ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 18:17:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Introduce gemm_softmax_gemm to codegen (#1542)

Browse Source 
* Introduce ck_host library and gemm_softmax_gemm.

* Minor refactor.

* Add descriptor to gemm_softmax_gemm.

* Bugfix.

* Revert ck_host library.

* fix clang format

---------

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>

[ROCm/composable_kernel commit: b5ca008d62]
This commit is contained in:
Mirza Halilčević
2025-02-11 17:07:24 +01:00
committed by GitHub
parent 8ed234da8c
commit af878d1bd8
12 changed files with 1071 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
codegen/include/ck/host/device_batched_gemm_softmax_gemm/operation.hpp Normal file
View File

@@ -0,0 +1,61 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <string>
#include "ck/host/types.hpp"
#include "ck/host/operation/gemm.hpp"
#include "ck/host/device_batched_gemm_softmax_gemm/problem.hpp"
namespace ck {
namespace host {
namespace device_batched_gemm_softmax_gemm {
// defines all values need for an instance of fwd conv
struct Operation_Xdl_CShuffle
{
// returns a vector of instances, only given fusion operators: will use default problem spec
static std::vector<std::vector<Operation_Xdl_CShuffle>>
CreateOperations(const std::string& prologue, const std::string& epilogue);
// returns a vector of instances, given a problem spec and fusion operators
static std::vector<Operation_Xdl_CShuffle>
CreateOperations(const Problem& prob, const std::string& prologue, const std::string& epilogue);
TensorDesc A{};
TensorDesc B{};
TensorDesc B1{};
TensorDesc C{};
DataType acc = DataType::Float;
DataType cs_type = DataType::Half;
std::string a_elem_op = PassThrough;
std::string b_elem_op = PassThrough;
std::string b1_elem_op = PassThrough;
std::string c_elem_op = PassThrough;
std::string acc_elem_op = Scale;
std::string prologue = "";
std::string epilogue = "";
std::string gemm_specialization = "ck::tensor_operation::device::GemmSpecialization::Default";
// tuning parameters
operation::TileDescGemmGemm tile_desc{};
operation::BlockTransferDesc a_block_transfer{};
operation::BlockTransferDesc b0_block_transfer{};
operation::BlockTransferDesc b1_block_transfer{};
operation::CShuffleDesc cshuffle{};
operation::CBlockTransferDesc c_block_transfer{};
bool mask_out_upper_triangle = false;
// functions to update fusion operators if provided
void update_prologue(const std::string& prologue);
void update_epilogue(const std::string& epilogue);
/**constexpr**/ bool
IsSupported(std::size_t MRaw_, std::size_t NRaw_, std::size_t KRaw_, std::size_t Gemm1NRaw_);
// returns a templated instance
Solution ToSolution() const;
};
} // namespace device_batched_gemm_softmax_gemm
} // namespace host
} // namespace ck

47
codegen/include/ck/host/device_batched_gemm_softmax_gemm/problem.hpp Normal file
View File

@@ -0,0 +1,47 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <string>
#include "ck/host/types.hpp"
namespace ck {
namespace host {
namespace device_batched_gemm_softmax_gemm {
// defines the problem specification for a GEMM operation
struct Problem
{
std::size_t M = 0;
std::size_t N = 0;
std::size_t K = 0;
std::size_t O = 0;
bool TransA = false;
bool TransB = false;
bool TransB1 = false;
bool TransC = false;
DataType ADataType = DataType::Half;
DataType BDataType = DataType::Half;
DataType B1DataType = DataType::Half;
DataType CDataType = DataType::Half;
std::string AElementOp = PassThrough;
std::string BElementOp = PassThrough;
std::string B1ElementOp = PassThrough;
std::string CElementOp = PassThrough;
std::string AccElementOp = Scale;
// returns the correct device op file for the operation
std::string GetIncludeHeader() const;
// returns a list of instances based on the problem spec and provided fusion operations
std::vector<Solution> GetSolutions(const std::string& arch,
const std::string& prologue,
const std::string& epilogue) const;
};
} // namespace device_batched_gemm_softmax_gemm
} // namespace host
} // namespace ck

2
codegen/include/ck/host/device_gemm_multiple_d/operation.hpp
View File

@@ -41,6 +41,8 @@ struct Operation_Xdl_CShuffle
operation::BlockTransferDesc b_block_transfer{};
operation::CShuffleDesc cshuffle{};
operation::CBlockTransferDesc c_block_transfer{};
LoopScheduler loop_scheduler{};
PipelineVersion pipeline_version{};
// functions to update fusion operators if provided
void update_prologue(const std::string& prologue);

20
codegen/include/ck/host/operation/gemm.hpp
View File

@@ -23,6 +23,26 @@ struct TileDesc
int n_Xdl_per_wave = 0;
int num_gemmk_prefetch_stage = 0;
};
struct TileDescGemmGemm
{
int block_size = 0;
int gemm01_m_per_block = 0;
int gemm0_n_per_block = 0;
int gemm0_k_per_block = 0;
int gemm1_n_per_block = 0;
int gemm1_k_per_block = 0;
int ak1 = 0;
int bk1 = 0;
int b1k1 = 0;
int m_per_XDL = 0;
int n_per_XDL = 0;
int gemm0_m_Xdl_per_wave = 0;
int gemm0_n_Xdl_per_wave = 0;
int gemm1_n_Xdl_per_wave = 0;
int num_gemmk_prefetch_stage = 0;
};
struct BlockTransferDesc
{
std::string thread_cluster_length = "";

15
codegen/include/ck/host/types.hpp
View File

@@ -66,6 +66,20 @@ enum class GemmType
};
std::string ToString(GemmType gt);
enum class LoopScheduler
{
Default,
Interwave,
};
std::string ToString(LoopScheduler ls);
enum class PipelineVersion
{
v1,
v2
};
std::string ToString(PipelineVersion pv);
struct TensorDesc
{
DataType element;
@@ -84,6 +98,7 @@ const std::string S = SequenceStr({xs...});
constexpr const char* PassThrough = "ck::tensor_operation::element_wise::PassThrough";
constexpr const char* Bilinear = "ck::tensor_operation::element_wise::Bilinear";
constexpr const char* Scale = "ck::tensor_operation::element_wise::Scale";
} // namespace host
} // namespace ck