File size: 11,254 Bytes
b024d42 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 | // SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "contraction_bilinear"
#define OP_DESC "CONTRACTION+Bilinear"
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg4: Number of dimension for M, N and K (one for all)\n"
<< "arg5: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg6: verification (0: no; 1: yes)\n"
<< "arg7: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg8: print tensor value (0: no; 1: yes)\n"
<< "arg9: time kernel (0: no, 1: yes)\n"
<< "arg10: alpha\n"
<< "arg11: beta\n"
<< "arg12 to 17/29: M0, M1, N0, N1, K0, K1\n"
<< "arg18/30 to 33/77: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_bilinear(int argc, char* argv[])
{
const bool default_strides = argc == 18 || 30;
if(argc != 34 && argc != 78 && !default_strides)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto compute_data_type = static_cast<ContractionComputeDataType>(std::stoi(argv[3]));
const ck::index_t NumDimMNK = std::stoi(argv[4]);
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const ck::index_t init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
const float alpha = std::stof(argv[10]);
const float beta = std::stof(argv[11]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 12;
collect_index_params(argv, M, dims_arg_num, NumDimMNK);
collect_index_params(argv, N, dims_arg_num + NumDimMNK, NumDimMNK);
collect_index_params(argv, K, dims_arg_num + NumDimMNK * 2, NumDimMNK);
std::vector<ck::index_t> StridesA(NumDimMNK * 2);
std::vector<ck::index_t> StridesB(NumDimMNK * 2);
std::vector<ck::index_t> StridesE(NumDimMNK * 2);
std::vector<ck::index_t> StridesD(NumDimMNK * 2);
if(!default_strides)
{
collect_index_params(argv, StridesA, dims_arg_num + NumDimMNK * 3, NumDimMNK * 2);
collect_index_params(argv, StridesB, dims_arg_num + NumDimMNK * 5, NumDimMNK * 2);
collect_index_params(argv, StridesE, dims_arg_num + NumDimMNK * 7, NumDimMNK * 2);
collect_index_params(argv, StridesD, dims_arg_num + NumDimMNK * 9, NumDimMNK * 2);
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
auto profile =
[&](auto a_layout, auto b_layout, auto cde_layout, auto type, auto compute_type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
using ComputeDataType = decltype(compute_type);
if(default_strides)
{
auto merge_dims = [](const std::vector<ck::index_t>& dims01,
const std::vector<ck::index_t>& dims23) {
std::vector<ck::index_t> dims_szt(dims01.begin(), dims01.end());
dims_szt.insert(dims_szt.end(), dims23.begin(), dims23.end());
return dims_szt;
};
assign_default_strides(a_layout, StridesA, merge_dims(M, K));
assign_default_strides(b_layout, StridesB, merge_dims(N, K));
assign_default_strides(cde_layout, StridesE, merge_dims(M, N));
assign_default_strides(cde_layout, StridesD, merge_dims(M, N));
}
if(NumDimMNK == 2)
{
bool pass = ck::profiler::profile_contraction_impl<2,
ALayout,
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
do_log,
time_kernel,
Bilinear{alpha, beta},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
}
else if(NumDimMNK == 6)
{
bool pass = ck::profiler::profile_contraction_impl<6,
ALayout,
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
do_log,
time_kernel,
Bilinear{alpha, beta},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
}
else
{
throw std::runtime_error("Not supported NumDimMNK");
return false;
}
};
auto run_profile_for_datatype = [&](auto type, auto compute_type) {
if(layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, type, compute_type);
}
return false;
};
if(data_type == ContractionDataType::F32_F32_F32_F32)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F32{}, F32{});
}
else if(compute_data_type == ContractionComputeDataType::F16)
{
return run_profile_for_datatype(F32{}, F16{});
}
else if(compute_data_type == ContractionComputeDataType::BF16)
{
return run_profile_for_datatype(F32{}, BF16{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F64_F64_F64_F64)
{
if(compute_data_type == ContractionComputeDataType::F64)
{
return run_profile_for_datatype(F64{}, F64{});
}
else if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F64{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F16_F16_F16_F16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::BF16_BF16_BF16_BF16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(BF16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_bilinear);
|