Skip to content

GitLab

Commit d3937687 authored by Artur Wojcik's avatar Artur Wojcik
Browse files

Merge branch 'develop' into uif2-initial

parents 87e80971 61cce232
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2226 additions and 2020 deletions
example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
View file @ d3937687
...@@ -3,7 +3,7 @@ ...@@ -3,7 +3,7 @@
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using DeviceConvFwdInstance = using DeviceConvFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InputLayout<NDimSpatial>, InputLayout<NDimSpatial>,
WeightLayout<NDimSpatial>, WeightLayout<NDimSpatial>,
......
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
View file @ d3937687
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = int8_t; using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
...@@ -33,7 +33,7 @@ template <ck::index_t NDimSpatial, ...@@ -33,7 +33,7 @@ template <ck::index_t NDimSpatial,
typename RequantScaleLayout, typename RequantScaleLayout,
typename OutLayout> typename OutLayout>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
View file @ d3937687
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = int8_t; using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
...@@ -31,7 +31,7 @@ template <ck::index_t NDimSpatial, ...@@ -31,7 +31,7 @@ template <ck::index_t NDimSpatial,
typename BiasLayout, typename BiasLayout,
typename OutLayout> typename OutLayout>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
View file @ d3937687
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = int8_t; using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
...@@ -31,7 +31,7 @@ template <ck::index_t NDimSpatial, ...@@ -31,7 +31,7 @@ template <ck::index_t NDimSpatial,
typename RequantScaleLayout, typename RequantScaleLayout,
typename OutLayout> typename OutLayout>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
View file @ d3937687
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
using InDataType = int8_t; using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
...@@ -26,7 +26,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio ...@@ -26,7 +26,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout> template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp
View file @ d3937687
...@@ -11,7 +11,7 @@ ...@@ -11,7 +11,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/library/utility/algorithm.hpp" #include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -47,7 +47,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio ...@@ -47,7 +47,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio
template <typename OutElementOp> template <typename OutElementOp>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/62_conv_fwd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp
View file @ d3937687
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/library/utility/algorithm.hpp" #include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -47,7 +47,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio ...@@ -47,7 +47,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio
template <typename OutElementOp> template <typename OutElementOp>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
example/62_conv_fwd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp
View file @ d3937687
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/library/utility/algorithm.hpp" #include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -44,7 +44,7 @@ template <typename DataType, ...@@ -44,7 +44,7 @@ template <typename DataType,
typename InElementOp, typename InElementOp,
typename WeiElementOp> typename WeiElementOp>
using DeviceGroupedConvNDMultiABFwdInstance = using DeviceGroupedConvNDMultiABFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp 0 → 100644
View file @ d3937687
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
/**
* \brief Grouped Convolution Forward
*
* \details
* input : input image A[G, N, C, Hi, Wi], A1[G, N, C, Hi, Wi]...
* input : weight B[G, K, C, Y, X], B1[G, K, C, Y, X]...
* input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
* output : output image E[G, N, K, Ho, Wo]
*
* C = a_op(A, A1...) * b_op(B, B1...)
* E = cde_op(C, D0, D1, ...)
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ALayout Input layout (also for a1, a2...).
* \tparam BLayout Weight layout (also for b1, b2...).
* \tparam DsLayout Ds layouts.
* \tparam ELayout Output layout.
* \tparam ADataType Input data type. Pass tuple if there is multiple A.
* \tparam BDataType Weight data type. Pass tuple if there is multiple B.
* \tparam DsDataType D data types.
* \tparam EDataType Output data type.
* \tparam AElementwiseOperation A elementwise operation.
* \tparam BElementwiseOperation B elementwise operation.
* \tparam CDEElementwiseOperation CDE elementwise operation.
* \tparam ComputeType Compute data type (default: ADataType, first if tuple passed).
*/
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputeType =
decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
Number<0>,
ADataType>())> // ComputeType is InputType by default (first
// in tuple for MultiAB), unpack if tuple was
// passed
struct DeviceGroupedConvFwdMultipleABD : public BaseOperator
{
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
static constexpr index_t NumATensor = GetNumABTensors<isMultiA, ADataType>();
static constexpr index_t NumBTensor = GetNumABTensors<isMultiB, BDataType>();
static constexpr index_t NumDTensor = DsDataType::Size();
static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
// If DataType is tuple, user has to pass std::array with pointers.
using APointers =
std::conditional_t<isMultiA, std::array<const void*, NumATensor>&, const void*>;
using BPointers =
std::conditional_t<isMultiB, std::array<const void*, NumBTensor>&, const void*>;
/**
* \brief Make argument pointer for grouped conv fwd.
*
* \param p_a A pointer to the input (std::array<const void*, NumA> with
pointers for multiple A).
* \param p_b A pointer to the weight (std::array<const void*, NumA> with
pointers for multiple B).
* \param p_ds A pointers to the Ds.
* \param p_e A pointers to the output.
* \param a_g_n_c_wis_lengths Input lengths [G, N, C, Spatial...] (for 3d).
* \param a_g_n_c_wis_strides Input strides [G, N, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_lengths Weight lengths [G, K, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_strides Weight strides [G, K, C, Spatial...] (for 3d).
* \param ds_g_n_k_wos_lengths Ds lengths [G, N, K, Spatial...] (for 3d).
* \param ds_g_n_k_wos_strides Ds strides [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_lengths Output lengths [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_strides Output strides [G, N, K, Spatial...] (for 3d).
* \param conv_filter_strides Convolution filter strides.
* \param conv_filter_dilations Convolution filter dilations.
* \param input_left_pads Input left paddings.
* \param input_right_pads Input right paddings.
* \param a_element_op A elementwise operation object.
* \param b_element_op B elementwise operation object.
* \param cde_element_op CDE elementwise operation object.
* \return Pointer to the argument.
*/
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
APointers p_a,
BPointers p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
View file @ d3937687
...@@ -3,30 +3,18 @@ ...@@ -3,30 +3,18 @@
#pragma once #pragma once
#include <array> #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
/** /**
* \brief Grouped Convolution Forward * \brief Grouped Convolution Forward
* *
* \details * \note This structure is deprecated (left for backwards compatibility). Please use
* input : input image A[G, N, C, Hi, Wi], A1[G, N, C, Hi, Wi]... * DeviceGroupedConvFwdMultipleABD.
* input : weight B[G, K, C, Y, X], B1[G, K, C, Y, X]...
* input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
* output : output image E[G, N, K, Ho, Wo]
*
* C = a_op(A, A1...) * b_op(B, B1...)
* E = cde_op(C, D0, D1, ...)
* *
* \tparam NDimSpatial Number of spatial dimensions. * \tparam NDimSpatial Number of spatial dimensions.
* \tparam ALayout Input layout (also for a1, a2...). * \tparam ALayout Input layout (also for a1, a2...).
...@@ -60,72 +48,19 @@ template <index_t NDimSpatial, ...@@ -60,72 +48,19 @@ template <index_t NDimSpatial,
ADataType>())> // ComputeType is InputType by default (first ADataType>())> // ComputeType is InputType by default (first
// in tuple for MultiAB), unpack if tuple was // in tuple for MultiAB), unpack if tuple was
// passed // passed
struct DeviceGroupedConvFwdMultipleD : public BaseOperator using DeviceGroupedConvFwdMultipleD = DeviceGroupedConvFwdMultipleABD<NDimSpatial,
{ ALayout,
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value; BLayout,
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value; DsLayout,
ELayout,
static constexpr index_t NumATensor = GetNumABTensors<isMultiA, ADataType>(); ADataType,
static constexpr index_t NumBTensor = GetNumABTensors<isMultiB, BDataType>(); BDataType,
static constexpr index_t NumDTensor = DsDataType::Size(); DsDataType,
EDataType,
static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor"); AElementwiseOperation,
BElementwiseOperation,
// If DataType is tuple, user has to pass std::array with pointers. CDEElementwiseOperation,
using APointers = ComputeType>;
std::conditional_t<isMultiA, std::array<const void*, NumATensor>&, const void*>;
using BPointers =
std::conditional_t<isMultiB, std::array<const void*, NumBTensor>&, const void*>;
/**
* \brief Make argument pointer for grouped conv fwd.
*
* \param p_a A pointer to the input (std::array<const void*, NumA> with
pointers for multiple A).
* \param p_b A pointer to the weight (std::array<const void*, NumA> with
pointers for multiple B).
* \param p_ds A pointers to the Ds.
* \param p_e A pointers to the output.
* \param a_g_n_c_wis_lengths Input lengths [G, N, C, Spatial...] (for 3d).
* \param a_g_n_c_wis_strides Input strides [G, N, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_lengths Weight lengths [G, K, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_strides Weight strides [G, K, C, Spatial...] (for 3d).
* \param ds_g_n_k_wos_lengths Ds lengths [G, N, K, Spatial...] (for 3d).
* \param ds_g_n_k_wos_strides Ds strides [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_lengths Output lengths [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_strides Output strides [G, N, K, Spatial...] (for 3d).
* \param conv_filter_strides Convolution filter strides.
* \param conv_filter_dilations Convolution filter dilations.
* \param input_left_pads Input left paddings.
* \param input_right_pads Input right paddings.
* \param a_element_op A elementwise operation object.
* \param b_element_op B elementwise operation object.
* \param cde_element_op CDE elementwise operation object.
* \return Pointer to the argument.
*/
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
APointers p_a,
BPointers p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ d3937687
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp" #include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp" #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
...@@ -216,18 +216,18 @@ template <index_t NDimSpatial, ...@@ -216,18 +216,18 @@ template <index_t NDimSpatial,
index_t CThreadTransferSrcDstVectorDim, index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector> index_t CThreadTransferDstScalarPerVector>
struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
: public DeviceGroupedConvFwdMultipleD<NDimSpatial, : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout, ALayout,
BLayout, BLayout,
DsLayout, DsLayout,
ELayout, ELayout,
ADataType, ADataType,
BDataType, BDataType,
DsDataType, DsDataType,
EDataType, EDataType,
AElementwiseOperation, AElementwiseOperation,
BElementwiseOperation, BElementwiseOperation,
CDEElementwiseOperation> CDEElementwiseOperation>
{ {
using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK; using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ d3937687
...@@ -834,7 +834,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle ...@@ -834,7 +834,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
// check if it's 1x1, stride=1 conv // check if it's 1x1, stride=1 conv
for(index_t i = 0; i < NDimSpatial; ++i) for(index_t i = 0; i < NDimSpatial; ++i)
{ {
const index_t X = arg.b_g_k_c_xs_lengths_[i + 2]; const index_t X = arg.b_g_k_c_xs_lengths_[i + 3];
const index_t ConvStride = arg.conv_filter_strides_[i]; const index_t ConvStride = arg.conv_filter_strides_[i];
const index_t LeftPad = arg.input_left_pads_[i]; const index_t LeftPad = arg.input_left_pads_[i];
const index_t RightPad = arg.input_right_pads_[i]; const index_t RightPad = arg.input_right_pads_[i];
...@@ -851,7 +851,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle ...@@ -851,7 +851,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
// check if it's 1x1 conv // check if it's 1x1 conv
for(index_t i = 0; i < NDimSpatial; ++i) for(index_t i = 0; i < NDimSpatial; ++i)
{ {
const index_t X = arg.b_g_k_c_xs_lengths_[i + 2]; const index_t X = arg.b_g_k_c_xs_lengths_[i + 3];
const index_t LeftPad = arg.input_left_pads_[i]; const index_t LeftPad = arg.input_left_pads_[i];
const index_t RightPad = arg.input_right_pads_[i]; const index_t RightPad = arg.input_right_pads_[i];
...@@ -1090,7 +1090,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle ...@@ -1090,7 +1090,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
auto str = std::stringstream(); auto str = std::stringstream();
// clang-format off // clang-format off
str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle" str << "DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle"
<< "<" << "<"
<< BlockSize << ", " << BlockSize << ", "
<< MPerBlock << ", " << MPerBlock << ", "
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
View file @ d3937687
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp" #include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp" #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
...@@ -92,18 +92,18 @@ template <index_t NDimSpatial, ...@@ -92,18 +92,18 @@ template <index_t NDimSpatial,
LoopScheduler LoopSched = make_default_loop_scheduler(), LoopScheduler LoopSched = make_default_loop_scheduler(),
ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1> ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
: public DeviceGroupedConvFwdMultipleD<NDimSpatial, : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout, ALayout,
BLayout, BLayout,
DsLayout, DsLayout,
ELayout, ELayout,
ADataType, ADataType,
BDataType, BDataType,
DsDataType, DsDataType,
EDataType, EDataType,
AElementwiseOperation, AElementwiseOperation,
BElementwiseOperation, BElementwiseOperation,
CDEElementwiseOperation> CDEElementwiseOperation>
{ {
using DeviceOp = DeviceGroupedConvFwdMultipleD_Wmma_CShuffle; using DeviceOp = DeviceGroupedConvFwdMultipleD_Wmma_CShuffle;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp
View file @ d3937687
...@@ -7,7 +7,7 @@ ...@@ -7,7 +7,7 @@
#include <memory> #include <memory>
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
...@@ -24,66 +24,66 @@ using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd; ...@@ -24,66 +24,66 @@ using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
// grouped conv3d forward multi AB scaleadd, NDHWGC/GKZYXC/NDHWGK // grouped conv3d forward multi AB scaleadd, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instances( void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC, NDHWGC,
GKZYXC, GKZYXC,
ck::Tuple<>, ck::Tuple<>,
NDHWGK, NDHWGK,
ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>,
ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>,
ck::Tuple<>, ck::Tuple<>,
BF16, BF16,
ScaleAdd, ScaleAdd,
ScaleAdd, ScaleAdd,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instances( void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC, NDHWGC,
GKZYXC, GKZYXC,
ck::Tuple<>, ck::Tuple<>,
NDHWGK, NDHWGK,
ck::Tuple<F16, F16>, ck::Tuple<F16, F16>,
ck::Tuple<F16, F16>, ck::Tuple<F16, F16>,
ck::Tuple<>, ck::Tuple<>,
F16, F16,
ScaleAdd, ScaleAdd,
ScaleAdd, ScaleAdd,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instances( void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC, NDHWGC,
GKZYXC, GKZYXC,
ck::Tuple<>, ck::Tuple<>,
NDHWGK, NDHWGK,
ck::Tuple<F32, F32>, ck::Tuple<F32, F32>,
ck::Tuple<F32, F32>, ck::Tuple<F32, F32>,
ck::Tuple<>, ck::Tuple<>,
F32, F32,
ScaleAdd, ScaleAdd,
ScaleAdd, ScaleAdd,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instances( void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC, NDHWGC,
GKZYXC, GKZYXC,
ck::Tuple<>, ck::Tuple<>,
NDHWGK, NDHWGK,
ck::Tuple<int8_t, int8_t>, ck::Tuple<int8_t, int8_t>,
ck::Tuple<int8_t, int8_t>, ck::Tuple<int8_t, int8_t>,
ck::Tuple<>, ck::Tuple<>,
int8_t, int8_t,
ScaleAdd, ScaleAdd,
ScaleAdd, ScaleAdd,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
...@@ -96,7 +96,7 @@ template <ck::index_t NumDimSpatial, ...@@ -96,7 +96,7 @@ template <ck::index_t NumDimSpatial,
typename DDataTypes, typename DDataTypes,
typename OutDataType, typename OutDataType,
typename ComputeType> typename ComputeType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
...@@ -111,19 +111,20 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -111,19 +111,20 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ComputeType>> ComputeType>>
{ {
using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp =
InLayout, DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
WeiLayout, InLayout,
DLayouts, WeiLayout,
OutLayout, DLayouts,
InDataType, OutLayout,
WeiDataType, InDataType,
DDataTypes, WeiDataType,
OutDataType, DDataTypes,
ck::tensor_operation::element_wise::ScaleAdd, OutDataType,
ck::tensor_operation::element_wise::ScaleAdd, ck::tensor_operation::element_wise::ScaleAdd,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::ScaleAdd,
ComputeType>; ck::tensor_operation::element_wise::PassThrough,
ComputeType>;
static auto GetInstances() static auto GetInstances()
{ {
......
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