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 | #ifndef LIBSPIRV_UTIL_HEX_FLOAT_H_ |
16 | #define LIBSPIRV_UTIL_HEX_FLOAT_H_ |
17 | |
18 | #include <cassert> |
19 | #include <cctype> |
20 | #include <cmath> |
21 | #include <cstdint> |
22 | #include <iomanip> |
23 | #include <limits> |
24 | #include <sstream> |
25 | |
26 | #include "bitutils.h" |
27 | |
28 | namespace spvutils { |
29 | |
30 | class Float16 { |
31 | public: |
32 | Float16(uint16_t v) : val(v) {} |
33 | Float16() {} |
34 | static bool isNan(const Float16& val) { |
35 | return ((val.val & 0x7C00) == 0x7C00) && ((val.val & 0x3FF) != 0); |
36 | } |
37 | // Returns true if the given value is any kind of infinity. |
38 | static bool isInfinity(const Float16& val) { |
39 | return ((val.val & 0x7C00) == 0x7C00) && ((val.val & 0x3FF) == 0); |
40 | } |
41 | Float16(const Float16& other) { val = other.val; } |
42 | uint16_t get_value() const { return val; } |
43 | |
44 | // Returns the maximum normal value. |
45 | static Float16 max() { return Float16(0x7bff); } |
46 | // Returns the lowest normal value. |
47 | static Float16 lowest() { return Float16(0xfbff); } |
48 | |
49 | private: |
50 | uint16_t val; |
51 | }; |
52 | |
53 | // To specialize this type, you must override uint_type to define |
54 | // an unsigned integer that can fit your floating point type. |
55 | // You must also add a isNan function that returns true if |
56 | // a value is Nan. |
57 | template <typename T> |
58 | struct FloatProxyTraits { |
59 | typedef void uint_type; |
60 | }; |
61 | |
62 | template <> |
63 | struct FloatProxyTraits<float> { |
64 | typedef uint32_t uint_type; |
65 | static bool isNan(float f) { return std::isnan(f); } |
66 | // Returns true if the given value is any kind of infinity. |
67 | static bool isInfinity(float f) { return std::isinf(f); } |
68 | // Returns the maximum normal value. |
69 | static float max() { return std::numeric_limits<float>::max(); } |
70 | // Returns the lowest normal value. |
71 | static float lowest() { return std::numeric_limits<float>::lowest(); } |
72 | }; |
73 | |
74 | template <> |
75 | struct FloatProxyTraits<double> { |
76 | typedef uint64_t uint_type; |
77 | static bool isNan(double f) { return std::isnan(f); } |
78 | // Returns true if the given value is any kind of infinity. |
79 | static bool isInfinity(double f) { return std::isinf(f); } |
80 | // Returns the maximum normal value. |
81 | static double max() { return std::numeric_limits<double>::max(); } |
82 | // Returns the lowest normal value. |
83 | static double lowest() { return std::numeric_limits<double>::lowest(); } |
84 | }; |
85 | |
86 | template <> |
87 | struct FloatProxyTraits<Float16> { |
88 | typedef uint16_t uint_type; |
89 | static bool isNan(Float16 f) { return Float16::isNan(f); } |
90 | // Returns true if the given value is any kind of infinity. |
91 | static bool isInfinity(Float16 f) { return Float16::isInfinity(f); } |
92 | // Returns the maximum normal value. |
93 | static Float16 max() { return Float16::max(); } |
94 | // Returns the lowest normal value. |
95 | static Float16 lowest() { return Float16::lowest(); } |
96 | }; |
97 | |
98 | // Since copying a floating point number (especially if it is NaN) |
99 | // does not guarantee that bits are preserved, this class lets us |
100 | // store the type and use it as a float when necessary. |
101 | template <typename T> |
102 | class FloatProxy { |
103 | public: |
104 | typedef typename FloatProxyTraits<T>::uint_type uint_type; |
105 | |
106 | // Since this is to act similar to the normal floats, |
107 | // do not initialize the data by default. |
108 | FloatProxy() {} |
109 | |
110 | // Intentionally non-explicit. This is a proxy type so |
111 | // implicit conversions allow us to use it more transparently. |
112 | FloatProxy(T val) { data_ = BitwiseCast<uint_type>(val); } |
113 | |
114 | // Intentionally non-explicit. This is a proxy type so |
115 | // implicit conversions allow us to use it more transparently. |
116 | FloatProxy(uint_type val) { data_ = val; } |
117 | |
118 | // This is helpful to have and is guaranteed not to stomp bits. |
119 | FloatProxy<T> operator-() const { |
120 | return static_cast<uint_type>(data_ ^ |
121 | (uint_type(0x1) << (sizeof(T) * 8 - 1))); |
122 | } |
123 | |
124 | // Returns the data as a floating point value. |
125 | T getAsFloat() const { return BitwiseCast<T>(data_); } |
126 | |
127 | // Returns the raw data. |
128 | uint_type data() const { return data_; } |
129 | |
130 | // Returns true if the value represents any type of NaN. |
131 | bool isNan() { return FloatProxyTraits<T>::isNan(getAsFloat()); } |
132 | // Returns true if the value represents any type of infinity. |
133 | bool isInfinity() { return FloatProxyTraits<T>::isInfinity(getAsFloat()); } |
134 | |
135 | // Returns the maximum normal value. |
136 | static FloatProxy<T> max() { |
137 | return FloatProxy<T>(FloatProxyTraits<T>::max()); |
138 | } |
139 | // Returns the lowest normal value. |
140 | static FloatProxy<T> lowest() { |
141 | return FloatProxy<T>(FloatProxyTraits<T>::lowest()); |
142 | } |
143 | |
144 | private: |
145 | uint_type data_; |
146 | }; |
147 | |
148 | template <typename T> |
149 | bool operator==(const FloatProxy<T>& first, const FloatProxy<T>& second) { |
150 | return first.data() == second.data(); |
151 | } |
152 | |
153 | // Reads a FloatProxy value as a normal float from a stream. |
154 | template <typename T> |
155 | std::istream& operator>>(std::istream& is, FloatProxy<T>& value) { |
156 | T float_val; |
157 | is >> float_val; |
158 | value = FloatProxy<T>(float_val); |
159 | return is; |
160 | } |
161 | |
162 | // This is an example traits. It is not meant to be used in practice, but will |
163 | // be the default for any non-specialized type. |
164 | template <typename T> |
165 | struct HexFloatTraits { |
166 | // Integer type that can store this hex-float. |
167 | typedef void uint_type; |
168 | // Signed integer type that can store this hex-float. |
169 | typedef void int_type; |
170 | // The numerical type that this HexFloat represents. |
171 | typedef void underlying_type; |
172 | // The type needed to construct the underlying type. |
173 | typedef void native_type; |
174 | // The number of bits that are actually relevant in the uint_type. |
175 | // This allows us to deal with, for example, 24-bit values in a 32-bit |
176 | // integer. |
177 | static const uint32_t num_used_bits = 0; |
178 | // Number of bits that represent the exponent. |
179 | static const uint32_t num_exponent_bits = 0; |
180 | // Number of bits that represent the fractional part. |
181 | static const uint32_t num_fraction_bits = 0; |
182 | // The bias of the exponent. (How much we need to subtract from the stored |
183 | // value to get the correct value.) |
184 | static const uint32_t exponent_bias = 0; |
185 | }; |
186 | |
187 | // Traits for IEEE float. |
188 | // 1 sign bit, 8 exponent bits, 23 fractional bits. |
189 | template <> |
190 | struct HexFloatTraits<FloatProxy<float>> { |
191 | typedef uint32_t uint_type; |
192 | typedef int32_t int_type; |
193 | typedef FloatProxy<float> underlying_type; |
194 | typedef float native_type; |
195 | static const uint_type num_used_bits = 32; |
196 | static const uint_type num_exponent_bits = 8; |
197 | static const uint_type num_fraction_bits = 23; |
198 | static const uint_type exponent_bias = 127; |
199 | }; |
200 | |
201 | // Traits for IEEE double. |
202 | // 1 sign bit, 11 exponent bits, 52 fractional bits. |
203 | template <> |
204 | struct HexFloatTraits<FloatProxy<double>> { |
205 | typedef uint64_t uint_type; |
206 | typedef int64_t int_type; |
207 | typedef FloatProxy<double> underlying_type; |
208 | typedef double native_type; |
209 | static const uint_type num_used_bits = 64; |
210 | static const uint_type num_exponent_bits = 11; |
211 | static const uint_type num_fraction_bits = 52; |
212 | static const uint_type exponent_bias = 1023; |
213 | }; |
214 | |
215 | // Traits for IEEE half. |
216 | // 1 sign bit, 5 exponent bits, 10 fractional bits. |
217 | template <> |
218 | struct HexFloatTraits<FloatProxy<Float16>> { |
219 | typedef uint16_t uint_type; |
220 | typedef int16_t int_type; |
221 | typedef uint16_t underlying_type; |
222 | typedef uint16_t native_type; |
223 | static const uint_type num_used_bits = 16; |
224 | static const uint_type num_exponent_bits = 5; |
225 | static const uint_type num_fraction_bits = 10; |
226 | static const uint_type exponent_bias = 15; |
227 | }; |
228 | |
229 | enum round_direction { |
230 | kRoundToZero, |
231 | kRoundToNearestEven, |
232 | kRoundToPositiveInfinity, |
233 | kRoundToNegativeInfinity |
234 | }; |
235 | |
236 | // Template class that houses a floating pointer number. |
237 | // It exposes a number of constants based on the provided traits to |
238 | // assist in interpreting the bits of the value. |
239 | template <typename T, typename Traits = HexFloatTraits<T>> |
240 | class HexFloat { |
241 | public: |
242 | typedef typename Traits::uint_type uint_type; |
243 | typedef typename Traits::int_type int_type; |
244 | typedef typename Traits::underlying_type underlying_type; |
245 | typedef typename Traits::native_type native_type; |
246 | |
247 | explicit HexFloat(T f) : value_(f) {} |
248 | |
249 | T value() const { return value_; } |
250 | void set_value(T f) { value_ = f; } |
251 | |
252 | // These are all written like this because it is convenient to have |
253 | // compile-time constants for all of these values. |
254 | |
255 | // Pass-through values to save typing. |
256 | static const uint32_t num_used_bits = Traits::num_used_bits; |
257 | static const uint32_t exponent_bias = Traits::exponent_bias; |
258 | static const uint32_t num_exponent_bits = Traits::num_exponent_bits; |
259 | static const uint32_t num_fraction_bits = Traits::num_fraction_bits; |
260 | |
261 | // Number of bits to shift left to set the highest relevant bit. |
262 | static const uint32_t top_bit_left_shift = num_used_bits - 1; |
263 | // How many nibbles (hex characters) the fractional part takes up. |
264 | static const uint32_t fraction_nibbles = (num_fraction_bits + 3) / 4; |
265 | // If the fractional part does not fit evenly into a hex character (4-bits) |
266 | // then we have to left-shift to get rid of leading 0s. This is the amount |
267 | // we have to shift (might be 0). |
268 | static const uint32_t num_overflow_bits = |
269 | fraction_nibbles * 4 - num_fraction_bits; |
270 | |
271 | // The representation of the fraction, not the actual bits. This |
272 | // includes the leading bit that is usually implicit. |
273 | static const uint_type fraction_represent_mask = |
274 | spvutils::SetBits<uint_type, 0, |
275 | num_fraction_bits + num_overflow_bits>::get; |
276 | |
277 | // The topmost bit in the nibble-aligned fraction. |
278 | static const uint_type fraction_top_bit = |
279 | uint_type(1) << (num_fraction_bits + num_overflow_bits - 1); |
280 | |
281 | // The least significant bit in the exponent, which is also the bit |
282 | // immediately to the left of the significand. |
283 | static const uint_type first_exponent_bit = uint_type(1) |
284 | << (num_fraction_bits); |
285 | |
286 | // The mask for the encoded fraction. It does not include the |
287 | // implicit bit. |
288 | static const uint_type fraction_encode_mask = |
289 | spvutils::SetBits<uint_type, 0, num_fraction_bits>::get; |
290 | |
291 | // The bit that is used as a sign. |
292 | static const uint_type sign_mask = uint_type(1) << top_bit_left_shift; |
293 | |
294 | // The bits that represent the exponent. |
295 | static const uint_type exponent_mask = |
296 | spvutils::SetBits<uint_type, num_fraction_bits, num_exponent_bits>::get; |
297 | |
298 | // How far left the exponent is shifted. |
299 | static const uint32_t exponent_left_shift = num_fraction_bits; |
300 | |
301 | // How far from the right edge the fraction is shifted. |
302 | static const uint32_t fraction_right_shift = |
303 | static_cast<uint32_t>(sizeof(uint_type) * 8) - num_fraction_bits; |
304 | |
305 | // The maximum representable unbiased exponent. |
306 | static const int_type max_exponent = |
307 | (exponent_mask >> num_fraction_bits) - exponent_bias; |
308 | // The minimum representable exponent for normalized numbers. |
309 | static const int_type min_exponent = -static_cast<int_type>(exponent_bias); |
310 | |
311 | // Returns the bits associated with the value. |
312 | uint_type getBits() const { return spvutils::BitwiseCast<uint_type>(value_); } |
313 | |
314 | // Returns the bits associated with the value, without the leading sign bit. |
315 | uint_type getUnsignedBits() const { |
316 | return static_cast<uint_type>(spvutils::BitwiseCast<uint_type>(value_) & |
317 | ~sign_mask); |
318 | } |
319 | |
320 | // Returns the bits associated with the exponent, shifted to start at the |
321 | // lsb of the type. |
322 | const uint_type getExponentBits() const { |
323 | return static_cast<uint_type>((getBits() & exponent_mask) >> |
324 | num_fraction_bits); |
325 | } |
326 | |
327 | // Returns the exponent in unbiased form. This is the exponent in the |
328 | // human-friendly form. |
329 | const int_type getUnbiasedExponent() const { |
330 | return static_cast<int_type>(getExponentBits() - exponent_bias); |
331 | } |
332 | |
333 | // Returns just the significand bits from the value. |
334 | const uint_type getSignificandBits() const { |
335 | return getBits() & fraction_encode_mask; |
336 | } |
337 | |
338 | // If the number was normalized, returns the unbiased exponent. |
339 | // If the number was denormal, normalize the exponent first. |
340 | const int_type getUnbiasedNormalizedExponent() const { |
341 | if ((getBits() & ~sign_mask) == 0) { // special case if everything is 0 |
342 | return 0; |
343 | } |
344 | int_type exp = getUnbiasedExponent(); |
345 | if (exp == min_exponent) { // We are in denorm land. |
346 | uint_type significand_bits = getSignificandBits(); |
347 | while ((significand_bits & (first_exponent_bit >> 1)) == 0) { |
348 | significand_bits = static_cast<uint_type>(significand_bits << 1); |
349 | exp = static_cast<int_type>(exp - 1); |
350 | } |
351 | significand_bits &= fraction_encode_mask; |
352 | } |
353 | return exp; |
354 | } |
355 | |
356 | // Returns the signficand after it has been normalized. |
357 | const uint_type getNormalizedSignificand() const { |
358 | int_type unbiased_exponent = getUnbiasedNormalizedExponent(); |
359 | uint_type significand = getSignificandBits(); |
360 | for (int_type i = unbiased_exponent; i <= min_exponent; ++i) { |
361 | significand = static_cast<uint_type>(significand << 1); |
362 | } |
363 | significand &= fraction_encode_mask; |
364 | return significand; |
365 | } |
366 | |
367 | // Returns true if this number represents a negative value. |
368 | bool isNegative() const { return (getBits() & sign_mask) != 0; } |
369 | |
370 | // Sets this HexFloat from the individual components. |
371 | // Note this assumes EVERY significand is normalized, and has an implicit |
372 | // leading one. This means that the only way that this method will set 0, |
373 | // is if you set a number so denormalized that it underflows. |
374 | // Do not use this method with raw bits extracted from a subnormal number, |
375 | // since subnormals do not have an implicit leading 1 in the significand. |
376 | // The significand is also expected to be in the |
377 | // lowest-most num_fraction_bits of the uint_type. |
378 | // The exponent is expected to be unbiased, meaning an exponent of |
379 | // 0 actually means 0. |
380 | // If underflow_round_up is set, then on underflow, if a number is non-0 |
381 | // and would underflow, we round up to the smallest denorm. |
382 | void setFromSignUnbiasedExponentAndNormalizedSignificand( |
383 | bool negative, int_type exponent, uint_type significand, |
384 | bool round_denorm_up) { |
385 | bool significand_is_zero = significand == 0; |
386 | |
387 | if (exponent <= min_exponent) { |
388 | // If this was denormalized, then we have to shift the bit on, meaning |
389 | // the significand is not zero. |
390 | significand_is_zero = false; |
391 | significand |= first_exponent_bit; |
392 | significand = static_cast<uint_type>(significand >> 1); |
393 | } |
394 | |
395 | while (exponent < min_exponent) { |
396 | significand = static_cast<uint_type>(significand >> 1); |
397 | ++exponent; |
398 | } |
399 | |
400 | if (exponent == min_exponent) { |
401 | if (significand == 0 && !significand_is_zero && round_denorm_up) { |
402 | significand = static_cast<uint_type>(0x1); |
403 | } |
404 | } |
405 | |
406 | uint_type new_value = 0; |
407 | if (negative) { |
408 | new_value = static_cast<uint_type>(new_value | sign_mask); |
409 | } |
410 | exponent = static_cast<int_type>(exponent + exponent_bias); |
411 | assert(exponent >= 0); |
412 | |
413 | // put it all together |
414 | exponent = static_cast<uint_type>((exponent << exponent_left_shift) & |
415 | exponent_mask); |
416 | significand = static_cast<uint_type>(significand & fraction_encode_mask); |
417 | new_value = static_cast<uint_type>(new_value | (exponent | significand)); |
418 | value_ = BitwiseCast<T>(new_value); |
419 | } |
420 | |
421 | // Increments the significand of this number by the given amount. |
422 | // If this would spill the significand into the implicit bit, |
423 | // carry is set to true and the significand is shifted to fit into |
424 | // the correct location, otherwise carry is set to false. |
425 | // All significands and to_increment are assumed to be within the bounds |
426 | // for a valid significand. |
427 | static uint_type incrementSignificand(uint_type significand, |
428 | uint_type to_increment, bool* carry) { |
429 | significand = static_cast<uint_type>(significand + to_increment); |
430 | *carry = false; |
431 | if (significand & first_exponent_bit) { |
432 | *carry = true; |
433 | // The implicit 1-bit will have carried, so we should zero-out the |
434 | // top bit and shift back. |
435 | significand = static_cast<uint_type>(significand & ~first_exponent_bit); |
436 | significand = static_cast<uint_type>(significand >> 1); |
437 | } |
438 | return significand; |
439 | } |
440 | |
441 | // These exist because MSVC throws warnings on negative right-shifts |
442 | // even if they are not going to be executed. Eg: |
443 | // constant_number < 0? 0: constant_number |
444 | // These convert the negative left-shifts into right shifts. |
445 | |
446 | template <typename int_type> |
447 | uint_type negatable_left_shift(int_type N, uint_type val) |
448 | { |
449 | if(N >= 0) |
450 | return val << N; |
451 | |
452 | return val >> -N; |
453 | } |
454 | |
455 | template <typename int_type> |
456 | uint_type negatable_right_shift(int_type N, uint_type val) |
457 | { |
458 | if(N >= 0) |
459 | return val >> N; |
460 | |
461 | return val << -N; |
462 | } |
463 | |
464 | // Returns the significand, rounded to fit in a significand in |
465 | // other_T. This is shifted so that the most significant |
466 | // bit of the rounded number lines up with the most significant bit |
467 | // of the returned significand. |
468 | template <typename other_T> |
469 | typename other_T::uint_type getRoundedNormalizedSignificand( |
470 | round_direction dir, bool* carry_bit) { |
471 | typedef typename other_T::uint_type other_uint_type; |
472 | static const int_type num_throwaway_bits = |
473 | static_cast<int_type>(num_fraction_bits) - |
474 | static_cast<int_type>(other_T::num_fraction_bits); |
475 | |
476 | static const uint_type last_significant_bit = |
477 | (num_throwaway_bits < 0) |
478 | ? 0 |
479 | : negatable_left_shift(num_throwaway_bits, 1u); |
480 | static const uint_type first_rounded_bit = |
481 | (num_throwaway_bits < 1) |
482 | ? 0 |
483 | : negatable_left_shift(num_throwaway_bits - 1, 1u); |
484 | |
485 | static const uint_type throwaway_mask_bits = |
486 | num_throwaway_bits > 0 ? num_throwaway_bits : 0; |
487 | static const uint_type throwaway_mask = |
488 | spvutils::SetBits<uint_type, 0, throwaway_mask_bits>::get; |
489 | |
490 | *carry_bit = false; |
491 | other_uint_type out_val = 0; |
492 | uint_type significand = getNormalizedSignificand(); |
493 | // If we are up-casting, then we just have to shift to the right location. |
494 | if (num_throwaway_bits <= 0) { |
495 | out_val = static_cast<other_uint_type>(significand); |
496 | uint_type shift_amount = static_cast<uint_type>(-num_throwaway_bits); |
497 | out_val = static_cast<other_uint_type>(out_val << shift_amount); |
498 | return out_val; |
499 | } |
500 | |
501 | // If every non-representable bit is 0, then we don't have any casting to |
502 | // do. |
503 | if ((significand & throwaway_mask) == 0) { |
504 | return static_cast<other_uint_type>( |
505 | negatable_right_shift(num_throwaway_bits, significand)); |
506 | } |
507 | |
508 | bool round_away_from_zero = false; |
509 | // We actually have to narrow the significand here, so we have to follow the |
510 | // rounding rules. |
511 | switch (dir) { |
512 | case kRoundToZero: |
513 | break; |
514 | case kRoundToPositiveInfinity: |
515 | round_away_from_zero = !isNegative(); |
516 | break; |
517 | case kRoundToNegativeInfinity: |
518 | round_away_from_zero = isNegative(); |
519 | break; |
520 | case kRoundToNearestEven: |
521 | // Have to round down, round bit is 0 |
522 | if ((first_rounded_bit & significand) == 0) { |
523 | break; |
524 | } |
525 | if (((significand & throwaway_mask) & ~first_rounded_bit) != 0) { |
526 | // If any subsequent bit of the rounded portion is non-0 then we round |
527 | // up. |
528 | round_away_from_zero = true; |
529 | break; |
530 | } |
531 | // We are exactly half-way between 2 numbers, pick even. |
532 | if ((significand & last_significant_bit) != 0) { |
533 | // 1 for our last bit, round up. |
534 | round_away_from_zero = true; |
535 | break; |
536 | } |
537 | break; |
538 | } |
539 | |
540 | if (round_away_from_zero) { |
541 | return static_cast<other_uint_type>( |
542 | negatable_right_shift(num_throwaway_bits, incrementSignificand( |
543 | significand, last_significant_bit, carry_bit))); |
544 | } else { |
545 | return static_cast<other_uint_type>( |
546 | negatable_right_shift(num_throwaway_bits, significand)); |
547 | } |
548 | } |
549 | |
550 | // Casts this value to another HexFloat. If the cast is widening, |
551 | // then round_dir is ignored. If the cast is narrowing, then |
552 | // the result is rounded in the direction specified. |
553 | // This number will retain Nan and Inf values. |
554 | // It will also saturate to Inf if the number overflows, and |
555 | // underflow to (0 or min depending on rounding) if the number underflows. |
556 | template <typename other_T> |
557 | void castTo(other_T& other, round_direction round_dir) { |
558 | other = other_T(static_cast<typename other_T::native_type>(0)); |
559 | bool negate = isNegative(); |
560 | if (getUnsignedBits() == 0) { |
561 | if (negate) { |
562 | other.set_value(-other.value()); |
563 | } |
564 | return; |
565 | } |
566 | uint_type significand = getSignificandBits(); |
567 | bool carried = false; |
568 | typename other_T::uint_type rounded_significand = |
569 | getRoundedNormalizedSignificand<other_T>(round_dir, &carried); |
570 | |
571 | int_type exponent = getUnbiasedExponent(); |
572 | if (exponent == min_exponent) { |
573 | // If we are denormal, normalize the exponent, so that we can encode |
574 | // easily. |
575 | exponent = static_cast<int_type>(exponent + 1); |
576 | for (uint_type check_bit = first_exponent_bit >> 1; check_bit != 0; |
577 | check_bit = static_cast<uint_type>(check_bit >> 1)) { |
578 | exponent = static_cast<int_type>(exponent - 1); |
579 | if (check_bit & significand) break; |
580 | } |
581 | } |
582 | |
583 | bool is_nan = |
584 | (getBits() & exponent_mask) == exponent_mask && significand != 0; |
585 | bool is_inf = |
586 | !is_nan && |
587 | ((exponent + carried) > static_cast<int_type>(other_T::exponent_bias) || |
588 | (significand == 0 && (getBits() & exponent_mask) == exponent_mask)); |
589 | |
590 | // If we are Nan or Inf we should pass that through. |
591 | if (is_inf) { |
592 | other.set_value(BitwiseCast<typename other_T::underlying_type>( |
593 | static_cast<typename other_T::uint_type>( |
594 | (negate ? other_T::sign_mask : 0) | other_T::exponent_mask))); |
595 | return; |
596 | } |
597 | if (is_nan) { |
598 | typename other_T::uint_type shifted_significand; |
599 | shifted_significand = static_cast<typename other_T::uint_type>( |
600 | negatable_left_shift( |
601 | static_cast<int_type>(other_T::num_fraction_bits) - |
602 | static_cast<int_type>(num_fraction_bits), significand)); |
603 | |
604 | // We are some sort of Nan. We try to keep the bit-pattern of the Nan |
605 | // as close as possible. If we had to shift off bits so we are 0, then we |
606 | // just set the last bit. |
607 | other.set_value(BitwiseCast<typename other_T::underlying_type>( |
608 | static_cast<typename other_T::uint_type>( |
609 | (negate ? other_T::sign_mask : 0) | other_T::exponent_mask | |
610 | (shifted_significand == 0 ? 0x1 : shifted_significand)))); |
611 | return; |
612 | } |
613 | |
614 | bool round_underflow_up = |
615 | isNegative() ? round_dir == kRoundToNegativeInfinity |
616 | : round_dir == kRoundToPositiveInfinity; |
617 | typedef typename other_T::int_type other_int_type; |
618 | // setFromSignUnbiasedExponentAndNormalizedSignificand will |
619 | // zero out any underflowing value (but retain the sign). |
620 | other.setFromSignUnbiasedExponentAndNormalizedSignificand( |
621 | negate, static_cast<other_int_type>(exponent), rounded_significand, |
622 | round_underflow_up); |
623 | return; |
624 | } |
625 | |
626 | private: |
627 | T value_; |
628 | |
629 | static_assert(num_used_bits == |
630 | Traits::num_exponent_bits + Traits::num_fraction_bits + 1, |
631 | "The number of bits do not fit" ); |
632 | static_assert(sizeof(T) == sizeof(uint_type), "The type sizes do not match" ); |
633 | }; |
634 | |
635 | // Returns 4 bits represented by the hex character. |
636 | inline uint8_t get_nibble_from_character(int character) { |
637 | const char* dec = "0123456789" ; |
638 | const char* lower = "abcdef" ; |
639 | const char* upper = "ABCDEF" ; |
640 | const char* p = nullptr; |
641 | if ((p = strchr(dec, character))) { |
642 | return static_cast<uint8_t>(p - dec); |
643 | } else if ((p = strchr(lower, character))) { |
644 | return static_cast<uint8_t>(p - lower + 0xa); |
645 | } else if ((p = strchr(upper, character))) { |
646 | return static_cast<uint8_t>(p - upper + 0xa); |
647 | } |
648 | |
649 | assert(false && "This was called with a non-hex character" ); |
650 | return 0; |
651 | } |
652 | |
653 | // Outputs the given HexFloat to the stream. |
654 | template <typename T, typename Traits> |
655 | std::ostream& operator<<(std::ostream& os, const HexFloat<T, Traits>& value) { |
656 | typedef HexFloat<T, Traits> HF; |
657 | typedef typename HF::uint_type uint_type; |
658 | typedef typename HF::int_type int_type; |
659 | |
660 | static_assert(HF::num_used_bits != 0, |
661 | "num_used_bits must be non-zero for a valid float" ); |
662 | static_assert(HF::num_exponent_bits != 0, |
663 | "num_exponent_bits must be non-zero for a valid float" ); |
664 | static_assert(HF::num_fraction_bits != 0, |
665 | "num_fractin_bits must be non-zero for a valid float" ); |
666 | |
667 | const uint_type bits = spvutils::BitwiseCast<uint_type>(value.value()); |
668 | const char* const sign = (bits & HF::sign_mask) ? "-" : "" ; |
669 | const uint_type exponent = static_cast<uint_type>( |
670 | (bits & HF::exponent_mask) >> HF::num_fraction_bits); |
671 | |
672 | uint_type fraction = static_cast<uint_type>((bits & HF::fraction_encode_mask) |
673 | << HF::num_overflow_bits); |
674 | |
675 | const bool is_zero = exponent == 0 && fraction == 0; |
676 | const bool is_denorm = exponent == 0 && !is_zero; |
677 | |
678 | // exponent contains the biased exponent we have to convert it back into |
679 | // the normal range. |
680 | int_type int_exponent = static_cast<int_type>(exponent - HF::exponent_bias); |
681 | // If the number is all zeros, then we actually have to NOT shift the |
682 | // exponent. |
683 | int_exponent = is_zero ? 0 : int_exponent; |
684 | |
685 | // If we are denorm, then start shifting, and decreasing the exponent until |
686 | // our leading bit is 1. |
687 | |
688 | if (is_denorm) { |
689 | while ((fraction & HF::fraction_top_bit) == 0) { |
690 | fraction = static_cast<uint_type>(fraction << 1); |
691 | int_exponent = static_cast<int_type>(int_exponent - 1); |
692 | } |
693 | // Since this is denormalized, we have to consume the leading 1 since it |
694 | // will end up being implicit. |
695 | fraction = static_cast<uint_type>(fraction << 1); // eat the leading 1 |
696 | fraction &= HF::fraction_represent_mask; |
697 | } |
698 | |
699 | uint_type fraction_nibbles = HF::fraction_nibbles; |
700 | // We do not have to display any trailing 0s, since this represents the |
701 | // fractional part. |
702 | while (fraction_nibbles > 0 && (fraction & 0xF) == 0) { |
703 | // Shift off any trailing values; |
704 | fraction = static_cast<uint_type>(fraction >> 4); |
705 | --fraction_nibbles; |
706 | } |
707 | |
708 | const auto saved_flags = os.flags(); |
709 | const auto saved_fill = os.fill(); |
710 | |
711 | os << sign << "0x" << (is_zero ? '0' : '1'); |
712 | if (fraction_nibbles) { |
713 | // Make sure to keep the leading 0s in place, since this is the fractional |
714 | // part. |
715 | os << "." << std::setw(static_cast<int>(fraction_nibbles)) |
716 | << std::setfill('0') << std::hex << fraction; |
717 | } |
718 | os << "p" << std::dec << (int_exponent >= 0 ? "+" : "" ) << int_exponent; |
719 | |
720 | os.flags(saved_flags); |
721 | os.fill(saved_fill); |
722 | |
723 | return os; |
724 | } |
725 | |
726 | // Returns true if negate_value is true and the next character on the |
727 | // input stream is a plus or minus sign. In that case we also set the fail bit |
728 | // on the stream and set the value to the zero value for its type. |
729 | template <typename T, typename Traits> |
730 | inline bool RejectParseDueToLeadingSign(std::istream& is, bool negate_value, |
731 | HexFloat<T, Traits>& value) { |
732 | if (negate_value) { |
733 | auto next_char = is.peek(); |
734 | if (next_char == '-' || next_char == '+') { |
735 | // Fail the parse. Emulate standard behaviour by setting the value to |
736 | // the zero value, and set the fail bit on the stream. |
737 | value = HexFloat<T, Traits>(typename HexFloat<T, Traits>::uint_type(0)); |
738 | is.setstate(std::ios_base::failbit); |
739 | return true; |
740 | } |
741 | } |
742 | return false; |
743 | } |
744 | |
745 | // Parses a floating point number from the given stream and stores it into the |
746 | // value parameter. |
747 | // If negate_value is true then the number may not have a leading minus or |
748 | // plus, and if it successfully parses, then the number is negated before |
749 | // being stored into the value parameter. |
750 | // If the value cannot be correctly parsed or overflows the target floating |
751 | // point type, then set the fail bit on the stream. |
752 | // TODO(dneto): Promise C++11 standard behavior in how the value is set in |
753 | // the error case, but only after all target platforms implement it correctly. |
754 | // In particular, the Microsoft C++ runtime appears to be out of spec. |
755 | template <typename T, typename Traits> |
756 | inline std::istream& ParseNormalFloat(std::istream& is, bool negate_value, |
757 | HexFloat<T, Traits>& value) { |
758 | if (RejectParseDueToLeadingSign(is, negate_value, value)) { |
759 | return is; |
760 | } |
761 | T val; |
762 | is >> val; |
763 | if (negate_value) { |
764 | val = -val; |
765 | } |
766 | value.set_value(val); |
767 | // In the failure case, map -0.0 to 0.0. |
768 | if (is.fail() && value.getUnsignedBits() == 0u) { |
769 | value = HexFloat<T, Traits>(typename HexFloat<T, Traits>::uint_type(0)); |
770 | } |
771 | if (val.isInfinity()) { |
772 | // Fail the parse. Emulate standard behaviour by setting the value to |
773 | // the closest normal value, and set the fail bit on the stream. |
774 | value.set_value((value.isNegative() || negate_value) ? T::lowest() |
775 | : T::max()); |
776 | is.setstate(std::ios_base::failbit); |
777 | } |
778 | return is; |
779 | } |
780 | |
781 | // Specialization of ParseNormalFloat for FloatProxy<Float16> values. |
782 | // This will parse the float as it were a 32-bit floating point number, |
783 | // and then round it down to fit into a Float16 value. |
784 | // The number is rounded towards zero. |
785 | // If negate_value is true then the number may not have a leading minus or |
786 | // plus, and if it successfully parses, then the number is negated before |
787 | // being stored into the value parameter. |
788 | // If the value cannot be correctly parsed or overflows the target floating |
789 | // point type, then set the fail bit on the stream. |
790 | // TODO(dneto): Promise C++11 standard behavior in how the value is set in |
791 | // the error case, but only after all target platforms implement it correctly. |
792 | // In particular, the Microsoft C++ runtime appears to be out of spec. |
793 | template <> |
794 | inline std::istream& |
795 | ParseNormalFloat<FloatProxy<Float16>, HexFloatTraits<FloatProxy<Float16>>>( |
796 | std::istream& is, bool negate_value, |
797 | HexFloat<FloatProxy<Float16>, HexFloatTraits<FloatProxy<Float16>>>& value) { |
798 | // First parse as a 32-bit float. |
799 | HexFloat<FloatProxy<float>> float_val(0.0f); |
800 | ParseNormalFloat(is, negate_value, float_val); |
801 | |
802 | // Then convert to 16-bit float, saturating at infinities, and |
803 | // rounding toward zero. |
804 | float_val.castTo(value, kRoundToZero); |
805 | |
806 | // Overflow on 16-bit behaves the same as for 32- and 64-bit: set the |
807 | // fail bit and set the lowest or highest value. |
808 | if (Float16::isInfinity(value.value().getAsFloat())) { |
809 | value.set_value(value.isNegative() ? Float16::lowest() : Float16::max()); |
810 | is.setstate(std::ios_base::failbit); |
811 | } |
812 | return is; |
813 | } |
814 | |
815 | // Reads a HexFloat from the given stream. |
816 | // If the float is not encoded as a hex-float then it will be parsed |
817 | // as a regular float. |
818 | // This may fail if your stream does not support at least one unget. |
819 | // Nan values can be encoded with "0x1.<not zero>p+exponent_bias". |
820 | // This would normally overflow a float and round to |
821 | // infinity but this special pattern is the exact representation for a NaN, |
822 | // and therefore is actually encoded as the correct NaN. To encode inf, |
823 | // either 0x0p+exponent_bias can be specified or any exponent greater than |
824 | // exponent_bias. |
825 | // Examples using IEEE 32-bit float encoding. |
826 | // 0x1.0p+128 (+inf) |
827 | // -0x1.0p-128 (-inf) |
828 | // |
829 | // 0x1.1p+128 (+Nan) |
830 | // -0x1.1p+128 (-Nan) |
831 | // |
832 | // 0x1p+129 (+inf) |
833 | // -0x1p+129 (-inf) |
834 | template <typename T, typename Traits> |
835 | std::istream& operator>>(std::istream& is, HexFloat<T, Traits>& value) { |
836 | using HF = HexFloat<T, Traits>; |
837 | using uint_type = typename HF::uint_type; |
838 | using int_type = typename HF::int_type; |
839 | |
840 | value.set_value(static_cast<typename HF::native_type>(0.f)); |
841 | |
842 | if (is.flags() & std::ios::skipws) { |
843 | // If the user wants to skip whitespace , then we should obey that. |
844 | while (std::isspace(is.peek())) { |
845 | is.get(); |
846 | } |
847 | } |
848 | |
849 | auto next_char = is.peek(); |
850 | bool negate_value = false; |
851 | |
852 | if (next_char != '-' && next_char != '0') { |
853 | return ParseNormalFloat(is, negate_value, value); |
854 | } |
855 | |
856 | if (next_char == '-') { |
857 | negate_value = true; |
858 | is.get(); |
859 | next_char = is.peek(); |
860 | } |
861 | |
862 | if (next_char == '0') { |
863 | is.get(); // We may have to unget this. |
864 | auto maybe_hex_start = is.peek(); |
865 | if (maybe_hex_start != 'x' && maybe_hex_start != 'X') { |
866 | is.unget(); |
867 | return ParseNormalFloat(is, negate_value, value); |
868 | } else { |
869 | is.get(); // Throw away the 'x'; |
870 | } |
871 | } else { |
872 | return ParseNormalFloat(is, negate_value, value); |
873 | } |
874 | |
875 | // This "looks" like a hex-float so treat it as one. |
876 | bool seen_p = false; |
877 | bool seen_dot = false; |
878 | uint_type fraction_index = 0; |
879 | |
880 | uint_type fraction = 0; |
881 | int_type exponent = HF::exponent_bias; |
882 | |
883 | // Strip off leading zeros so we don't have to special-case them later. |
884 | while ((next_char = is.peek()) == '0') { |
885 | is.get(); |
886 | } |
887 | |
888 | bool is_denorm = |
889 | true; // Assume denorm "representation" until we hear otherwise. |
890 | // NB: This does not mean the value is actually denorm, |
891 | // it just means that it was written 0. |
892 | bool bits_written = false; // Stays false until we write a bit. |
893 | while (!seen_p && !seen_dot) { |
894 | // Handle characters that are left of the fractional part. |
895 | if (next_char == '.') { |
896 | seen_dot = true; |
897 | } else if (next_char == 'p') { |
898 | seen_p = true; |
899 | } else if (::isxdigit(next_char)) { |
900 | // We know this is not denormalized since we have stripped all leading |
901 | // zeroes and we are not a ".". |
902 | is_denorm = false; |
903 | int number = get_nibble_from_character(next_char); |
904 | for (int i = 0; i < 4; ++i, number <<= 1) { |
905 | uint_type write_bit = (number & 0x8) ? 0x1 : 0x0; |
906 | if (bits_written) { |
907 | // If we are here the bits represented belong in the fractional |
908 | // part of the float, and we have to adjust the exponent accordingly. |
909 | fraction = static_cast<uint_type>( |
910 | fraction | |
911 | static_cast<uint_type>( |
912 | write_bit << (HF::top_bit_left_shift - fraction_index++))); |
913 | exponent = static_cast<int_type>(exponent + 1); |
914 | } |
915 | bits_written |= write_bit != 0; |
916 | } |
917 | } else { |
918 | // We have not found our exponent yet, so we have to fail. |
919 | is.setstate(std::ios::failbit); |
920 | return is; |
921 | } |
922 | is.get(); |
923 | next_char = is.peek(); |
924 | } |
925 | bits_written = false; |
926 | while (seen_dot && !seen_p) { |
927 | // Handle only fractional parts now. |
928 | if (next_char == 'p') { |
929 | seen_p = true; |
930 | } else if (::isxdigit(next_char)) { |
931 | int number = get_nibble_from_character(next_char); |
932 | for (int i = 0; i < 4; ++i, number <<= 1) { |
933 | uint_type write_bit = (number & 0x8) ? 0x01 : 0x00; |
934 | bits_written |= write_bit != 0; |
935 | if (is_denorm && !bits_written) { |
936 | // Handle modifying the exponent here this way we can handle |
937 | // an arbitrary number of hex values without overflowing our |
938 | // integer. |
939 | exponent = static_cast<int_type>(exponent - 1); |
940 | } else { |
941 | fraction = static_cast<uint_type>( |
942 | fraction | |
943 | static_cast<uint_type>( |
944 | write_bit << (HF::top_bit_left_shift - fraction_index++))); |
945 | } |
946 | } |
947 | } else { |
948 | // We still have not found our 'p' exponent yet, so this is not a valid |
949 | // hex-float. |
950 | is.setstate(std::ios::failbit); |
951 | return is; |
952 | } |
953 | is.get(); |
954 | next_char = is.peek(); |
955 | } |
956 | |
957 | bool seen_sign = false; |
958 | int8_t exponent_sign = 1; |
959 | int_type written_exponent = 0; |
960 | while (true) { |
961 | if ((next_char == '-' || next_char == '+')) { |
962 | if (seen_sign) { |
963 | is.setstate(std::ios::failbit); |
964 | return is; |
965 | } |
966 | seen_sign = true; |
967 | exponent_sign = (next_char == '-') ? -1 : 1; |
968 | } else if (::isdigit(next_char)) { |
969 | // Hex-floats express their exponent as decimal. |
970 | written_exponent = static_cast<int_type>(written_exponent * 10); |
971 | written_exponent = |
972 | static_cast<int_type>(written_exponent + (next_char - '0')); |
973 | } else { |
974 | break; |
975 | } |
976 | is.get(); |
977 | next_char = is.peek(); |
978 | } |
979 | |
980 | written_exponent = static_cast<int_type>(written_exponent * exponent_sign); |
981 | exponent = static_cast<int_type>(exponent + written_exponent); |
982 | |
983 | bool is_zero = is_denorm && (fraction == 0); |
984 | if (is_denorm && !is_zero) { |
985 | fraction = static_cast<uint_type>(fraction << 1); |
986 | exponent = static_cast<int_type>(exponent - 1); |
987 | } else if (is_zero) { |
988 | exponent = 0; |
989 | } |
990 | |
991 | if (exponent <= 0 && !is_zero) { |
992 | fraction = static_cast<uint_type>(fraction >> 1); |
993 | fraction |= static_cast<uint_type>(1) << HF::top_bit_left_shift; |
994 | } |
995 | |
996 | fraction = (fraction >> HF::fraction_right_shift) & HF::fraction_encode_mask; |
997 | |
998 | const int_type max_exponent = |
999 | SetBits<uint_type, 0, HF::num_exponent_bits>::get; |
1000 | |
1001 | // Handle actual denorm numbers |
1002 | while (exponent < 0 && !is_zero) { |
1003 | fraction = static_cast<uint_type>(fraction >> 1); |
1004 | exponent = static_cast<int_type>(exponent + 1); |
1005 | |
1006 | fraction &= HF::fraction_encode_mask; |
1007 | if (fraction == 0) { |
1008 | // We have underflowed our fraction. We should clamp to zero. |
1009 | is_zero = true; |
1010 | exponent = 0; |
1011 | } |
1012 | } |
1013 | |
1014 | // We have overflowed so we should be inf/-inf. |
1015 | if (exponent > max_exponent) { |
1016 | exponent = max_exponent; |
1017 | fraction = 0; |
1018 | } |
1019 | |
1020 | uint_type output_bits = static_cast<uint_type>( |
1021 | static_cast<uint_type>(negate_value ? 1 : 0) << HF::top_bit_left_shift); |
1022 | output_bits |= fraction; |
1023 | |
1024 | uint_type shifted_exponent = static_cast<uint_type>( |
1025 | static_cast<uint_type>(exponent << HF::exponent_left_shift) & |
1026 | HF::exponent_mask); |
1027 | output_bits |= shifted_exponent; |
1028 | |
1029 | T output_float = spvutils::BitwiseCast<T>(output_bits); |
1030 | value.set_value(output_float); |
1031 | |
1032 | return is; |
1033 | } |
1034 | |
1035 | // Writes a FloatProxy value to a stream. |
1036 | // Zero and normal numbers are printed in the usual notation, but with |
1037 | // enough digits to fully reproduce the value. Other values (subnormal, |
1038 | // NaN, and infinity) are printed as a hex float. |
1039 | template <typename T> |
1040 | std::ostream& operator<<(std::ostream& os, const FloatProxy<T>& value) { |
1041 | auto float_val = value.getAsFloat(); |
1042 | switch (std::fpclassify(float_val)) { |
1043 | case FP_ZERO: |
1044 | case FP_NORMAL: { |
1045 | auto saved_precision = os.precision(); |
1046 | os.precision(std::numeric_limits<T>::digits10); |
1047 | os << float_val; |
1048 | os.precision(saved_precision); |
1049 | } break; |
1050 | default: |
1051 | os << HexFloat<FloatProxy<T>>(value); |
1052 | break; |
1053 | } |
1054 | return os; |
1055 | } |
1056 | |
1057 | template <> |
1058 | inline std::ostream& operator<<<Float16>(std::ostream& os, |
1059 | const FloatProxy<Float16>& value) { |
1060 | os << HexFloat<FloatProxy<Float16>>(value); |
1061 | return os; |
1062 | } |
1063 | } |
1064 | |
1065 | #endif // LIBSPIRV_UTIL_HEX_FLOAT_H_ |
1066 | |