Skip to content

GitLab

Commit 8f722700 authored by Chao Liu's avatar Chao Liu
Browse files

update ckprofiler for conv bwd data

parent a711f90b
Hide whitespace changes
Inline Side-by-side
Showing with 86 additions and 379 deletions
library/src/tensor_operation_instance/gpu/convnd_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp deleted 100644 → 0
View file @ a711f90b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_data_xdl_ndhwc_kzyxc_ndhwk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances =
std::tuple<
// clang-format off
//#############################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Num| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#############################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Dim| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#############################################################################| | | | | Operation| Operation| Operation| Specialization|Spatial| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#############################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>
// clang-format on
>;
using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances =
std::tuple<
// clang-format off
//#############################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Num| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#############################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Dim| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#############################################################################| | | | | Operation| Operation| Operation| Specialization|Spatial| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#############################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>
// clang-format on
>;
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances{});
add_device_operation_instances(
instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/convnd_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp deleted 100644 → 0
View file @ a711f90b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_data_xdl_ndhwc_kzyxc_ndhwk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DataType = int8_t;
using AccType = int32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances =
std::tuple<
// clang-format off
//#############################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Num| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#############################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Dim| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#############################################################################| | | | | Operation| Operation| Operation| Specialization|Spatial| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#############################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 256, 4, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 128, 4, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 64, 128, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 64, 128, 4, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 32, 128, 4, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 32, 64, 4, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 64, 64, 4, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 256, 128, 64, 4, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 128, 128, 32, 4, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 3, 64, 64, 32, 4, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>
// clang-format on
>;
using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances =
std::tuple<
// clang-format off
//##############################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Num| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##############################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Dim| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##############################################################################| | | | | Operation| Operation| Operation| Specialization|Spatial| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##############################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 256, 4, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 128, 4, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 64, 128, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 64, 64, 4, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 128, 64, 4, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 256, 64, 128, 4, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 128, 32, 4, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 128, 32, 128, 4, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 64, 32, 4, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 3, 64, 32, 64, 4, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>
// clang-format on
>;
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances{});
add_device_operation_instances(
instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 8f722700
......@@ -17,7 +17,7 @@ set(PROFILER_SOURCE
src/profile_conv_fwd.cpp
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
# src/profile_conv_bwd_data.cpp
src/profile_conv_bwd_data.cpp
src/profile_conv_bwd_weight.cpp
src/profile_reduce.cpp
src/profile_normalization.cpp
......@@ -38,11 +38,13 @@ target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv1d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv3d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
#target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_conv1d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_conv3d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_normalization_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
profiler/include/profile_conv_bwd_data_impl.hpp
View file @ 8f722700
......@@ -8,153 +8,17 @@
#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
using DeviceConvBwdDataNoOpPtr = ck::tensor_operation::device::instance::DeviceConvBwdDataNoOpPtr;
template <typename InDataType, typename WeiDataType, typename OutDataType>
void get_device_conv_bwd_data_op_ptr(
InDataType, WeiDataType, OutDataType, std::vector<DeviceConvBwdDataNoOpPtr>&, int)
{
std::cout << "can not find device conv bwd data" << std::endl;
exit(1);
}
template <>
void get_device_conv_bwd_data_op_ptr(
F32, F32, F32, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
F16, F16, F16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
BF16, BF16, BF16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
INT8, INT8, INT8, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
break;
default: break;
}
}
template <typename DataType>
void show_data_nhwc_layout(Tensor<DataType>& nhwc)
......@@ -182,7 +46,7 @@ void show_data_nhwc_layout(Tensor<DataType>& nhwc)
std::cout << "]";
}
template <int NDimSpatial,
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename OutLayout,
......@@ -329,15 +193,22 @@ bool profile_conv_bwd_data_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
get_device_conv_bwd_data_op_ptr(
InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_avg_time = 0;
......@@ -347,34 +218,35 @@ bool profile_conv_bwd_data_impl(int do_verification,
// profile device Conv instances
bool pass = true;
for(auto& conv_ptr : conv_ptrs)
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.output_spatial_lengths_,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.output_spatial_lengths_,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// reset input to zero
// for conv bwd data, some input tensor element are zero, but not written by kernel,
// need to set zero
in_device_buf.SetZero();
std::string op_name = conv_ptr->GetTypeString();
std::string op_name = op_ptr->GetTypeString();
auto invoker_ptr = op_ptr->MakeInvokerPointer();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -423,10 +295,15 @@ bool profile_conv_bwd_data_impl(int do_verification,
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
std::cout << "Best configuration parameters:"
<< "\nname: " << best_op_name << "\navg_time: " << best_avg_time
<< "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
return pass;
}
......
profiler/include/profile_conv_bwd_weight_impl.hpp
View file @ 8f722700
......@@ -65,6 +65,14 @@ bool profile_conv_bwd_weight_impl(int do_verification,
const ck::tensor_operation::device::ConvParams& params,
ck::index_t split_k)
{
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
// make host tensor descritpor
auto f_nhwc_host_tensor_descriptor =
[](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
......@@ -156,14 +164,6 @@ bool profile_conv_bwd_weight_impl(int do_verification,
output.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weight_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
......@@ -227,9 +227,6 @@ bool profile_conv_bwd_weight_impl(int do_verification,
for(auto& op_ptr : op_ptrs)
{
// using atomic, so need to reset input, setzero is done in invoker
wei_device_buf.SetZero();
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
......@@ -251,6 +248,9 @@ bool profile_conv_bwd_weight_impl(int do_verification,
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// using atomic add, so need to reset input
wei_device_buf.SetZero();
std::string op_name = op_ptr->GetTypeString();
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
profiler/src/profile_conv_bwd_data.cpp
View file @ 8f722700
......@@ -26,10 +26,12 @@ enum struct ConvDataType
static void print_helper_msg()
{
// clang-format-off
std::cout
<< "arg1: tensor operation (conv_bwd_data: Convolution Backward Data)\n"
<< "arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight bf16, Output bf16\n"
<< " 3: Input int8, Weight int8, Output int8)\n"
<< "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
<< " 1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
"K])\n"
......@@ -47,7 +49,6 @@ static void print_helper_msg()
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
// clang-format-on
}
ck::tensor_operation::device::ConvParams
......@@ -134,6 +135,11 @@ int profile_conv_bwd_data(int argc, char* argv[])
const auto params = parse_conv_params(num_dim_spatial, 9, argv);
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
using NWC = ck::tensor_layout::convolution::NWC;
using NHWC = ck::tensor_layout::convolution::NHWC;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
......
profiler/src/profile_conv_bwd_weight.cpp
View file @ 8f722700
......@@ -18,9 +18,9 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
};
static void print_helper_msg()
......@@ -29,7 +29,7 @@ static void print_helper_msg()
<< "arg1: tensor operation (conv_bwd_weight: Convolution Backward Weight\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight fp32, Output bf16\n"
<< " 2: Input bf16, Weight fp32, Output bf16)\n"
<< "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
<< " 1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, K]\n"
<< "arg4: verification (0: no, 1: yes)\n"
......
profiler/src/profile_conv_fwd.cpp
View file @ 8f722700
......@@ -26,10 +26,12 @@ enum struct ConvDataType
static void print_helper_msg()
{
// clang-format-off
std::cout
<< "arg1: tensor operation (conv_fwd: Convolution Forward)\n"
<< "arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight bf16, Output bf16\n"
<< " 3: Input int8, Weight int8, Output int8)\n"
<< "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
<< " 1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
"K])\n"
......@@ -47,7 +49,6 @@ static void print_helper_msg()
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
// clang-format-on
}
ck::tensor_operation::device::ConvParams
......
profiler/src/profiler.cpp
View file @ 8f722700
......@@ -15,7 +15,7 @@ int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
// int profile_conv_bwd_data(int, char*[]);
int profile_conv_bwd_data(int, char*[]);
int profile_conv_bwd_weight(int, char*[]);
int profile_normalization(int, char*[]);
int profile_reduce(int, char*[]);
......@@ -98,12 +98,10 @@ int main(int argc, char* argv[])
{
return profile_conv_fwd_bias_relu_add(argc, argv);
}
#if 0
else if(strcmp(argv[1], "conv_bwd_data") == 0)
{
return profile_conv_bwd_data(argc, argv);
}
#endif
else if(strcmp(argv[1], "conv_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
......
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