mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-05 06:01:23 +00:00
0db7c8f0b2
* Add Gemm fp8xint4 example and kernel, function pass. * Init Gemm_fp8xint4 Bpreshuffle * Added gemm_fp8xint4_Bpreshuffle files, function not checked yet * General fix. * fp8xint4 bpreshuffle function pass * fix. * init b preshuffle dequant in VGPR. * fix bug, function pass. * move b thread dequant copy to blockwise. * fix bug, function now passes. * modified the tile size to 256, 128x128x128. * fixed a bug. * Initial int4 moe, compile pass, function not check. * fix bug in moe_gemm1.cpp, now function pass. * test expert = 8 and function pass. * Added moe_pk_i4_gemm2, function pass. * Added b preshuffle pipeline v3 support. * fixed merge issue. fp8xint4 and fp8xint4_bpreshuffle function pass. * Split the blockwise pipeline for fp8xint4. * commit missing files * opt gemm2 to 2x2 wave * fix swizzle = false * update int4 moe with latest input changes. * update tile size. * enable pipeline v3. * fix nswizzle = true * commit a version for compiler debug. * Updated transfer_v3r1_gather to support pk_i4_t type. * for int4 moe2 for type_convert support. * remove some values between mfma instructions. * fix int4 moe * Updated transfer_v3r1_gather to support pk_i4_t type. * i4 support lds multiple shuffle * fixed int4 moe tflops calculation. * Modified CshuffleCShuffleMXdlPerWavePerShuffle to 1 to suit C multiple shuffle * updated gemm2. * change int4 moe example names * fix and format code. * format. * format codes. * update fp8xint4 example tile size. * add <unordered_map> header * fixed. * format. * Added conditional compilation for int4 -> fp8 conversion kernels --------- Co-authored-by: mtgu0705 <mtgu@amd.com> Co-authored-by: coderfeli <coderfeli@163.com>
395 lines
12 KiB
C++
395 lines
12 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#pragma once
|
|
|
|
#include <cstdlib>
|
|
#include <iostream>
|
|
#include <initializer_list>
|
|
#include <numeric>
|
|
#include <unordered_map>
|
|
|
|
#include "ck/ck.hpp"
|
|
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
|
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
|
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
|
#include "ck/utility/data_type.hpp"
|
|
|
|
#include "ck/library/utility/check_err.hpp"
|
|
#include "ck/library/utility/device_memory.hpp"
|
|
#include "ck/library/utility/fill.hpp"
|
|
#include "ck/library/utility/host_tensor.hpp"
|
|
#include "ck/library/utility/host_tensor_generator.hpp"
|
|
#include "ck/library/utility/literals.hpp"
|
|
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
|
#include "ck/library/reference_tensor_operation/gpu/reference_gemm.hpp"
|
|
|
|
struct ProblemSize final
|
|
{
|
|
ck::index_t M = 3840;
|
|
ck::index_t N = 4096;
|
|
ck::index_t K = 4096;
|
|
|
|
ck::index_t StrideA = -1;
|
|
ck::index_t StrideB = -1;
|
|
ck::index_t StrideC = -1;
|
|
};
|
|
|
|
struct ProblemSizeStreamK final
|
|
{
|
|
ck::index_t M = 3840;
|
|
ck::index_t N = 4096;
|
|
ck::index_t K = 4096;
|
|
|
|
ck::index_t StrideA = -1;
|
|
ck::index_t StrideB = -1;
|
|
ck::index_t StrideC = -1;
|
|
|
|
ck::index_t NumSKBlocks = -1; // number of stream-k blocks
|
|
};
|
|
struct ProblemSizeStreamK_universal final
|
|
{
|
|
ck::index_t M = 3840;
|
|
ck::index_t N = 4096;
|
|
ck::index_t K = 4096;
|
|
|
|
ck::index_t StrideA = -1;
|
|
ck::index_t StrideB = -1;
|
|
ck::index_t StrideC = -1;
|
|
|
|
ck::index_t Grid_size = -1; // defaults to max occupancy
|
|
ck::index_t Streamk_sel = 1; // defaults to 1-tile SK
|
|
};
|
|
|
|
struct ProblemSizeSplitK final
|
|
{
|
|
ck::index_t M = 3840;
|
|
ck::index_t N = 4096;
|
|
ck::index_t K = 4096;
|
|
|
|
ck::index_t StrideA = -1;
|
|
ck::index_t StrideB = -1;
|
|
ck::index_t StrideC = -1;
|
|
|
|
ck::index_t KBatch = 1;
|
|
};
|
|
|
|
struct ExecutionConfig final
|
|
{
|
|
// 0 - no verification, 1 - CPU, 2 - GPU, 3 - CPU + GPU
|
|
int do_verification = 1;
|
|
int init_method = 2;
|
|
bool time_kernel = false;
|
|
};
|
|
|
|
template <ck::index_t... Is>
|
|
using S = ck::Sequence<Is...>;
|
|
|
|
using Row = ck::tensor_layout::gemm::RowMajor;
|
|
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
|
|
|
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
|
|
|
template <typename ProblemType>
|
|
bool parse_cmd_args(int, char*[], ProblemType&, ExecutionConfig&)
|
|
{
|
|
return false;
|
|
}
|
|
|
|
template <>
|
|
bool parse_cmd_args<ProblemSize>(int argc,
|
|
char* argv[],
|
|
ProblemSize& problem_size,
|
|
ExecutionConfig& config)
|
|
{
|
|
if(argc == 1)
|
|
{
|
|
// use default case
|
|
}
|
|
else if(argc == 4)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
}
|
|
else if(argc == 10)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
|
|
problem_size.M = std::stoi(argv[4]);
|
|
problem_size.N = std::stoi(argv[5]);
|
|
problem_size.K = std::stoi(argv[6]);
|
|
|
|
problem_size.StrideA = std::stoi(argv[7]);
|
|
problem_size.StrideB = std::stoi(argv[8]);
|
|
problem_size.StrideC = std::stoi(argv[9]);
|
|
}
|
|
else
|
|
{
|
|
std::cerr << "arg1: verification (0=no, 1=CPU, 2=GPU, 3=CPU and GPU)" << std::endl
|
|
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
|
|
<< std::endl
|
|
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
|
|
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl;
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
template <>
|
|
bool parse_cmd_args<ProblemSizeStreamK_universal>(int argc,
|
|
char* argv[],
|
|
ProblemSizeStreamK_universal& problem_size,
|
|
ExecutionConfig& config)
|
|
{
|
|
if(argc == 1)
|
|
{
|
|
// use default case
|
|
}
|
|
else if(argc == 4)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
}
|
|
else if(argc >= 10)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
|
|
problem_size.M = std::stoi(argv[4]);
|
|
problem_size.N = std::stoi(argv[5]);
|
|
problem_size.K = std::stoi(argv[6]);
|
|
|
|
problem_size.StrideA = std::stoi(argv[7]);
|
|
problem_size.StrideB = std::stoi(argv[8]);
|
|
problem_size.StrideC = std::stoi(argv[9]);
|
|
|
|
if(argc >= 11)
|
|
{
|
|
problem_size.Streamk_sel = std::stoi(argv[10]);
|
|
problem_size.Grid_size = std::stoi(argv[11]);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cerr
|
|
<< "arg1: verification (0=no, 1=CPU, 2=GPU, 3=CPU and GPU)" << std::endl
|
|
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)" << std::endl
|
|
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
|
|
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
|
|
<< "arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
|
|
<< "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
template <>
|
|
bool parse_cmd_args<ProblemSizeStreamK>(int argc,
|
|
char* argv[],
|
|
ProblemSizeStreamK& problem_size,
|
|
ExecutionConfig& config)
|
|
{
|
|
if(argc == 1)
|
|
{
|
|
// use default case
|
|
}
|
|
else if(argc == 4)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
}
|
|
else if(argc >= 10)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
|
|
problem_size.M = std::stoi(argv[4]);
|
|
problem_size.N = std::stoi(argv[5]);
|
|
problem_size.K = std::stoi(argv[6]);
|
|
|
|
problem_size.StrideA = std::stoi(argv[7]);
|
|
problem_size.StrideB = std::stoi(argv[8]);
|
|
problem_size.StrideC = std::stoi(argv[9]);
|
|
|
|
if(argc >= 11)
|
|
{
|
|
problem_size.NumSKBlocks = std::stoi(argv[10]);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cerr << "arg1: verification (0=no, 1=CPU, 2=GPU, 3=CPU and GPU)" << std::endl
|
|
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
|
|
<< std::endl
|
|
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
|
|
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
|
|
<< "arg10: stream-k select (0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
|
|
<< "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
template <>
|
|
bool parse_cmd_args<ProblemSizeSplitK>(int argc,
|
|
char* argv[],
|
|
ProblemSizeSplitK& problem_size,
|
|
ExecutionConfig& config)
|
|
{
|
|
if(argc == 1)
|
|
{
|
|
// use default case
|
|
}
|
|
else if(argc == 4)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
}
|
|
else if(argc >= 10)
|
|
{
|
|
config.do_verification = std::stoi(argv[1]);
|
|
config.init_method = std::stoi(argv[2]);
|
|
config.time_kernel = std::stoi(argv[3]);
|
|
|
|
problem_size.M = std::stoi(argv[4]);
|
|
problem_size.N = std::stoi(argv[5]);
|
|
problem_size.K = std::stoi(argv[6]);
|
|
|
|
problem_size.StrideA = std::stoi(argv[7]);
|
|
problem_size.StrideB = std::stoi(argv[8]);
|
|
problem_size.StrideC = std::stoi(argv[9]);
|
|
|
|
if(argc >= 11)
|
|
{
|
|
problem_size.KBatch = std::stoi(argv[10]);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cerr << "arg1: verification (0=no, 1=CPU, 2=GPU, 3=CPU and GPU)" << std::endl
|
|
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
|
|
<< std::endl
|
|
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
|
|
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
|
|
<< "arg10: KBatch" << std::endl;
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
template <typename DataType>
|
|
inline __host__ __device__ constexpr double get_rtol()
|
|
{
|
|
if constexpr(std::is_same_v<DataType, float>)
|
|
{
|
|
return 1e-3;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, double>)
|
|
{
|
|
return 1e-6;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::half_t>)
|
|
{
|
|
return 1e-3;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
|
|
{
|
|
return 5e-2;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, int32_t>)
|
|
{
|
|
return 1e-1;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, int8_t>)
|
|
{
|
|
return 1e-1;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
|
|
{
|
|
return 1e-1; // 240 and 224 are acceptable
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
|
|
{
|
|
return 1.5e-1; // 57344 and 49152 are acceptable
|
|
}
|
|
else
|
|
{
|
|
return 1e-3;
|
|
}
|
|
}
|
|
|
|
template <typename DataType>
|
|
inline __host__ __device__ constexpr double get_atol()
|
|
{
|
|
if constexpr(std::is_same_v<DataType, float>)
|
|
{
|
|
return 1e-3;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, double>)
|
|
{
|
|
return 1e-6;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::half_t>)
|
|
{
|
|
return 1e-3;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
|
|
{
|
|
return 5e-2;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, int32_t>)
|
|
{
|
|
return 1e-1;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, int8_t>)
|
|
{
|
|
return 1e-1;
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
|
|
{
|
|
return 16.1; // 240 and 224 are acceptable
|
|
}
|
|
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
|
|
{
|
|
return 8192.1; // 57344 and 49152 are acceptable
|
|
}
|
|
else
|
|
{
|
|
return 1e-3;
|
|
}
|
|
}
|
|
|
|
float i4_to_f32_gfx9(uint8_t i4)
|
|
{
|
|
static std::unordered_map<uint8_t, float> u = {{0b1000, -0.5000f},
|
|
{0b1001, -0.4375f},
|
|
{0b1010, -0.3750f},
|
|
{0b1011, -0.3125f},
|
|
{0b1100, -0.2500f},
|
|
{0b1101, -0.1875f},
|
|
{0b1110, -0.1250f},
|
|
{0b1111, -0.0625f},
|
|
{0b0, +0.0000f},
|
|
{0b1, +0.0625f},
|
|
{0b10, +0.1250f},
|
|
{0b11, +0.1875f},
|
|
{0b100, +0.2500f},
|
|
{0b101, +0.3125f},
|
|
{0b110, +0.3750f},
|
|
{0b111, +0.4375f}};
|
|
|
|
return u[i4];
|
|
}
|