1 // Copyright (c) 2015-2016 The Khronos Group Inc.
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
7 // http://www.apache.org/licenses/LICENSE-2.0
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.
22 #include <gmock/gmock.h>
23 #include "SPIRV/hex_float.h"
27 using spvutils::BitwiseCast;
28 using spvutils::Float16;
29 using spvutils::FloatProxy;
30 using spvutils::HexFloat;
31 using spvutils::ParseNormalFloat;
33 // In this file "encode" means converting a number into a string,
34 // and "decode" means converting a string into a number.
37 ::testing::TestWithParam<std::pair<FloatProxy<float>, std::string>>;
38 using DecodeHexFloatTest =
39 ::testing::TestWithParam<std::pair<std::string, FloatProxy<float>>>;
41 ::testing::TestWithParam<std::pair<FloatProxy<double>, std::string>>;
42 using DecodeHexDoubleTest =
43 ::testing::TestWithParam<std::pair<std::string, FloatProxy<double>>>;
45 // Hex-encodes a float value.
47 std::string EncodeViaHexFloat(const T& value) {
49 ss << spvutils::HexFloat<T>(value);
53 // The following two tests can't be DRY because they take different parameter
56 TEST_P(HexFloatTest, EncodeCorrectly) {
57 EXPECT_THAT(EncodeViaHexFloat(GetParam().first), Eq(GetParam().second));
60 TEST_P(HexDoubleTest, EncodeCorrectly) {
61 EXPECT_THAT(EncodeViaHexFloat(GetParam().first), Eq(GetParam().second));
64 // Decodes a hex-float string.
66 FloatProxy<T> Decode(const std::string& str) {
67 spvutils::HexFloat<FloatProxy<T>> decoded(0.f);
68 EXPECT_TRUE((std::stringstream(str) >> decoded).eof());
69 return decoded.value();
72 TEST_P(HexFloatTest, DecodeCorrectly) {
73 EXPECT_THAT(Decode<float>(GetParam().second), Eq(GetParam().first));
76 TEST_P(HexDoubleTest, DecodeCorrectly) {
77 EXPECT_THAT(Decode<double>(GetParam().second), Eq(GetParam().first));
80 INSTANTIATE_TEST_CASE_P(
81 Float32Tests, HexFloatTest,
82 ::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({
94 {-0.75f, "-0x1.8p-1"},
100 {-1024.f, "-0x1p+10"},
101 {1024.f + 8.f, "0x1.02p+10"},
102 {-1024.f - 8.f, "-0x1.02p+10"},
105 {1.0f / 512.f, "0x1p-9"},
106 {1.0f / -512.f, "-0x1p-9"},
107 {1.0f / 1024.f, "0x1p-10"},
108 {1.0f / -1024.f, "-0x1p-10"},
109 {1.0f / 1024.f + 1.0f / 8.f, "0x1.02p-3"},
110 {1.0f / -1024.f - 1.0f / 8.f, "-0x1.02p-3"},
113 {float(ldexp(1.0f, -126)), "0x1p-126"},
114 {float(ldexp(-1.0f, -126)), "-0x1p-126"},
116 // Denormalized values
117 {float(ldexp(1.0f, -127)), "0x1p-127"},
118 {float(ldexp(1.0f, -127) / 2.0f), "0x1p-128"},
119 {float(ldexp(1.0f, -127) / 4.0f), "0x1p-129"},
120 {float(ldexp(1.0f, -127) / 8.0f), "0x1p-130"},
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"},
126 {float(ldexp(1.0, -127) + (ldexp(1.0, -127) / 2.0f)), "0x1.8p-127"},
127 {float(ldexp(1.0, -127) / 2.0 + (ldexp(1.0, -127) / 4.0f)),
132 INSTANTIATE_TEST_CASE_P(
133 Float32NanTests, HexFloatTest,
134 ::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({
135 // Various NAN and INF cases
136 {uint32_t(0xFF800000), "-0x1p+128"}, // -inf
137 {uint32_t(0x7F800000), "0x1p+128"}, // inf
138 {uint32_t(0xFFC00000), "-0x1.8p+128"}, // -nan
139 {uint32_t(0xFF800100), "-0x1.0002p+128"}, // -nan
140 {uint32_t(0xFF800c00), "-0x1.0018p+128"}, // -nan
141 {uint32_t(0xFF80F000), "-0x1.01ep+128"}, // -nan
142 {uint32_t(0xFFFFFFFF), "-0x1.fffffep+128"}, // -nan
143 {uint32_t(0x7FC00000), "0x1.8p+128"}, // +nan
144 {uint32_t(0x7F800100), "0x1.0002p+128"}, // +nan
145 {uint32_t(0x7f800c00), "0x1.0018p+128"}, // +nan
146 {uint32_t(0x7F80F000), "0x1.01ep+128"}, // +nan
147 {uint32_t(0x7FFFFFFF), "0x1.fffffep+128"}, // +nan
150 INSTANTIATE_TEST_CASE_P(
151 Float64Tests, HexDoubleTest,
153 std::vector<std::pair<FloatProxy<double>, std::string>>({
165 {-0.75, "-0x1.8p-1"},
171 {-1024., "-0x1p+10"},
172 {1024. + 8., "0x1.02p+10"},
173 {-1024. - 8., "-0x1.02p+10"},
175 // Large outside the range of normal floats
176 {ldexp(1.0, 128), "0x1p+128"},
177 {ldexp(1.0, 129), "0x1p+129"},
178 {ldexp(-1.0, 128), "-0x1p+128"},
179 {ldexp(-1.0, 129), "-0x1p+129"},
180 {ldexp(1.0, 128) + ldexp(1.0, 90), "0x1.0000000004p+128"},
181 {ldexp(1.0, 129) + ldexp(1.0, 120), "0x1.008p+129"},
182 {ldexp(-1.0, 128) + ldexp(1.0, 90), "-0x1.fffffffff8p+127"},
183 {ldexp(-1.0, 129) + ldexp(1.0, 120), "-0x1.ffp+128"},
186 {1.0 / 512., "0x1p-9"},
187 {1.0 / -512., "-0x1p-9"},
188 {1.0 / 1024., "0x1p-10"},
189 {1.0 / -1024., "-0x1p-10"},
190 {1.0 / 1024. + 1.0 / 8., "0x1.02p-3"},
191 {1.0 / -1024. - 1.0 / 8., "-0x1.02p-3"},
193 // Small outside the range of normal floats
194 {ldexp(1.0, -128), "0x1p-128"},
195 {ldexp(1.0, -129), "0x1p-129"},
196 {ldexp(-1.0, -128), "-0x1p-128"},
197 {ldexp(-1.0, -129), "-0x1p-129"},
198 {ldexp(1.0, -128) + ldexp(1.0, -90), "0x1.0000000004p-90"},
199 {ldexp(1.0, -129) + ldexp(1.0, -120), "0x1.008p-120"},
200 {ldexp(-1.0, -128) + ldexp(1.0, -90), "0x1.fffffffff8p-91"},
201 {ldexp(-1.0, -129) + ldexp(1.0, -120), "0x1.ffp-121"},
204 {ldexp(1.0, -1022), "0x1p-1022"},
205 {ldexp(-1.0, -1022), "-0x1p-1022"},
207 // Denormalized values
208 {ldexp(1.0, -1023), "0x1p-1023"},
209 {ldexp(1.0, -1023) / 2.0, "0x1p-1024"},
210 {ldexp(1.0, -1023) / 4.0, "0x1p-1025"},
211 {ldexp(1.0, -1023) / 8.0, "0x1p-1026"},
212 {ldexp(-1.0, -1024), "-0x1p-1024"},
213 {ldexp(-1.0, -1024) / 2.0, "-0x1p-1025"},
214 {ldexp(-1.0, -1024) / 4.0, "-0x1p-1026"},
215 {ldexp(-1.0, -1024) / 8.0, "-0x1p-1027"},
217 {ldexp(1.0, -1023) + (ldexp(1.0, -1023) / 2.0), "0x1.8p-1023"},
218 {ldexp(1.0, -1023) / 2.0 + (ldexp(1.0, -1023) / 4.0),
223 INSTANTIATE_TEST_CASE_P(
224 Float64NanTests, HexDoubleTest,
225 ::testing::ValuesIn(std::vector<
226 std::pair<FloatProxy<double>, std::string>>({
227 // Various NAN and INF cases
228 {uint64_t(0xFFF0000000000000LL), "-0x1p+1024"}, //-inf
229 {uint64_t(0x7FF0000000000000LL), "0x1p+1024"}, //+inf
230 {uint64_t(0xFFF8000000000000LL), "-0x1.8p+1024"}, // -nan
231 {uint64_t(0xFFF0F00000000000LL), "-0x1.0fp+1024"}, // -nan
232 {uint64_t(0xFFF0000000000001LL), "-0x1.0000000000001p+1024"}, // -nan
233 {uint64_t(0xFFF0000300000000LL), "-0x1.00003p+1024"}, // -nan
234 {uint64_t(0xFFFFFFFFFFFFFFFFLL), "-0x1.fffffffffffffp+1024"}, // -nan
235 {uint64_t(0x7FF8000000000000LL), "0x1.8p+1024"}, // +nan
236 {uint64_t(0x7FF0F00000000000LL), "0x1.0fp+1024"}, // +nan
237 {uint64_t(0x7FF0000000000001LL), "0x1.0000000000001p+1024"}, // -nan
238 {uint64_t(0x7FF0000300000000LL), "0x1.00003p+1024"}, // -nan
239 {uint64_t(0x7FFFFFFFFFFFFFFFLL), "0x1.fffffffffffffp+1024"}, // -nan
242 TEST(HexFloatStreamTest, OperatorLeftShiftPreservesFloatAndFill) {
244 s << std::setw(4) << std::oct << std::setfill('x') << 8 << " "
245 << FloatProxy<float>(uint32_t(0xFF800100)) << " " << std::setw(4) << 9;
246 EXPECT_THAT(s.str(), Eq(std::string("xx10 -0x1.0002p+128 xx11")));
249 TEST(HexDoubleStreamTest, OperatorLeftShiftPreservesFloatAndFill) {
251 s << std::setw(4) << std::oct << std::setfill('x') << 8 << " "
252 << FloatProxy<double>(uint64_t(0x7FF0F00000000000LL)) << " " << std::setw(4)
254 EXPECT_THAT(s.str(), Eq(std::string("xx10 0x1.0fp+1024 xx11")));
257 TEST_P(DecodeHexFloatTest, DecodeCorrectly) {
258 EXPECT_THAT(Decode<float>(GetParam().first), Eq(GetParam().second));
261 TEST_P(DecodeHexDoubleTest, DecodeCorrectly) {
262 EXPECT_THAT(Decode<double>(GetParam().first), Eq(GetParam().second));
265 INSTANTIATE_TEST_CASE_P(
266 Float32DecodeTests, DecodeHexFloatTest,
267 ::testing::ValuesIn(std::vector<std::pair<std::string, FloatProxy<float>>>({
272 // flush to zero cases
273 {"0x1p-500", 0.f}, // Exponent underflows.
275 {"0x0.00000000001p-126", 0.f}, // Fraction causes underflow.
276 {"-0x0.0000000001p-127", -0.f},
277 {"-0x0.01p-142", -0.f}, // Fraction causes additional underflow.
278 {"0x0.01p-142", 0.f},
280 // Some floats that do not encode the same way as they decode.
287 INSTANTIATE_TEST_CASE_P(
288 Float32DecodeInfTests, DecodeHexFloatTest,
289 ::testing::ValuesIn(std::vector<std::pair<std::string, FloatProxy<float>>>({
291 {"-0x1p+128", uint32_t(0xFF800000)}, // -inf
292 {"0x32p+127", uint32_t(0x7F800000)}, // inf
293 {"0x32p+500", uint32_t(0x7F800000)}, // inf
294 {"-0x32p+127", uint32_t(0xFF800000)}, // -inf
297 INSTANTIATE_TEST_CASE_P(
298 Float64DecodeTests, DecodeHexDoubleTest,
300 std::vector<std::pair<std::string, FloatProxy<double>>>({
305 // flush to zero cases
306 {"0x1p-5000", 0.}, // Exponent underflows.
308 {"0x0.0000000000000001p-1023", 0.}, // Fraction causes underflow.
309 {"-0x0.000000000000001p-1024", -0.},
310 {"-0x0.01p-1090", -0.f}, // Fraction causes additional underflow.
311 {"0x0.01p-1090", 0.},
313 // Some floats that do not encode the same way as they decode.
320 INSTANTIATE_TEST_CASE_P(
321 Float64DecodeInfTests, DecodeHexDoubleTest,
323 std::vector<std::pair<std::string, FloatProxy<double>>>({
325 {"-0x1p+1024", uint64_t(0xFFF0000000000000)}, // -inf
326 {"0x32p+1023", uint64_t(0x7FF0000000000000)}, // inf
327 {"0x32p+5000", uint64_t(0x7FF0000000000000)}, // inf
328 {"-0x32p+1023", uint64_t(0xFFF0000000000000)}, // -inf
331 TEST(FloatProxy, ValidConversion) {
332 EXPECT_THAT(FloatProxy<float>(1.f).getAsFloat(), Eq(1.0f));
333 EXPECT_THAT(FloatProxy<float>(32.f).getAsFloat(), Eq(32.0f));
334 EXPECT_THAT(FloatProxy<float>(-1.f).getAsFloat(), Eq(-1.0f));
335 EXPECT_THAT(FloatProxy<float>(0.f).getAsFloat(), Eq(0.0f));
336 EXPECT_THAT(FloatProxy<float>(-0.f).getAsFloat(), Eq(-0.0f));
337 EXPECT_THAT(FloatProxy<float>(1.2e32f).getAsFloat(), Eq(1.2e32f));
339 EXPECT_TRUE(std::isinf(FloatProxy<float>(uint32_t(0xFF800000)).getAsFloat()));
340 EXPECT_TRUE(std::isinf(FloatProxy<float>(uint32_t(0x7F800000)).getAsFloat()));
341 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFFC00000)).getAsFloat()));
342 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF800100)).getAsFloat()));
343 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF800c00)).getAsFloat()));
344 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFF80F000)).getAsFloat()));
345 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0xFFFFFFFF)).getAsFloat()));
346 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7FC00000)).getAsFloat()));
347 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7F800100)).getAsFloat()));
348 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7f800c00)).getAsFloat()));
349 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7F80F000)).getAsFloat()));
350 EXPECT_TRUE(std::isnan(FloatProxy<float>(uint32_t(0x7FFFFFFF)).getAsFloat()));
352 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800000)).data(), Eq(0xFF800000u));
353 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F800000)).data(), Eq(0x7F800000u));
354 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFFC00000)).data(), Eq(0xFFC00000u));
355 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800100)).data(), Eq(0xFF800100u));
356 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF800c00)).data(), Eq(0xFF800c00u));
357 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFF80F000)).data(), Eq(0xFF80F000u));
358 EXPECT_THAT(FloatProxy<float>(uint32_t(0xFFFFFFFF)).data(), Eq(0xFFFFFFFFu));
359 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7FC00000)).data(), Eq(0x7FC00000u));
360 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F800100)).data(), Eq(0x7F800100u));
361 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7f800c00)).data(), Eq(0x7f800c00u));
362 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7F80F000)).data(), Eq(0x7F80F000u));
363 EXPECT_THAT(FloatProxy<float>(uint32_t(0x7FFFFFFF)).data(), Eq(0x7FFFFFFFu));
366 TEST(FloatProxy, Nan) {
367 EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFFC00000)).isNan());
368 EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF800100)).isNan());
369 EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF800c00)).isNan());
370 EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFF80F000)).isNan());
371 EXPECT_TRUE(FloatProxy<float>(uint32_t(0xFFFFFFFF)).isNan());
372 EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7FC00000)).isNan());
373 EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7F800100)).isNan());
374 EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7f800c00)).isNan());
375 EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7F80F000)).isNan());
376 EXPECT_TRUE(FloatProxy<float>(uint32_t(0x7FFFFFFF)).isNan());
379 TEST(FloatProxy, Negation) {
380 EXPECT_THAT((-FloatProxy<float>(1.f)).getAsFloat(), Eq(-1.0f));
381 EXPECT_THAT((-FloatProxy<float>(0.f)).getAsFloat(), Eq(-0.0f));
383 EXPECT_THAT((-FloatProxy<float>(-1.f)).getAsFloat(), Eq(1.0f));
384 EXPECT_THAT((-FloatProxy<float>(-0.f)).getAsFloat(), Eq(0.0f));
386 EXPECT_THAT((-FloatProxy<float>(32.f)).getAsFloat(), Eq(-32.0f));
387 EXPECT_THAT((-FloatProxy<float>(-32.f)).getAsFloat(), Eq(32.0f));
389 EXPECT_THAT((-FloatProxy<float>(1.2e32f)).getAsFloat(), Eq(-1.2e32f));
390 EXPECT_THAT((-FloatProxy<float>(-1.2e32f)).getAsFloat(), Eq(1.2e32f));
393 (-FloatProxy<float>(std::numeric_limits<float>::infinity())).getAsFloat(),
394 Eq(-std::numeric_limits<float>::infinity()));
395 EXPECT_THAT((-FloatProxy<float>(-std::numeric_limits<float>::infinity()))
397 Eq(std::numeric_limits<float>::infinity()));
400 // Test conversion of FloatProxy values to strings.
402 // In previous cases, we always wrapped the FloatProxy value in a HexFloat
403 // before conversion to a string. In the following cases, the FloatProxy
404 // decides for itself whether to print as a regular number or as a hex float.
406 using FloatProxyFloatTest =
407 ::testing::TestWithParam<std::pair<FloatProxy<float>, std::string>>;
408 using FloatProxyDoubleTest =
409 ::testing::TestWithParam<std::pair<FloatProxy<double>, std::string>>;
411 // Converts a float value to a string via a FloatProxy.
412 template <typename T>
413 std::string EncodeViaFloatProxy(const T& value) {
414 std::stringstream ss;
419 // Converts a floating point string so that the exponent prefix
420 // is 'e', and the exponent value does not have leading zeros.
421 // The Microsoft runtime library likes to write things like "2.5E+010".
422 // Convert that to "2.5e+10".
423 // We don't care what happens to strings that are not floating point
425 std::string NormalizeExponentInFloatString(std::string in) {
427 // Reserve one spot for the terminating null, even when the sscanf fails.
428 std::vector<char> prefix(in.size() + 1);
431 int exponent; // in base 10
432 if ((4 == std::sscanf(in.c_str(), "%[-+.0123456789]%c%c%d", prefix.data(), &e,
433 &plus_or_minus, &exponent)) &&
434 (e == 'e' || e == 'E') &&
435 (plus_or_minus == '-' || plus_or_minus == '+')) {
436 // It looks like a floating point value with exponent.
437 std::stringstream out;
438 out << prefix.data() << 'e' << plus_or_minus << exponent;
446 TEST(NormalizeFloat, Sample) {
447 EXPECT_THAT(NormalizeExponentInFloatString(""), Eq(""));
448 EXPECT_THAT(NormalizeExponentInFloatString("1e-12"), Eq("1e-12"));
449 EXPECT_THAT(NormalizeExponentInFloatString("1E+14"), Eq("1e+14"));
450 EXPECT_THAT(NormalizeExponentInFloatString("1e-0012"), Eq("1e-12"));
451 EXPECT_THAT(NormalizeExponentInFloatString("1.263E+014"), Eq("1.263e+14"));
454 // The following two tests can't be DRY because they take different parameter
456 TEST_P(FloatProxyFloatTest, EncodeCorrectly) {
458 NormalizeExponentInFloatString(EncodeViaFloatProxy(GetParam().first)),
459 Eq(GetParam().second));
462 TEST_P(FloatProxyDoubleTest, EncodeCorrectly) {
464 NormalizeExponentInFloatString(EncodeViaFloatProxy(GetParam().first)),
465 Eq(GetParam().second));
468 INSTANTIATE_TEST_CASE_P(
469 Float32Tests, FloatProxyFloatTest,
470 ::testing::ValuesIn(std::vector<std::pair<FloatProxy<float>, std::string>>({
478 // Still normal numbers, but with large magnitude exponents.
479 {float(ldexp(1.f, 126)), "8.50706e+37"},
480 {float(ldexp(-1.f, -126)), "-1.17549e-38"},
482 // denormalized values are printed as hex floats.
483 {float(ldexp(1.0f, -127)), "0x1p-127"},
484 {float(ldexp(1.5f, -128)), "0x1.8p-128"},
485 {float(ldexp(1.25, -129)), "0x1.4p-129"},
486 {float(ldexp(1.125, -130)), "0x1.2p-130"},
487 {float(ldexp(-1.0f, -127)), "-0x1p-127"},
488 {float(ldexp(-1.0f, -128)), "-0x1p-128"},
489 {float(ldexp(-1.0f, -129)), "-0x1p-129"},
490 {float(ldexp(-1.5f, -130)), "-0x1.8p-130"},
493 {FloatProxy<float>(uint32_t(0xFFC00000)), "-0x1.8p+128"},
494 {FloatProxy<float>(uint32_t(0xFF800100)), "-0x1.0002p+128"},
496 {std::numeric_limits<float>::infinity(), "0x1p+128"},
497 {-std::numeric_limits<float>::infinity(), "-0x1p+128"},
500 INSTANTIATE_TEST_CASE_P(
501 Float64Tests, FloatProxyDoubleTest,
503 std::vector<std::pair<FloatProxy<double>, std::string>>({
509 // Large outside the range of normal floats
510 {ldexp(1.0, 128), "3.40282366920938e+38"},
511 {ldexp(1.5, 129), "1.02084710076282e+39"},
512 {ldexp(-1.0, 128), "-3.40282366920938e+38"},
513 {ldexp(-1.5, 129), "-1.02084710076282e+39"},
515 // Small outside the range of normal floats
516 {ldexp(1.5, -129), "2.20405190779179e-39"},
517 {ldexp(-1.5, -129), "-2.20405190779179e-39"},
520 {ldexp(1.0, -1022), "2.2250738585072e-308"},
521 {ldexp(-1.0, -1022), "-2.2250738585072e-308"},
523 // Denormalized values
524 {ldexp(1.125, -1023), "0x1.2p-1023"},
525 {ldexp(-1.375, -1024), "-0x1.6p-1024"},
528 {uint64_t(0x7FF8000000000000LL), "0x1.8p+1024"},
529 {uint64_t(0xFFF0F00000000000LL), "-0x1.0fp+1024"},
532 {std::numeric_limits<double>::infinity(), "0x1p+1024"},
533 {-std::numeric_limits<double>::infinity(), "-0x1p+1024"},
537 // double is used so that unbiased_exponent can be used with the output
538 // of ldexp directly.
539 int32_t unbiased_exponent(double f) {
540 return spvutils::HexFloat<spvutils::FloatProxy<float>>(
541 static_cast<float>(f)).getUnbiasedNormalizedExponent();
544 int16_t unbiased_half_exponent(uint16_t f) {
545 return spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>(f)
546 .getUnbiasedNormalizedExponent();
549 TEST(HexFloatOperationTest, UnbiasedExponent) {
551 EXPECT_EQ(0, unbiased_exponent(ldexp(1.0f, 0)));
552 EXPECT_EQ(-32, unbiased_exponent(ldexp(1.0f, -32)));
553 EXPECT_EQ(42, unbiased_exponent(ldexp(1.0f, 42)));
554 EXPECT_EQ(125, unbiased_exponent(ldexp(1.0f, 125)));
556 EXPECT_EQ(128, unbiased_exponent(ldexp(1.0f, 256)));
558 EXPECT_EQ(-100, unbiased_exponent(ldexp(1.0f, -100)));
559 EXPECT_EQ(-127, unbiased_exponent(ldexp(1.0f, -127))); // First denorm
560 EXPECT_EQ(-128, unbiased_exponent(ldexp(1.0f, -128)));
561 EXPECT_EQ(-129, unbiased_exponent(ldexp(1.0f, -129)));
562 EXPECT_EQ(-140, unbiased_exponent(ldexp(1.0f, -140)));
563 // Smallest representable number
564 EXPECT_EQ(-126 - 23, unbiased_exponent(ldexp(1.0f, -126 - 23)));
565 // Should get rounded to 0 first.
566 EXPECT_EQ(0, unbiased_exponent(ldexp(1.0f, -127 - 23)));
569 // The exponent is represented in the bits 0x7C00
571 EXPECT_EQ(0, unbiased_half_exponent(0x3C00));
572 EXPECT_EQ(3, unbiased_half_exponent(0x4800));
573 EXPECT_EQ(-1, unbiased_half_exponent(0x3800));
574 EXPECT_EQ(-14, unbiased_half_exponent(0x0400));
575 EXPECT_EQ(16, unbiased_half_exponent(0x7C00));
576 EXPECT_EQ(10, unbiased_half_exponent(0x6400));
578 // Smallest representable number
579 EXPECT_EQ(-24, unbiased_half_exponent(0x0001));
582 // Creates a float that is the sum of 1/(2 ^ fractions[i]) for i in factions
583 float float_fractions(const std::vector<uint32_t>& fractions) {
585 for(int32_t i: fractions) {
586 f += std::ldexp(1.0f, -i);
591 // Returns the normalized significand of a HexFloat<FloatProxy<float>>
592 // that was created by calling float_fractions with the input fractions,
593 // raised to the power of exp.
594 uint32_t normalized_significand(const std::vector<uint32_t>& fractions, uint32_t exp) {
595 return spvutils::HexFloat<spvutils::FloatProxy<float>>(
596 static_cast<float>(ldexp(float_fractions(fractions), exp)))
597 .getNormalizedSignificand();
600 // Sets the bits from MSB to LSB of the significand part of a float.
601 // For example 0 would set the bit 23 (counting from LSB to MSB),
602 // and 1 would set the 22nd bit.
603 uint32_t bits_set(const std::vector<uint32_t>& bits) {
604 const uint32_t top_bit = 1u << 22u;
606 for(uint32_t i: bits) {
612 // The same as bits_set but for a Float16 value instead of 32-bit floating
614 uint16_t half_bits_set(const std::vector<uint32_t>& bits) {
615 const uint32_t top_bit = 1u << 9u;
617 for(uint32_t i: bits) {
620 return static_cast<uint16_t>(val);
623 TEST(HexFloatOperationTest, NormalizedSignificand) {
624 // For normalized numbers (the following) it should be a simple matter
625 // of getting rid of the top implicit bit
626 EXPECT_EQ(bits_set({}), normalized_significand({0}, 0));
627 EXPECT_EQ(bits_set({0}), normalized_significand({0, 1}, 0));
628 EXPECT_EQ(bits_set({0, 1}), normalized_significand({0, 1, 2}, 0));
629 EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 0));
630 EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 32));
631 EXPECT_EQ(bits_set({1}), normalized_significand({0, 2}, 126));
633 // For denormalized numbers we expect the normalized significand to
634 // shift as if it were normalized. This means, in practice that the
635 // top_most set bit will be cut off. Looks very similar to above (on purpose)
636 EXPECT_EQ(bits_set({}), normalized_significand({0}, -127));
637 EXPECT_EQ(bits_set({3}), normalized_significand({0, 4}, -128));
638 EXPECT_EQ(bits_set({3}), normalized_significand({0, 4}, -127));
639 EXPECT_EQ(bits_set({}), normalized_significand({22}, -127));
640 EXPECT_EQ(bits_set({0}), normalized_significand({21, 22}, -127));
643 // Returns the 32-bit floating point value created by
644 // calling setFromSignUnbiasedExponentAndNormalizedSignificand
645 // on a HexFloat<FloatProxy<float>>
646 float set_from_sign(bool negative, int32_t unbiased_exponent,
647 uint32_t significand, bool round_denorm_up) {
648 spvutils::HexFloat<spvutils::FloatProxy<float>> f(0.f);
649 f.setFromSignUnbiasedExponentAndNormalizedSignificand(
650 negative, unbiased_exponent, significand, round_denorm_up);
651 return f.value().getAsFloat();
654 TEST(HexFloatOperationTests,
655 SetFromSignUnbiasedExponentAndNormalizedSignificand) {
657 EXPECT_EQ(1.f, set_from_sign(false, 0, 0, false));
659 // Tests insertion of various denormalized numbers with and without round up.
660 EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), set_from_sign(false, -149, 0, false));
661 EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), set_from_sign(false, -149, 0, true));
662 EXPECT_EQ(0.f, set_from_sign(false, -150, 1, false));
663 EXPECT_EQ(static_cast<float>(ldexp(1.f, -149)), set_from_sign(false, -150, 1, true));
665 EXPECT_EQ(ldexp(1.0f, -127), set_from_sign(false, -127, 0, false));
666 EXPECT_EQ(ldexp(1.0f, -128), set_from_sign(false, -128, 0, false));
667 EXPECT_EQ(float_fractions({0, 1, 2, 5}),
668 set_from_sign(false, 0, bits_set({0, 1, 4}), false));
669 EXPECT_EQ(ldexp(float_fractions({0, 1, 2, 5}), -32),
670 set_from_sign(false, -32, bits_set({0, 1, 4}), false));
671 EXPECT_EQ(ldexp(float_fractions({0, 1, 2, 5}), -128),
672 set_from_sign(false, -128, bits_set({0, 1, 4}), false));
674 // The negative cases from above.
675 EXPECT_EQ(-1.f, set_from_sign(true, 0, 0, false));
676 EXPECT_EQ(-ldexp(1.0, -127), set_from_sign(true, -127, 0, false));
677 EXPECT_EQ(-ldexp(1.0, -128), set_from_sign(true, -128, 0, false));
678 EXPECT_EQ(-float_fractions({0, 1, 2, 5}),
679 set_from_sign(true, 0, bits_set({0, 1, 4}), false));
680 EXPECT_EQ(-ldexp(float_fractions({0, 1, 2, 5}), -32),
681 set_from_sign(true, -32, bits_set({0, 1, 4}), false));
682 EXPECT_EQ(-ldexp(float_fractions({0, 1, 2, 5}), -128),
683 set_from_sign(true, -128, bits_set({0, 1, 4}), false));
686 TEST(HexFloatOperationTests, NonRounding) {
687 // Rounding from 32-bit hex-float to 32-bit hex-float should be trivial,
688 // except in the denorm case which is a bit more complex.
689 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
690 bool carry_bit = false;
692 spvutils::round_direction rounding[] = {
693 spvutils::round_direction::kToZero,
694 spvutils::round_direction::kToNearestEven,
695 spvutils::round_direction::kToPositiveInfinity,
696 spvutils::round_direction::kToNegativeInfinity};
698 // Everything fits, so this should be straight-forward
699 for (spvutils::round_direction round : rounding) {
700 EXPECT_EQ(bits_set({}), HF(0.f).getRoundedNormalizedSignificand<HF>(
702 EXPECT_FALSE(carry_bit);
704 EXPECT_EQ(bits_set({0}),
705 HF(float_fractions({0, 1}))
706 .getRoundedNormalizedSignificand<HF>(round, &carry_bit));
707 EXPECT_FALSE(carry_bit);
709 EXPECT_EQ(bits_set({1, 3}),
710 HF(float_fractions({0, 2, 4}))
711 .getRoundedNormalizedSignificand<HF>(round, &carry_bit));
712 EXPECT_FALSE(carry_bit);
716 HF(static_cast<float>(-ldexp(float_fractions({0, 1, 2, 5}), -128)))
717 .getRoundedNormalizedSignificand<HF>(round, &carry_bit));
718 EXPECT_FALSE(carry_bit);
721 bits_set({0, 1, 4, 22}),
722 HF(static_cast<float>(float_fractions({0, 1, 2, 5, 23})))
723 .getRoundedNormalizedSignificand<HF>(round, &carry_bit));
724 EXPECT_FALSE(carry_bit);
728 using RD = spvutils::round_direction;
729 struct RoundSignificandCase {
731 std::pair<int16_t, bool> expected_results;
732 spvutils::round_direction round;
735 using HexFloatRoundTest =
736 ::testing::TestWithParam<RoundSignificandCase>;
738 TEST_P(HexFloatRoundTest, RoundDownToFP16) {
739 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
740 using HF16 = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>;
742 HF input_value(GetParam().source_float);
743 bool carry_bit = false;
744 EXPECT_EQ(GetParam().expected_results.first,
745 input_value.getRoundedNormalizedSignificand<HF16>(
746 GetParam().round, &carry_bit));
747 EXPECT_EQ(carry_bit, GetParam().expected_results.second);
751 INSTANTIATE_TEST_CASE_P(F32ToF16, HexFloatRoundTest,
752 ::testing::ValuesIn(std::vector<RoundSignificandCase>(
754 {float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToZero},
755 {float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToNearestEven},
756 {float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToPositiveInfinity},
757 {float_fractions({0}), std::make_pair(half_bits_set({}), false), RD::kToNegativeInfinity},
758 {float_fractions({0, 1}), std::make_pair(half_bits_set({0}), false), RD::kToZero},
760 {float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToZero},
761 {float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity},
762 {float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity},
763 {float_fractions({0, 1, 11}), std::make_pair(half_bits_set({0}), false), RD::kToNearestEven},
765 {float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToZero},
766 {float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 8}), false), RD::kToPositiveInfinity},
767 {float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNegativeInfinity},
768 {float_fractions({0, 1, 10, 11}), std::make_pair(half_bits_set({0, 8}), false), RD::kToNearestEven},
770 {float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToZero},
771 {float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity},
772 {float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity},
773 {float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven},
775 {-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToZero},
776 {-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0}), false), RD::kToPositiveInfinity},
777 {-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNegativeInfinity},
778 {-float_fractions({0, 1, 11, 12}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven},
780 {float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0}), false), RD::kToZero},
781 {float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity},
782 {float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity},
783 {float_fractions({0, 1, 11, 22}), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven},
786 {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},
787 {float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({}), true), RD::kToPositiveInfinity},
788 {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},
789 {float_fractions({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}), std::make_pair(half_bits_set({}), true), RD::kToNearestEven},
791 // Cases where original number was denorm. Note: this should have no effect
792 // the number is pre-normalized.
793 {static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -128)), std::make_pair(half_bits_set({0}), false), RD::kToZero},
794 {static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -129)), std::make_pair(half_bits_set({0, 9}), false), RD::kToPositiveInfinity},
795 {static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -131)), std::make_pair(half_bits_set({0}), false), RD::kToNegativeInfinity},
796 {static_cast<float>(ldexp(float_fractions({0, 1, 11, 13}), -130)), std::make_pair(half_bits_set({0, 9}), false), RD::kToNearestEven},
800 struct UpCastSignificandCase {
801 uint16_t source_half;
802 uint32_t expected_result;
805 using HexFloatRoundUpSignificandTest =
806 ::testing::TestWithParam<UpCastSignificandCase>;
807 TEST_P(HexFloatRoundUpSignificandTest, Widening) {
808 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
809 using HF16 = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>;
810 bool carry_bit = false;
812 spvutils::round_direction rounding[] = {
813 spvutils::round_direction::kToZero,
814 spvutils::round_direction::kToNearestEven,
815 spvutils::round_direction::kToPositiveInfinity,
816 spvutils::round_direction::kToNegativeInfinity};
818 // Everything fits, so everything should just be bit-shifts.
819 for (spvutils::round_direction round : rounding) {
821 HF16 input_value(GetParam().source_half);
823 GetParam().expected_result,
824 input_value.getRoundedNormalizedSignificand<HF>(round, &carry_bit))
826 << input_value.getRoundedNormalizedSignificand<HF>(round, &carry_bit)
827 << " 0x" << GetParam().expected_result;
828 EXPECT_FALSE(carry_bit);
832 INSTANTIATE_TEST_CASE_P(F16toF32, HexFloatRoundUpSignificandTest,
833 // 0xFC00 of the source 16-bit hex value cover the sign and the exponent.
834 // They are ignored for this test.
835 ::testing::ValuesIn(std::vector<UpCastSignificandCase>(
843 struct DownCastTest {
845 uint16_t expected_half;
846 std::vector<spvutils::round_direction> directions;
849 std::string get_round_text(spvutils::round_direction direction) {
850 #define CASE(round_direction) \
851 case round_direction: \
852 return #round_direction
855 CASE(spvutils::round_direction::kToZero);
856 CASE(spvutils::round_direction::kToPositiveInfinity);
857 CASE(spvutils::round_direction::kToNegativeInfinity);
858 CASE(spvutils::round_direction::kToNearestEven);
864 using HexFloatFP32To16Tests = ::testing::TestWithParam<DownCastTest>;
866 TEST_P(HexFloatFP32To16Tests, NarrowingCasts) {
867 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
868 using HF16 = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>;
869 HF f(GetParam().source_float);
870 for (auto round : GetParam().directions) {
872 f.castTo(half, round);
873 EXPECT_EQ(GetParam().expected_half, half.value().getAsFloat().get_value())
874 << get_round_text(round) << " " << std::hex
875 << spvutils::BitwiseCast<uint32_t>(GetParam().source_float)
876 << " cast to: " << half.value().getAsFloat().get_value();
880 const uint16_t positive_infinity = 0x7C00;
881 const uint16_t negative_infinity = 0xFC00;
883 INSTANTIATE_TEST_CASE_P(F32ToF16, HexFloatFP32To16Tests,
884 ::testing::ValuesIn(std::vector<DownCastTest>(
886 // Exactly representable as half.
887 {0.f, 0x0, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
888 {-0.f, 0x8000, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
889 {1.0f, 0x3C00, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
890 {-1.0f, 0xBC00, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
892 {float_fractions({0, 1, 10}) , 0x3E01, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
893 {-float_fractions({0, 1, 10}) , 0xBE01, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
894 {static_cast<float>(ldexp(float_fractions({0, 1, 10}), 3)), 0x4A01, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
895 {static_cast<float>(-ldexp(float_fractions({0, 1, 10}), 3)), 0xCA01, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
899 {static_cast<float>(ldexp(1.0f, -25)), 0x0, {RD::kToZero, RD::kToNegativeInfinity, RD::kToNearestEven}},
900 {static_cast<float>(ldexp(1.0f, -25)), 0x1, {RD::kToPositiveInfinity}},
901 {static_cast<float>(-ldexp(1.0f, -25)), 0x8000, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNearestEven}},
902 {static_cast<float>(-ldexp(1.0f, -25)), 0x8001, {RD::kToNegativeInfinity}},
903 {static_cast<float>(ldexp(1.0f, -24)), 0x1, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
906 {static_cast<float>(ldexp(1.0f, 16)), positive_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
907 {static_cast<float>(ldexp(1.0f, 18)), positive_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
908 {static_cast<float>(ldexp(1.3f, 16)), positive_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
909 {static_cast<float>(-ldexp(1.0f, 16)), negative_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
910 {static_cast<float>(-ldexp(1.0f, 18)), negative_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
911 {static_cast<float>(-ldexp(1.3f, 16)), negative_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
913 // Transfer of Infinities
914 {std::numeric_limits<float>::infinity(), positive_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
915 {-std::numeric_limits<float>::infinity(), negative_infinity, {RD::kToZero, RD::kToPositiveInfinity, RD::kToNegativeInfinity, RD::kToNearestEven}},
917 // Nans are below because we cannot test for equality.
921 uint16_t source_half;
922 float expected_float;
925 using HexFloatFP16To32Tests = ::testing::TestWithParam<UpCastCase>;
926 TEST_P(HexFloatFP16To32Tests, WideningCasts) {
927 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
928 using HF16 = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>;
929 HF16 f(GetParam().source_half);
931 spvutils::round_direction rounding[] = {
932 spvutils::round_direction::kToZero,
933 spvutils::round_direction::kToNearestEven,
934 spvutils::round_direction::kToPositiveInfinity,
935 spvutils::round_direction::kToNegativeInfinity};
937 // Everything fits, so everything should just be bit-shifts.
938 for (spvutils::round_direction round : rounding) {
940 f.castTo(flt, round);
941 EXPECT_EQ(GetParam().expected_float, flt.value().getAsFloat())
942 << get_round_text(round) << " " << std::hex
943 << spvutils::BitwiseCast<uint16_t>(GetParam().source_half)
944 << " cast to: " << flt.value().getAsFloat();
948 INSTANTIATE_TEST_CASE_P(F16ToF32, HexFloatFP16To32Tests,
949 ::testing::ValuesIn(std::vector<UpCastCase>(
955 {0x3F00, float_fractions({0, 1, 2})},
956 {0xBF00, -float_fractions({0, 1, 2})},
957 {0x3F01, float_fractions({0, 1, 2, 10})},
958 {0xBF01, -float_fractions({0, 1, 2, 10})},
961 {0x0001, static_cast<float>(ldexp(1.0, -24))},
962 {0x0002, static_cast<float>(ldexp(1.0, -23))},
963 {0x8001, static_cast<float>(-ldexp(1.0, -24))},
964 {0x8011, static_cast<float>(-ldexp(1.0, -20) + -ldexp(1.0, -24))},
967 {0x7C00, std::numeric_limits<float>::infinity()},
968 {0xFC00, -std::numeric_limits<float>::infinity()},
971 TEST(HexFloatOperationTests, NanTests) {
972 using HF = spvutils::HexFloat<spvutils::FloatProxy<float>>;
973 using HF16 = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>>;
974 spvutils::round_direction rounding[] = {
975 spvutils::round_direction::kToZero,
976 spvutils::round_direction::kToNearestEven,
977 spvutils::round_direction::kToPositiveInfinity,
978 spvutils::round_direction::kToNegativeInfinity};
980 // Everything fits, so everything should just be bit-shifts.
981 for (spvutils::round_direction round : rounding) {
984 HF(std::numeric_limits<float>::quiet_NaN()).castTo(f16, round);
985 EXPECT_TRUE(f16.value().isNan());
986 HF(std::numeric_limits<float>::signaling_NaN()).castTo(f16, round);
987 EXPECT_TRUE(f16.value().isNan());
989 HF16(0x7C01).castTo(f, round);
990 EXPECT_TRUE(f.value().isNan());
991 HF16(0x7C11).castTo(f, round);
992 EXPECT_TRUE(f.value().isNan());
993 HF16(0xFC01).castTo(f, round);
994 EXPECT_TRUE(f.value().isNan());
995 HF16(0x7C10).castTo(f, round);
996 EXPECT_TRUE(f.value().isNan());
997 HF16(0xFF00).castTo(f, round);
998 EXPECT_TRUE(f.value().isNan());
1002 // A test case for parsing good and bad HexFloat<FloatProxy<T>> literals.
1003 template <typename T>
1004 struct FloatParseCase {
1005 std::string literal;
1007 bool expect_success;
1008 HexFloat<FloatProxy<T>> expected_value;
1011 using ParseNormalFloatTest = ::testing::TestWithParam<FloatParseCase<float>>;
1013 TEST_P(ParseNormalFloatTest, Samples) {
1014 std::stringstream input(GetParam().literal);
1015 HexFloat<FloatProxy<float>> parsed_value(0.0f);
1016 ParseNormalFloat(input, GetParam().negate_value, parsed_value);
1017 EXPECT_NE(GetParam().expect_success, input.fail())
1018 << " literal: " << GetParam().literal
1019 << " negate: " << GetParam().negate_value;
1020 if (GetParam().expect_success) {
1021 EXPECT_THAT(parsed_value.value(), Eq(GetParam().expected_value.value()))
1022 << " literal: " << GetParam().literal
1023 << " negate: " << GetParam().negate_value;
1027 // Returns a FloatParseCase with expected failure.
1028 template <typename T>
1029 FloatParseCase<T> BadFloatParseCase(std::string literal, bool negate_value,
1031 HexFloat<FloatProxy<T>> proxy_expected_value(expected_value);
1032 return FloatParseCase<T>{literal, negate_value, false, proxy_expected_value};
1035 // Returns a FloatParseCase that should successfully parse to a given value.
1036 template <typename T>
1037 FloatParseCase<T> GoodFloatParseCase(std::string literal, bool negate_value,
1039 HexFloat<FloatProxy<T>> proxy_expected_value(expected_value);
1040 return FloatParseCase<T>{literal, negate_value, true, proxy_expected_value};
1043 INSTANTIATE_TEST_CASE_P(
1044 FloatParse, ParseNormalFloatTest,
1045 ::testing::ValuesIn(std::vector<FloatParseCase<float>>{
1046 // Failing cases due to trivially incorrect syntax.
1047 BadFloatParseCase("abc", false, 0.0f),
1048 BadFloatParseCase("abc", true, 0.0f),
1051 GoodFloatParseCase("0", false, 0.0f),
1052 GoodFloatParseCase("0.0", false, 0.0f),
1053 GoodFloatParseCase("-0.0", false, -0.0f),
1054 GoodFloatParseCase("2.0", false, 2.0f),
1055 GoodFloatParseCase("-2.0", false, -2.0f),
1056 GoodFloatParseCase("+2.0", false, 2.0f),
1057 // Cases with negate_value being true.
1058 GoodFloatParseCase("0.0", true, -0.0f),
1059 GoodFloatParseCase("2.0", true, -2.0f),
1061 // When negate_value is true, we should not accept a
1062 // leading minus or plus.
1063 BadFloatParseCase("-0.0", true, 0.0f),
1064 BadFloatParseCase("-2.0", true, 0.0f),
1065 BadFloatParseCase("+0.0", true, 0.0f),
1066 BadFloatParseCase("+2.0", true, 0.0f),
1068 // Overflow is an error for 32-bit float parsing.
1069 BadFloatParseCase("1e40", false, FLT_MAX),
1070 BadFloatParseCase("1e40", true, -FLT_MAX),
1071 BadFloatParseCase("-1e40", false, -FLT_MAX),
1072 // We can't have -1e40 and negate_value == true since
1073 // that represents an original case of "--1e40" which
1077 using ParseNormalFloat16Test =
1078 ::testing::TestWithParam<FloatParseCase<Float16>>;
1080 TEST_P(ParseNormalFloat16Test, Samples) {
1081 std::stringstream input(GetParam().literal);
1082 HexFloat<FloatProxy<Float16>> parsed_value(0);
1083 ParseNormalFloat(input, GetParam().negate_value, parsed_value);
1084 EXPECT_NE(GetParam().expect_success, input.fail())
1085 << " literal: " << GetParam().literal
1086 << " negate: " << GetParam().negate_value;
1087 if (GetParam().expect_success) {
1088 EXPECT_THAT(parsed_value.value(), Eq(GetParam().expected_value.value()))
1089 << " literal: " << GetParam().literal
1090 << " negate: " << GetParam().negate_value;
1094 INSTANTIATE_TEST_CASE_P(
1095 Float16Parse, ParseNormalFloat16Test,
1096 ::testing::ValuesIn(std::vector<FloatParseCase<Float16>>{
1097 // Failing cases due to trivially incorrect syntax.
1098 BadFloatParseCase<Float16>("abc", false, uint16_t{0}),
1099 BadFloatParseCase<Float16>("abc", true, uint16_t{0}),
1102 GoodFloatParseCase<Float16>("0", false, uint16_t{0}),
1103 GoodFloatParseCase<Float16>("0.0", false, uint16_t{0}),
1104 GoodFloatParseCase<Float16>("-0.0", false, uint16_t{0x8000}),
1105 GoodFloatParseCase<Float16>("2.0", false, uint16_t{0x4000}),
1106 GoodFloatParseCase<Float16>("-2.0", false, uint16_t{0xc000}),
1107 GoodFloatParseCase<Float16>("+2.0", false, uint16_t{0x4000}),
1108 // Cases with negate_value being true.
1109 GoodFloatParseCase<Float16>("0.0", true, uint16_t{0x8000}),
1110 GoodFloatParseCase<Float16>("2.0", true, uint16_t{0xc000}),
1112 // When negate_value is true, we should not accept a leading minus or
1114 BadFloatParseCase<Float16>("-0.0", true, uint16_t{0}),
1115 BadFloatParseCase<Float16>("-2.0", true, uint16_t{0}),
1116 BadFloatParseCase<Float16>("+0.0", true, uint16_t{0}),
1117 BadFloatParseCase<Float16>("+2.0", true, uint16_t{0}),
1120 // A test case for detecting infinities.
1121 template <typename T>
1122 struct OverflowParseCase {
1124 bool expect_success;
1128 using FloatProxyParseOverflowFloatTest =
1129 ::testing::TestWithParam<OverflowParseCase<float>>;
1131 TEST_P(FloatProxyParseOverflowFloatTest, Sample) {
1132 std::istringstream input(GetParam().input);
1133 HexFloat<FloatProxy<float>> value(0.0f);
1135 EXPECT_NE(GetParam().expect_success, input.fail());
1136 if (GetParam().expect_success) {
1137 EXPECT_THAT(value.value().getAsFloat(), GetParam().expected_value);
1141 INSTANTIATE_TEST_CASE_P(
1142 FloatOverflow, FloatProxyParseOverflowFloatTest,
1143 ::testing::ValuesIn(std::vector<OverflowParseCase<float>>({
1145 {"0.0", true, 0.0f},
1146 {"1.0", true, 1.0f},
1147 {"1e38", true, 1e38f},
1148 {"-1e38", true, -1e38f},
1149 {"1e40", false, FLT_MAX},
1150 {"-1e40", false, -FLT_MAX},
1151 {"1e400", false, FLT_MAX},
1152 {"-1e400", false, -FLT_MAX},
1155 using FloatProxyParseOverflowDoubleTest =
1156 ::testing::TestWithParam<OverflowParseCase<double>>;
1158 TEST_P(FloatProxyParseOverflowDoubleTest, Sample) {
1159 std::istringstream input(GetParam().input);
1160 HexFloat<FloatProxy<double>> value(0.0);
1162 EXPECT_NE(GetParam().expect_success, input.fail());
1163 if (GetParam().expect_success) {
1164 EXPECT_THAT(value.value().getAsFloat(), Eq(GetParam().expected_value));
1168 INSTANTIATE_TEST_CASE_P(
1169 DoubleOverflow, FloatProxyParseOverflowDoubleTest,
1170 ::testing::ValuesIn(std::vector<OverflowParseCase<double>>({
1174 {"1e38", true, 1e38},
1175 {"-1e38", true, -1e38},
1176 {"1e40", true, 1e40},
1177 {"-1e40", true, -1e40},
1178 {"1e400", false, DBL_MAX},
1179 {"-1e400", false, -DBL_MAX},
1182 using FloatProxyParseOverflowFloat16Test =
1183 ::testing::TestWithParam<OverflowParseCase<uint16_t>>;
1185 TEST_P(FloatProxyParseOverflowFloat16Test, Sample) {
1186 std::istringstream input(GetParam().input);
1187 HexFloat<FloatProxy<Float16>> value(0);
1189 EXPECT_NE(GetParam().expect_success, input.fail()) << " literal: "
1190 << GetParam().input;
1191 if (GetParam().expect_success) {
1192 EXPECT_THAT(value.value().data(), Eq(GetParam().expected_value))
1193 << " literal: " << GetParam().input;
1197 INSTANTIATE_TEST_CASE_P(
1198 Float16Overflow, FloatProxyParseOverflowFloat16Test,
1199 ::testing::ValuesIn(std::vector<OverflowParseCase<uint16_t>>({
1200 {"0", true, uint16_t{0}},
1201 {"0.0", true, uint16_t{0}},
1202 {"1.0", true, uint16_t{0x3c00}},
1203 // Overflow for 16-bit float is an error, and returns max or
1205 {"1e38", false, uint16_t{0x7bff}},
1206 {"1e40", false, uint16_t{0x7bff}},
1207 {"1e400", false, uint16_t{0x7bff}},
1208 {"-1e38", false, uint16_t{0xfbff}},
1209 {"-1e40", false, uint16_t{0xfbff}},
1210 {"-1e400", false, uint16_t{0xfbff}},
1213 TEST(FloatProxy, Max) {
1214 EXPECT_THAT(FloatProxy<Float16>::max().getAsFloat().get_value(),
1215 Eq(uint16_t{0x7bff}));
1216 EXPECT_THAT(FloatProxy<float>::max().getAsFloat(),
1217 Eq(std::numeric_limits<float>::max()));
1218 EXPECT_THAT(FloatProxy<double>::max().getAsFloat(),
1219 Eq(std::numeric_limits<double>::max()));
1222 TEST(FloatProxy, Lowest) {
1223 EXPECT_THAT(FloatProxy<Float16>::lowest().getAsFloat().get_value(),
1224 Eq(uint16_t{0xfbff}));
1225 EXPECT_THAT(FloatProxy<float>::lowest().getAsFloat(),
1226 Eq(std::numeric_limits<float>::lowest()));
1227 EXPECT_THAT(FloatProxy<double>::lowest().getAsFloat(),
1228 Eq(std::numeric_limits<double>::lowest()));
1231 // TODO(awoloszyn): Add fp16 tests and HexFloatTraits.
1232 } // anonymous namespace