Skip to content

GitLab

Commit e2594ff7 authored by rocking's avatar rocking
Browse files

devicePool2DFwd Inherit devicePool3DFwd

parent 156423d6
Show whitespace changes
Inline Side-by-side
Showing with 156 additions and 385 deletions
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ e2594ff7
......@@ -9,7 +9,7 @@
#include "ck/utility/reduction_enums.hpp"
#include "ck/utility/reduction_functions_accumulate.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -45,8 +45,7 @@ bool pool_test(bool do_verification,
ck::index_t in_right_pad_w)
{
using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
InDataType, // InDataType
ck::tensor_operation::device::DevicePool2dFwdImpl<InDataType, // InDataType
OutDataType, // OutDataType
IndexDataType, // IndexDataType
ComputeDataType, // ComputeDataType
......@@ -57,7 +56,8 @@ bool pool_test(bool do_verification,
1, // ReduceKThreadClusterSize
4, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
4>; // InSrcOutDstVectorSize
1, // InSrcOutDstVectorSize
false>; // IsFastestDimReduced
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
......
example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
View file @ e2594ff7
......@@ -7,7 +7,7 @@
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_index_pool_bwd_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......@@ -44,19 +44,19 @@ bool maxpool_bwd_test(bool do_verification,
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
InDataType, // InDataType
ck::tensor_operation::device::DevicePool2dFwdImpl<InDataType, // InDataType
OutDataType, // OutDataType
IndexDataType, // IndexDataType
ComputeDataType, // ComputeDataType
ck::ReduceTensorOp::MAX,
true, // OutputIndex
true,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
4, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
1>; // InSrcOutDstVectorSize
1, // InSrcOutDstVectorSize
false>; // IsFastestDimReduced
using DeviceMaxPoolBwdInstance = ck::tensor_operation::device::
DeviceIndexPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>;
......
include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_impl.hpp 0 → 100644
View file @ e2594ff7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#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/impl/device_pool3d_fwd_impl.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"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InDataType,
typename OutDataType,
typename IndexDataType, // enable if OutputIndex == true
typename ComputeDataType,
ck::ReduceTensorOp ReduceOpId,
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,
bool IsFastestDimReduced>
struct DevicePool2dFwdImpl : public DevicePool3dFwdImpl<InDataType,
OutDataType,
IndexDataType,
ComputeDataType,
ReduceOpId,
OutputIndex,
BlockSize,
ReduceMThreadClusterSize,
ReduceKThreadClusterSize,
ReduceMThreadSliceSize,
ReduceKThreadSliceSize,
InSrcOutDstVectorSize,
IsFastestDimReduced>
{
using DevicePool3D = DevicePool3dFwdImpl<InDataType,
OutDataType,
IndexDataType,
ComputeDataType,
ReduceOpId,
OutputIndex,
BlockSize,
ReduceMThreadClusterSize,
ReduceKThreadClusterSize,
ReduceMThreadSliceSize,
ReduceKThreadSliceSize,
InSrcOutDstVectorSize,
IsFastestDimReduced>;
std::unique_ptr<BaseArgument>
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) override
{
static constexpr index_t InOutRank = 4;
static constexpr index_t WindowRank = 2;
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");
// NCHW to NCDHW
input_lengths.insert(input_lengths.begin() + 2, 1);
output_lengths.insert(output_lengths.begin() + 2, 1);
input_stride.insert(input_stride.begin() + 2, 0);
output_stride.insert(output_stride.begin() + 2, 0);
indices_stride.insert(indices_stride.begin() + 2, 0);
// YX to ZYX
window_lengths.insert(window_lengths.begin(), 1);
window_strides.insert(window_strides.begin(), 0);
input_left_pads.insert(input_left_pads.begin(), 0);
input_right_pads.insert(input_right_pads.begin(), 0);
pooling_dims = {2, 3, 4};
return DevicePool3D::MakeArgumentPointer(p_in_dev,
p_out_dev,
p_out_indices_dev,
input_lengths,
window_lengths,
output_lengths,
input_stride,
output_stride,
indices_stride,
window_strides,
input_left_pads,
input_right_pads,
pooling_dims);
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp deleted 100644 → 0
View file @ 156423d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#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_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"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InDataType,
typename OutDataType,
typename IndexDataType, // enable if OutputIndex == true
typename ComputeDataType,
ck::ReduceTensorOp ReduceOpId,
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 DevicePoolFwd<4, 2, InDataType, OutDataType, IndexDataType, ReduceOpId, OutputIndex>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr index_t InOutRank = 4;
static constexpr index_t WindowRank = 2;
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
static constexpr index_t InSrcOutDstVectorDim =
0; // for NHWC, the dim C is the vector Dim for both input and output in memory, which is
// not reduced.
static constexpr ck::index_t ReduceM_BlockTileSize =
ReduceMThreadClusterSize * ReduceMThreadSliceSize;
static constexpr ck::index_t ReduceK_BlockTileSize =
ReduceKThreadClusterSize * ReduceKThreadSliceSize;
static auto MakeABGridDescriptor_A_M_K_B_M(ck::index_t N,
ck::index_t C,
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];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t Y = window_spatial_lengths[0];
const index_t X = window_spatial_lengths[1];
const index_t ConvStrideH = window_strides[0];
const index_t ConvStrideW = window_strides[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const index_t ReduceMRaw = N * Ho * Wo * C;
const index_t ReduceMPad =
math::integer_least_multiple(ReduceMRaw, ReduceM_BlockTileSize) - ReduceMRaw;
const index_t ReduceKRaw = Y * X;
const index_t ReduceKPad =
math::integer_least_multiple(ReduceKRaw, ReduceK_BlockTileSize) - ReduceKRaw;
// A[ReduceM, ReduceK]
const auto in_grid_desc_n_hi_wi_c =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_grid_desc_n_hip_wip_c = transform_tensor_descriptor(
in_grid_desc_n_hi_wi_c,
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_grid_desc_n_y_ho_x_wo_c = transform_tensor_descriptor(
in_grid_desc_n_hip_wip_c,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(I1, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(I1, 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_grid_desc_reducemraw_reducekraw =
transform_tensor_descriptor(in_grid_desc_n_y_ho_x_wo_c,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo, C)),
make_merge_transform(make_tuple(Y, X))),
make_tuple(Sequence<0, 2, 4, 5>{}, Sequence<1, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_grid_desc_reducem_reducek = transform_tensor_descriptor(
in_grid_desc_reducemraw_reducekraw,
make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad),
make_right_pad_transform(ReduceKRaw, ReduceKPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
// B[ReduceM]
const auto out_grid_desc_reducemraw =
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo * C));
const auto out_grid_desc_reducem = transform_tensor_descriptor(
out_grid_desc_reducemraw,
make_tuple(make_right_pad_transform(ReduceMRaw, ReduceMPad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
return make_tuple(in_grid_desc_reducem_reducek, out_grid_desc_reducem);
}
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])>;
// TODO
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_dev,
OutDataType* p_out_dev,
IndexDataType* p_out_indices_dev,
ck::index_t N,
ck::index_t C,
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},
a_grid_desc_m_k_{},
b_grid_desc_m_{}
{
const auto descs = MakeABGridDescriptor_A_M_K_B_M(N,
C,
input_spatial_lengths,
window_spatial_lengths,
output_spatial_lengths,
window_strides,
input_left_pads,
input_right_pads);
a_grid_desc_m_k_ = descs[I0];
b_grid_desc_m_ = descs[I1];
invariant_lowest_length_ = C;
reduce_lowest_length_ = window_spatial_lengths[1];
int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
std::tie(in_element_op_, acc_element_op_) =
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
}
const InDataType* p_in_dev_;
OutDataType* p_out_dev_;
IndexDataType* p_out_indices_dev_;
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_M b_grid_desc_m_;
InElementwiseOperation in_element_op_;
AccElementwiseOperation acc_element_op_;
// for checking vector load/store
ck::index_t invariant_lowest_length_;
ck::index_t reduce_lowest_length_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
using gridwise_reduce =
GridwiseReduction_mk_to_m_threadwise<InDataType,
OutDataType,
ComputeDataType,
IndexDataType,
AGridDesc_M_K,
BGridDesc_M,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
false, // propagate_nan
BlockSize,
ReduceMThreadSliceSize,
ReduceKThreadSliceSize,
InSrcOutDstVectorDim,
InSrcOutDstVectorSize,
InSrcOutDstVectorSize>;
const auto kernel =
kernel_reduce_threadwise<gridwise_reduce,
OutputIndex,
true, // pooling need to return global index
false, // don't have index input
InDataType,
OutDataType,
ComputeDataType,
IndexDataType,
AGridDesc_M_K,
BGridDesc_M,
InElementwiseOperation,
AccElementwiseOperation>;
ck::index_t ReduceM = arg.a_grid_desc_m_k_.GetLength(I0);
const index_t grid_size = (ReduceM / ReduceM_BlockTileSize);
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.a_grid_desc_m_k_,
arg.b_grid_desc_m_,
arg.in_element_op_,
arg.acc_element_op_,
float(1),
arg.p_in_dev_,
nullptr,
float(0),
arg.p_out_dev_,
arg.p_out_indices_dev_);
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
if(pArg->invariant_lowest_length_ % InSrcOutDstVectorSize != 0)
{
return (false);
}
return (true);
}
std::unique_ptr<BaseArgument>
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>, // 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<IndexDataType*>(p_out_indices_dev),
N,
C,
input_spatial_lengths,
window_lengths,
output_spatial_lengths,
window_strides,
input_left_pads,
input_right_pads);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<" << BlockSize << ",";
str << "M_C" << ReduceMThreadClusterSize << "_S" << ReduceMThreadSliceSize << ",";
str << "K_C" << ReduceKThreadClusterSize << "_S" << ReduceKThreadSliceSize << ",";
str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_impl.hpp
View file @ e2594ff7
......@@ -312,7 +312,7 @@ struct DevicePool3dFwdImpl
return true;
}
std::unique_ptr<BaseArgument>
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_dev,
void* p_out_dev,
void* p_out_indices_dev,
......
library/src/tensor_operation_instance/gpu/pool_fwd/pool_fwd_instance_common.hpp
View file @ e2594ff7
......@@ -28,9 +28,9 @@ template <typename InDataType,
using device_pool2d_fwd_nhwc_instances =
// clang-format off
std::tuple <
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
DevicePool2dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1, false>,
DevicePool2dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2, false>,
DevicePool2dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4, false>
// clang-format on
>;
......@@ -43,9 +43,9 @@ template <typename InDataType,
using device_pool3d_fwd_ndhwc_instances =
// clang-format off
std::tuple <
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1, false>,
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2, false>,
DevicePool3dFwdImpl<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4, false>
// clang-format on
>;
......
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