Skip to content

GitLab

Commit 6ef4e211 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into contraction

parents b0a2afb9 9e4429f9
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1582 additions and 1560 deletions
include/ck/tensor_operation/gpu/device/device_batched_gemm_xdl.hpp
View file @ 6ef4e211
......@@ -10,7 +10,7 @@
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
#include "ck/device_utility/device_prop.hpp"
......@@ -113,7 +113,7 @@ __global__ void
ignore = c_element_op;
ignore = compute_ptr_offset_of_batch;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
#endif
}
template <typename ADataType,
......@@ -151,8 +151,15 @@ template <typename ADataType,
bool BBlockLdsAddExtraN,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceBatchedGemmXdl
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -334,26 +341,26 @@ struct DeviceBatchedGemmXdl
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t Batch,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount)
CElementwiseOperation c_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
BatchCount_(BatchCount),
Batch_(Batch),
a_grid_desc_k0_m_k1_{
DeviceBatchedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA)},
b_grid_desc_k0_n_k1_{
DeviceBatchedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB)},
c_grid_desc_m_n_{DeviceBatchedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC)},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
compute_ptr_offset_of_batch_{
type_convert<index_t>(a_grid_desc_k0_m_k1_.GetElementSpaceSize()),
type_convert<index_t>(b_grid_desc_k0_n_k1_.GetElementSpaceSize()),
type_convert<index_t>(c_grid_desc_m_n_.GetElementSpaceSize())},
compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideC},
block_2_ctile_map_{
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01)},
M01_{M01},
......@@ -376,7 +383,7 @@ struct DeviceBatchedGemmXdl
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
index_t BatchCount_;
index_t Batch_;
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
......@@ -420,7 +427,7 @@ struct DeviceBatchedGemmXdl
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
......@@ -451,7 +458,7 @@ struct DeviceBatchedGemmXdl
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.BatchCount_,
arg.Batch_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
......@@ -485,7 +492,7 @@ struct DeviceBatchedGemmXdl
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.BatchCount_,
arg.Batch_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
......@@ -536,10 +543,13 @@ struct DeviceBatchedGemmXdl
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t Batch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount)
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_b,
......@@ -550,12 +560,15 @@ struct DeviceBatchedGemmXdl
StrideA,
StrideB,
StrideC,
BatchStrideA,
BatchStrideB,
BatchStrideC,
Batch,
1,
1,
a_element_op,
b_element_op,
c_element_op,
BatchCount};
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -570,10 +583,13 @@ struct DeviceBatchedGemmXdl
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t Batch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount) override
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -584,12 +600,15 @@ struct DeviceBatchedGemmXdl
StrideA,
StrideB,
StrideC,
BatchStrideA,
BatchStrideB,
BatchStrideC,
Batch,
1,
1,
a_element_op,
b_element_op,
c_element_op,
BatchCount);
c_element_op);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/device_binary_elementwise.hpp
View file @ 6ef4e211
......@@ -9,6 +9,7 @@
#include "ck/device_utility/device_prop.hpp"
#include "ck/device_utility/kernel_launch.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp"
namespace ck {
......@@ -25,7 +26,7 @@ template <typename ADataType,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector>
struct DeviceBinaryElementwise : public BaseOperator
struct DeviceBinaryElementwise : public DeviceElementwise<2, 1, NDim, ElementwiseFunctor>
{
static constexpr auto I0 = Number<0>{};
......@@ -198,27 +199,30 @@ struct DeviceBinaryElementwise : public BaseOperator
return true;
};
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, 2> p_inputs,
std::array<void*, 1> p_outputs,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
ElementwiseFunctor functor)
std::vector<std::vector<index_t>> input_strides,
std::vector<std::vector<index_t>> output_strides,
ElementwiseFunctor functor) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
return std::make_unique<Argument>(static_cast<const ADataType*>(p_inputs[0]),
static_cast<const BDataType*>(p_inputs[1]),
static_cast<CDataType*>(p_outputs[0]),
lengths,
a_strides,
b_strides,
c_strides,
input_strides[0],
input_strides[1],
output_strides[0],
functor);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
......@@ -226,7 +230,11 @@ struct DeviceBinaryElementwise : public BaseOperator
// clang-format off
str << "DeviceBinaryElementwise"
<< "<"
<< "NDim = " << NDim
<< "MPerThread = " << MPerThread
<< "AScalarPerVector = " << AScalarPerVector
<< "BScalarPerVector = " << BScalarPerVector
<< "CScalarPerVector = " << CScalarPerVector
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ 6ef4e211
......@@ -871,7 +871,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvFwdSpecializationStr(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ 6ef4e211
......@@ -711,7 +711,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvFwdSpecializationStr(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ 6ef4e211
......@@ -1033,7 +1033,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvFwdSpecializationStr(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_elementwise.hpp 0 → 100644
View file @ 6ef4e211
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <ck::index_t NumInputTensor,
ck::index_t NumOutputTensor,
index_t NDim,
typename ElementwiseFunctor>
struct DeviceElementwise : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumInputTensor> p_inputs,
std::array<void*, NumOutputTensor> p_outputs,
std::vector<index_t> lengths,
std::vector<std::vector<index_t>> input_strides,
std::vector<std::vector<index_t>> output_strides,
ElementwiseFunctor functor) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <ck::index_t NumInputTensor,
ck::index_t NumOutputTensor,
index_t NDim,
typename ElementwiseFunctor>
using DeviceElementwisePtr =
std::unique_ptr<DeviceElementwise<NumInputTensor, NumOutputTensor, NDim, ElementwiseFunctor>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm.hpp
View file @ 6ef4e211
......@@ -18,12 +18,19 @@ struct GemmShape
ck::index_t StrideA, StrideB, StrideC;
};
template <typename AElementwiseOperation,
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGemm : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
ck::index_t M,
......@@ -34,17 +41,29 @@ struct DeviceGemm : public BaseOperator
ck::index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) = 0;
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmPtr = std::unique_ptr<
DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
using DeviceGemmPtr = std::unique_ptr<DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>>;
template <typename AElementwiseOperation,
typename BElementwiseOperation,
......
include/ck/tensor_operation/gpu/device/device_gemm_bias.hpp → include/ck/tensor_operation/gpu/device/device_gemm_bias_c_permute.hpp
View file @ 6ef4e211
......@@ -3,33 +3,45 @@
#pragma once
#include <iostream>
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
struct DEGridDesc_M0_M1_M2_N0_N1
{
ck::index_t M0_, M1_, M2_, N0_, N1_;
ck::index_t stride_M0_, stride_M1_, stride_M2_, stride_N0_, stride_N1_;
};
// input : A[M, K], B[K, N],
// input : D[M, N], ...
// output : E[M, N]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D)
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGemmBias : public BaseOperator
typename CDEElementwiseOperation>
struct DeviceGemmBiasCPermute : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_bias,
void* p_c,
const void* p_d,
void* p_e,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
DEGridDesc_M0_M1_M2_N0_N1 d_gride_desc,
DEGridDesc_M0_M1_M2_N0_N1 e_gride_desc,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) = 0;
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
......@@ -37,8 +49,8 @@ struct DeviceGemmBias : public BaseOperator
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmBiasPtr = std::unique_ptr<
DeviceGemmBias<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
using DeviceGemmBiasCPermutePtr = std::unique_ptr<
DeviceGemmBiasCPermute<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_gemm_dl.hpp
View file @ 6ef4e211
......@@ -64,8 +64,16 @@ template <
is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
bool> = false>
struct DeviceGemmDl
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
struct DeviceGemmDl : public DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -534,8 +542,7 @@ struct DeviceGemmDl
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t /* KBatch */ = 1) override
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......
include/ck/tensor_operation/gpu/device/device_gemm_xdl.hpp
View file @ 6ef4e211
......@@ -57,8 +57,15 @@ template <typename ADataType,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
ck::index_t NumPrefetch = 1>
struct DeviceGemmXdl
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
struct DeviceGemmXdl : public DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -487,8 +494,7 @@ struct DeviceGemmXdl
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t /* KBatch */ = 1) override
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......
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