| 1 |
// Copyright (c) 2015-2016 The Khronos Group Inc. |
| 2 |
// |
| 3 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 |
// you may not use this file except in compliance with the License. |
| 5 |
// You may obtain a copy of the License at |
| 6 |
// |
| 7 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 8 |
// |
| 9 |
// Unless required by applicable law or agreed to in writing, software |
| 10 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 11 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 |
// See the License for the specific language governing permissions and |
| 13 |
// limitations under the License. |
| 14 |
|
| 15 |
#include <cfloat> |
| 16 |
#include <cmath> |
| 17 |
#include <cstdio> |
| 18 |
#include <limits> |
| 19 |
#include <sstream> |
| 20 |
#include <string> |
| 21 |
#include <tuple> |
| 22 |
#include <utility> |
| 23 |
#include <vector> |
| 24 |
|
| 25 |
#include "gmock/gmock.h" |
| 26 |
#include "source/util/hex_float.h" |
| 27 |
#include "test/unit_spirv.h" |
| 28 |
|
| 29 |
namespace spvtools { |
| 30 |
namespace utils { |
| 31 |
namespace { |
| 32 |
|
| 33 |
using ::testing::Eq; |
| 34 |
|
| 35 |
// In this file "encode" means converting a number into a string, |
| 36 |
// and "decode" means converting a string into a number. |
| 37 |
|
| 38 |
using HexFloatTest = |
| 39 |
::testing::TestWithParam<std::pair<FloatProxy<float>, std::string>>; |
| 40 |
using DecodeHexFloatTest = |
| 41 |
::testing::TestWithParam<std::pair<std::string, FloatProxy<float>>>; |
| 42 |
using HexDoubleTest = |
| 43 |
::testing::TestWithParam<std::pair<FloatProxy<double>, std::string>>; |
| 44 |
using DecodeHexDoubleTest = |
| 45 |
::testing::TestWithParam<std::pair<std::string, FloatProxy<double>>>; |
| 46 |
using RoundTripFloatTest = ::testing::TestWithParam<float>; |
| 47 |
using RoundTripDoubleTest = ::testing::TestWithParam<double>; |
| 48 |
|
| 49 |
// Hex-encodes a float value. |
| 50 |
template <typename T> |
| 51 |
std::string EncodeViaHexFloat(const T& value) { |
| 52 |
std::stringstream ss; |
| 53 |
ss << HexFloat<T>(value); |
| 54 |
return ss.str(); |
| 55 |
} |
| 56 |
|
| 57 |
// The following two tests can't be DRY because they take different parameter |
| 58 |
// types. |
| 59 |
|
| 60 |
TEST_P(HexFloatTest, EncodeCorrectly) { |
| 61 |
EXPECT_THAT(EncodeViaHexFloat(GetParam().first), Eq(GetParam().second)); |
| 62 |
} |
| 63 |
|
| 64 |
TEST_P(HexDoubleTest, EncodeCorrectly) { |
| 65 |
EXPECT_THAT(EncodeViaHexFloat(GetParam().first), Eq(GetParam().second)); |
| 66 |
} |
| 67 |
|
| 68 |
// Decodes a hex-float string. |
| 69 |
template <typename T> |
| 70 |
FloatProxy<T> Decode(const std::string& str) { |
| 71 |
HexFloat<FloatProxy<T>> decoded(0.f); |
| 72 |
EXPECT_TRUE((std::stringstream(str) >> decoded).eof()); |
| 73 |
return decoded.value(); |
| 74 |
} |
| 75 |
|
| 76 |
TEST_P(HexFloatTest, DecodeCorrectly) { |
| 77 |
EXPECT_THAT(Decode<float>(GetParam().second), Eq(GetParam().first)); |
| 78 |
} |
| 79 |
|
| 80 |
TEST_P(HexDoubleTest, DecodeCorrectly) { |
| 81 |
EXPECT_THAT(Decode<double>(GetParam().second), Eq(GetParam().first)); |
| 82 |
} |
| 83 |
|
| 84 |
INSTANTIATE_TEST_SUITE_P( |
| 85 |
Float32Tests, HexFloatTest, |
| 86 |
::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({ |
| 87 |
{0.f, "0x0p+0"}, |
| 88 |
{1.f, "0x1p+0"}, |
| 89 |
{2.f, "0x1p+1"}, |
| 90 |
{3.f, "0x1.8p+1"}, |
| 91 |
{0.5f, "0x1p-1"}, |
| 92 |
{0.25f, "0x1p-2"}, |
| 93 |
{0.75f, "0x1.8p-1"}, |
| 94 |
{-0.f, "-0x0p+0"}, |
| 95 |
{-1.f, "-0x1p+0"}, |
| 96 |
{-0.5f, "-0x1p-1"}, |
| 97 |
{-0.25f, "-0x1p-2"}, |
| 98 |
{-0.75f, "-0x1.8p-1"}, |
| 99 |
|
| 100 |
// Larger numbers |
| 101 |
{512.f, "0x1p+9"}, |
| 102 |
{-512.f, "-0x1p+9"}, |
| 103 |
{1024.f, "0x1p+10"}, |
| 104 |
{-1024.f, "-0x1p+10"}, |
| 105 |
{1024.f + 8.f, "0x1.02p+10"}, |
| 106 |
{-1024.f - 8.f, "-0x1.02p+10"}, |
| 107 |
|
| 108 |
// Small numbers |
| 109 |
{1.0f / 512.f, "0x1p-9"}, |
| 110 |
{1.0f / -512.f, "-0x1p-9"}, |
| 111 |
{1.0f / 1024.f, "0x1p-10"}, |
| 112 |
{1.0f / -1024.f, "-0x1p-10"}, |
| 113 |
{1.0f / 1024.f + 1.0f / 8.f, "0x1.02p-3"}, |
| 114 |
{1.0f / -1024.f - 1.0f / 8.f, "-0x1.02p-3"}, |
| 115 |
|
| 116 |
// lowest non-denorm |
| 117 |
{float(ldexp(1.0f, -126)), "0x1p-126"}, |
| 118 |
{float(ldexp(-1.0f, -126)), "-0x1p-126"}, |
| 119 |
|
| 120 |
// Denormalized values |
| 121 |
{float(ldexp(1.0f, -127)), "0x1p-127"}, |
| 122 |
{float(ldexp(1.0f, -127) / 2.0f), "0x1p-128"}, |
| 123 |
{float(ldexp(1.0f, -127) / 4.0f), "0x1p-129"}, |
| 124 |
{float(ldexp(1.0f, -127) / 8.0f), "0x1p-130"}, |
| 125 |
{float(ldexp(-1.0f, -127)), "-0x1p-127"}, |
| 126 |
{float(ldexp(-1.0f, -127) / 2.0f), "-0x1p-128"}, |
| 127 |
{float(ldexp(-1.0f, -127) / 4.0f), "-0x1p-129"}, |
| 128 |
{float(ldexp(-1.0f, -127) / 8.0f), "-0x1p-130"}, |
| 129 |
|
| 130 |
{float(ldexp(1.0, -127) + (ldexp(1.0, -127) / 2.0f)), "0x1.8p-127"}, |
| 131 |
{float(ldexp(1.0, -127) / 2.0 + (ldexp(1.0, -127) / 4.0f)), |
| 132 |
"0x1.8p-128"}, |
| 133 |
|
| 134 |
}))); |
| 135 |
|
| 136 |
INSTANTIATE_TEST_SUITE_P( |
| 137 |
Float32NanTests, HexFloatTest, |
| 138 |
::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({ |
| 139 |
// Various NAN and INF cases |
| 140 |
{uint32_t(0xFF800000), "-0x1p+128"}, // -inf |
| 141 |
{uint32_t(0x7F800000), "0x1p+128"}, // inf |
| 142 |
{uint32_t(0xFFC00000), "-0x1.8p+128"}, // -nan |
| 143 |
{uint32_t(0xFF800100), "-0x1.0002p+128"}, // -nan |
| 144 |
{uint32_t(0xFF800c00), "-0x1.0018p+128"}, // -nan |
| 145 |
{uint32_t(0xFF80F000), "-0x1.01ep+128"}, // -nan |
| 146 |
{uint32_t(0xFFFFFFFF), "-0x1.fffffep+128"}, // -nan |
| 147 |
{uint32_t(0x7FC00000), "0x1.8p+128"}, // +nan |
| 148 |
{uint32_t(0x7F800100), "0x1.0002p+128"}, // +nan |
| 149 |
{uint32_t(0x7f800c00), "0x1.0018p+128"}, // +nan |
| 150 |
{uint32_t(0x7F80F000), "0x1.01ep+128"}, // +nan |
| 151 |
{uint32_t(0x7FFFFFFF), "0x1.fffffep+128"}, // +nan |
| 152 |
}))); |
| 153 |
|
| 154 |
INSTANTIATE_TEST_SUITE_P( |
| 155 |
Float64Tests, HexDoubleTest, |
| 156 |
::testing::ValuesIn( |
| 157 |
std::vector<std::pair<FloatProxy<double>, std::string>>({ |
| 158 |
{0., "0x0p+0"}, |
| 159 |
{1., "0x1p+0"}, |
| 160 |
{2., "0x1p+1"}, |
| 161 |
{3., "0x1.8p+1"}, |
| 162 |
{0.5, "0x1p-1"}, |
| 163 |
{0.25, "0x1p-2"}, |
| 164 |
{0.75, "0x1.8p-1"}, |
| 165 |
{-0., "-0x0p+0"}, |
| 166 |
{-1., "-0x1p+0"}, |
| 167 |
{-0.5, "-0x1p-1"}, |
| 168 |
{-0.25, "-0x1p-2"}, |
| 169 |
{-0.75, "-0x1.8p-1"}, |
| 170 |
|
| 171 |
// Larger numbers |
| 172 |
{512., "0x1p+9"}, |
| 173 |
{-512., "-0x1p+9"}, |
| 174 |
{1024., "0x1p+10"}, |
| 175 |
{-1024., "-0x1p+10"}, |
| 176 |
{1024. + 8., "0x1.02p+10"}, |
| 177 |
{-1024. - 8., "-0x1.02p+10"}, |
| 178 |
|
| 179 |
// Large outside the range of normal floats |
| 180 |
{ldexp(1.0, 128), "0x1p+128"}, |
| 181 |
{ldexp(1.0, 129), "0x1p+129"}, |
| 182 |
{ldexp(-1.0, 128), "-0x1p+128"}, |
| 183 |
{ldexp(-1.0, 129), "-0x1p+129"}, |
| 184 |
{ldexp(1.0, 128) + ldexp(1.0, 90), "0x1.0000000004p+128"}, |
| 185 |
{ldexp(1.0, 129) + ldexp(1.0, 120), "0x1.008p+129"}, |
| 186 |
{ldexp(-1.0, 128) + ldexp(1.0, 90), "-0x1.fffffffff8p+127"}, |
| 187 |
{ldexp(-1.0, 129) + ldexp(1.0, 120), "-0x1.ffp+128"}, |
| 188 |
|
| 189 |
// Small numbers |
| 190 |
{1.0 / 512., "0x1p-9"}, |
| 191 |
{1.0 / -512., "-0x1p-9"}, |
| 192 |
{1.0 / 1024., "0x1p-10"}, |
| 193 |
{1.0 / -1024., "-0x1p-10"}, |
| 194 |
{1.0 / 1024. + 1.0 / 8., "0x1.02p-3"}, |
| 195 |
{1.0 / -1024. - 1.0 / 8., "-0x1.02p-3"}, |
| 196 |
|
| 197 |
// Small outside the range of normal floats |
| 198 |
{ldexp(1.0, -128), "0x1p-128"}, |
| 199 |
{ldexp(1.0, -129), "0x1p-129"}, |
| 200 |
{ldexp(-1.0, -128), "-0x1p-128"}, |
| 201 |
{ldexp(-1.0, -129), "-0x1p-129"}, |
| 202 |
{ldexp(1.0, -128) + ldexp(1.0, -90), "0x1.0000000004p-90"}, |
| 203 |
{ldexp(1.0, -129) + ldexp(1.0, -120), "0x1.008p-120"}, |
| 204 |
{ldexp(-1.0, -128) + ldexp(1.0, -90), "0x1.fffffffff8p-91"}, |
| 205 |
{ldexp(-1.0, -129) + ldexp(1.0, -120), "0x1.ffp-121"}, |
| 206 |
|
| 207 |
// lowest non-denorm |
| 208 |
{ldexp(1.0, -1022), "0x1p-1022"}, |
| 209 |
{ldexp(-1.0, -1022), "-0x1p-1022"}, |
| 210 |
|
| 211 |
// Denormalized values |
| 212 |
{ldexp(1.0, -1023), "0x1p-1023"}, |
| 213 |
{ldexp(1.0, -1023) / 2.0, "0x1p-1024"}, |
| 214 |
{ldexp(1.0, -1023) / 4.0, "0x1p-1025"}, |
| 215 |
{ldexp(1.0, -1023) / 8.0, "0x1p-1026"}, |
| 216 |
{ldexp(-1.0, -1024), "-0x1p-1024"}, |
| 217 |
{ldexp(-1.0, -1024) / 2.0, "-0x1p-1025"}, |
| 218 |
{ldexp(-1.0, -1024) / 4.0, "-0x1p-1026"}, |
| 219 |
{ldexp(-1.0, -1024) / 8.0, "-0x1p-1027"}, |
| 220 |
|
| 221 |
{ldexp(1.0, -1023) + (ldexp(1.0, -1023) / 2.0), "0x1.8p-1023"}, |
| 222 |
{ldexp(1.0, -1023) / 2.0 + (ldexp(1.0, -1023) / 4.0), |
| 223 |
"0x1.8p-1024"}, |
| 224 |
|
| 225 |
}))); |
| 226 |
|
| 227 |
INSTANTIATE_TEST_SUITE_P( |
| 228 |
Float64NanTests, HexDoubleTest, |
| 229 |
::testing::ValuesIn(std::vector< |
| 230 |
std::pair<FloatProxy<double>, std::string>>({ |
| 231 |
// Various NAN and INF cases |
| 232 |
{uint64_t(0xFFF0000000000000LL), "-0x1p+1024"}, // -inf |
| 233 |
{uint64_t(0x7FF0000000000000LL), "0x1p+1024"}, // +inf |
| 234 |
{uint64_t(0xFFF8000000000000LL), "-0x1.8p+1024"}, // -nan |
| 235 |
{uint64_t(0xFFF0F00000000000LL), "-0x1.0fp+1024"}, // -nan |
| 236 |
{uint64_t(0xFFF0000000000001LL), "-0x1.0000000000001p+1024"}, // -nan |
| 237 |
{uint64_t(0xFFF0000300000000LL), "-0x1.00003p+1024"}, // -nan |
| 238 |
{uint64_t(0xFFFFFFFFFFFFFFFFLL), "-0x1.fffffffffffffp+1024"}, // -nan |
| 239 |
{uint64_t(0x7FF8000000000000LL), "0x1.8p+1024"}, // +nan |
| 240 |
{uint64_t(0x7FF0F00000000000LL), "0x1.0fp+1024"}, // +nan |
| 241 |
{uint64_t(0x7FF0000000000001LL), "0x1.0000000000001p+1024"}, // -nan |
| 242 |
{uint64_t(0x7FF0000300000000LL), "0x1.00003p+1024"}, // -nan |
| 243 |
{uint64_t(0x7FFFFFFFFFFFFFFFLL), "0x1.fffffffffffffp+1024"}, // -nan |
| 244 |
}))); |
| 245 |
|
| 246 |
// Tests that encoding a value and decoding it again restores |
| 247 |
// the same value. |
| 248 |
TEST_P(RoundTripFloatTest, CanStoreAccurately) { |
| 249 |
std::stringstream ss; |
| 250 |
ss << FloatProxy<float>(GetParam()); |
| 251 |
ss.seekg(0); |
| 252 |
FloatProxy<float> res; |
| 253 |
ss >> res; |
| 254 |
EXPECT_THAT(GetParam(), Eq(res.getAsFloat())); |
| 255 |
} |
| 256 |
|
| 257 |
TEST_P(RoundTripDoubleTest, CanStoreAccurately) { |
| 258 |
std::stringstream ss; |
| 259 |
ss << FloatProxy<double>(GetParam()); |
| 260 |
ss.seekg(0); |
| 261 |
FloatProxy<double> res; |
| 262 |
ss >> res; |
| 263 |
EXPECT_THAT(GetParam(), Eq(res.getAsFloat())); |
| 264 |
} |
| 265 |
|
| 266 |
INSTANTIATE_TEST_SUITE_P( |
| 267 |
Float32StoreTests, RoundTripFloatTest, |
| 268 |
::testing::ValuesIn(std::vector<float>( |
| 269 |
{// Value requiring more than 6 digits of precision to be |
| 270 |
// represented accurately. |
| 271 |
3.0000002f}))); |
| 272 |
|
| 273 |
INSTANTIATE_TEST_SUITE_P( |
| 274 |
Float64StoreTests, RoundTripDoubleTest, |
| 275 |
::testing::ValuesIn(std::vector<double>( |
| 276 |
{// Value requiring more than 15 digits of precision to be |
| 277 |
// represented accurately. |
| 278 |
1.5000000000000002}))); |
| 279 |
|
| 280 |
TEST(HexFloatStreamTest, OperatorLeftShiftPreservesFloatAndFill) { |
| 281 |
std::stringstream s; |
| 282 |
s << std::setw(4) << std::oct << std::setfill('x') << 8 << " " |
| 283 |
<< FloatProxy<float>(uint32_t(0xFF800100)) << " " << std::setw(4) << 9; |
| 284 |
EXPECT_THAT(s.str(), Eq(std::string("xx10 -0x1.0002p+128 xx11"))); |
| 285 |
} |
| 286 |
|
| 287 |
TEST(HexDoubleStreamTest, OperatorLeftShiftPreservesFloatAndFill) { |
| 288 |
std::stringstream s; |
| 289 |
s << std::setw(4) << std::oct << std::setfill('x') << 8 << " " |
| 290 |
<< FloatProxy<double>(uint64_t(0x7FF0F00000000000LL)) << " " << std::setw(4) |
| 291 |
<< 9; |
| 292 |
EXPECT_THAT(s.str(), Eq(std::string("xx10 0x1.0fp+1024 xx11"))); |
| 293 |
} |
| 294 |
|
| 295 |
TEST_P(DecodeHexFloatTest, DecodeCorrectly) { |
| 296 |
EXPECT_THAT(Decode<float>(GetParam().first), Eq(GetParam().second)); |
| 297 |
} |
| 298 |
|
| 299 |
TEST_P(DecodeHexDoubleTest, DecodeCorrectly) { |
| 300 |
EXPECT_THAT(Decode<double>(GetParam().first), Eq(GetParam().second)); |
| 301 |
} |
| 302 |
|
| 303 |
INSTANTIATE_TEST_SUITE_P( |
| 304 |
Float32DecodeTests, DecodeHexFloatTest, |
| 305 |
::testing::ValuesIn(std::vector<std::pair<std::string, FloatProxy<float>>>({ |
| 306 |
{"0x0p+000", 0.f}, |
| 307 |
{"0x0p0", 0.f}, |
| 308 |
{"0x0p-0", 0.f}, |
| 309 |
|
| 310 |
// flush to zero cases |
| 311 |
{"0x1p-500", 0.f}, // Exponent underflows. |
| 312 |
{"-0x1p-500", -0.f}, |
| 313 |
{"0x0.00000000001p-126", 0.f}, // Fraction causes underflow. |
| 314 |
{"-0x0.0000000001p-127", -0.f}, |
| 315 |
{"-0x0.01p-142", -0.f}, // Fraction causes additional underflow. |
| 316 |
{"0x0.01p-142", 0.f}, |
| 317 |
|
| 318 |
// Some floats that do not encode the same way as they decode. |
| 319 |
{"0x2p+0", 2.f}, |
| 320 |
{"0xFFp+0", 255.f}, |
| 321 |
{"0x0.8p+0", 0.5f}, |
| 322 |
{"0x0.4p+0", 0.25f}, |
| 323 |
}))); |
| 324 |
|
| 325 |
INSTANTIATE_TEST_SUITE_P( |
| 326 |
Float32DecodeInfTests, DecodeHexFloatTest, |
| 327 |
::testing::ValuesIn(std::vector<std::pair<std::string, FloatProxy<float>>>({ |
| 328 |
// inf cases |
| 329 |
{"-0x1p+128", uint32_t(0xFF800000)}, // -inf |
| 330 |
{"0x32p+127", uint32_t(0x7F800000)}, // inf |
| 331 |
{"0x32p+500", uint32_t(0x7F800000)}, // inf |
| 332 |
{"-0x32p+127", uint32_t(0xFF800000)}, // -inf |
| 333 |
}))); |
| 334 |
|
| 335 |
INSTANTIATE_TEST_SUITE_P( |
| 336 |
Float64DecodeTests, DecodeHexDoubleTest, |
| 337 |
::testing::ValuesIn( |
| 338 |
std::vector<std::pair<std::string, FloatProxy<double>>>({ |
| 339 |
{"0x0p+000", 0.}, |
| 340 |
{"0x0p0", 0.}, |
| 341 |
{"0x0p-0", 0.}, |
| 342 |
|
| 343 |
// flush to zero cases |
| 344 |
{"0x1p-5000", 0.}, // Exponent underflows. |
| 345 |
{"-0x1p-5000", -0.}, |
| 346 |
{"0x0.0000000000000001p-1023", 0.}, // Fraction causes underflow. |
| 347 |
{"-0x0.000000000000001p-1024", -0.}, |
| 348 |
{"-0x0.01p-1090", -0.f}, // Fraction causes additional underflow. |
| 349 |
{"0x0.01p-1090", 0.}, |
| 350 |
|
| 351 |
// Some floats that do not encode the same way as they decode. |
| 352 |
{"0x2p+0", 2.}, |
| 353 |
{"0xFFp+0", 255.}, |
| 354 |
{"0x0.8p+0", 0.5}, |
| 355 |
{"0x0.4p+0", 0.25}, |
| 356 |
}))); |
| 357 |
|
| 358 |
INSTANTIATE_TEST_SUITE_P( |
| 359 |
Float64DecodeInfTests, DecodeHexDoubleTest, |
| 360 |
::testing::ValuesIn( |
| 361 |
std::vector<std::pair<std::string, FloatProxy<double>>>({ |
| 362 |
// inf cases |
| 363 |
{"-0x1p+1024", uint64_t(0xFFF0000000000000)}, // -inf |
| 364 |
{"0x32p+1023", uint64_t(0x7FF0000000000000)}, // inf |
| 365 |
{"0x32p+5000", uint64_t(0x7FF0000000000000)}, // inf |
| 366 |
{"-0x32p+1023", uint64_t(0xFFF0000000000000)}, // -inf |
| 367 |
}))); |
| 368 |
|
| 369 |
TEST(FloatProxy, ValidConversion) { |
| 370 |
EXPECT_THAT(FloatProxy<float>(1.f).getAsFloat(), Eq(1.0f)); |
| 371 |
EXPECT_THAT(FloatProxy<float>(32.f).getAsFloat(), Eq(32.0f)); |
| 372 |
EXPECT_THAT(FloatProxy<float>(-1.f).getAsFloat(), Eq(-1.0f)); |
| 373 |
EXPECT_THAT(FloatProxy<float>(0.f).getAsFloat(), Eq(0.0f)); |
| 374 |
EXPECT_THAT(FloatProxy<float>(-0.f).getAsFloat(), Eq(-0.0f)); |
| 375 |
EXPECT_THAT(FloatProxy<float>(1.2e32f).getAsFloat(), Eq(1.2e32f)); |
| 376 |
|
| 377 |
EXPECT_TRUE(std::isinf(FloatProxy<float>(uint32_t(0xFF800000)).getAsFloat())); |
| 378 |
EXPECT_TRUE(std::isinf(FloatProxy<float>(uint32_t(0x7F800000)).getAsFloat())); |
| 379 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFFC00000)).getAsFloat())); |
| 380 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF800100)).getAsFloat())); |
| 381 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF800c00)).getAsFloat())); |
| 382 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF80F000)).getAsFloat())); |
| 383 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFFFFFFFF)).getAsFloat())); |
| 384 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7FC00000)).getAsFloat())); |
| 385 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7F800100)).getAsFloat())); |
| 386 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7f800c00)).getAsFloat())); |
| 387 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7F80F000)).getAsFloat())); |
| 388 |
EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7FFFFFFF)).getAsFloat())); |
| 389 |
|
| 390 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800000)).data(), Eq(0xFF800000u)); |
| 391 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F800000)).data(), Eq(0x7F800000u)); |
| 392 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFFC00000)).data(), Eq(0xFFC00000u)); |
| 393 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800100)).data(), Eq(0xFF800100u)); |
| 394 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800c00)).data(), Eq(0xFF800c00u)); |
| 395 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF80F000)).data(), Eq(0xFF80F000u)); |
| 396 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0xFFFFFFFF)).data(), Eq(0xFFFFFFFFu)); |
| 397 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7FC00000)).data(), Eq(0x7FC00000u)); |
| 398 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F800100)).data(), Eq(0x7F800100u)); |
| 399 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7f800c00)).data(), Eq(0x7f800c00u)); |
| 400 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F80F000)).data(), Eq(0x7F80F000u)); |
| 401 |
EXPECT_THAT(FloatProxy<float>(uint32_t(0x7FFFFFFF)).data(), Eq(0x7FFFFFFFu)); |
| 402 |
} |
| 403 |
|
| 404 |
TEST(FloatProxy, Nan) { |
| 405 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFFC00000)).isNan()); |
| 406 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF800100)).isNan()); |
| 407 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF800c00)).isNan()); |
| 408 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF80F000)).isNan()); |
| 409 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFFFFFFFF)).isNan()); |
| 410 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7FC00000)).isNan()); |
| 411 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7F800100)).isNan()); |
| 412 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7f800c00)).isNan()); |
| 413 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7F80F000)).isNan()); |
| 414 |
EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7FFFFFFF)).isNan()); |
| 415 |
} |
| 416 |
|
| 417 |
TEST(FloatProxy, Negation) { |
| 418 |
EXPECT_THAT((-FloatProxy<float>(1.f)).getAsFloat(), Eq(-1.0f)); |
| 419 |
EXPECT_THAT((-FloatProxy<float>(0.f)).getAsFloat(), Eq(-0.0f)); |
| 420 |
|
| 421 |
EXPECT_THAT((-FloatProxy<float>(-1.f)).getAsFloat(), Eq(1.0f)); |
| 422 |
EXPECT_THAT((-FloatProxy<float>(-0.f)).getAsFloat(), Eq(0.0f)); |
| 423 |
|
| 424 |
EXPECT_THAT((-FloatProxy<float>(32.f)).getAsFloat(), Eq(-32.0f)); |
| 425 |
EXPECT_THAT((-FloatProxy<float>(-32.f)).getAsFloat(), Eq(32.0f)); |
| 426 |
|
| 427 |
EXPECT_THAT((-FloatProxy<float>(1.2e32f)).getAsFloat(), Eq(-1.2e32f)); |
| 428 |
EXPECT_THAT((-FloatProxy<float>(-1.2e32f)).getAsFloat(), Eq(1.2e32f)); |
| 429 |
|
| 430 |
EXPECT_THAT( |
| 431 |
(-FloatProxy<float>(std::numeric_limits<float>::infinity())).getAsFloat(), |
| 432 |
Eq(-std::numeric_limits<float>::infinity())); |
| 433 |
EXPECT_THAT((-FloatProxy<float>(-std::numeric_limits<float>::infinity())) |
| 434 |
.getAsFloat(), |
| 435 |
Eq(std::numeric_limits<float>::infinity())); |
| 436 |
} |
| 437 |
|
| 438 |
// Test conversion of FloatProxy values to strings. |
| 439 |
// |
| 440 |
// In previous cases, we always wrapped the FloatProxy value in a HexFloat |
| 441 |
// before conversion to a string. In the following cases, the FloatProxy |
| 442 |
// decides for itself whether to print as a regular number or as a hex float. |
| 443 |
|
| 444 |
using FloatProxyFloatTest = |
| 445 |
::testing::TestWithParam<std::pair<FloatProxy<float>, std::string>>; |
| 446 |
using FloatProxyDoubleTest = |
| 447 |
::testing::TestWithParam<std::pair<FloatProxy<double>, std::string>>; |
| 448 |
|
| 449 |
// Converts a float value to a string via a FloatProxy. |
| 450 |
template <typename T> |
| 451 |
std::string EncodeViaFloatProxy(const T& value) { |
| 452 |
std::stringstream ss; |
| 453 |
ss << value; |
| 454 |
return ss.str(); |
| 455 |
} |
| 456 |
|
| 457 |
// Converts a floating point string so that the exponent prefix |
| 458 |
// is 'e', and the exponent value does not have leading zeros. |
| 459 |
// The Microsoft runtime library likes to write things like "2.5E+010". |
| 460 |
// Convert that to "2.5e+10". |
| 461 |
// We don't care what happens to strings that are not floating point |
| 462 |
// strings. |
| 463 |
std::string NormalizeExponentInFloatString(std::string in) { |
| 464 |
std::string result; |
| 465 |
// Reserve one spot for the terminating null, even when the sscanf fails. |
| 466 |
std::vector<char> prefix(in.size() + 1); |
| 467 |
char e; |
| 468 |
char plus_or_minus; |
| 469 |
int exponent; // in base 10 |
| 470 |
if ((4 == std::sscanf(in.c_str(), "%[-+.0123456789]%c%c%d", prefix.data(), &e, |
| 471 |
&plus_or_minus, &exponent)) && |
| 472 |
(e == 'e' || e == 'E') && |
| 473 |
(plus_or_minus == '-' || plus_or_minus == '+')) { |
| 474 |
// It looks like a floating point value with exponent. |
| 475 |
std::stringstream out; |
| 476 |
out << prefix.data() << 'e' << plus_or_minus << exponent; |
| 477 |
result = out.str(); |
| 478 |
} else { |
| 479 |
result = in; |
| 480 |
} |
| 481 |
return result; |
| 482 |
} |
| 483 |
|
| 484 |
TEST(NormalizeFloat, Sample) { |
| 485 |
EXPECT_THAT(NormalizeExponentInFloatString(""), Eq("")); |
| 486 |
EXPECT_THAT(NormalizeExponentInFloatString("1e-12"), Eq("1e-12")); |
| 487 |
EXPECT_THAT(NormalizeExponentInFloatString("1E+14"), Eq("1e+14")); |
| 488 |
EXPECT_THAT(NormalizeExponentInFloatString("1e-0012"), Eq("1e-12")); |
| 489 |
EXPECT_THAT(NormalizeExponentInFloatString("1.263E+014"), Eq("1.263e+14")); |
| 490 |
} |
| 491 |
|
| 492 |
// The following two tests can't be DRY because they take different parameter |
| 493 |
// types. |
| 494 |
TEST_P(FloatProxyFloatTest, EncodeCorrectly) { |
| 495 |
EXPECT_THAT( |
| 496 |
NormalizeExponentInFloatString(EncodeViaFloatProxy(GetParam().first)), |
| 497 |
Eq(GetParam().second)); |
| 498 |
} |
| 499 |
|
| 500 |
TEST_P(FloatProxyDoubleTest, EncodeCorrectly) { |
| 501 |
EXPECT_THAT( |
| 502 |
NormalizeExponentInFloatString(EncodeViaFloatProxy(GetParam().first)), |
| 503 |
Eq(GetParam().second)); |
| 504 |
} |
| 505 |
|
| 506 |
INSTANTIATE_TEST_SUITE_P( |
| 507 |
Float32Tests, FloatProxyFloatTest, |
| 508 |
::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({ |
| 509 |
// Zero |
| 510 |
{0.f, "0"}, |
| 511 |
// Normal numbers |
| 512 |
{1.f, "1"}, |
| 513 |
{-0.25f, "-0.25"}, |
| 514 |
{1000.0f, "1000"}, |
| 515 |
|
| 516 |
// Still normal numbers, but with large magnitude exponents. |
| 517 |
{float(ldexp(1.f, 126)), "8.50705917e+37"}, |
| 518 |
{float(ldexp(-1.f, -126)), "-1.17549435e-38"}, |
| 519 |
|
| 520 |
// denormalized values are printed as hex floats. |
| 521 |
{float(ldexp(1.0f, -127)), "0x1p-127"}, |
| 522 |
{float(ldexp(1.5f, -128)), "0x1.8p-128"}, |
| 523 |
{float(ldexp(1.25, -129)), "0x1.4p-129"}, |
| 524 |
{float(ldexp(1.125, -130)), "0x1.2p-130"}, |
| 525 |
{float(ldexp(-1.0f, -127)), "-0x1p-127"}, |
| 526 |
{float(ldexp(-1.0f, -128)), "-0x1p-128"}, |
| 527 |
{float(ldexp(-1.0f, -129)), "-0x1p-129"}, |
| 528 |
{float(ldexp(-1.5f, -130)), "-0x1.8p-130"}, |
| 529 |
|
| 530 |
// NaNs |
| 531 |
{FloatProxy<float>(uint32_t(0xFFC00000)), "-0x1.8p+128"}, |
| 532 |
{FloatProxy<float>(uint32_t(0xFF800100)), "-0x1.0002p+128"}, |
| 533 |
|
| 534 |
{std::numeric_limits<float>::infinity(), "0x1p+128"}, |
| 535 |
{-std::numeric_limits<float>::infinity(), "-0x1p+128"}, |
| 536 |
}))); |
| 537 |
|
| 538 |
INSTANTIATE_TEST_SUITE_P( |
| 539 |
Float64Tests, FloatProxyDoubleTest, |
| 540 |
::testing::ValuesIn( |
| 541 |
std::vector<std::pair<FloatProxy<double>, std::string>>({ |
| 542 |
{0., "0"}, |
| 543 |
{1., "1"}, |
| 544 |
{-0.25, "-0.25"}, |
| 545 |
{1000.0, "1000"}, |
| 546 |
|
| 547 |
// Large outside the range of normal floats |
| 548 |
{ldexp(1.0, 128), "3.4028236692093846e+38"}, |
| 549 |
{ldexp(1.5, 129), "1.0208471007628154e+39"}, |
| 550 |
{ldexp(-1.0, 128), "-3.4028236692093846e+38"}, |
| 551 |
{ldexp(-1.5, 129), "-1.0208471007628154e+39"}, |
| 552 |
|
| 553 |
// Small outside the range of normal floats |
| 554 |
{ldexp(1.5, -129), "2.2040519077917891e-39"}, |
| 555 |
{ldexp(-1.5, -129), "-2.2040519077917891e-39"}, |
| 556 |
|
| 557 |
// lowest non-denorm |
| 558 |
{ldexp(1.0, -1022), "2.2250738585072014e-308"}, |
| 559 |
{ldexp(-1.0, -1022), "-2.2250738585072014e-308"}, |
| 560 |
|
| 561 |
// Denormalized values |
| 562 |
{ldexp(1.125, -1023), "0x1.2p-1023"}, |
| 563 |
{ldexp(-1.375, -1024), "-0x1.6p-1024"}, |
| 564 |
|
| 565 |
// NaNs |
| 566 |
{uint64_t(0x7FF8000000000000LL), "0x1.8p+1024"}, |
| 567 |
{uint64_t(0xFFF0F00000000000LL), "-0x1.0fp+1024"}, |
| 568 |
|
| 569 |
// Infinity |
| 570 |
{std::numeric_limits<double>::infinity(), "0x1p+1024"}, |
| 571 |
{-std::numeric_limits<double>::infinity(), "-0x1p+1024"}, |
| 572 |
|
| 573 |
}))); |
| 574 |
|
| 575 |
// double is used so that unbiased_exponent can be used with the output |
| 576 |
// of ldexp directly. |
| 577 |
int32_t unbiased_exponent(double f) { |
| 578 |
return HexFloat<FloatProxy<float>>(static_cast<float>(f)) |
| 579 |
.getUnbiasedNormalizedExponent(); |
| 580 |
} |
| 581 |
|
| 582 |
int16_t unbiased_half_exponent(uint16_t f) { |
| 583 |
return HexFloat<FloatProxy<Float16>>(f).getUnbiasedNormalizedExponent(); |
| 584 |
} |
| 585 |
|
| 586 |
TEST(HexFloatOperationTest, UnbiasedExponent) { |
| 587 |
// Float cases |
| 588 |
EXPECT_EQ(0, unbiased_exponent(ldexp(1.0f, 0))); |
| 589 |
EXPECT_EQ(-32, unbiased_exponent(ldexp(1.0f, -32))); |
| 590 |
EXPECT_EQ(42, unbiased_exponent(ldexp(1.0f, 42))); |
| 591 |
EXPECT_EQ(125, unbiased_exponent(ldexp(1.0f, 125))); |
| 592 |
|
| 593 |
EXPECT_EQ(128, |
| 594 |
HexFloat<FloatProxy<float>>(std::numeric_limits<float>::infinity()) |
| 595 |
.getUnbiasedNormalizedExponent()); |
| 596 |
|
| 597 |
EXPECT_EQ(-100, unbiased_exponent(ldexp(1.0f, -100))); |
| 598 |
EXPECT_EQ(-127, unbiased_exponent(ldexp(1.0f, -127))); // First denorm |
| 599 |
EXPECT_EQ(-128, unbiased_exponent(ldexp(1.0f, -128))); |
| 600 |
EXPECT_EQ(-129, unbiased_exponent(ldexp(1.0f, -129))); |
| 601 |
EXPECT_EQ(-140, unbiased_exponent(ldexp(1.0f, -140))); |
| 602 |
// Smallest representable number |
| 603 |
EXPECT_EQ(-126 - 23, unbiased_exponent(ldexp(1.0f, -126 - 23))); |
| 604 |
// Should get rounded to 0 first. |
| 605 |
EXPECT_EQ(0, unbiased_exponent(ldexp(1.0f, -127 - 23))); |
| 606 |
|
| 607 |
// Float16 cases |
| 608 |
// The exponent is represented in the bits 0x7C00 |
| 609 |
// The offset is -15 |
| 610 |
EXPECT_EQ(0, unbiased_half_exponent(0x3C00)); |
| 611 |
EXPECT_EQ(3, unbiased_half_exponent(0x4800)); |
| 612 |
EXPECT_EQ(-1, unbiased_half_exponent(0x3800)); |
| 613 |
EXPECT_EQ(-14, unbiased_half_exponent(0x0400)); |
| 614 |
EXPECT_EQ(16, unbiased_half_exponent(0x7C00)); |
| 615 |
EXPECT_EQ(10, unbiased_half_exponent(0x6400)); |
| 616 |
|
| 617 |
// Smallest representable number |
| 618 |
EXPECT_EQ(-24, unbiased_half_exponent(0x0001)); |
| 619 |
} |
| 620 |
|
| 621 |
// Creates a float that is the sum of 1/(2 ^ fractions[i]) for i in factions |
| 622 |
float float_fractions(const std::vector<uint32_t>& fractions) { |
| 623 |
float f = 0; |
| 624 |
for (int32_t i : fractions) { |
| 625 |
f += std::ldexp(1.0f, -i); |
| 626 |
} |
| 627 |
return f; |
| 628 |
} |
| 629 |
|
| 630 |
// Returns the normalized significand of a HexFloat<FloatProxy<float>> |
| 631 |
// that was created by calling float_fractions with the input fractions, |
| 632 |
// raised to the power of exp. |
| 633 |
uint32_t normalized_significand(const std::vector<uint32_t>& fractions, |
| 634 |
uint32_t exp) { |
| 635 |
return HexFloat<FloatProxy<float>>( |
| 636 |
static_cast<float>(ldexp(float_fractions(fractions), exp))) |
| 637 |
.getNormalizedSignificand(); |
| 638 |
} |
| 639 |
|
| 640 |
// Sets the bits from MSB to LSB of the significand part of a float. |
| 641 |
// For example 0 would set the bit 23 (counting from LSB to MSB), |
| 642 |
// and 1 would set the 22nd bit. |
| 643 |
uint32_t bits_set(const std::vector<uint32_t>& bits) { |
| 644 |
const uint32_t top_bit = 1u << 22u; |
| 645 |
uint32_t val = 0; |
| 646 |
for (uint32_t i : bits) { |
| 647 |
val |= top_bit >> i; |
| 648 |
} |
| 649 |
return val; |
| 650 |
} |
| 651 |
|
| 652 |
// The same as bits_set but for a Float16 value instead of 32-bit floating |
| 653 |
// point. |
| 654 |
uint16_t half_bits_set(const std::vector<uint32_t>& bits) { |
| 655 |
const uint32_t top_bit = 1u << 9u; |
| 656 |
uint32_t val = 0; |
| 657 |
for (uint32_t i : bits) { |
| 658 |
val |= top_bit >> i; |
| 659 |
} |
| 660 |
return static_cast<uint16_t>(val); |
| 661 |
} |
| 662 |
|
| 663 |
TEST(HexFloatOperationTest, NormalizedSignificand) { |
| 664 |
// For normalized numbers (the following) it should be a simple matter |
| 665 |
// of getting rid of the top implicit bit |
| 666 |
EXPECT_EQ(bits_set({}), normalized_significand({0}, 0)); |
| 667 |
EXPECT_EQ(bits_set({0}), normalized_significand({0, 1}, 0)); |
| 668 |
EXPECT_EQ(bits_set({0, 1}), normalized_significand({0, 1, 2}, 0)); |
| 669 |
EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 0)); |
| 670 |
EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 32)); |
| 671 |
EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 126)); |
| 672 |
|
| 673 |
// For denormalized numbers we expect the normalized significand to |
| 674 |
// shift as if it were normalized. This means, in practice that the |
| 675 |
// top_most set bit will be cut off. Looks very similar to above (on purpose) |
| 676 |
EXPECT_EQ(bits_set({}), |
| 677 |
normalized_significand({0}, static_cast<uint32_t>(-127))); |
| 678 |
EXPECT_EQ(bits_set({3}), |
| 679 |
normalized_significand({0, 4}, static_cast<uint32_t>(-128))); |
| 680 |
EXPECT_EQ(bits_set({3}), |
| 681 |
normalized_significand({0, 4}, static_cast<uint32_t>(-127))); |
| 682 |
EXPECT_EQ(bits_set({}), |
| 683 |
normalized_significand({22}, static_cast<uint32_t>(-127))); |
| 684 |
EXPECT_EQ(bits_set({0}), |
| 685 |
normalized_significand({21, 22}, static_cast<uint32_t>(-127))); |
| 686 |
} |
| 687 |
|
| 688 |
// Returns the 32-bit floating point value created by |
| 689 |
// calling setFromSignUnbiasedExponentAndNormalizedSignificand |
| 690 |
// on a HexFloat<FloatProxy<float>> |
| 691 |
float set_from_sign(bool negative, int32_t unbiased_exponent, |
| 692 |
uint32_t significand, bool round_denorm_up) { |
| 693 |
HexFloat<FloatProxy<float>> f(0.f); |
| 694 |
f.setFromSignUnbiasedExponentAndNormalizedSignificand( |
| 695 |
negative, unbiased_exponent, significand, round_denorm_up); |
| 696 |
return f.value().getAsFloat(); |
| 697 |
} |
| 698 |
|
| 699 |
TEST(HexFloatOperationTests, |
| 700 |
SetFromSignUnbiasedExponentAndNormalizedSignificand) { |
| 701 |
EXPECT_EQ(1.f, set_from_sign(false, 0, 0, false)); |
| 702 |
|
| 703 |
// Tests insertion of various denormalized numbers with and without round up. |
| 704 |
EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), |
| 705 |
set_from_sign(false, -149, 0, false)); |
| 706 |
EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), |
| 707 |
set_from_sign(false, -149, 0, true)); |
| 708 |
EXPECT_EQ(0.f, set_from_sign(false, -150, 1, false)); |
| 709 |
EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), |
| 710 |
set_from_sign(false, -150, 1, true)); |
| 711 |
|
| 712 |
EXPECT_EQ(ldexp(1.0f, -127), set_from_sign(false, -127, 0, false)); |
| 713 |
EXPECT_EQ(ldexp(1.0f, -128), set_from_sign(false, -128, 0, false)); |
| 714 |
EXPECT_EQ(float_fractions({0, 1, 2, 5}), |
| 715 |
set_from_sign(false, 0, bits_set({0, 1, 4}), false)); |
| 716 |
EXPECT_EQ(ldexp(float_fractions({0, 1, 2, 5}), -32), |
| 717 |
set_from_sign(false, -32, bits_set({0, 1, 4}), false)); |
| 718 |
EXPECT_EQ(ldexp(float_fractions({0, 1, 2, 5}), -128), |
| 719 |
set_from_sign(false, -128, bits_set({0, 1, 4}), false)); |
| 720 |
|
| 721 |
// The negative cases from above. |
| 722 |
EXPECT_EQ(-1.f, set_from_sign(true, 0, 0, false)); |
| 723 |
EXPECT_EQ(-ldexp(1.0, -127), set_from_sign(true, -127, 0, false)); |
| 724 |
EXPECT_EQ(-ldexp(1.0, -128), set_from_sign(true, -128, 0, false)); |
| 725 |
EXPECT_EQ(-float_fractions({0, 1, 2, 5}), |
| 726 |
set_from_sign(true, 0, bits_set({0, 1, 4}), false)); |
| 727 |
EXPECT_EQ(-ldexp(float_fractions({0, 1, 2, 5}), -32), |
| 728 |
set_from_sign(true, -32, bits_set({0, 1, 4}), false)); |
| 729 |
EXPECT_EQ(-ldexp(float_fractions({0, 1, 2, 5}), -128), |
| 730 |
set_from_sign(true, -128, bits_set({0, 1, 4}), false)); |
| 731 |
} |
| 732 |
|
| 733 |
TEST(HexFloatOperationTests, NonRounding) { |
| 734 |
// Rounding from 32-bit hex-float to 32-bit hex-float should be trivial, |
| 735 |
// except in the denorm case which is a bit more complex. |
| 736 |
using HF = HexFloat<FloatProxy<float>>; |
| 737 |
bool carry_bit = false; |
| 738 |
|
| 739 |
round_direction rounding[] = {round_direction::kToZero, |
| 740 |
round_direction::kToNearestEven, |
| 741 |
round_direction::kToPositiveInfinity, |
| 742 |
round_direction::kToNegativeInfinity}; |
| 743 |
|
| 744 |
// Everything fits, so this should be straight-forward |
| 745 |
for (round_direction round : rounding) { |
| 746 |
EXPECT_EQ(bits_set({}), |
| 747 |
HF(0.f).getRoundedNormalizedSignificand<HF>(round, &carry_bit)); |
| 748 |
EXPECT_FALSE(carry_bit); |
| 749 |
|
| 750 |
EXPECT_EQ(bits_set({0}), |
| 751 |
HF(float_fractions({0, 1})) |
| 752 |
.getRoundedNormalizedSignificand<HF>(round, &carry_bit)); |
| 753 |
EXPECT_FALSE(carry_bit); |
| 754 |
|
| 755 |
EXPECT_EQ(bits_set({1, 3}), |
| 756 |
HF(float_fractions({0, 2, 4})) |
| 757 |
.getRoundedNormalizedSignificand<HF>(round, &carry_bit)); |
| 758 |
EXPECT_FALSE(carry_bit); |
| 759 |
|
| 760 |
EXPECT_EQ( |
| 761 |
bits_set({0, 1, 4}), |
| 762 |
HF(static_cast<float>(-ldexp(float_fractions({0, 1, 2, 5}), -128))) |
| 763 |
.getRoundedNormalizedSignificand<HF>(round, &carry_bit)); |
| 764 |
EXPECT_FALSE(carry_bit); |
| 765 |
|
| 766 |
EXPECT_EQ(bits_set({0, 1, 4, 22}), |
| 767 |
HF(static_cast<float>(float_fractions({0, 1, 2, 5, 23}))) |
| 768 |
.getRoundedNormalizedSignificand<HF>(round, &carry_bit)); |
| 769 |
EXPECT_FALSE(carry_bit); |
| 770 |
} |
| 771 |
} |
| 772 |
|
| 773 |
using RD = round_direction; |
| 774 |
struct RoundSignificandCase { |
| 775 |
float source_float; |
| 776 |
std::pair<int16_t, bool> expected_results; |
| 777 |
round_direction round; |
| 778 |
}; |
| 779 |
|
| 780 |
using HexFloatRoundTest = ::testing::TestWithParam<RoundSignificandCase>; |
| 781 |
|
| 782 |
TEST_P(HexFloatRoundTest, RoundDownToFP16) { |
| 783 |
using HF = HexFloat<FloatProxy<float>>; |
| 784 |
using HF16 = HexFloat<FloatProxy<Float16>>; |
| 785 |
|
| 786 |
HF input_value(GetParam().source_float); |
| 787 |
bool carry_bit = false; |
| 788 |
EXPECT_EQ(GetParam().expected_results.first, |
| 789 |
input_value.getRoundedNormalizedSignificand<HF16>(GetParam().round, |
| 790 |
&carry_bit)); |
| 791 |
EXPECT_EQ(carry_bit, GetParam().expected_results.second); |
| 792 |
} |
| 793 |
|
| 794 |
// clang-format off |
| 795 |
INSTANTIATE_TEST_SUITE_P(F32ToF16, HexFloatRoundTest, |
| 796 |
::testing::ValuesIn(std::vector<RoundSignificandCase>( |
| 797 |
{ |
| 798 |
{float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToZero}, |
| 799 |
{float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToNearestEven}, |
| 800 |
{float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToPositiveInfinity}, |
| 801 |
{float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToNegativeInfinity}, |
| 802 |
{float_fractions({0, 1}), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 803 |
|
| 804 |
{float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 805 |
{float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity}, |
| 806 |
{float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity}, |
| 807 |
{float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToNearestEven}, |
| 808 |
|
| 809 |
{float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToZero}, |
| 810 |
{float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 8}), false), RD::kToPositiveInfinity}, |
| 811 |
{float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNegativeInfinity}, |
| 812 |
{float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 8}), false), RD::kToNearestEven}, |
| 813 |
|
| 814 |
{float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 815 |
{float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity}, |
| 816 |
{float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity}, |
| 817 |
{float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven}, |
| 818 |
|
| 819 |
{-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 820 |
{-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToPositiveInfinity}, |
| 821 |
{-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNegativeInfinity}, |
| 822 |
{-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven}, |
| 823 |
|
| 824 |
{float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 825 |
{float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity}, |
| 826 |
{float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity}, |
| 827 |
{float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven}, |
| 828 |
|
| 829 |
// Carries |
| 830 |
{float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}), false), RD::kToZero}, |
| 831 |
{float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({}), true), RD::kToPositiveInfinity}, |
| 832 |
{float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}), false), RD::kToNegativeInfinity}, |
| 833 |
{float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({}), true), RD::kToNearestEven}, |
| 834 |
|
| 835 |
// Cases where original number was denorm. Note: this should have no effect |
| 836 |
// the number is pre-normalized. |
| 837 |
{static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -128)), std::make_pair(half_bits_set({0}), false), RD::kToZero}, |
| 838 |
{static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -129)), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity}, |
| 839 |
{static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -131)), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity}, |
| 840 |
{static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -130)), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven}, |
| 841 |
}))); |
| 842 |
// clang-format on |
| 843 |
|
| 844 |
struct UpCastSignificandCase { |
| 845 |
uint16_t source_half; |
| 846 |
uint32_t expected_result; |
| 847 |
}; |
| 848 |
|
| 849 |
using HexFloatRoundUpSignificandTest = |
| 850 |
::testing::TestWithParam<UpCastSignificandCase>; |
| 851 |
TEST_P(HexFloatRoundUpSignificandTest, Widening) { |
| 852 |
using HF = HexFloat<FloatProxy<float>>; |
| 853 |
using HF16 = HexFloat<FloatProxy<Float16>>; |
| 854 |
bool carry_bit = false; |
| 855 |
|
| 856 |
round_direction rounding[] = {round_direction::kToZero, |
| 857 |
round_direction::kToNearestEven, |
| 858 |
round_direction::kToPositiveInfinity, |
| 859 |
round_direction::kToNegativeInfinity}; |
| 860 |
|
| 861 |
// Everything fits, so everything should just be bit-shifts. |
| 862 |
for (round_direction round : rounding) { |
| 863 |
carry_bit = false; |
| 864 |
HF16 input_value(GetParam().source_half); |
| 865 |
EXPECT_EQ( |
| 866 |
GetParam().expected_result, |
| 867 |
input_value.getRoundedNormalizedSignificand<HF>(round, &carry_bit)) |
| 868 |
<< std::hex << "0x" |
| 869 |
<< input_value.getRoundedNormalizedSignificand<HF>(round, &carry_bit) |
| 870 |
<< " 0x" << GetParam().expected_result; |
| 871 |
EXPECT_FALSE(carry_bit); |
| 872 |
} |
| 873 |
} |
| 874 |
|
| 875 |
INSTANTIATE_TEST_SUITE_P( |
| 876 |
F16toF32, HexFloatRoundUpSignificandTest, |
| 877 |
// 0xFC00 of the source 16-bit hex value cover the sign and the exponent. |
| 878 |
// They are ignored for this test. |
| 879 |
::testing::ValuesIn(std::vector<UpCastSignificandCase>({ |
| 880 |
{0x3F00, 0x600000}, |
| 881 |
{0x0F00, 0x600000}, |
| 882 |
{0x0F01, 0x602000}, |
| 883 |
{0x0FFF, 0x7FE000}, |
| 884 |
}))); |
| 885 |
|
| 886 |
struct DownCastTest { |
| 887 |
float source_float; |
| 888 |
uint16_t expected_half; |
| 889 |
std::vector<round_direction> directions; |
| 890 |
}; |
| 891 |
|
| 892 |
std::string get_round_text(round_direction direction) { |
| 893 |
#define CASE(round_direction) \ |
| 894 |
case round_direction: \ |
| 895 |
return #round_direction |
| 896 |
|
| 897 |
switch (direction) { |
| 898 |
CASE(round_direction::kToZero); |
| 899 |
CASE(round_direction::kToPositiveInfinity); |
| 900 |
CASE(round_direction::kToNegativeInfinity); |
| 901 |
CASE(round_direction::kToNearestEven); |
| 902 |
} |
| 903 |
#undef CASE |
| 904 |
return ""; |
| 905 |
} |
| 906 |
|
| 907 |
using HexFloatFP32To16Tests = ::testing::TestWithParam<DownCastTest>; |
| 908 |
|
| 909 |
TEST_P(HexFloatFP32To16Tests, NarrowingCasts) { |
| 910 |
using HF = HexFloat<FloatProxy<float>>; |
| 911 |
using HF16 = HexFloat<FloatProxy<Float16>>; |
| 912 |
HF f(GetParam().source_float); |
| 913 |
for (auto round : GetParam().directions) { |
| 914 |
HF16 half(0); |
| 915 |
f.castTo(half, round); |
| 916 |
EXPECT_EQ(GetParam().expected_half, half.value().getAsFloat().get_value()) |
| 917 |
<< get_round_text(round) << " " << std::hex |
| 918 |
<< BitwiseCast<uint32_t>(GetParam().source_float) |
| 919 |
<< " cast to: " << half.value().getAsFloat().get_value(); |
| 920 |
} |
| 921 |
} |
| 922 |
|
| 923 |
const uint16_t positive_infinity = 0x7C00; |
| 924 |
const uint16_t negative_infinity = 0xFC00; |
| 925 |
|
| 926 |
INSTANTIATE_TEST_SUITE_P( |
| 927 |
F32ToF16, HexFloatFP32To16Tests, |
| 928 |
::testing::ValuesIn(std::vector<DownCastTest>({ |
| 929 |
// Exactly representable as half. |
| 930 |
{0.f, |
| 931 |
0x0, |
| 932 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 933 |
RD::kToNearestEven}}, |
| 934 |
{-0.f, |
| 935 |
0x8000, |
| 936 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 937 |
RD::kToNearestEven}}, |
| 938 |
{1.0f, |
| 939 |
0x3C00, |
| 940 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 941 |
RD::kToNearestEven}}, |
| 942 |
{-1.0f, |
| 943 |
0xBC00, |
| 944 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 945 |
RD::kToNearestEven}}, |
| 946 |
|
| 947 |
{float_fractions({0, 1, 10}), |
| 948 |
0x3E01, |
| 949 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 950 |
RD::kToNearestEven}}, |
| 951 |
{-float_fractions({0, 1, 10}), |
| 952 |
0xBE01, |
| 953 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 954 |
RD::kToNearestEven}}, |
| 955 |
{static_cast<float>(ldexp(float_fractions({0, 1, 10}), 3)), |
| 956 |
0x4A01, |
| 957 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 958 |
RD::kToNearestEven}}, |
| 959 |
{static_cast<float>(-ldexp(float_fractions({0, 1, 10}), 3)), |
| 960 |
0xCA01, |
| 961 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 962 |
RD::kToNearestEven}}, |
| 963 |
|
| 964 |
// Underflow |
| 965 |
{static_cast<float>(ldexp(1.0f, -25)), |
| 966 |
0x0, |
| 967 |
{RD::kToZero, RD::kToNegativeInfinity, RD::kToNearestEven}}, |
| 968 |
{static_cast<float>(ldexp(1.0f, -25)), 0x1, {RD::kToPositiveInfinity}}, |
| 969 |
{static_cast<float>(-ldexp(1.0f, -25)), |
| 970 |
0x8000, |
| 971 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNearestEven}}, |
| 972 |
{static_cast<float>(-ldexp(1.0f, -25)), |
| 973 |
0x8001, |
| 974 |
{RD::kToNegativeInfinity}}, |
| 975 |
{static_cast<float>(ldexp(1.0f, -24)), |
| 976 |
0x1, |
| 977 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 978 |
RD::kToNearestEven}}, |
| 979 |
|
| 980 |
// Overflow |
| 981 |
{static_cast<float>(ldexp(1.0f, 16)), |
| 982 |
positive_infinity, |
| 983 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 984 |
RD::kToNearestEven}}, |
| 985 |
{static_cast<float>(ldexp(1.0f, 18)), |
| 986 |
positive_infinity, |
| 987 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 988 |
RD::kToNearestEven}}, |
| 989 |
{static_cast<float>(ldexp(1.3f, 16)), |
| 990 |
positive_infinity, |
| 991 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 992 |
RD::kToNearestEven}}, |
| 993 |
{static_cast<float>(-ldexp(1.0f, 16)), |
| 994 |
negative_infinity, |
| 995 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 996 |
RD::kToNearestEven}}, |
| 997 |
{static_cast<float>(-ldexp(1.0f, 18)), |
| 998 |
negative_infinity, |
| 999 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 1000 |
RD::kToNearestEven}}, |
| 1001 |
{static_cast<float>(-ldexp(1.3f, 16)), |
| 1002 |
negative_infinity, |
| 1003 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 1004 |
RD::kToNearestEven}}, |
| 1005 |
|
| 1006 |
// Transfer of Infinities |
| 1007 |
{std::numeric_limits<float>::infinity(), |
| 1008 |
positive_infinity, |
| 1009 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 1010 |
RD::kToNearestEven}}, |
| 1011 |
{-std::numeric_limits<float>::infinity(), |
| 1012 |
negative_infinity, |
| 1013 |
{RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, |
| 1014 |
RD::kToNearestEven}}, |
| 1015 |
|
| 1016 |
// Nans are below because we cannot test for equality. |
| 1017 |
}))); |
| 1018 |
|
| 1019 |
struct UpCastCase { |
| 1020 |
uint16_t source_half; |
| 1021 |
float expected_float; |
| 1022 |
}; |
| 1023 |
|
| 1024 |
using HexFloatFP16To32Tests = ::testing::TestWithParam<UpCastCase>; |
| 1025 |
TEST_P(HexFloatFP16To32Tests, WideningCasts) { |
| 1026 |
using HF = HexFloat<FloatProxy<float>>; |
| 1027 |
using HF16 = HexFloat<FloatProxy<Float16>>; |
| 1028 |
HF16 f(GetParam().source_half); |
| 1029 |
|
| 1030 |
round_direction rounding[] = {round_direction::kToZero, |
| 1031 |
round_direction::kToNearestEven, |
| 1032 |
round_direction::kToPositiveInfinity, |
| 1033 |
round_direction::kToNegativeInfinity}; |
| 1034 |
|
| 1035 |
// Everything fits, so everything should just be bit-shifts. |
| 1036 |
for (round_direction round : rounding) { |
| 1037 |
HF flt(0.f); |
| 1038 |
f.castTo(flt, round); |
| 1039 |
EXPECT_EQ(GetParam().expected_float, flt.value().getAsFloat()) |
| 1040 |
<< get_round_text(round) << " " << std::hex |
| 1041 |
<< BitwiseCast<uint16_t>(GetParam().source_half) |
| 1042 |
<< " cast to: " << flt.value().getAsFloat(); |
| 1043 |
} |
| 1044 |
} |
| 1045 |
|
| 1046 |
INSTANTIATE_TEST_SUITE_P( |
| 1047 |
F16ToF32, HexFloatFP16To32Tests, |
| 1048 |
::testing::ValuesIn(std::vector<UpCastCase>({ |
| 1049 |
{0x0000, 0.f}, |
| 1050 |
{0x8000, -0.f}, |
| 1051 |
{0x3C00, 1.0f}, |
| 1052 |
{0xBC00, -1.0f}, |
| 1053 |
{0x3F00, float_fractions({0, 1, 2})}, |
| 1054 |
{0xBF00, -float_fractions({0, 1, 2})}, |
| 1055 |
{0x3F01, float_fractions({0, 1, 2, 10})}, |
| 1056 |
{0xBF01, -float_fractions({0, 1, 2, 10})}, |
| 1057 |
|
| 1058 |
// denorm |
| 1059 |
{0x0001, static_cast<float>(ldexp(1.0, -24))}, |
| 1060 |
{0x0002, static_cast<float>(ldexp(1.0, -23))}, |
| 1061 |
{0x8001, static_cast<float>(-ldexp(1.0, -24))}, |
| 1062 |
{0x8011, static_cast<float>(-ldexp(1.0, -20) + -ldexp(1.0, -24))}, |
| 1063 |
|
| 1064 |
// inf |
| 1065 |
{0x7C00, std::numeric_limits<float>::infinity()}, |
| 1066 |
{0xFC00, -std::numeric_limits<float>::infinity()}, |
| 1067 |
}))); |
| 1068 |
|
| 1069 |
TEST(HexFloatOperationTests, NanTests) { |
| 1070 |
using HF = HexFloat<FloatProxy<float>>; |
| 1071 |
using HF16 = HexFloat<FloatProxy<Float16>>; |
| 1072 |
round_direction rounding[] = {round_direction::kToZero, |
| 1073 |
round_direction::kToNearestEven, |
| 1074 |
round_direction::kToPositiveInfinity, |
| 1075 |
round_direction::kToNegativeInfinity}; |
| 1076 |
|
| 1077 |
// Everything fits, so everything should just be bit-shifts. |
| 1078 |
for (round_direction round : rounding) { |
| 1079 |
HF16 f16(0); |
| 1080 |
HF f(0.f); |
| 1081 |
HF(std::numeric_limits<float>::quiet_NaN()).castTo(f16, round); |
| 1082 |
EXPECT_TRUE(f16.value().isNan()); |
| 1083 |
HF(std::numeric_limits<float>::signaling_NaN()).castTo(f16, round); |
| 1084 |
EXPECT_TRUE(f16.value().isNan()); |
| 1085 |
|
| 1086 |
HF16(0x7C01).castTo(f, round); |
| 1087 |
EXPECT_TRUE(f.value().isNan()); |
| 1088 |
HF16(0x7C11).castTo(f, round); |
| 1089 |
EXPECT_TRUE(f.value().isNan()); |
| 1090 |
HF16(0xFC01).castTo(f, round); |
| 1091 |
EXPECT_TRUE(f.value().isNan()); |
| 1092 |
HF16(0x7C10).castTo(f, round); |
| 1093 |
EXPECT_TRUE(f.value().isNan()); |
| 1094 |
HF16(0xFF00).castTo(f, round); |
| 1095 |
EXPECT_TRUE(f.value().isNan()); |
| 1096 |
} |
| 1097 |
} |
| 1098 |
|
| 1099 |
// A test case for parsing good and bad HexFloat<FloatProxy<T>> literals. |
| 1100 |
template <typename T> |
| 1101 |
struct FloatParseCase { |
| 1102 |
std::string literal; |
| 1103 |
bool negate_value; |
| 1104 |
bool expect_success; |
| 1105 |
HexFloat<FloatProxy<T>> expected_value; |
| 1106 |
}; |
| 1107 |
|
| 1108 |
using ParseNormalFloatTest = ::testing::TestWithParam<FloatParseCase<float>>; |
| 1109 |
|
| 1110 |
TEST_P(ParseNormalFloatTest, Samples) { |
| 1111 |
std::stringstream input(GetParam().literal); |
| 1112 |
HexFloat<FloatProxy<float>> parsed_value(0.0f); |
| 1113 |
ParseNormalFloat(input, GetParam().negate_value, parsed_value); |
| 1114 |
EXPECT_NE(GetParam().expect_success, input.fail()) |
| 1115 |
<< " literal: " << GetParam().literal |
| 1116 |
<< " negate: " << GetParam().negate_value; |
| 1117 |
if (GetParam().expect_success) { |
| 1118 |
EXPECT_THAT(parsed_value.value(), Eq(GetParam().expected_value.value())) |
| 1119 |
<< " literal: " << GetParam().literal |
| 1120 |
<< " negate: " << GetParam().negate_value; |
| 1121 |
} |
| 1122 |
} |
| 1123 |
|
| 1124 |
// Returns a FloatParseCase with expected failure. |
| 1125 |
template <typename T> |
| 1126 |
FloatParseCase<T> BadFloatParseCase(std::string literal, bool negate_value, |
| 1127 |
T expected_value) { |
| 1128 |
HexFloat<FloatProxy<T>> proxy_expected_value(expected_value); |
| 1129 |
return FloatParseCase<T>{literal, negate_value, false, proxy_expected_value}; |
| 1130 |
} |
| 1131 |
|
| 1132 |
// Returns a FloatParseCase that should successfully parse to a given value. |
| 1133 |
template <typename T> |
| 1134 |
FloatParseCase<T> GoodFloatParseCase(std::string literal, bool negate_value, |
| 1135 |
T expected_value) { |
| 1136 |
HexFloat<FloatProxy<T>> proxy_expected_value(expected_value); |
| 1137 |
return FloatParseCase<T>{literal, negate_value, true, proxy_expected_value}; |
| 1138 |
} |
| 1139 |
|
| 1140 |
INSTANTIATE_TEST_SUITE_P( |
| 1141 |
FloatParse, ParseNormalFloatTest, |
| 1142 |
::testing::ValuesIn(std::vector<FloatParseCase<float>>{ |
| 1143 |
// Failing cases due to trivially incorrect syntax. |
| 1144 |
BadFloatParseCase("abc", false, 0.0f), |
| 1145 |
BadFloatParseCase("abc", true, 0.0f), |
| 1146 |
|
| 1147 |
// Valid cases. |
| 1148 |
GoodFloatParseCase("0", false, 0.0f), |
| 1149 |
GoodFloatParseCase("0.0", false, 0.0f), |
| 1150 |
GoodFloatParseCase("-0.0", false, -0.0f), |
| 1151 |
GoodFloatParseCase("2.0", false, 2.0f), |
| 1152 |
GoodFloatParseCase("-2.0", false, -2.0f), |
| 1153 |
GoodFloatParseCase("+2.0", false, 2.0f), |
| 1154 |
// Cases with negate_value being true. |
| 1155 |
GoodFloatParseCase("0.0", true, -0.0f), |
| 1156 |
GoodFloatParseCase("2.0", true, -2.0f), |
| 1157 |
|
| 1158 |
// When negate_value is true, we should not accept a |
| 1159 |
// leading minus or plus. |
| 1160 |
BadFloatParseCase("-0.0", true, 0.0f), |
| 1161 |
BadFloatParseCase("-2.0", true, 0.0f), |
| 1162 |
BadFloatParseCase("+0.0", true, 0.0f), |
| 1163 |
BadFloatParseCase("+2.0", true, 0.0f), |
| 1164 |
|
| 1165 |
// Overflow is an error for 32-bit float parsing. |
| 1166 |
BadFloatParseCase("1e40", false, FLT_MAX), |
| 1167 |
BadFloatParseCase("1e40", true, -FLT_MAX), |
| 1168 |
BadFloatParseCase("-1e40", false, -FLT_MAX), |
| 1169 |
// We can't have -1e40 and negate_value == true since |
| 1170 |
// that represents an original case of "--1e40" which |
| 1171 |
// is invalid. |
| 1172 |
})); |
| 1173 |
|
| 1174 |
using ParseNormalFloat16Test = |
| 1175 |
::testing::TestWithParam<FloatParseCase<Float16>>; |
| 1176 |
|
| 1177 |
TEST_P(ParseNormalFloat16Test, Samples) { |
| 1178 |
std::stringstream input(GetParam().literal); |
| 1179 |
HexFloat<FloatProxy<Float16>> parsed_value(0); |
| 1180 |
ParseNormalFloat(input, GetParam().negate_value, parsed_value); |
| 1181 |
EXPECT_NE(GetParam().expect_success, input.fail()) |
| 1182 |
<< " literal: " << GetParam().literal |
| 1183 |
<< " negate: " << GetParam().negate_value; |
| 1184 |
if (GetParam().expect_success) { |
| 1185 |
EXPECT_THAT(parsed_value.value(), Eq(GetParam().expected_value.value())) |
| 1186 |
<< " literal: " << GetParam().literal |
| 1187 |
<< " negate: " << GetParam().negate_value; |
| 1188 |
} |
| 1189 |
} |
| 1190 |
|
| 1191 |
INSTANTIATE_TEST_SUITE_P( |
| 1192 |
Float16Parse, ParseNormalFloat16Test, |
| 1193 |
::testing::ValuesIn(std::vector<FloatParseCase<Float16>>{ |
| 1194 |
// Failing cases due to trivially incorrect syntax. |
| 1195 |
BadFloatParseCase<Float16>("abc", false, uint16_t{0}), |
| 1196 |
BadFloatParseCase<Float16>("abc", true, uint16_t{0}), |
| 1197 |
|
| 1198 |
// Valid cases. |
| 1199 |
GoodFloatParseCase<Float16>("0", false, uint16_t{0}), |
| 1200 |
GoodFloatParseCase<Float16>("0.0", false, uint16_t{0}), |
| 1201 |
GoodFloatParseCase<Float16>("-0.0", false, uint16_t{0x8000}), |
| 1202 |
GoodFloatParseCase<Float16>("2.0", false, uint16_t{0x4000}), |
| 1203 |
GoodFloatParseCase<Float16>("-2.0", false, uint16_t{0xc000}), |
| 1204 |
GoodFloatParseCase<Float16>("+2.0", false, uint16_t{0x4000}), |
| 1205 |
// Cases with negate_value being true. |
| 1206 |
GoodFloatParseCase<Float16>("0.0", true, uint16_t{0x8000}), |
| 1207 |
GoodFloatParseCase<Float16>("2.0", true, uint16_t{0xc000}), |
| 1208 |
|
| 1209 |
// When negate_value is true, we should not accept a leading minus or |
| 1210 |
// plus. |
| 1211 |
BadFloatParseCase<Float16>("-0.0", true, uint16_t{0}), |
| 1212 |
BadFloatParseCase<Float16>("-2.0", true, uint16_t{0}), |
| 1213 |
BadFloatParseCase<Float16>("+0.0", true, uint16_t{0}), |
| 1214 |
BadFloatParseCase<Float16>("+2.0", true, uint16_t{0}), |
| 1215 |
})); |
| 1216 |
|
| 1217 |
// A test case for detecting infinities. |
| 1218 |
template <typename T> |
| 1219 |
struct OverflowParseCase { |
| 1220 |
std::string input; |
| 1221 |
bool expect_success; |
| 1222 |
T expected_value; |
| 1223 |
}; |
| 1224 |
|
| 1225 |
using FloatProxyParseOverflowFloatTest = |
| 1226 |
::testing::TestWithParam<OverflowParseCase<float>>; |
| 1227 |
|
| 1228 |
TEST_P(FloatProxyParseOverflowFloatTest, Sample) { |
| 1229 |
std::istringstream input(GetParam().input); |
| 1230 |
HexFloat<FloatProxy<float>> value(0.0f); |
| 1231 |
input >> value; |
| 1232 |
EXPECT_NE(GetParam().expect_success, input.fail()); |
| 1233 |
if (GetParam().expect_success) { |
| 1234 |
EXPECT_THAT(value.value().getAsFloat(), GetParam().expected_value); |
| 1235 |
} |
| 1236 |
} |
| 1237 |
|
| 1238 |
INSTANTIATE_TEST_SUITE_P( |
| 1239 |
FloatOverflow, FloatProxyParseOverflowFloatTest, |
| 1240 |
::testing::ValuesIn(std::vector<OverflowParseCase<float>>({ |
| 1241 |
{"0", true, 0.0f}, |
| 1242 |
{"0.0", true, 0.0f}, |
| 1243 |
{"1.0", true, 1.0f}, |
| 1244 |
{"1e38", true, 1e38f}, |
| 1245 |
{"-1e38", true, -1e38f}, |
| 1246 |
{"1e40", false, FLT_MAX}, |
| 1247 |
{"-1e40", false, -FLT_MAX}, |
| 1248 |
{"1e400", false, FLT_MAX}, |
| 1249 |
{"-1e400", false, -FLT_MAX}, |
| 1250 |
}))); |
| 1251 |
|
| 1252 |
using FloatProxyParseOverflowDoubleTest = |
| 1253 |
::testing::TestWithParam<OverflowParseCase<double>>; |
| 1254 |
|
| 1255 |
TEST_P(FloatProxyParseOverflowDoubleTest, Sample) { |
| 1256 |
std::istringstream input(GetParam().input); |
| 1257 |
HexFloat<FloatProxy<double>> value(0.0); |
| 1258 |
input >> value; |
| 1259 |
EXPECT_NE(GetParam().expect_success, input.fail()); |
| 1260 |
if (GetParam().expect_success) { |
| 1261 |
EXPECT_THAT(value.value().getAsFloat(), Eq(GetParam().expected_value)); |
| 1262 |
} |
| 1263 |
} |
| 1264 |
|
| 1265 |
INSTANTIATE_TEST_SUITE_P( |
| 1266 |
DoubleOverflow, FloatProxyParseOverflowDoubleTest, |
| 1267 |
::testing::ValuesIn(std::vector<OverflowParseCase<double>>({ |
| 1268 |
{"0", true, 0.0}, |
| 1269 |
{"0.0", true, 0.0}, |
| 1270 |
{"1.0", true, 1.0}, |
| 1271 |
{"1e38", true, 1e38}, |
| 1272 |
{"-1e38", true, -1e38}, |
| 1273 |
{"1e40", true, 1e40}, |
| 1274 |
{"-1e40", true, -1e40}, |
| 1275 |
{"1e400", false, DBL_MAX}, |
| 1276 |
{"-1e400", false, -DBL_MAX}, |
| 1277 |
}))); |
| 1278 |
|
| 1279 |
using FloatProxyParseOverflowFloat16Test = |
| 1280 |
::testing::TestWithParam<OverflowParseCase<uint16_t>>; |
| 1281 |
|
| 1282 |
TEST_P(FloatProxyParseOverflowFloat16Test, Sample) { |
| 1283 |
std::istringstream input(GetParam().input); |
| 1284 |
HexFloat<FloatProxy<Float16>> value(0); |
| 1285 |
input >> value; |
| 1286 |
EXPECT_NE(GetParam().expect_success, input.fail()) |
| 1287 |
<< " literal: " << GetParam().input; |
| 1288 |
if (GetParam().expect_success) { |
| 1289 |
EXPECT_THAT(value.value().data(), Eq(GetParam().expected_value)) |
| 1290 |
<< " literal: " << GetParam().input; |
| 1291 |
} |
| 1292 |
} |
| 1293 |
|
| 1294 |
INSTANTIATE_TEST_SUITE_P( |
| 1295 |
Float16Overflow, FloatProxyParseOverflowFloat16Test, |
| 1296 |
::testing::ValuesIn(std::vector<OverflowParseCase<uint16_t>>({ |
| 1297 |
{"0", true, uint16_t{0}}, |
| 1298 |
{"0.0", true, uint16_t{0}}, |
| 1299 |
{"1.0", true, uint16_t{0x3c00}}, |
| 1300 |
// Overflow for 16-bit float is an error, and returns max or |
| 1301 |
// lowest value. |
| 1302 |
{"1e38", false, uint16_t{0x7bff}}, |
| 1303 |
{"1e40", false, uint16_t{0x7bff}}, |
| 1304 |
{"1e400", false, uint16_t{0x7bff}}, |
| 1305 |
{"-1e38", false, uint16_t{0xfbff}}, |
| 1306 |
{"-1e40", false, uint16_t{0xfbff}}, |
| 1307 |
{"-1e400", false, uint16_t{0xfbff}}, |
| 1308 |
}))); |
| 1309 |
|
| 1310 |
TEST(FloatProxy, Max) { |
| 1311 |
EXPECT_THAT(FloatProxy<Float16>::max().getAsFloat().get_value(), |
| 1312 |
Eq(uint16_t{0x7bff})); |
| 1313 |
EXPECT_THAT(FloatProxy<float>::max().getAsFloat(), |
| 1314 |
Eq(std::numeric_limits<float>::max())); |
| 1315 |
EXPECT_THAT(FloatProxy<double>::max().getAsFloat(), |
| 1316 |
Eq(std::numeric_limits<double>::max())); |
| 1317 |
} |
| 1318 |
|
| 1319 |
TEST(FloatProxy, Lowest) { |
| 1320 |
EXPECT_THAT(FloatProxy<Float16>::lowest().getAsFloat().get_value(), |
| 1321 |
Eq(uint16_t{0xfbff})); |
| 1322 |
EXPECT_THAT(FloatProxy<float>::lowest().getAsFloat(), |
| 1323 |
Eq(std::numeric_limits<float>::lowest())); |
| 1324 |
EXPECT_THAT(FloatProxy<double>::lowest().getAsFloat(), |
| 1325 |
Eq(std::numeric_limits<double>::lowest())); |
| 1326 |
} |
| 1327 |
|
| 1328 |
// TODO(awoloszyn): Add fp16 tests and HexFloatTraits. |
| 1329 |
} // namespace |
| 1330 |
} // namespace utils |
| 1331 |
} // namespace spvtools |