Skip to content

GitLab

Unverified Commit d1db6a0c authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Absolute include path (#281) 

* ad gelu and fast_gelu

* added GeLU and fast GeLU

* clean up

* add gemm+fastgelu example

* add gemm+gelu instances

* update profiler

* clean up

* clean up

* adding gemm+bias+activation

* clean

* adding bias

* clean

* adding gemm multiple d

* debugging

* add gemm bias add fastgelu

* rename, clean

* refactoring; add readme

* refactor

* refactor

* refactor

* refactor

* refactor

* refactor

* fix

* fix

* update example

* update example

* rename

* update example

* add ckProfiler

* clean

* clean

* clean

* clean

* add client app example

* update readme

* delete obselete files

* remove old client app

* delete old file

* cleaning

* clean

* remove half

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path for all examples

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* fix header path

* revert client app example

* clean build

* fix build

* temporary disable client test on Jenkins

* clean

* clean

* clean
parent a49115b9
Show whitespace changes
Inline Side-by-side
Showing with 75 additions and 1917 deletions
library/include/ck/library/obselete_driver_offline/driver_convolution_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_CONVOLUTION_FORWARD_IMPLICIT_GEMM_V5R1_DLOPS_NC0HWc1_KC0YXC1_NK0HWK1_HPP
#define DRIVER_CONVOLUTION_FORWARD_IMPLICIT_GEMM_V5R1_DLOPS_NC0HWc1_KC0YXC1_NK0HWK1_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_dlops_v3.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::index_t E1_,
ck::index_t E2_,
ck::index_t K2_,
ck::index_t KPerBlock,
ck::index_t HoPerBlock,
ck::index_t WoPerBlock,
ck::index_t E1PerBlock,
ck::index_t KPerThread,
ck::index_t HoPerThread,
ck::index_t WoPerThread,
ck::index_t EPerThread,
typename ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
typename ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
ck::index_t ABlockTransferSrcScalarPerVector_E2,
ck::index_t ABlockTransferDstScalarPerVector_E2,
ck::index_t BThreadTransferSrcScalarPerVector_E2,
ck::index_t CThreadTransferDstScalarPerVector_K,
ck::ActivTypeEnum activ_type>
struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_outpad
{
template <typename... Wei,
typename... In,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
__host__ float Run(const ck::TensorDescriptor<Wei...>& wei_k_c0_y_x_c1_global_desc,
const ck::TensorDescriptor<In...>& in_n_c0_hi_wi_c1_global_desc,
const ck::TensorDescriptor<Out...>& out_n_k0_ho_wo_k1_global_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatC* __restrict__ p_bias_grid,
FloatC* __restrict__ p_c_grid,
const int nrepeat) const
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
const auto N = in_n_c0_hi_wi_c1_global_desc.GetLength(I0);
const auto C0 = in_n_c0_hi_wi_c1_global_desc.GetLength(I1);
const auto Hi = in_n_c0_hi_wi_c1_global_desc.GetLength(I2);
const auto Wi = in_n_c0_hi_wi_c1_global_desc.GetLength(I3);
// const auto C1 = in_n_c0_hi_wi_c1_global_desc.GetLength(I4);
const auto K0 = out_n_k0_ho_wo_k1_global_desc.GetLength(I1);
const auto Ho = out_n_k0_ho_wo_k1_global_desc.GetLength(I2);
const auto Wo = out_n_k0_ho_wo_k1_global_desc.GetLength(I3);
const auto K1 = out_n_k0_ho_wo_k1_global_desc.GetLength(I4);
const auto K = wei_k_c0_y_x_c1_global_desc.GetLength(I0);
const auto Y = wei_k_c0_y_x_c1_global_desc.GetLength(I2);
const auto X = wei_k_c0_y_x_c1_global_desc.GetLength(I3);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
#if CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR
const auto Hop = Number<(Ho + HoPerBlock - 1) / HoPerBlock * HoPerBlock>{};
const auto Wop = Number<(Wo + WoPerBlock - 1) / WoPerBlock * WoPerBlock>{};
#else
const auto Hop = (Ho + HoPerBlock - 1) / HoPerBlock * HoPerBlock;
const auto Wop = (Wo + WoPerBlock - 1) / WoPerBlock * WoPerBlock;
#endif
const auto OutRightPadH = Hop - Ho;
const auto OutRightPadW = Wop - Wo;
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0] + OutRightPadH * ConvStrideH;
const auto InRightPadW = in_right_pads[I1] + OutRightPadW * ConvStrideW;
const auto E = C0 * Y * X;
constexpr auto E1 = Number<E1_>{};
constexpr auto E2 = Number<E2_>{};
constexpr auto K2 = Number<K2_>{};
const auto E0 = E / E1;
// weight tensor
const auto a_e_k_e2_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, C0 * Y * X, E2)),
make_tuple(make_pass_through_transform(K),
make_pass_through_transform(C0 * Y * X),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}, Sequence<2>{}));
const auto a_e0_e1_k_e2_grid_desc =
transform_tensor_descriptor(a_e_k_e2_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)),
make_pass_through_transform(K),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
// input tensor
const auto in_n_c0_hip_wip_e2_global_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, E2)),
make_tuple(make_pass_through_transform(N),
make_pass_through_transform(C0),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(E2)),
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_c0_y_ho_x_wo_e2_global_desc = transform_tensor_descriptor(
in_n_c0_hip_wip_e2_global_desc,
make_tuple(
make_pass_through_transform(N),
make_pass_through_transform(C0),
make_embed_transform(make_tuple(Y, Hop), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wop), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}, Sequence<6>{}));
const auto in_e_n_ho_wo_e2_grid_desc = transform_tensor_descriptor(
in_n_c0_y_ho_x_wo_e2_global_desc,
make_tuple(make_merge_transform(make_tuple(C0, Y, X)),
make_pass_through_transform(N),
make_pass_through_transform(Hop),
make_pass_through_transform(Wop),
make_pass_through_transform(E2)),
make_tuple(
Sequence<1, 2, 4>{}, Sequence<0>{}, Sequence<3>{}, Sequence<5>{}, Sequence<6>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto b_e0_e1_n_ho_wo_e2_grid_desc = transform_tensor_descriptor(
in_e_n_ho_wo_e2_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)),
make_pass_through_transform(N),
make_pass_through_transform(Hop),
make_pass_through_transform(Wop),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}, Sequence<5>{}));
// output tensor
const auto c_k_n_hop_wop_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1)),
make_tuple(make_merge_transform(make_tuple(K0, K1)),
make_pass_through_transform(N),
make_pad_transform(Ho, I0, OutRightPadH),
make_pad_transform(Wo, I0, OutRightPadW)),
make_tuple(Sequence<1, 4>{}, Sequence<0>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
std::cerr << "Hop = " << Hop << " Wop = " << Wop << std::endl;
if(!((K % KPerBlock) == 0 && (Hop % HoPerBlock) == 0 && (Wop % WoPerBlock) == 0 &&
(E1 % E1PerBlock) == 0))
{
throw std::runtime_error("wrong! GEMM size no divisible");
}
// clang-format off
// hack to control index calculation when iterating over a_e0_e1_k_e2_global tensor
constexpr auto a_e0_e1_k_e2_global_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
constexpr auto a_e0_e1_k_e2_global_move_slice_window_step_hack =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
// hack to control index calculation when iterating over b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global tensor
constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks =
make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})
);
constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
// hack to control index calculation when iterating over c_k0_k1_n_h0_h1_h2_w0_w1_w2_global tensor
constexpr auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks =
make_tuple(make_tuple(Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
// clang-format on
// GEMM
using GridwiseGemm = GridwiseGemmDlops_km_kn_mn_v3<
BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum::Set,
decltype(a_e0_e1_k_e2_grid_desc),
decltype(b_e0_e1_n_ho_wo_e2_grid_desc),
decltype(c_k_n_hop_wop_grid_desc),
decltype(c_k_n_hop_wop_grid_desc),
E1,
E2,
K2,
KPerBlock,
HoPerBlock,
WoPerBlock,
E1PerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
Sequence<2, 3, 0, 1, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>, // E0, E1, N, H0, H1, H2, W0, W1, W2, E2
9,
BThreadTransferSrcScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8>, // K0, K1, N, H0, H1, H2, W0, W1, W2
1,
CThreadTransferDstScalarPerVector_K,
decltype(a_e0_e1_k_e2_global_step_hacks),
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks),
decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack)>;
const auto a_e0_e1_k0_k1_e2_grid_desc =
GridwiseGemm::MakeAE0E1K0K1E2GridDescriptor(a_e0_e1_k_e2_grid_desc);
const auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc =
GridwiseGemm::MakeBE0E1NH0H1H2W0W1W2E2GridDescriptor(b_e0_e1_n_ho_wo_e2_grid_desc);
const auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc =
GridwiseGemm::MakeCK0K1NH0H1H2W0W1W2GridDescriptor(c_k_n_hop_wop_grid_desc);
using AGridDesc_E0_E1_K0_K1_E2 = decltype(a_e0_e1_k0_k1_e2_grid_desc);
using BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 =
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc);
using CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 = decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
const auto grid_size = (K / KPerBlock) * (Hop / HoPerBlock) * (Wop / WoPerBlock) * N;
const bool has_main_e0_block_loop = E0 > 1;
std::cerr << "has_main_e0_block_loop = " << has_main_e0_block_loop << std::endl;
const auto cblockid_to_k_n_h_w_block_cluster_adaptor =
GridwiseGemm::MakeCBlockIdToKNHoWoBlockClusterAdaptor(c_k_n_hop_wop_grid_desc);
using CBlockIdToBlockClusterAdaptor_K_N_H_W =
decltype(cblockid_to_k_n_h_w_block_cluster_adaptor);
float ave_time = 0;
if(has_main_e0_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K0_K1_E2>,
remove_reference_t<BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2>,
remove_reference_t<CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_H_W>,
true,
activ_type>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_bias_grid,
p_c_grid,
a_e0_e1_k0_k1_e2_grid_desc,
b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
cblockid_to_k_n_h_w_block_cluster_adaptor);
}
else
{
const auto kernel =
kernel_gemm_dlops_v3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K0_K1_E2>,
remove_reference_t<BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2>,
remove_reference_t<CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_H_W>,
false,
activ_type>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_bias_grid,
p_c_grid,
a_e0_e1_k0_k1_e2_grid_desc,
b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
cblockid_to_k_n_h_w_block_cluster_adaptor);
}
return ave_time;
}
};
#endif
library/include/ck/library/obselete_driver_offline/driver_convolution_maxpool_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_CONVOLUTION_MAXPOOL_FORWARD_IMPLICIT_GEMM_V5R1_DLOPS_NC0HWc1_KC0YXC1_NK0HWK1_HPP
#define DRIVER_CONVOLUTION_MAXPOOL_FORWARD_IMPLICIT_GEMM_V5R1_DLOPS_NC0HWc1_KC0YXC1_NK0HWK1_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_dlops_v3.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::index_t E1_,
ck::index_t E2_,
ck::index_t K2_,
ck::index_t KPerBlock,
ck::index_t HoPerBlock,
ck::index_t WoPerBlock,
ck::index_t E1PerBlock,
ck::index_t KPerThread,
ck::index_t HoPerThread,
ck::index_t WoPerThread,
ck::index_t EPerThread,
typename ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
typename ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
ck::index_t ABlockTransferSrcScalarPerVector_E2,
ck::index_t ABlockTransferDstScalarPerVector_E2,
ck::index_t BThreadTransferSrcScalarPerVector_E2,
ck::index_t CThreadTransferDstScalarPerVector_K,
ck::ActivTypeEnum activ_type>
struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_maxpool
{
template <typename... Wei,
typename... In,
typename... MaxPool,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
__host__ float Run(const ck::TensorDescriptor<Wei...>& wei_k_c0_y_x_c1_global_desc,
const ck::TensorDescriptor<In...>& in_n_c0_hi_wi_c1_global_desc,
const ck::TensorDescriptor<Out...>& out_n_k0_ho_wo_k1_global_desc,
const ck::TensorDescriptor<MaxPool...>& max_n_k0_hx_wx_k1_global_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatC* __restrict__ p_bias_grid,
FloatC* __restrict__ p_c_grid,
FloatC* __restrict__ p_d_grid,
const int nrepeat) const
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
const auto N = in_n_c0_hi_wi_c1_global_desc.GetLength(I0);
const auto C0 = in_n_c0_hi_wi_c1_global_desc.GetLength(I1);
const auto Hi = in_n_c0_hi_wi_c1_global_desc.GetLength(I2);
const auto Wi = in_n_c0_hi_wi_c1_global_desc.GetLength(I3);
// const auto C1 = in_n_c0_hi_wi_c1_global_desc.GetLength(I4);
const auto K0 = out_n_k0_ho_wo_k1_global_desc.GetLength(I1);
const auto Ho = out_n_k0_ho_wo_k1_global_desc.GetLength(I2);
const auto Wo = out_n_k0_ho_wo_k1_global_desc.GetLength(I3);
const auto K1 = out_n_k0_ho_wo_k1_global_desc.GetLength(I4);
const auto Hx = max_n_k0_hx_wx_k1_global_desc.GetLength(I2);
const auto Wx = max_n_k0_hx_wx_k1_global_desc.GetLength(I3);
const auto K = wei_k_c0_y_x_c1_global_desc.GetLength(I0);
const auto Y = wei_k_c0_y_x_c1_global_desc.GetLength(I2);
const auto X = wei_k_c0_y_x_c1_global_desc.GetLength(I3);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
#if CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR
const auto Hop = Number<(Ho + HoPerBlock - 1) / HoPerBlock * HoPerBlock>{};
const auto Wop = Number<(Wo + WoPerBlock - 1) / WoPerBlock * WoPerBlock>{};
const auto OutRightPadH = Hop - Ho;
const auto OutRightPadW = Wop - Wo;
const auto OutRightPadHx = Number<OutRightPadH / 2>{};
const auto OutRightPadWx = Number<OutRightPadW / 2>{};
#else
const auto Hop = (Ho + HoPerBlock - 1) / HoPerBlock * HoPerBlock;
const auto Wop = (Wo + WoPerBlock - 1) / WoPerBlock * WoPerBlock;
const auto OutRightPadH = Hop - Ho;
const auto OutRightPadW = Wop - Wo;
const auto OutRightPadHx = OutRightPadH / 2;
const auto OutRightPadWx = OutRightPadW / 2;
#endif
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0] + OutRightPadH * ConvStrideH;
const auto InRightPadW = in_right_pads[I1] + OutRightPadW * ConvStrideW;
const auto E = C0 * Y * X;
constexpr auto E1 = Number<E1_>{};
constexpr auto E2 = Number<E2_>{};
constexpr auto K2 = Number<K2_>{};
const auto E0 = E / E1;
// weight tensor
const auto a_e_k_e2_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, C0 * Y * X, E2)),
make_tuple(make_pass_through_transform(K),
make_pass_through_transform(C0 * Y * X),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}, Sequence<2>{}));
const auto a_e0_e1_k_e2_grid_desc =
transform_tensor_descriptor(a_e_k_e2_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)),
make_pass_through_transform(K),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
// input tensor
const auto in_n_c0_hip_wip_e2_global_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, E2)),
make_tuple(make_pass_through_transform(N),
make_pass_through_transform(C0),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(E2)),
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_c0_y_ho_x_wo_e2_global_desc = transform_tensor_descriptor(
in_n_c0_hip_wip_e2_global_desc,
make_tuple(
make_pass_through_transform(N),
make_pass_through_transform(C0),
make_embed_transform(make_tuple(Y, Hop), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wop), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}, Sequence<6>{}));
const auto in_e_n_ho_wo_e2_grid_desc = transform_tensor_descriptor(
in_n_c0_y_ho_x_wo_e2_global_desc,
make_tuple(make_merge_transform(make_tuple(C0, Y, X)),
make_pass_through_transform(N),
make_pass_through_transform(Hop),
make_pass_through_transform(Wop),
make_pass_through_transform(E2)),
make_tuple(
Sequence<1, 2, 4>{}, Sequence<0>{}, Sequence<3>{}, Sequence<5>{}, Sequence<6>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto b_e0_e1_n_ho_wo_e2_grid_desc = transform_tensor_descriptor(
in_e_n_ho_wo_e2_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)),
make_pass_through_transform(N),
make_pass_through_transform(Hop),
make_pass_through_transform(Wop),
make_pass_through_transform(E2)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}, Sequence<5>{}));
// output tensor
const auto c_k_n_hop_wop_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1)),
make_tuple(make_merge_transform(make_tuple(K0, K1)),
make_pass_through_transform(N),
make_pad_transform(Ho, I0, OutRightPadH),
make_pad_transform(Wo, I0, OutRightPadW)),
make_tuple(Sequence<1, 4>{}, Sequence<0>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
// max tensor
const auto d_k_n_hx_wx_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Hx, Wx, K1)),
make_tuple(make_merge_transform(make_tuple(K0, K1)),
make_pass_through_transform(N),
make_pad_transform(Hx, I0, OutRightPadHx),
make_pad_transform(Wx, I0, OutRightPadWx)),
make_tuple(Sequence<1, 4>{}, Sequence<0>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
std::cerr << "Hop = " << Hop << " Wop = " << Wop << std::endl;
if(!((K % KPerBlock) == 0 && (Hop % HoPerBlock) == 0 && (Wop % WoPerBlock) == 0 &&
(E1 % E1PerBlock) == 0))
{
throw std::runtime_error("wrong! GEMM size no divisible");
}
// clang-format off
// hack to control index calculation when iterating over a_e0_e1_k_e2_global tensor
constexpr auto a_e0_e1_k_e2_global_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
constexpr auto a_e0_e1_k_e2_global_move_slice_window_step_hack =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
// hack to control index calculation when iterating over b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global tensor
constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks =
make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})
);
constexpr auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
constexpr auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks =
make_tuple(make_tuple(Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
constexpr auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks =
make_tuple(make_tuple(Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
// clang-format on
// GEMM
using GridwiseGemm = GridwiseGemmDlops_km_kn_mn_v3<
BlockSize,
FloatAB,
FloatAcc,
FloatC,
InMemoryDataOperationEnum::Set,
decltype(a_e0_e1_k_e2_grid_desc),
decltype(b_e0_e1_n_ho_wo_e2_grid_desc),
decltype(c_k_n_hop_wop_grid_desc),
decltype(d_k_n_hx_wx_grid_desc),
E1,
E2,
K2,
KPerBlock,
HoPerBlock,
WoPerBlock,
E1PerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
Sequence<2, 3, 0, 1, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>, // E0, E1, N, H0, H1, H2, W0, W1, W2, E2
9,
BThreadTransferSrcScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy, which will be fused
// with MoveSrcSliceWindow() to save addr computation
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8>, // K0, K1, N, H0, H1, I2, H2, W0, W1, I2, W2
1,
CThreadTransferDstScalarPerVector_K,
decltype(a_e0_e1_k_e2_global_step_hacks),
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_step_hacks),
decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks),
decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack)>;
const auto a_e0_e1_k0_k1_e2_grid_desc =
GridwiseGemm::MakeAE0E1K0K1E2GridDescriptor(a_e0_e1_k_e2_grid_desc);
const auto b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc =
GridwiseGemm::MakeBE0E1NH0H1H2W0W1W2E2GridDescriptor(b_e0_e1_n_ho_wo_e2_grid_desc);
const auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc =
GridwiseGemm::MakeCK0K1NH0H1H2W0W1W2GridDescriptor(c_k_n_hop_wop_grid_desc);
const auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc =
GridwiseGemm::MakeDK0K1NH0H1HxW0W1WxGridDescriptorMaxPool(d_k_n_hx_wx_grid_desc);
using AGridDesc_E0_E1_K0_K1_E2 = decltype(a_e0_e1_k0_k1_e2_grid_desc);
using BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 =
decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc);
using CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2 = decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc);
using DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx = decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc);
const auto grid_size = (K / KPerBlock) * (Hop / HoPerBlock) * (Wop / WoPerBlock) * N;
const bool has_main_e0_block_loop = E0 > 1;
std::cerr << "has_main_e0_block_loop = " << has_main_e0_block_loop << std::endl;
const auto cblockid_to_k_n_h_w_block_cluster_adaptor =
GridwiseGemm::MakeCBlockIdToKNHoWoBlockClusterAdaptor(c_k_n_hop_wop_grid_desc);
using CBlockIdToBlockClusterAdaptor_K_N_H_W =
decltype(cblockid_to_k_n_h_w_block_cluster_adaptor);
float ave_time = 0;
if(has_main_e0_block_loop)
{
const auto kernel = kernel_gemm_dlops_v3_maxpool<
GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K0_K1_E2>,
remove_reference_t<BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2>,
remove_reference_t<CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>,
remove_reference_t<DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_H_W>,
true,
activ_type>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_bias_grid,
p_c_grid,
p_d_grid,
a_e0_e1_k0_k1_e2_grid_desc,
b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
cblockid_to_k_n_h_w_block_cluster_adaptor);
}
else
{
const auto kernel = kernel_gemm_dlops_v3_maxpool<
GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K0_K1_E2>,
remove_reference_t<BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2>,
remove_reference_t<CGridDesc_K0_K1_N_H0_H1_H2_W0_W1_W2>,
remove_reference_t<DGridDesc_K0_K1_N_H0_H1_Hx_W0_W1_Wx>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_H_W>,
false,
activ_type>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_bias_grid,
p_c_grid,
p_d_grid,
a_e0_e1_k0_k1_e2_grid_desc,
b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_grid_desc,
c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc,
d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc,
cblockid_to_k_n_h_w_block_cluster_adaptor);
}
return ave_time;
}
};
#endif
library/include/ck/library/obselete_driver_offline/driver_gemm_dlops_v1r2.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_GEMM_DLOPS_V1R2
#define DRIVER_GEMM_DLOPS_V1R2
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_dlops_v1r2.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AKMGridDesc,
typename BKNGridDesc,
typename CMNGridDesc,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t M1PerThread,
ck::index_t N1PerThread,
ck::index_t KPerThread,
ck::index_t M1N1ThreadClusterM10,
ck::index_t M1N1ThreadClusterN10,
ck::index_t M1N1ThreadClusterM11,
ck::index_t M1N1ThreadClusterN11,
typename ABlockTransferThreadSliceLengths_K_M0_M1,
typename ABlockTransferThreadClusterLengths_K_M0_M1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_M1,
bool AThreadTransferSrcResetCoordinateAfterRun,
typename BBlockTransferThreadSliceLengths_K_N0_N1,
typename BBlockTransferThreadClusterLengths_K_N0_N1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_N1,
bool BThreadTransferSrcResetCoordinateAfterRun,
typename CThreadTransferSrcDstAccessOrder,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks>
__host__ float driver_gemm_dlops_v1r2(const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
const AKMGridDesc& a_k_m_grid_desc,
const BKNGridDesc& b_k_n_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
ck::index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
// GEMM
using GridwiseGemm = GridwiseGemmDlops_km_kn_mn_v1r2<BlockSize,
FloatAB,
FloatAcc,
FloatC,
CGlobalMemoryDataOperation,
AKMGridDesc,
BKNGridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
M1PerThread,
N1PerThread,
KPerThread,
M1N1ThreadClusterM10,
M1N1ThreadClusterN10,
M1N1ThreadClusterM11,
M1N1ThreadClusterN11,
ABlockTransferThreadSliceLengths_K_M0_M1,
ABlockTransferThreadClusterLengths_K_M0_M1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_M1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K_N0_N1,
BBlockTransferThreadClusterLengths_K_N0_N1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_N1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
const auto M = a_k_m_grid_desc.GetLength(I1);
const auto N = b_k_n_grid_desc.GetLength(I1);
const auto K = a_k_m_grid_desc.GetLength(I0);
if(!GridwiseGemm::CheckValidity(a_k_m_grid_desc, b_k_n_grid_desc, c_m_n_grid_desc))
{
throw std::runtime_error("wrong! GridwiseGemmDlops_km_kn_mn_v1r2 has invalid setting");
}
const auto a_k_m0_m1_grid_desc = GridwiseGemm::MakeAKM0M1GridDescriptor(a_k_m_grid_desc);
const auto b_k_n0_n1_grid_desc = GridwiseGemm::MakeBKN0N1GridDescriptor(b_k_n_grid_desc);
using AKM0M1GridDesc = decltype(a_k_m0_m1_grid_desc);
using BKN0N1GridDesc = decltype(b_k_n0_n1_grid_desc);
// c_m0_m10_m11_n0_n10_n11_grid_desc
const auto c_m0_m10_m11_n0_n10_n11_grid_desc =
GridwiseGemm::MakeCM0M10M11N0N10N11GridDescriptor(c_m_n_grid_desc);
using CM0M10M11N0N10N11GridDesc = decltype(c_m0_m10_m11_n0_n10_n11_grid_desc);
// cblockid_to_m0_n0_block_cluster_adaptor
const auto cblockid_to_m0_n0_block_cluster_adaptor =
GridwiseGemm::MakeCBlockIdToM0N0BlockClusterAdaptor(c_m_n_grid_desc);
using CBlockIdToM0N0BlockClusterAdaptor = decltype(cblockid_to_m0_n0_block_cluster_adaptor);
const index_t grid_size = GridwiseGemm::CalculateGridSize(M, N);
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K);
const bool has_double_tail_k_block_loop = GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K);
{
std::cout << "a_k_m0_m1_grid_desc{" << a_k_m0_m1_grid_desc.GetLength(I0) << ", "
<< a_k_m0_m1_grid_desc.GetLength(I1) << ", " << a_k_m0_m1_grid_desc.GetLength(I2)
<< "}" << std::endl;
std::cout << "b_k_n0_n1_grid_desc{" << b_k_n0_n1_grid_desc.GetLength(I0) << ", "
<< b_k_n0_n1_grid_desc.GetLength(I1) << ", " << b_k_n0_n1_grid_desc.GetLength(I2)
<< "}" << std::endl;
std::cout << "c_m0_m10_m11_n0_n10_n11_grid_desc{ "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I0) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I1) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I2) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I3) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I4) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I5) << "}" << std::endl;
}
float ave_time = 0;
if(has_main_k_block_loop && has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AKM0M1GridDesc>,
remove_reference_t<BKN0N1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
true,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k_m0_m1_grid_desc,
b_k_n0_n1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AKM0M1GridDesc>,
remove_reference_t<BKN0N1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
true,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k_m0_m1_grid_desc,
b_k_n0_n1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AKM0M1GridDesc>,
remove_reference_t<BKN0N1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
false,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k_m0_m1_grid_desc,
b_k_n0_n1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else
{
const auto kernel =
kernel_gemm_dlops_v1r2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AKM0M1GridDesc>,
remove_reference_t<BKN0N1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
false,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k_m0_m1_grid_desc,
b_k_n0_n1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
return ave_time;
}
#endif
library/include/ck/library/obselete_driver_offline/driver_gemm_dlops_v1r3.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_GEMM_DLOPS_V1R3
#define DRIVER_GEMM_DLOPS_V1R3
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_dlops_v1r3.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AK0MK1GridDesc,
typename BK0NK1GridDesc,
typename CMNGridDesc,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t M1PerThread,
ck::index_t N1PerThread,
ck::index_t KPerThread,
typename M1N1ThreadClusterM1Xs,
typename M1N1ThreadClusterN1Xs,
typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
typename ABlockTransferSrcVectorTensorContiguousDimOrder,
typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
typename BBlockTransferSrcVectorTensorContiguousDimOrder,
typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
typename CThreadTransferSrcDstAccessOrder,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks>
__host__ float driver_gemm_dlops_v1r3(const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
const AK0MK1GridDesc& a_k0_m_k1_grid_desc,
const BK0NK1GridDesc& b_k0_n_k1_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
ck::index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
// GEMM
using GridwiseGemm =
GridwiseGemmDlops_km_kn_mn_v1r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
CGlobalMemoryDataOperation,
AK0MK1GridDesc,
BK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
M1PerThread,
N1PerThread,
KPerThread,
M1N1ThreadClusterM1Xs,
M1N1ThreadClusterN1Xs,
ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
ABlockTransferSrcVectorTensorContiguousDimOrder,
ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
BBlockTransferSrcVectorTensorContiguousDimOrder,
BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks>;
const auto M = a_k0_m_k1_grid_desc.GetLength(I1);
const auto N = b_k0_n_k1_grid_desc.GetLength(I1);
const auto K0 = a_k0_m_k1_grid_desc.GetLength(I0);
if(!GridwiseGemm::CheckValidity(a_k0_m_k1_grid_desc, b_k0_n_k1_grid_desc, c_m_n_grid_desc))
{
throw std::runtime_error("wrong! GridwiseGemmDlops_km_kn_mn_v1r3 has invalid setting");
}
const auto a_k0_m0_m1_k1_grid_desc =
GridwiseGemm::MakeAK0M0M1K1GridDescriptor(a_k0_m_k1_grid_desc);
const auto b_k0_n0_n1_k1_grid_desc =
GridwiseGemm::MakeBK0N0N1K1GridDescriptor(b_k0_n_k1_grid_desc);
using AK0M0M1K1GridDesc = decltype(a_k0_m0_m1_k1_grid_desc);
using BK0N0N1K1GridDesc = decltype(b_k0_n0_n1_k1_grid_desc);
// c_m0_m10_m11_n0_n10_n11_grid_desc
const auto c_m0_m10_m11_n0_n10_n11_grid_desc =
GridwiseGemm::MakeCM0M10M11N0N10N11GridDescriptor(c_m_n_grid_desc);
using CM0M10M11N0N10N11GridDesc = decltype(c_m0_m10_m11_n0_n10_n11_grid_desc);
// cblockid_to_m0_n0_block_cluster_adaptor
const auto cblockid_to_m0_n0_block_cluster_adaptor =
GridwiseGemm::MakeCBlockIdToM0N0BlockClusterAdaptor(c_m_n_grid_desc);
using CBlockIdToM0N0BlockClusterAdaptor = decltype(cblockid_to_m0_n0_block_cluster_adaptor);
const index_t grid_size = GridwiseGemm::CalculateGridSize(M, N);
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
const bool has_double_tail_k_block_loop = GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
{
std::cout << "a_k0_m0_m1_k1_grid_desc{" << a_k0_m0_m1_k1_grid_desc.GetLength(I0) << ", "
<< a_k0_m0_m1_k1_grid_desc.GetLength(I1) << ", "
<< a_k0_m0_m1_k1_grid_desc.GetLength(I2) << ", "
<< a_k0_m0_m1_k1_grid_desc.GetLength(I3) << "}" << std::endl;
std::cout << "b_k0_n0_n1_k1_grid_desc{" << b_k0_n0_n1_k1_grid_desc.GetLength(I0) << ", "
<< b_k0_n0_n1_k1_grid_desc.GetLength(I1) << ", "
<< b_k0_n0_n1_k1_grid_desc.GetLength(I2) << ", "
<< b_k0_n0_n1_k1_grid_desc.GetLength(I3) << "}" << std::endl;
std::cout << "c_m0_m10_m11_n0_n10_n11_grid_desc{ "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I0) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I1) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I2) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I3) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I4) << ", "
<< c_m0_m10_m11_n0_n10_n11_grid_desc.GetLength(I5) << "}" << std::endl;
}
float ave_time = 0;
if(has_main_k_block_loop && has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AK0M0M1K1GridDesc>,
remove_reference_t<BK0N0N1K1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
true,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k0_m0_m1_k1_grid_desc,
b_k0_n0_n1_k1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AK0M0M1K1GridDesc>,
remove_reference_t<BK0N0N1K1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
true,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k0_m0_m1_k1_grid_desc,
b_k0_n0_n1_k1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
{
const auto kernel =
kernel_gemm_dlops_v1r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AK0M0M1K1GridDesc>,
remove_reference_t<BK0N0N1K1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
false,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k0_m0_m1_k1_grid_desc,
b_k0_n0_n1_k1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
else
{
const auto kernel =
kernel_gemm_dlops_v1r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AK0M0M1K1GridDesc>,
remove_reference_t<BK0N0N1K1GridDesc>,
remove_reference_t<CM0M10M11N0N10N11GridDesc>,
remove_reference_t<CBlockIdToM0N0BlockClusterAdaptor>,
false,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_k0_m0_m1_k1_grid_desc,
b_k0_n0_n1_k1_grid_desc,
c_m0_m10_m11_n0_n10_n11_grid_desc,
cblockid_to_m0_n0_block_cluster_adaptor);
}
return ave_time;
}
#endif
library/include/ck/library/obselete_driver_offline/driver_gemm_xdlops_v2r3.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_GEMM_XDLOPS_V2R3_HPP
#define DRIVER_GEMM_XDLOPS_V2R3_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
#include "element_wise_operation.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K,
typename CMNGridDesc,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t K1,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadSliceLengths_K0_M_K1,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
typename BBlockTransferThreadSliceLengths_K0_N_K1,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
typename CThreadTransferSrcDstAccessOrder,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks,
bool CAccessOrderMRepeatNRepeat,
bool ABlockLdsAddExtraM,
bool BBlockLdsAddExtraN>
__host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K& b_grid_desc_k0_n_k1,
const CMNGridDesc& c_grid_desc_m_n,
ck::index_t M01,
ck::index_t N01,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
ck::index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
using ElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
using GridwiseGemm =
GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
FloatAB,
FloatAcc,
FloatC,
CGlobalMemoryDataOperation,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K,
CMNGridDesc,
ElementwiseOperation,
ElementwiseOperation,
ElementwiseOperation,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector>;
{
std::cout << "a_grid_desc_k0_m_k1{" << a_grid_desc_k0_m_k1.GetLength(I0) << ", "
<< a_grid_desc_k0_m_k1.GetLength(I1) << ", " << a_grid_desc_k0_m_k1.GetLength(I2)
<< "}" << std::endl;
std::cout << "b_grid_desc_k0_n_k1{" << b_grid_desc_k0_n_k1.GetLength(I0) << ", "
<< b_grid_desc_k0_n_k1.GetLength(I1) << ", " << b_grid_desc_k0_n_k1.GetLength(I2)
<< "}" << std::endl;
std::cout << "c_grid_desc_m_n{ " << c_grid_desc_m_n.GetLength(I0) << ", "
<< c_grid_desc_m_n.GetLength(I1) << "}" << std::endl;
}
if(!GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1, b_grid_desc_k0_n_k1, c_grid_desc_m_n, M01, N01))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
}
const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 = decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
const auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n, M01, N01);
using Block2CTileMap = decltype(block_2_ctile_map);
const index_t grid_size = GridwiseGemm::CalculateGridSize(c_grid_desc_m_n);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0;
auto element_op_ = ElementwiseOperation{};
if(has_main_k0_block_loop)
{
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_K0_M_K1>,
remove_reference_t<BGridDesc_K0_N_K>,
remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
ElementwiseOperation,
ElementwiseOperation,
ElementwiseOperation,
remove_reference_t<Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
element_op_,
element_op_,
element_op_,
block_2_ctile_map);
}
else
{
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_K0_M_K1>,
remove_reference_t<BGridDesc_K0_N_K>,
remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
ElementwiseOperation,
ElementwiseOperation,
ElementwiseOperation,
remove_reference_t<Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
element_op_,
element_op_,
element_op_,
block_2_ctile_map);
}
return ave_time;
}
#endif
library/include/ck/library/obselete_driver_offline/driver_gemm_xdlops_v2r4.hpp deleted 100644 → 0
View file @ a49115b9
#ifndef DRIVER_GEMM_XDLOPS_V2R4
#define DRIVER_GEMM_XDLOPS_V2R4
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r4.hpp"
template <ck::index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
ck::InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename ABK0MK1GridDesc,
typename BBK0NK1GridDesc,
typename CMNGridDesc,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t K1,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadSliceLengths_K0_M_K1,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
typename BBlockTransferThreadSliceLengths_K0_N_K1,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
typename CThreadTransferSrcDstAccessOrder,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks,
bool CAccessOrderMRepeatNRepeat,
bool ABlockLdsAddExtraM,
bool BBlockLdsAddExtraN>
__host__ float driver_gemm_xdlops_v2r4(const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
ck::index_t M01,
ck::index_t N01,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
ck::index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
using GridwiseGemm =
GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4<BlockSize,
FloatAB,
FloatAcc,
FloatC,
CGlobalMemoryDataOperation,
ABK0MK1GridDesc,
BBK0NK1GridDesc,
CMNGridDesc,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
K1,
MRepeat,
NRepeat,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
BThreadTransferSrcResetCoordinateAfterRun,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
AGridStepHacks,
BGridStepHacks,
CGridStepHacks,
AGridMoveSliceWindowStepHacks,
BGridMoveSliceWindowStepHacks,
CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
{
std::cout << "a_b_k0_m_k1_grid_desc{" << a_b_k0_m_k1_grid_desc.GetLength(I0) << ", "
<< a_b_k0_m_k1_grid_desc.GetLength(I1) << ", "
<< a_b_k0_m_k1_grid_desc.GetLength(I2) << ", "
<< a_b_k0_m_k1_grid_desc.GetLength(I3) << "}" << std::endl;
std::cout << "b_b_k0_n_k1_grid_desc{" << b_b_k0_n_k1_grid_desc.GetLength(I0) << ", "
<< b_b_k0_n_k1_grid_desc.GetLength(I1) << ", "
<< b_b_k0_n_k1_grid_desc.GetLength(I2) << ", "
<< b_b_k0_n_k1_grid_desc.GetLength(I3) << "}" << std::endl;
std::cout << "c_m_n_grid_desc{ " << c_m_n_grid_desc.GetLength(I0) << ", "
<< c_m_n_grid_desc.GetLength(I1) << "}" << std::endl;
}
if(!GridwiseGemm::CheckValidity(
a_b_k0_m_k1_grid_desc, b_b_k0_n_k1_grid_desc, c_m_n_grid_desc, M01, N01))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r4 has invalid setting");
}
const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
using CM0N0M1N1M2M3M4N2GridDesc = decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
const auto c_block_cluster_adaptor =
GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc, M01, N01, KBatch);
using CBlockClusterAdaptor = decltype(c_block_cluster_adaptor);
const index_t grid_size = GridwiseGemm::CalculateGridSize(c_m_n_grid_desc, KBatch);
{
std::cout << "gridSize : " << grid_size << std::endl;
}
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0;
if(has_main_k0_block_loop)
{
const auto kernel = kernel_gemm_xdlops_v2r4<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<ABK0MK1GridDesc>,
remove_reference_t<BBK0NK1GridDesc>,
remove_reference_t<CM0N0M1N1M2M3M4N2GridDesc>,
remove_reference_t<CBlockClusterAdaptor>,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r4<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<ABK0MK1GridDesc>,
remove_reference_t<BBK0NK1GridDesc>,
remove_reference_t<CM0N0M1N1M2M3M4N2GridDesc>,
remove_reference_t<CBlockClusterAdaptor>,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
p_a_grid,
p_b_grid,
p_c_grid,
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor);
}
return ave_time;
}
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
View file @ d1db6a0c
#ifndef REFERENCE_BATCHED_GEMM_HPP
#define REFERENCE_BATCHED_GEMM_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -132,4 +132,3 @@ struct ReferenceBatchedGemm : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
View file @ d1db6a0c
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2022 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp
View file @ d1db6a0c
......@@ -2,8 +2,9 @@
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View file @ d1db6a0c
#ifndef REFERENCE_CONV_BWD_DATA_HPP
#define REFERENCE_CONV_BWD_DATA_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -351,4 +352,3 @@ struct ReferenceConvBwdData : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ d1db6a0c
......@@ -4,9 +4,8 @@
#include <type_traits>
#include <sstream>
#include "stream_config.hpp"
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
View file @ d1db6a0c
#ifndef REFERENCE_CONV_FWD_BIAS_ACTIVATION_HPP
#define REFERENCE_CONV_FWD_BIAS_ACTIVATION_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -187,4 +187,3 @@ struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
View file @ d1db6a0c
#ifndef REFERENCE_CONV2D_FWD_BIAS_ACTIVATION_ADD_HPP
#define REFERENCE_CONV2D_FWD_BIAS_ACTIVATION_ADD_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -195,4 +195,3 @@ struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ d1db6a0c
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
View file @ d1db6a0c
#ifndef REFERENCE_GEMM_BIAS_BIAS_2D_HPP
#define REFERENCE_GEMM_BIAS_BIAS_2D_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -131,4 +131,3 @@ struct ReferenceGemmBias2D : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
View file @ d1db6a0c
#ifndef REFERENCE_GEMM_BIAS_ACTIVATION_HPP
#define REFERENCE_GEMM_BIAS_ACTIVATION_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -134,4 +135,3 @@ struct ReferenceGemmBiasActivation : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
View file @ d1db6a0c
#ifndef REFERENCE_GEMM_BIAS_ACTIVATION_ADD_HPP
#define REFERENCE_GEMM_BIAS_ACTIVATION_ADD_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
......@@ -142,4 +143,3 @@ struct ReferenceGemmBiasActivationAdd : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
View file @ d1db6a0c
#pragma once
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include "device_base.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
View file @ d1db6a0c
#ifndef DEVICE_REDUCE_INSTANTCE_HPP
#define DEVICE_REDUCE_INSTANTCE_HPP
#pragma once
#include "device_reduce_instance_blockwise_f16_f16_f16.hpp"
#include "device_reduce_instance_blockwise_f16_f32_f16.hpp"
#include "device_reduce_instance_blockwise_f32_f32_f32.hpp"
#include "device_reduce_instance_blockwise_f32_f64_f32.hpp"
#include "device_reduce_instance_blockwise_f64_f64_f64.hpp"
#include "device_reduce_instance_blockwise_i8_i8_i8.hpp"
#include "device_reduce_instance_blockwise_i8_i32_i8.hpp"
#include "device_reduce_instance_blockwise_b16_f32_b16.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f16_f32_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f32_f32_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f32_f64_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f64_f64_f64.hpp"
#include "device_reduce_instance_multiblock_atomic_add_b16_f32_f32.hpp"
#include "device_reduce_instance_threadwise_f16_f16_f16.hpp"
#include "device_reduce_instance_threadwise_f16_f32_f16.hpp"
#include "device_reduce_instance_threadwise_f32_f32_f32.hpp"
#include "device_reduce_instance_threadwise_f32_f64_f32.hpp"
#include "device_reduce_instance_threadwise_f64_f64_f64.hpp"
#include "device_reduce_instance_threadwise_i8_i8_i8.hpp"
#include "device_reduce_instance_threadwise_i8_i32_i8.hpp"
#include "device_reduce_instance_threadwise_b16_f32_b16.hpp"
#endif
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.hpp"
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
View file @ d1db6a0c
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_HPP
#pragma once
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_impl_common.hpp"
#include "device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
namespace ck {
namespace tensor_operation {
......@@ -175,7 +174,4 @@ void add_device_reduce_instance_blockwise(
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#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