,
p_src + 6 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 7 * k_n1_per_block,
p_src + 7 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
i_n0_itr -= 8;
p_dst += 8 * k_n1_per_block;
p_src += 8 * SrcStride_K_N1;
}
if(i_n0_itr & 4)
{
avx2_util::memcpy32_avx2(p_dst + 0 * k_n1_per_block,
p_src + 0 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 1 * k_n1_per_block,
p_src + 1 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 2 * k_n1_per_block,
p_src + 2 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 3 * k_n1_per_block,
p_src + 3 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
p_dst += 4 * k_n1_per_block;
p_src += 4 * SrcStride_K_N1;
}
if(i_n0_itr & 2)
{
avx2_util::memcpy32_avx2(p_dst + 0 * k_n1_per_block,
p_src + 0 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
avx2_util::memcpy32_avx2(p_dst + 1 * k_n1_per_block,
p_src + 1 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
p_dst += 2 * k_n1_per_block;
p_src += 2 * SrcStride_K_N1;
}
if(i_n0_itr & 1)
{
avx2_util::memcpy32_avx2(p_dst + 0 * k_n1_per_block,
p_src + 0 * SrcStride_K_N1,
k_n1_per_block,
element_op_);
}
}
}
// src_slice_origin_step_idx need to be known at compile-time, for performance reason
void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& src_slice_origin_step_idx)
{
ck::index_t move_n0 = src_slice_origin_step_idx[Number<0>{}];
ck::index_t move_k = src_slice_origin_step_idx[Number<1>{}];
ck::index_t move_n1 = src_slice_origin_step_idx[Number<2>{}];
// i_gemm_k += move_k;
// printf("wei move:%d\n", move_k); fflush(stdout);
src_offset += move_n0 * GemmK * GemmN1 + move_k * GemmN1 + move_n1;
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
void MoveDstSliceWindow(const DstDesc&, const Index&) {}
private:
const ElementwiseOperation element_op_;
ck::index_t i_gemm_n;
// ck::index_t i_gemm_k;
// ck::index_t GemmN0;
ck::index_t GemmN1;
ck::index_t GemmN;
ck::index_t GemmK;
intptr_t src_offset;
};
template <typename SrcData,
typename DstData,
typename SrcDesc,
......
library/src/tensor_operation_instance/cpu/conv2d_fwd/CMakeLists.txt
View file @ ad09ebdb
# device_conv2d_fwd_cpu_instance
set(DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE
device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_instance.cpp
device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_instance.cpp
)
add_library(device_conv2d_fwd_cpu_instance SHARED ${DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE})
target_compile_features(device_conv2d_fwd_cpu_instance PUBLIC)
......
library/src/tensor_operation_instance/cpu/conv2d_fwd/device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_instance.cpp 0 → 100644
View file @ ad09ebdb
#include <stdlib.h>
#include "convolution_forward_specialization_cpu.hpp"
#include "config.hpp"
#include "device_convnd_fwd_avx2_nhwc_kyxck8_nhwk.hpp"
#include "element_wise_operation_cpu.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
namespace device_conv2d_fwd_avx2_instance {
using InType = float;
using WeiType = float;
using OutType = float;
using AccType = float;
using InLayout = ck::tensor_layout::gemm::RowMajor; // NHWC
using WeiLayout = ck::tensor_layout::gemm::ColumnMajor; // KYXCK8
static constexpr bool NonTemporalStore = false;
using PT = ck::tensor_operation::cpu::element_wise::PassThrough;
using Relu = ck::tensor_operation::cpu::element_wise::Relu;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Default;
static constexpr auto ConvFwd1x1P0 =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0;
static constexpr auto DefaultGemmKLoop =
ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::DefaultGemmKLoop;
static constexpr auto GemmKLoopOverC =
ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC;
static constexpr auto LoopOver_MNK = ck::tensor_operation::cpu::device::LoopOver_MNK;
static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver_MKN;
// clang-format off
#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(a_elem_op, b_elem_op, c_elem_op, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf) \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>
// clang-format on
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 128, 256, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 240, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 256, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 768, 320, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 896, 352, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true, true, false)>;
// clang-format on
// use this in single thread, but gemm_n is not multiple of 8
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 128, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 240, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 768, 320, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 896, 352, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true, true, true)>;
// clang-format on
// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
// time no local c is better...)
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 48, 24, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 72, 16, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 72, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 96, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 96, 64, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 120, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 120, 64, 128, 6, 16, true, true, true),
// DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 128, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 240, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 512, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 768, 320, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 896, 352, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true, true, true)>;
// clang-format on
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_relu_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 256, 128, 64, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 256, 128, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 128, 256, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 240, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 256, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 768, 320, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 896, 352, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true, true, false)>;
// clang-format on
// use this in single thread, but gemm_n is not multiple of 8
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_relu_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 256, 128, 64, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 256, 128, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 128, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 240, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 768, 320, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 896, 352, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true, true, true)>;
// clang-format on
// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
// time no local c is better...)
using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 48, 24, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 72, 16, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 72, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 96, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 96, 64, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 120, 32, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 120, 64, 128, 6, 16, true, true, true),
// DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 256, 128, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 128, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 240, 128, 4, 24, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 512, 256, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 768, 320, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 896, 352, 128, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true, true, true)>;
// clang-format on
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(
std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_instances{});
}
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(
std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances{});
}
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(
std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances{});
}
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(
std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_relu_instances{});
}
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(
std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_relu_instances{});
}
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(
std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances{});
}
} // namespace device_conv2d_fwd_avx2_instance
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
profiler/include/profile_conv_fwd_cpu_impl.hpp
View file @ ad09ebdb
......@@ -33,6 +33,24 @@ void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c_relu(
void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
} // namespace device_conv2d_fwd_avx2_instance
} // namespace device
} // namespace cpu
......
test/convnd_fwd_cpu/conv2d_fwd_cpu.cpp
View file @ ad09ebdb
......@@ -16,7 +16,11 @@
#define TEST_FUSION_PASSTHROUGH 0
#define TEST_FUSION_RELU 1
#define TEST_FUSION TEST_FUSION_RELU
#define TEST_FUSION TEST_FUSION_PASSTHROUGH
#define TEST_LAYOUT_NHWC_KYXC_NHWK 0
#define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1
#define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXCK8_NHWK
using F32 = float;
using F16 = ck::half_t;
......@@ -48,6 +52,24 @@ void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c_relu(
void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, Relu>>& instances);
} // namespace device_conv2d_fwd_avx2_instance
} // namespace device
} // namespace cpu
......@@ -115,6 +137,31 @@ check_out(const Tensor<T>& ref, const Tensor<T>& result, double nrms, int per_pi
float calculate_gflops() {}
template <typename T>
void transpose_kyxc_2_kyxc8k(Tensor<T>& dst,
const Tensor<T>& src,
ck::index_t K,
ck::index_t Y,
ck::index_t X,
ck::index_t C)
{
ck::index_t batch = K / 8;
ck::index_t row = 8;
ck::index_t col = C * Y * X;
for(auto i_b = 0; i_b < batch; i_b++)
{
for(auto i_r = 0; i_r < row; i_r++)
{
for(auto i_c = 0; i_c < col; i_c++)
{
ck::index_t src_idx = i_b * row * col + i_r * col + i_c;
ck::index_t dst_idx = i_b * col * row + i_c * row + i_r;
dst.mData[dst_idx] = src.mData[src_idx];
}
}
}
}
int main(int argc, char* argv[])
{
int data_type = 0;
......@@ -213,6 +260,10 @@ int main(int argc, char* argv[])
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
Tensor<WeiDataType> wei_k_c_y_x_k8(
f_host_tensor_descriptor(K, C, Y, X)); // TODO: This is only to hold data
#endif
Tensor<OutDataType> out_n_k_ho_wo_host_result(f_host_tensor_descriptor(N, K, Ho, Wo));
Tensor<OutDataType> out_n_k_ho_wo_device_result(f_host_tensor_descriptor(N, K, Ho, Wo));
......@@ -296,8 +347,13 @@ int main(int argc, char* argv[])
AVX2_DATA_ALIGNMENT);
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
#endif
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
transpose_kyxc_2_kyxc8k(wei_k_c_y_x_k8, wei_k_c_y_x, K, Y, X, C);
wei_device_buf.ToDevice(wei_k_c_y_x_k8.mData.data());
#endif
// get host result
{
auto ref_conv = ReferenceConvFwdInstance{};
......@@ -334,6 +390,7 @@ int main(int argc, char* argv[])
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK
#if TEST_FUSION == TEST_FUSION_PASSTHROUGH
if(omp_get_max_threads() > 1)
{
......@@ -369,6 +426,45 @@ int main(int argc, char* argv[])
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c_relu(conv_ptrs);
}
#endif
#endif
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
#if TEST_FUSION == TEST_FUSION_PASSTHROUGH
if(omp_get_max_threads() > 1)
{
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(conv_ptrs);
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(conv_ptrs);
}
else
{
if(K % 8 == 0)
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(conv_ptrs);
else
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(conv_ptrs);
}
#endif
#if TEST_FUSION == TEST_FUSION_RELU
if(omp_get_max_threads() > 1)
{
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(conv_ptrs);
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(conv_ptrs);
}
else
{
if(K % 8 == 0)
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(conv_ptrs);
else
ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(conv_ptrs);
}
#endif
#endif
}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment