ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

revert contraction deviceop changes

Browse Source
This commit is contained in:
root
2024-04-01 00:26:01 +00:00
parent 38f109c460
commit 943199a991
5 changed files with 23 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

64
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
View File

@@ -656,26 +656,16 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
{
bool valid_as_access = true;
static_for<0, NumATensor, 1>{}([&](auto i) {
// vector memory access of A: could be on M or AK1 dimension
if constexpr(ABlockTransferSrcVectorDim == 1)
const bool valid_a_vector_size =
arg.as_max_read_elems_[i] % ABlockTransferSrcScalarPerVector == 0;
const bool valid_a_access_dim_m =
ABlockTransferSrcVectorDim == 1 && arg.as_mz_consecutive_[i];
const bool valid_a_access_dim_k =
ABlockTransferSrcVectorDim == 2 && arg.as_kz_consecutive_[i];
const bool valid_a_access_dim = valid_a_access_dim_m || valid_a_access_dim_k;
if(!(valid_a_vector_size && valid_a_access_dim))
{
if(!(arg.a_mz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I1) %
ABlockTransferSrcScalarPerVector ==
0) &&
ABlockTransferSrcScalarPerVector != 1)
{
all_valid = false;
}
}
else
{
if(!(arg.a_kz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I2) %
ABlockTransferSrcScalarPerVector ==
0) &&
ABlockTransferSrcScalarPerVector != 1)
{
all_valid = false;
}
valid_as_access = false;
}
});
if(!valid_as_access)
@@ -694,23 +684,7 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
const bool valid_b_access_dim = valid_b_access_dim_n || valid_b_access_dim_k;
if(!(valid_b_vector_size && valid_b_access_dim))
{
if(!(arg.b_nz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I1) %
BBlockTransferSrcScalarPerVector ==
0) &&
BBlockTransferSrcScalarPerVector != 1)
{
all_valid = false;
}
}
else
{
if(!(arg.b_kz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I2) %
BBlockTransferSrcScalarPerVector ==
0) &&
BBlockTransferSrcScalarPerVector != 1)
{
all_valid = false;
}
valid_bs_access = false;
}
});
if(!valid_bs_access)
@@ -720,22 +694,16 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
bool valid_ds_access = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
if(!(arg.ds_nz_stride_[i] == 1 &&
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
CDEBlockTransferScalarPerVector_NPerBlock ==
0) &&
CDEBlockTransferScalarPerVector_NPerBlock != 1)
const bool valid_d_vector_size =
arg.ds_max_read_elems_[i] % CDEBlockTransferScalarPerVector_NPerBlock == 0;
// Vector read of Ds is always on N dimension.
const bool valid_d_access_dim = arg.ds_nz_consecutive_[i];
if(!(valid_d_vector_size && valid_d_access_dim))
{
valid_ds_access = false;
}
});
// vector memory access of E: always on NPerBlock dimension
if(!(arg.e_nz_stride_ == 1 &&
arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
CDEBlockTransferScalarPerVector_NPerBlock ==
0) &&
CDEBlockTransferScalarPerVector_NPerBlock != 1)
if(!valid_ds_access)
{
return false;
}

81
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
View File

@@ -169,78 +169,6 @@ struct DeviceGemmMultipleABD_Xdl_CShuffle : public DeviceGemmMultipleABD<AsLayou
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
#if 0
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideAs)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, AsLayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideAs, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, AsLayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideAs));
}
}();
return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
}
static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideBs)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BsLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideBs));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BsLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideBs, I1));
}
}();
return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
{
const auto e_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideE));
}
}();
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
},
Number<NumDTensor>{});
}
#endif
using ComputeDataType = EDataType;
// GridwiseGemm
@@ -424,15 +352,6 @@ struct DeviceGemmMultipleABD_Xdl_CShuffle : public DeviceGemmMultipleABD<AsLayou
}
}
void Print() const
{
// std::cout << "A[M, K]: " << as_grid_desc_m_k_ << std::endl;
// std::cout << "B[N, K]: " << bs_grid_desc_n_k_ << std::endl;
// static_for<0, NumDTensor, 1>{}(
//[&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
// std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
}
// private:
// pointers
typename GridwiseGemm::AsGridPointer p_as_grid_;

33
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View File

@@ -10,38 +10,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
namespace ck {
// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
// and sometimes useless instructions:
// 1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
// instead
// 2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
// tensor coordinate instead
// 3. Don't use a pointer to VGPR buffer, use vector instead
namespace detail {
// TODO: How to fix this? It uses an struct instead of lambda because lambda
// doesn't have constructor
template <index_t VectorDim, index_t ScalarPerVector>
struct lambda_scalar_per_access
{
__host__ __device__ constexpr auto operator()(index_t i) const
{
return (i == VectorDim) ? ScalarPerVector : 1;
}
};
template <index_t VectorDim>
struct lambda_scalar_step_in_vector
{
__host__ __device__ constexpr auto operator()(index_t i) const
{
return (i == VectorDim) ? 1 : 0;
}
};
} // namespace detail
// Assume:
// 1. src:
// 1. SrcDesc is known at compile-time

37
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -7,44 +7,13 @@
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor/static_tensor.hpp"
#include "ck/utility/is_detected.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
namespace ck {
namespace detail {
// TODO: How to fix this? It uses an struct instead of lambda because lambda
// doesn't have constructor
template <index_t SrcVectorDim,
index_t SrcScalarPerVector,
index_t DstVectorDim,
index_t DstScalarPerVector>
struct lambda_scalar_per_access_for_src_and_dst
{
__host__ __device__ constexpr auto operator()(index_t i) const
{
if(i == SrcVectorDim && i == DstVectorDim)
{
return math::lcm(SrcScalarPerVector, DstScalarPerVector);
}
else if(i == SrcVectorDim)
{
return SrcScalarPerVector;
}
else if(i == DstVectorDim)
{
return DstScalarPerVector;
}
else
{
return 1;
}
}
};
} // namespace detail
// Assume:
// 1. src_desc and dst_desc are not known at compile-time
// 2. SrcBuffer and DstBuffer are DynamicBuffer

35
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
View File

@@ -10,40 +10,9 @@
#include "ck/utility/is_detected.hpp"
#include "ck/tensor/static_tensor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
namespace ck {
namespace detail {
// TODO: How to fix this? It uses an struct instead of lambda because lambda
// doesn't have constructor
template <index_t SrcVectorDim,
index_t SrcScalarPerVector,
index_t DstVectorDim,
index_t DstScalarPerVector>
struct lambda_scalar_per_access_for_src_and_dst
{
__host__ __device__ constexpr auto operator()(index_t i) const
{
if(i == SrcVectorDim && i == DstVectorDim)
{
return math::lcm(SrcScalarPerVector, DstScalarPerVector);
}
else if(i == SrcVectorDim)
{
return SrcScalarPerVector;
}
else if(i == DstVectorDim)
{
return DstScalarPerVector;
}
else
{
return 1;
}
}
};
} // namespace detail
// Thread-level multi-source, multi-destination tensor slice data movement
// Assume:
// 1. All sources and destinations are DynamicBuffer