ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-21 05:19:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Wmma support for multiple ABD GEMM (#2803)

Browse Source 
* multi_abd wmma support:

 - Add multiple A and B support to multiple D implementation (gridwise level)
 - Add multi_abd GEMM (device level)
 - Add instances (xdl parity)
 - Add tests (both xdl and wmma)
 - Add examples
 - Add ckProfiler support (both xdl and wmma)

* Fix bug in device print function

* Fix unused template parameter

* Fix batched gemm for multiABD gridwise implementation

* Fix gemm_universal_reduce with multiABDs gridwise implementation

---------

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

[ROCm/composable_kernel commit: 3d29bff2f0]
This commit is contained in:
Enrico Degregori
2025-09-23 03:49:06 +02:00
committed by GitHub
parent 7122e27fd8
commit 0b149c8695
38 changed files with 5343 additions and 312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

180
profiler/src/profile_gemm_multi_abd.cpp Normal file
View File

@@ -0,0 +1,180 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_multi_abd_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
BF16_I8_BF16_BF16, // 0
};
enum struct GemmElementOp
{
PASS_THROUGH, // 0
MULTIPLY, // 1
ADD, // 2
FASTGELU, // 3
ADD_FASTGELU, // 4
MULTIPLY_ADD, // 5
MULTIPLY_FASTGELU, // 6
MULTIPLY_ADD_FASTGELU, // 7
};
#define OP_NAME "gemm_multi_abd"
#define OP_DESC "GEMM_Multiple_ABD"
int profile_gemm_multi_abd(int argc, char* argv[])
{
if(argc != 18)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: bf16@int8/bf16->bf16;)\n");
printf("arg3: matrix layout (0: E[m, n] = A[m, k] * B[k, n];\n");
printf(" 1: E[m, n] = A[m, k] * B[n, k];\n");
printf(" 2: E[m, n] = A[k, m] * B[k, n];\n");
printf(" 3: E[m, n] = A[k, m] * B[n, k])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8: number of As (1)\n");
printf("arg9: number of Bs (1/2)\n");
printf("arg10: number of Ds (0/1/2)\n");
printf("arg11 to 17: M, N, K, StrideA, StrideB, StrideE, StrideD\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int num_as = std::stoi(argv[8]);
const int num_bs = std::stoi(argv[9]);
const int num_ds = std::stoi(argv[10]);
const int M = std::stoi(argv[11]);
const int N = std::stoi(argv[12]);
const int K = std::stoi(argv[13]);
const int StrideA = std::stoi(argv[14]);
const int StrideB = std::stoi(argv[15]);
const int StrideE = std::stoi(argv[16]);
const int StrideD = std::stoi(argv[17]);
using F32 = float;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
auto profile = [&](auto b_layout, auto b_element_op, auto cde_element_op, auto num_d_tensor) {
using ADataType = BF16;
using B0DataType = I8;
using B1DataType = BF16;
using DDataType = BF16;
using EDataType = BF16;
using ALayout = Row;
using BLayout = decltype(b_layout);
using DLayout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = decltype(b_element_op);
using CDEElementOp = decltype(cde_element_op);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD = ck::is_same_v<DLayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
constexpr auto NumberDTensor = decltype(num_d_tensor){};
// Only num_d_tensor == 0 and 1 are supported
using DsDataType = typename std::
conditional<(NumberDTensor == 0), ck::Tuple<>, ck::Tuple<DDataType>>::type;
using DsLayout =
typename std::conditional<(NumberDTensor == 0), ck::Tuple<>, ck::Tuple<DLayout>>::type;
bool pass = ck::profiler::profile_gemm_multi_abd_impl<ck::Tuple<ADataType>,
ck::Tuple<B0DataType, B1DataType>,
F32,
DsDataType,
EDataType,
ck::Tuple<ALayout>,
ck::Tuple<BLayout, BLayout>,
DsLayout,
ELayout,
AElementOp,
BElementOp,
CDEElementOp>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD < 0) ? DefaultStrideD : StrideD,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
// num_as == 1 is only supported
if(data_type != GemmDataType::BF16_I8_BF16_BF16 || num_as != 1)
{
std::cout << "The provided input parameters are not supported" << std::endl;
return 1;
}
// Supported configurations
if(layout == GemmMatrixLayout::MK_KN_MN && num_bs == 2 && num_ds == 1)
{
return profile(Row{}, Multiply{}, AddFastGelu{}, ck::Number<1>{});
}
else if(layout == GemmMatrixLayout::MK_KN_MN && num_bs == 2 && num_ds == 0)
{
return profile(Row{}, Multiply{}, FastGelu{}, ck::Number<0>{});
}
else if(layout == GemmMatrixLayout::MK_NK_MN && num_bs == 2 && num_ds == 1)
{
return profile(Col{}, Multiply{}, AddFastGelu{}, ck::Number<1>{});
}
else if(layout == GemmMatrixLayout::MK_NK_MN && num_bs == 2 && num_ds == 0)
{
return profile(Col{}, Multiply{}, FastGelu{}, ck::Number<0>{});
}
std::cout << "The provided input parameters are not supported" << std::endl;
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_multi_abd);