| 1 |
// Copyright Advanced Micro Devices, Inc., or its affiliates. |
| 2 |
// SPDX-License-Identifier: MIT |
| 3 |
|
| 4 |
#include "testing_origami.hpp" |
| 5 |
#include <cctype> |
| 6 |
#include <fstream> |
| 7 |
#include <iostream> |
| 8 |
#include <vector> |
| 9 |
|
| 10 |
#ifndef _WIN32 |
| 11 |
#include <unistd.h> |
| 12 |
#include <limits.h> |
| 13 |
#else |
| 14 |
#include <windows.h> |
| 15 |
#endif |
| 16 |
|
| 17 |
// Returns the directory path where the current executable resides. |
| 18 |
// Adds a trailing slash ('/' on Linux, '\' on Windows) for easy file concatenation. |
| 19 |
std::string getExecutableDir() { |
| 20 |
#ifndef _WIN32 |
| 21 |
// Linux branch |
| 22 |
|
| 23 |
char result[PATH_MAX]; // Buffer to store the path |
| 24 |
ssize_t count = readlink("/proc/self/exe", result, PATH_MAX); |
| 25 |
// readlink reads the symbolic link /proc/self/exe, which points to the current executable |
| 26 |
|
| 27 |
if (count == -1) { |
| 28 |
// If readlink fails, return empty string |
| 29 |
return ""; |
| 30 |
} |
| 31 |
|
| 32 |
result[count] = '\0'; // Null-terminate the buffer |
| 33 |
std::string fullPath(result); // Convert to std::string |
| 34 |
|
| 35 |
// Find the position of the last slash ('/') in the path |
| 36 |
// This separates the directory from the binary name |
| 37 |
size_t pos = fullPath.find_last_of('/'); |
| 38 |
|
| 39 |
// Extract the directory portion |
| 40 |
std::string dir = (pos != std::string::npos) ? fullPath.substr(0, pos) : fullPath; |
| 41 |
|
| 42 |
// Ensure the directory string ends with a slash |
| 43 |
if (!dir.empty() && dir.back() != '/') |
| 44 |
dir += '/'; |
| 45 |
|
| 46 |
return dir; |
| 47 |
|
| 48 |
#else |
| 49 |
// Windows branch |
| 50 |
|
| 51 |
char path[MAX_PATH]; // Buffer to store the path |
| 52 |
DWORD length = GetModuleFileNameA(NULL, path, MAX_PATH); |
| 53 |
// GetModuleFileNameA returns the full path of the current executable |
| 54 |
|
| 55 |
if (length == 0) { |
| 56 |
// Failed to get the executable path |
| 57 |
return ""; |
| 58 |
} |
| 59 |
|
| 60 |
std::string fullPath(path, length); // Convert to std::string |
| 61 |
|
| 62 |
// Find the position of the last backslash ('\') or forward slash ('/') |
| 63 |
size_t pos = fullPath.find_last_of("\\/"); |
| 64 |
|
| 65 |
// Extract the directory portion |
| 66 |
std::string dir = (pos != std::string::npos) ? fullPath.substr(0, pos) : fullPath; |
| 67 |
|
| 68 |
// Ensure the directory string ends with a backslash |
| 69 |
if (!dir.empty() && dir.back() != '\\' && dir.back() != '/') |
| 70 |
dir += '\\'; |
| 71 |
|
| 72 |
return dir; |
| 73 |
#endif |
| 74 |
} |
| 75 |
|
| 76 |
// Parse origami_gtest.yaml to get the test data |
| 77 |
std::vector<MyTestData> parseYamlManually(const std::string& filename) |
| 78 |
{ |
| 79 |
std::string YamlfullPath = getExecutableDir() + filename; |
| 80 |
std::ifstream file(YamlfullPath); |
| 81 |
if(!file) |
| 82 |
{ |
| 83 |
std::cerr << "Failed to open file: " << YamlfullPath << std::endl; |
| 84 |
return {}; |
| 85 |
} |
| 86 |
|
| 87 |
std::string line; |
| 88 |
std::vector<MyTestData> tests; |
| 89 |
MyTestData current; |
| 90 |
enum class State |
| 91 |
{ |
| 92 |
None, |
| 93 |
Inputs |
| 94 |
} state |
| 95 |
= State::None; |
| 96 |
int line_number = 0; |
| 97 |
|
| 98 |
while(std::getline(file, line)) |
| 99 |
{ |
| 100 |
line_number++; |
| 101 |
line.erase(0, line.find_first_not_of(" \t\r\n")); |
| 102 |
line.erase(line.find_last_not_of(" \t\r\n") + 1); |
| 103 |
|
| 104 |
if(line.empty() || line[0] == '#') |
| 105 |
continue; |
| 106 |
|
| 107 |
if(line.rfind("- name:", 0) == 0) |
| 108 |
{ |
| 109 |
if(!current.name.empty()) |
| 110 |
tests.push_back(current); |
| 111 |
current = MyTestData{}; |
| 112 |
current.name = line.substr(7); |
| 113 |
current.name.erase(0, current.name.find_first_not_of(" \t")); |
| 114 |
state = State::None; |
| 115 |
} |
| 116 |
else if(line.rfind("inputs:", 0) == 0) |
| 117 |
{ |
| 118 |
state = State::Inputs; |
| 119 |
} |
| 120 |
else if(state == State::Inputs && line.rfind("- {", 0) == 0) |
| 121 |
{ |
| 122 |
std::string inner = line.substr(3); |
| 123 |
if(!inner.empty() && inner.back() == '}') |
| 124 |
inner.pop_back(); |
| 125 |
|
| 126 |
std::map<std::string, int> values; |
| 127 |
std::optional<int> expected; |
| 128 |
std::optional<int> expected_gt; |
| 129 |
std::optional<int> expected_lt; |
| 130 |
|
| 131 |
std::stringstream ss(inner); |
| 132 |
std::string pair; |
| 133 |
while(std::getline(ss, pair, ',')) |
| 134 |
{ |
| 135 |
auto colon = pair.find(':'); |
| 136 |
if(colon == std::string::npos) |
| 137 |
continue; |
| 138 |
|
| 139 |
std::string key = pair.substr(0, colon); |
| 140 |
std::string val = pair.substr(colon + 1); |
| 141 |
key.erase(0, key.find_first_not_of(" \t")); |
| 142 |
key.erase(key.find_last_not_of(" \t") + 1); |
| 143 |
val.erase(0, val.find_first_not_of(" \t")); |
| 144 |
val.erase(val.find_last_not_of(" \t") + 1); |
| 145 |
|
| 146 |
try |
| 147 |
{ |
| 148 |
int num = std::stoi(val); |
| 149 |
if(key == "expected") |
| 150 |
expected = num; |
| 151 |
else if(key == "expected_gt") |
| 152 |
expected_gt = num; |
| 153 |
else if(key == "expected_lt") |
| 154 |
expected_lt = num; |
| 155 |
else |
| 156 |
values[key] = num; |
| 157 |
} |
| 158 |
catch(...) |
| 159 |
{ |
| 160 |
std::cerr << "Invalid number in line " << line_number << ": " << val |
| 161 |
<< std::endl; |
| 162 |
} |
| 163 |
} |
| 164 |
|
| 165 |
current.inputs.push_back(InputWithExpected{values, expected, expected_gt, expected_lt}); |
| 166 |
} |
| 167 |
} |
| 168 |
|
| 169 |
if(!current.name.empty()) |
| 170 |
tests.push_back(current); |
| 171 |
|
| 172 |
return tests; |
| 173 |
} |
| 174 |
|
| 175 |
TEST_P(AnalyticalGtest, DynamicDispatch) |
| 176 |
{ |
| 177 |
const MyTestData& test = GetParam(); |
| 178 |
|
| 179 |
const std::string gpuArchNumber = std::to_string(test.inputs[0].values.at("gpu_arch")); |
| 180 |
auto gpuArchEnum = origami::hardware_t::arch_name_to_enum("gfx" + gpuArchNumber); |
| 181 |
|
| 182 |
//TODO: Hardcoding numbers for gfx942. Future archs could be added here with if else loop. |
| 183 |
auto gpuInfo = origami::hardware_t( |
| 184 |
gpuArchEnum, 304, 65536, 8, 1.0, 1.0, 1.0, 4000000, 1.0, 1, std::make_tuple(0, 0.015, 0)); |
| 185 |
|
| 186 |
if(test.name == "ComputeLoads") |
| 187 |
{ |
| 188 |
for(const auto& input_case : test.inputs) |
| 189 |
{ |
| 190 |
ComputeLoads(input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.expected); |
| 191 |
} |
| 192 |
} |
| 193 |
else if(test.name == "EstimateL2Hit") |
| 194 |
{ |
| 195 |
for(const auto& input_case : test.inputs) |
| 196 |
{ |
| 197 |
EstimateL2Hit(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 198 |
input_case.values.at("MT_M"), input_case.values.at("MT_N"), input_case.values.at("MT_K"), input_case.values.at("element_size"), |
| 199 |
input_case.values.at("splittingFactor"),input_case.expected_gt, input_case.expected_lt); |
| 200 |
} |
| 201 |
} |
| 202 |
else if(test.name == "ComputeNumMatrixInstructions") |
| 203 |
{ |
| 204 |
for(const auto& input_case : test.inputs) |
| 205 |
{ |
| 206 |
ComputeNumMatrixInstructions(gpuInfo, input_case.values.at("MT_M"), input_case.values.at("MT_N"), input_case.values.at("MT_K"), |
| 207 |
input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), input_case.expected); |
| 208 |
} |
| 209 |
} |
| 210 |
else if(test.name == "ComputeMTComputeLatency") |
| 211 |
{ |
| 212 |
for(const auto& input_case : test.inputs) |
| 213 |
{ |
| 214 |
ComputeMTComputeLatency(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), |
| 215 |
input_case.values.at("transA"), input_case.values.at("transB"),input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 216 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 217 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.expected, input_case.expected_gt); |
| 218 |
} |
| 219 |
} |
| 220 |
else if(test.name == "ComputeMemoryLatency") |
| 221 |
{ |
| 222 |
for(const auto& input_case : test.inputs) |
| 223 |
{ |
| 224 |
ComputeMemoryLatency(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 225 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 226 |
input_case.values.at("MT_K"), input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), |
| 227 |
input_case.values.at("mx_block_size"), input_case.values.at("wgm"), input_case.values.at("numActiveCUs"), input_case.values.at("splittingFactor")); |
| 228 |
} |
| 229 |
} |
| 230 |
else if(test.name == "ComputeTileLatency") |
| 231 |
{ |
| 232 |
for(const auto& input_case : test.inputs) |
| 233 |
{ |
| 234 |
ComputeTileLatency(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 235 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 236 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 237 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), |
| 238 |
input_case.values.at("mx_block_size"), input_case.values.at("wgm"), input_case.values.at("numActiveCUs"), input_case.values.at("splittingFactor")); |
| 239 |
} |
| 240 |
} |
| 241 |
else if(test.name == "ComputeWaveLatency") |
| 242 |
{ |
| 243 |
for(const auto& input_case : test.inputs) |
| 244 |
{ |
| 245 |
ComputeWaveLatency(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 246 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 247 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 248 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), |
| 249 |
input_case.values.at("mx_block_size"), input_case.values.at("wgm"), input_case.values.at("numActiveCUs"), input_case.values.at("splittingFactor")); |
| 250 |
} |
| 251 |
} |
| 252 |
else if(test.name == "ComputeTotalLatency") |
| 253 |
{ |
| 254 |
for(const auto& input_case : test.inputs) |
| 255 |
{ |
| 256 |
ComputeTotalLatency(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 257 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 258 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 259 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), |
| 260 |
input_case.values.at("mx_block_size"), input_case.values.at("wgm"), input_case.values.at("splittingFactor")); |
| 261 |
} |
| 262 |
} |
| 263 |
else if(test.name == "ComputePerfGflops") |
| 264 |
{ |
| 265 |
for(const auto& input_case : test.inputs) |
| 266 |
{ |
| 267 |
ComputePerfGflops(input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 268 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 269 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 270 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), |
| 271 |
input_case.values.at("WGM")); |
| 272 |
} |
| 273 |
} |
| 274 |
else if(test.name == "EstimateMallHit") |
| 275 |
{ |
| 276 |
for(const auto& input_case : test.inputs) |
| 277 |
{ |
| 278 |
EstimateMallHit(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 279 |
input_case.values.at("MT_M"), input_case.values.at("MT_N"), input_case.values.at("MT_K"), input_case.values.at("numActiveCUs"), |
| 280 |
input_case.values.at("splittingFactor"), input_case.expected_gt); |
| 281 |
} |
| 282 |
} |
| 283 |
else if(test.name == "CheckLDSCapacity") |
| 284 |
{ |
| 285 |
for(const auto& input_case : test.inputs) |
| 286 |
{ |
| 287 |
CheckLDSCapacity(gpuInfo, input_case.values.at("MT_M"), input_case.values.at("MT_N"), input_case.values.at("MT_K"), input_case.values.at("element_size")); |
| 288 |
} |
| 289 |
} |
| 290 |
else if(test.name == "HardwareArchEnum") |
| 291 |
{ |
| 292 |
for(const auto& input_case : test.inputs) |
| 293 |
{ |
| 294 |
HardwareArchEnum(std::to_string(test.inputs[0].values.at("gpu_arch"))); |
| 295 |
} |
| 296 |
} |
| 297 |
else if(test.name == "BestGridSize") |
| 298 |
{ |
| 299 |
for(const auto& input_case : test.inputs) |
| 300 |
{ |
| 301 |
BestGridSize(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 302 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("MT_M"), input_case.values.at("MT_N"), |
| 303 |
input_case.values.at("MT_K"), input_case.values.at("MI_M"), input_case.values.at("MI_N"), input_case.values.at("MI_K"), |
| 304 |
input_case.values.at("element_size_A"), input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), |
| 305 |
input_case.values.at("mx_block_size"), input_case.values.at("H_L2"), input_case.values.at("WGM"), |
| 306 |
input_case.values.at("biggest_allowable_split"), input_case.expected_gt); |
| 307 |
} |
| 308 |
} |
| 309 |
else if(test.name == "BestMacroTileSize") |
| 310 |
{ |
| 311 |
for(const auto& input_case : test.inputs) |
| 312 |
{ |
| 313 |
BestMacroTileSize(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 314 |
input_case.values.at("transA"), input_case.values.at("transB"), input_case.values.at("element_size_A"), |
| 315 |
input_case.values.at("element_size_B"), input_case.values.at("element_size_out"), input_case.values.at("mx_block_size"), |
| 316 |
input_case.values.at("H_L2"), input_case.values.at("WGM")); |
| 317 |
} |
| 318 |
} |
| 319 |
else if(test.name == "BestWGM") |
| 320 |
{ |
| 321 |
for(const auto& input_case : test.inputs) |
| 322 |
{ |
| 323 |
BestWGM(gpuInfo, input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), input_case.values.at("batch"), |
| 324 |
input_case.values.at("MT_M"), input_case.values.at("MT_N"), input_case.values.at("MT_K"), input_case.values.at("MI_M"), |
| 325 |
input_case.values.at("MI_N"), input_case.values.at("MI_K"), input_case.values.at("element_size"), input_case.values.at("H_L2")); |
| 326 |
} |
| 327 |
} |
| 328 |
else if(test.name == "UtilsTFlopsFromLatency") |
| 329 |
{ |
| 330 |
for(const auto& input_case : test.inputs) |
| 331 |
{ |
| 332 |
UtilsTFlopsFromLatency(input_case.values.at("M"), input_case.values.at("N"), input_case.values.at("K"), |
| 333 |
input_case.values.at("latency_cycles"), input_case.values.at("clock_GHz")); |
| 334 |
} |
| 335 |
} |
| 336 |
else |
| 337 |
{ |
| 338 |
FAIL() << "Unknown test name: " << test.name; |
| 339 |
} |
| 340 |
} |
| 341 |
|
| 342 |
// Instantiate tests using manual parser |
| 343 |
INSTANTIATE_TEST_SUITE_P(AnalyticalYamlTests, |
| 344 |
AnalyticalGtest, |
| 345 |
::testing::ValuesIn(parseYamlManually("origami_gtest.yaml")), |
| 346 |
[](const ::testing::TestParamInfo<MyTestData>& info) { |
| 347 |
std::string name = info.param.name; |
| 348 |
for(auto& c : name) |
| 349 |
if(!std::isalnum(c)) |
| 350 |
c = '_'; |
| 351 |
return name; |
| 352 |
}); |
| 353 |
|