Skip to content

GitLab

Unverified Commit 9697ad4e authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into add_int8_wmma_example_instance

parents 1c97db8a 582e31e8
Expand all Show whitespace changes
Inline Side-by-side
Showing with 764 additions and 2068 deletions
include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ 1c97db8a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// GemmM = K
// GemmN = N * Ho * Wo
// GemmK = C * Y * X
template <typename... Wei,
typename... In,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value>
__host__ __device__ constexpr auto
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = K;
const auto GemmN = N * Ho * Wo;
const auto GemmK = C * Y * X;
const auto GemmK0 = GemmK / GemmK1;
// weight tensor
const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto wei_gemmk0_gemmm_gemmk1_grid_desc =
transform_tensor_descriptor(wei_gemmk_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// input tensor
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_gemmk_gemmn_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_gemmk0_gemmn_gemmk1_grid_desc =
transform_tensor_descriptor(in_gemmk_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// output tensor
const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ 1c97db8a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// A: in
// B: wei
// C: out
// GemmM = N * Ho * Wo
// GemmN = K
// GemmK = Y * X * C
template <typename... In,
typename... Wei,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value>
__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = N * Ho * Wo;
const auto GemmN = K;
const auto GemmK = Y * X * C;
const auto GemmK0 = GemmK / GemmK1;
// A: input tensor
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_gemmk_gemmm_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_gemmk0_gemmm_gemmk1_grid_desc =
transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// B: weight tensor
const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
transform_tensor_descriptor(wei_gemmk_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// C: output tensor
const auto out_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
wei_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ 1c97db8a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// GemmM0 = 1
// GemmM1 = K
// GemmN0 = N0
// GemmN1 = (N / N0) * Ho * Wo
// GemmK0 = (C / C0) * Y * X
// GemmK1 = C0
template <typename... Wei,
typename... In,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
typename N0Type,
typename C0Type>
__host__ __device__ constexpr auto
transform_forward_convolution_into_contraction_v6r1_nchw_kcyx_nkhw_pad(
const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const N0Type& N0,
const C0Type& C0)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
const auto X = wei_k_c_y_x_grid_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];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto N1 = N / N0;
const auto C1 = C / C0;
// weight tensor
const auto wei_gk0_gm0_gm1_gk1_grid_desc =
transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
make_tuple(make_unmerge_transform(make_tuple(I1, K)),
make_unmerge_transform(make_tuple(C0, C1 * Y * X))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1, 2>{}, Sequence<3, 0>{}));
// input tensor
const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
in_n_c_hi_wi_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pass_through_transform(C),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n0_n1_c0_c1_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
in_n_c_hip_wip_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
make_unmerge_transform(make_tuple(C0, C1)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}, Sequence<6, 7>{}));
const auto in_gk0_gn0_gn1_gk1_grid_desc = transform_tensor_descriptor(
in_n0_n1_c0_c1_y_ho_x_wo_grid_desc,
make_tuple(make_merge_transform(make_tuple(C1, Y, X)),
make_pass_through_transform(N0),
make_merge_transform(make_tuple(N1, Ho, Wo)),
make_pass_through_transform(C0)),
make_tuple(Sequence<3, 4, 6>{}, Sequence<0>{}, Sequence<1, 5, 7>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
// output tensor
const auto out_n_k_howo_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo));
const auto out_n0_n1_1_k_howo_grid_desc =
transform_tensor_descriptor(out_n_k_howo_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
make_unmerge_transform(make_tuple(I1, K)),
make_pass_through_transform(Ho * Wo)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}, Sequence<4>{}));
const auto out_gm0_gm1_gn0_gn1_grid_desc = transform_tensor_descriptor(
out_n0_n1_1_k_howo_grid_desc,
make_tuple(make_pass_through_transform(I1),
make_pass_through_transform(K),
make_pass_through_transform(N0),
make_merge_transform_v2_magic_division(make_tuple(N1, Ho * Wo))),
make_tuple(Sequence<2>{}, Sequence<3>{}, Sequence<0>{}, Sequence<1, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
return make_tuple(
wei_gk0_gm0_gm1_gk1_grid_desc, in_gk0_gn0_gn1_gk1_grid_desc, out_gm0_gm1_gn0_gn1_grid_desc);
}
} // namespace ck
#endif
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 9697ad4e
......@@ -135,7 +135,8 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__))
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__) || defined(__gfx1100__) || \
defined(__gfx1101__) || defined(__gfx1102__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -711,7 +712,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx940"))
ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
View file @ 9697ad4e
......@@ -106,7 +106,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -600,7 +601,9 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
namespace ctc = tensor_layout::convolution;
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
View file @ 9697ad4e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
......
include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp → include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
View file @ 9697ad4e
......@@ -3,8 +3,7 @@
#pragma once
#include <iostream>
#include <array>
#include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/utility/reduction_enums.hpp"
......@@ -13,28 +12,33 @@ namespace ck {
namespace tensor_operation {
namespace device {
template <ck::ReduceTensorOp ReduceOpId>
struct DevicePool2dFwd : public BaseOperator
template <index_t InOutRank,
index_t WindowRank,
typename InDataType,
typename OutDataType,
typename IndexDataType,
ReduceTensorOp ReduceOpId,
bool OutputIndex>
struct DevicePoolFwd : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* in_dev,
void* out_dev,
void* out_indices_dev,
ck::index_t N,
ck::index_t C,
std::array<ck::index_t, 2> input_spatial_lengths,
std::array<ck::index_t, 2> window_spatial_lengths,
std::array<ck::index_t, 2> output_spatial_lengths,
std::array<ck::index_t, 2> window_strides,
std::array<ck::index_t, 2> input_left_pads,
std::array<ck::index_t, 2> input_right_pads) = 0;
MakeArgumentPointer(const void* p_in_dev,
void* p_out_dev,
void* p_out_indices_dev,
std::vector<ck::index_t> input_lengths,
std::vector<ck::index_t> window_lengths,
std::vector<ck::index_t> output_lengths,
std::vector<ck::index_t> input_stride,
std::vector<ck::index_t> output_stride,
std::vector<ck::index_t> indices_stride,
std::vector<ck::index_t> window_strides,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
std::vector<ck::index_t> pooling_dims) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <ck::ReduceTensorOp ReduceOpId>
using DevicePool2dFwdPtr = std::unique_ptr<DevicePool2dFwd<ReduceOpId>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
View file @ 9697ad4e
......@@ -1393,7 +1393,9 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
static bool IsSupportedArgument(const Argument& arg)
{
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
View file @ 9697ad4e
......@@ -485,7 +485,9 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030")
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
View file @ 9697ad4e
......@@ -51,7 +51,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__))
defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
defined(__gfx1101__) || defined(__gfx1102__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
......@@ -553,7 +554,8 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
{
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx940")
ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102")
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ 9697ad4e
......@@ -73,6 +73,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
// TODO: should be exposed as Tparams.
static constexpr index_t NumGemmKPrefetchStage = 1;
static constexpr LoopScheduler LoopSched = make_default_loop_scheduler();
static constexpr PipelineVersion PipelineVer = PipelineVersion::v2;
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
......@@ -85,6 +90,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
NumGemmKPrefetchStage,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -112,7 +118,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVer>;
using Argument = typename GridwiseGemm::Argument;
using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
......@@ -257,7 +265,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
StrideC,
GridwiseGemm::CalculateMPadded(M),
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch};
}
......@@ -290,7 +298,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
StrideC,
GridwiseGemm::CalculateMPadded(M),
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch);
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp
View file @ 9697ad4e
......@@ -1027,7 +1027,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
static bool IsSupportedArgument(const Argument& arg)
{
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
View file @ 9697ad4e
......@@ -40,7 +40,8 @@ __global__ void
const CDEElementwiseOperation cde_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx1030__))
defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ 9697ad4e
......@@ -85,7 +85,7 @@ template <typename ALayout,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
ck::index_t NumPrefetch,
ck::index_t NumGemmKPrefetchStage,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
......@@ -152,6 +152,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
BElementwiseOperation,
CDEElementwiseOperation,
GemmSpec,
NumGemmKPrefetchStage,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -179,7 +180,9 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CDEBlockTransferScalarPerVector_NPerBlock,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVersion::v2>;
using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
using Block2ETileMapKSplit =
......@@ -265,8 +268,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
const index_t k0 = GridwiseGemm::CalculateK0(K, K_BATCH);
const auto c_grid_desc_m_n =
GridwiseGemm::MakeCGridDescriptor_M_N(M, N, m_padded, n_padded, stride_c);
const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, stride_c);
const auto local_b2c_tile_map =
Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
......@@ -319,8 +321,8 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
const index_t k0 = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(
karg.M, karg.N, karg.MPadded, karg.NPadded, karg.StrideC);
const auto c_grid_desc_m_n =
GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
const auto local_b2c_tile_map =
Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
......@@ -501,6 +503,11 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
if((ck::type_convert<ck::index_t>(arg.gemm_kernel_args_.size()) +
arg.skipped_group_count_) != arg.group_count_)
{
#if DEBUG_LOG
std::cout << "The group count is not equal to sum of skipped groups "
"and kernel args size!"
<< std::endl;
#endif // DEBUG_LOG
return false;
}
......@@ -509,14 +516,15 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
{
const auto& a = arg.gemm_kernel_args_[i].karg_;
bool group_arg_valid = GridwiseGemm::CheckValidity(a);
#if DEBUG_LOG
if(not group_arg_valid)
{
std::cout << "[" << __func__ << "] group id: " << i << " is not supported!\n";
#if DEBUG_LOG
std::cout << "[" << __func__ << "] group id: " << i
<< " has invalid GridwiseGemm settings!" << std::endl;
a.Print();
}
#endif // DEBUG_LOG
supported &= group_arg_valid;
}
supported = supported && group_arg_valid;
}
return supported;
}
......
include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
View file @ 9697ad4e
......@@ -9,7 +9,7 @@
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -20,16 +20,18 @@ namespace device {
template <typename InDataType,
typename OutDataType,
typename AccDataType,
typename IndexDataType, // enable if OutputIndex == true
typename ComputeDataType,
ck::ReduceTensorOp ReduceOpId,
bool OuputIndex,
bool OutputIndex,
ck::index_t BlockSize,
ck::index_t ReduceMThreadClusterSize,
ck::index_t ReduceKThreadClusterSize,
ck::index_t ReduceMThreadSliceSize,
ck::index_t ReduceKThreadSliceSize,
ck::index_t InSrcOutDstVectorSize>
struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd<ReduceOpId>
struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C
: public DevicePoolFwd<4, 2, InDataType, OutDataType, IndexDataType, ReduceOpId, OutputIndex>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -38,7 +40,8 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
using IndexDataType = int32_t;
static constexpr index_t InOutRank = 4;
static constexpr index_t WindowRank = 2;
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
......@@ -59,12 +62,12 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
static auto MakeABGridDescriptor_A_M_K_B_M(ck::index_t N,
ck::index_t C,
std::array<ck::index_t, 2> input_spatial_lengths,
std::array<ck::index_t, 2> window_spatial_lengths,
std::array<ck::index_t, 2> output_spatial_lengths,
std::array<ck::index_t, 2> window_strides,
std::array<ck::index_t, 2> input_left_pads,
std::array<ck::index_t, 2> input_right_pads)
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> window_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> window_strides,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
......@@ -141,9 +144,7 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
return make_tuple(in_grid_desc_reducem_reducek, out_grid_desc_reducem);
}
using ABGridDescs = decltype(
MakeABGridDescriptor_A_M_K_B_M(1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
using ABGridDescs = decltype(MakeABGridDescriptor_A_M_K_B_M(1, 1, {}, {}, {}, {}, {}, {}));
using AGridDesc_M_K = remove_cvref_t<decltype(ABGridDescs{}[I0])>;
using BGridDesc_M = remove_cvref_t<decltype(ABGridDescs{}[I1])>;
......@@ -152,15 +153,15 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
{
Argument(const InDataType* p_in_dev,
OutDataType* p_out_dev,
int* p_out_indices_dev,
IndexDataType* p_out_indices_dev,
ck::index_t N,
ck::index_t C,
std::array<ck::index_t, 2>& input_spatial_lengths,
std::array<ck::index_t, 2>& window_spatial_lengths,
std::array<ck::index_t, 2>& output_spatial_lengths,
std::array<ck::index_t, 2>& window_strides,
std::array<ck::index_t, 2>& input_left_pads,
std::array<ck::index_t, 2>& input_right_pads)
std::vector<ck::index_t>& input_spatial_lengths,
std::vector<ck::index_t>& window_spatial_lengths,
std::vector<ck::index_t>& output_spatial_lengths,
std::vector<ck::index_t>& window_strides,
std::vector<ck::index_t>& input_left_pads,
std::vector<ck::index_t>& input_right_pads)
: p_in_dev_{p_in_dev},
p_out_dev_{p_out_dev},
p_out_indices_dev_{p_out_indices_dev},
......@@ -190,7 +191,7 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
const InDataType* p_in_dev_;
OutDataType* p_out_dev_;
int* p_out_indices_dev_;
IndexDataType* p_out_indices_dev_;
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_M b_grid_desc_m_;
InElementwiseOperation in_element_op_;
......@@ -208,7 +209,7 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
using gridwise_reduce =
GridwiseReduction_mk_to_m_threadwise<InDataType,
OutDataType,
AccDataType,
ComputeDataType,
IndexDataType,
AGridDesc_M_K,
BGridDesc_M,
......@@ -224,12 +225,14 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
InSrcOutDstVectorSize,
InSrcOutDstVectorSize>;
const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
OuputIndex,
const auto kernel =
kernel_reduce_threadwise<gridwise_reduce,
OutputIndex,
true, // pooling need to return global index
false, // don't have index input
InDataType,
OutDataType,
AccDataType,
ComputeDataType,
IndexDataType,
AGridDesc_M_K,
BGridDesc_M,
......@@ -280,22 +283,42 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
MakeArgumentPointer(const void* p_in_dev,
void* p_out_dev,
void* p_out_indices_dev,
ck::index_t N,
ck::index_t C,
std::array<ck::index_t, 2> input_spatial_lengths,
std::array<ck::index_t, 2> window_spatial_lengths,
std::array<ck::index_t, 2> output_spatial_lengths,
std::array<ck::index_t, 2> window_strides,
std::array<ck::index_t, 2> input_left_pads,
std::array<ck::index_t, 2> input_right_pads) override
std::vector<ck::index_t> input_lengths,
std::vector<ck::index_t> window_lengths,
std::vector<ck::index_t> output_lengths,
std::vector<ck::index_t>, // Suppose tensor layout = NHWC
std::vector<ck::index_t>, // Suppose tensor layout = NHWC
std::vector<ck::index_t>, // Suppose tensor layout = NHWC
std::vector<ck::index_t> window_strides,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
std::vector<ck::index_t> pooling_dims) override
{
if(input_lengths.size() != InOutRank || window_lengths.size() != WindowRank ||
input_lengths.size() != InOutRank || window_strides.size() != WindowRank ||
input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
throw std::runtime_error("dimension is incorrect");
if(pooling_dims != std::vector<ck::index_t>{2, 3})
throw std::runtime_error("pooling_dims only support {2, 3} in pool2d so far");
index_t N = input_lengths[0];
index_t C = input_lengths[1];
index_t Hi = input_lengths[2];
index_t Wi = input_lengths[3];
index_t Ho = output_lengths[2];
index_t Wo = output_lengths[3];
std::vector<ck::index_t> input_spatial_lengths = {Hi, Wi};
std::vector<ck::index_t> output_spatial_lengths = {Ho, Wo};
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_dev),
static_cast<OutDataType*>(p_out_dev),
static_cast<int*>(p_out_indices_dev),
static_cast<IndexDataType*>(p_out_indices_dev),
N,
C,
input_spatial_lengths,
window_spatial_lengths,
window_lengths,
output_spatial_lengths,
window_strides,
input_left_pads,
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
View file @ 9697ad4e
......@@ -28,6 +28,7 @@ template <typename InDataType,
typename AccElementwiseOperation,
bool PropagateNan,
bool OutputIndex,
bool TransformIndexKtoGlobal,
bool HaveIndexInputIfOutputIndex,
index_t BlockSize,
index_t MThreadSliceSize,
......@@ -260,6 +261,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InDataType,
const auto kernel = kernel_reduce_threadwise<GridwiseReduce,
OutputIndex,
TransformIndexKtoGlobal,
HaveIndexInput,
InDataType,
OutDataType,
......
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