1 /*
2 * Copyright © 2020 Collabora Ltd.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #ifndef NIR_CONVERSION_BUILDER_H
25 #define NIR_CONVERSION_BUILDER_H
26
27 #include "util/u_math.h"
28 #include "nir_builder.h"
29 #include "nir_builtin_builder.h"
30
31 #ifdef __cplusplus
32 extern "C" {
33 #endif
34
35 static inline nir_def *
nir_round_float_to_int(nir_builder * b,nir_def * src,nir_rounding_mode round)36 nir_round_float_to_int(nir_builder *b, nir_def *src,
37 nir_rounding_mode round)
38 {
39 switch (round) {
40 case nir_rounding_mode_ru:
41 return nir_fceil(b, src);
42
43 case nir_rounding_mode_rd:
44 return nir_ffloor(b, src);
45
46 case nir_rounding_mode_rtne:
47 return nir_fround_even(b, src);
48
49 case nir_rounding_mode_undef:
50 case nir_rounding_mode_rtz:
51 break;
52 }
53 unreachable("unexpected rounding mode");
54 }
55
56 static inline nir_def *
nir_round_float_to_float(nir_builder * b,nir_def * src,unsigned dest_bit_size,nir_rounding_mode round)57 nir_round_float_to_float(nir_builder *b, nir_def *src,
58 unsigned dest_bit_size,
59 nir_rounding_mode round)
60 {
61 unsigned src_bit_size = src->bit_size;
62 if (dest_bit_size > src_bit_size)
63 return src; /* No rounding is needed for an up-convert */
64
65 nir_op low_conv = nir_type_conversion_op(nir_type_float | src_bit_size,
66 nir_type_float | dest_bit_size,
67 nir_rounding_mode_undef);
68 nir_op high_conv = nir_type_conversion_op(nir_type_float | dest_bit_size,
69 nir_type_float | src_bit_size,
70 nir_rounding_mode_undef);
71
72 switch (round) {
73 case nir_rounding_mode_ru: {
74 /* If lower-precision conversion results in a lower value, push it
75 * up one ULP. */
76 nir_def *lower_prec =
77 nir_build_alu(b, low_conv, src, NULL, NULL, NULL);
78 nir_def *roundtrip =
79 nir_build_alu(b, high_conv, lower_prec, NULL, NULL, NULL);
80 nir_def *cmp = nir_flt(b, roundtrip, src);
81 nir_def *inf = nir_imm_floatN_t(b, INFINITY, dest_bit_size);
82 return nir_bcsel(b, cmp, nir_nextafter(b, lower_prec, inf), lower_prec);
83 }
84 case nir_rounding_mode_rd: {
85 /* If lower-precision conversion results in a higher value, push it
86 * down one ULP. */
87 nir_def *lower_prec =
88 nir_build_alu(b, low_conv, src, NULL, NULL, NULL);
89 nir_def *roundtrip =
90 nir_build_alu(b, high_conv, lower_prec, NULL, NULL, NULL);
91 nir_def *cmp = nir_flt(b, src, roundtrip);
92 nir_def *neg_inf = nir_imm_floatN_t(b, -INFINITY, dest_bit_size);
93 return nir_bcsel(b, cmp, nir_nextafter(b, lower_prec, neg_inf), lower_prec);
94 }
95 case nir_rounding_mode_rtz:
96 return nir_bcsel(b, nir_flt_imm(b, src, 1),
97 nir_round_float_to_float(b, src, dest_bit_size,
98 nir_rounding_mode_ru),
99 nir_round_float_to_float(b, src, dest_bit_size,
100 nir_rounding_mode_rd));
101 case nir_rounding_mode_rtne:
102 case nir_rounding_mode_undef:
103 break;
104 }
105 unreachable("unexpected rounding mode");
106 }
107
108 static inline nir_def *
nir_round_int_to_float(nir_builder * b,nir_def * src,nir_alu_type src_type,unsigned dest_bit_size,nir_rounding_mode round)109 nir_round_int_to_float(nir_builder *b, nir_def *src,
110 nir_alu_type src_type,
111 unsigned dest_bit_size,
112 nir_rounding_mode round)
113 {
114 /* We only care whether or not its signed */
115 src_type = nir_alu_type_get_base_type(src_type);
116
117 unsigned mantissa_bits;
118 switch (dest_bit_size) {
119 case 16:
120 mantissa_bits = 10;
121 break;
122 case 32:
123 mantissa_bits = 23;
124 break;
125 case 64:
126 mantissa_bits = 52;
127 break;
128 default:
129 unreachable("Unsupported bit size");
130 }
131
132 if (src->bit_size < mantissa_bits)
133 return src;
134
135 if (src_type == nir_type_int) {
136 nir_def *sign =
137 nir_i2b(b, nir_ishr(b, src, nir_imm_int(b, src->bit_size - 1)));
138 nir_def *abs = nir_iabs(b, src);
139 nir_def *positive_rounded =
140 nir_round_int_to_float(b, abs, nir_type_uint, dest_bit_size, round);
141 nir_def *max_positive =
142 nir_imm_intN_t(b, (1ull << (src->bit_size - 1)) - 1, src->bit_size);
143 switch (round) {
144 case nir_rounding_mode_rtz:
145 return nir_bcsel(b, sign, nir_ineg(b, positive_rounded),
146 positive_rounded);
147 break;
148 case nir_rounding_mode_ru:
149 return nir_bcsel(b, sign,
150 nir_ineg(b, nir_round_int_to_float(b, abs, nir_type_uint, dest_bit_size, nir_rounding_mode_rd)),
151 nir_umin(b, positive_rounded, max_positive));
152 break;
153 case nir_rounding_mode_rd:
154 return nir_bcsel(b, sign,
155 nir_ineg(b,
156 nir_umin(b, max_positive,
157 nir_round_int_to_float(b, abs, nir_type_uint, dest_bit_size, nir_rounding_mode_ru))),
158 positive_rounded);
159 case nir_rounding_mode_rtne:
160 case nir_rounding_mode_undef:
161 break;
162 }
163 unreachable("unexpected rounding mode");
164 } else {
165 nir_def *mantissa_bit_size = nir_imm_int(b, mantissa_bits);
166 nir_def *msb = nir_imax(b, nir_ufind_msb(b, src), mantissa_bit_size);
167 nir_def *bits_to_lose = nir_isub(b, msb, mantissa_bit_size);
168 nir_def *one = nir_imm_intN_t(b, 1, src->bit_size);
169 nir_def *adjust = nir_ishl(b, one, bits_to_lose);
170 nir_def *mask = nir_inot(b, nir_isub(b, adjust, one));
171 nir_def *truncated = nir_iand(b, src, mask);
172 switch (round) {
173 case nir_rounding_mode_rtz:
174 case nir_rounding_mode_rd:
175 return truncated;
176 break;
177 case nir_rounding_mode_ru:
178 return nir_bcsel(b, nir_ieq(b, src, truncated),
179 src, nir_uadd_sat(b, truncated, adjust));
180 case nir_rounding_mode_rtne:
181 case nir_rounding_mode_undef:
182 break;
183 }
184 unreachable("unexpected rounding mode");
185 }
186 }
187
188 /** Returns true if the representable range of a contains the representable
189 * range of b.
190 */
191 static inline bool
nir_alu_type_range_contains_type_range(nir_alu_type a,nir_alu_type b)192 nir_alu_type_range_contains_type_range(nir_alu_type a, nir_alu_type b)
193 {
194 /* Split types from bit sizes */
195 nir_alu_type a_base_type = nir_alu_type_get_base_type(a);
196 nir_alu_type b_base_type = nir_alu_type_get_base_type(b);
197 unsigned a_bit_size = nir_alu_type_get_type_size(a);
198 unsigned b_bit_size = nir_alu_type_get_type_size(b);
199
200 /* This requires sized types */
201 assert(a_bit_size > 0 && b_bit_size > 0);
202
203 if (a_base_type == b_base_type && a_bit_size >= b_bit_size)
204 return true;
205
206 if (a_base_type == nir_type_int && b_base_type == nir_type_uint &&
207 a_bit_size > b_bit_size)
208 return true;
209
210 /* 16-bit floats fit in 32-bit integers */
211 if (a_base_type == nir_type_int && a_bit_size >= 32 &&
212 b == nir_type_float16)
213 return true;
214
215 /* All signed or unsigned ints can fit in float or above. A uint8 can fit
216 * in a float16.
217 */
218 if (a_base_type == nir_type_float && b_base_type != nir_type_float &&
219 (a_bit_size >= 32 || b_bit_size == 8))
220 return true;
221
222 return false;
223 }
224
225 /**
226 * Retrieves limits used for clamping a value of the src type into
227 * the widest representable range of the dst type via cmp + bcsel
228 */
229 static inline void
nir_get_clamp_limits(nir_builder * b,nir_alu_type src_type,nir_alu_type dest_type,nir_def ** low,nir_def ** high)230 nir_get_clamp_limits(nir_builder *b,
231 nir_alu_type src_type,
232 nir_alu_type dest_type,
233 nir_def **low, nir_def **high)
234 {
235 /* Split types from bit sizes */
236 nir_alu_type src_base_type = nir_alu_type_get_base_type(src_type);
237 nir_alu_type dest_base_type = nir_alu_type_get_base_type(dest_type);
238 unsigned src_bit_size = nir_alu_type_get_type_size(src_type);
239 unsigned dest_bit_size = nir_alu_type_get_type_size(dest_type);
240 assert(dest_bit_size != 0 && src_bit_size != 0);
241
242 *low = NULL;
243 *high = NULL;
244
245 /* limits of the destination type, expressed in the source type */
246 switch (dest_base_type) {
247 case nir_type_int: {
248 int64_t ilow, ihigh;
249 if (dest_bit_size == 64) {
250 ilow = INT64_MIN;
251 ihigh = INT64_MAX;
252 } else {
253 ilow = -(1ll << (dest_bit_size - 1));
254 ihigh = (1ll << (dest_bit_size - 1)) - 1;
255 }
256
257 if (src_base_type == nir_type_int) {
258 *low = nir_imm_intN_t(b, ilow, src_bit_size);
259 *high = nir_imm_intN_t(b, ihigh, src_bit_size);
260 } else if (src_base_type == nir_type_uint) {
261 assert(src_bit_size >= dest_bit_size);
262 *high = nir_imm_intN_t(b, ihigh, src_bit_size);
263 } else {
264 *low = nir_imm_floatN_t(b, ilow, src_bit_size);
265 *high = nir_imm_floatN_t(b, ihigh, src_bit_size);
266 }
267 break;
268 }
269 case nir_type_uint: {
270 uint64_t uhigh = dest_bit_size == 64 ? ~0ull : (1ull << dest_bit_size) - 1;
271 if (src_base_type != nir_type_float) {
272 /* for uint->uint conversions, no need to clamp negatives */
273 if (src_base_type != nir_type_uint)
274 *low = nir_imm_intN_t(b, 0, src_bit_size);
275
276 if (src_base_type == nir_type_uint || src_bit_size > dest_bit_size)
277 *high = nir_imm_intN_t(b, uhigh, src_bit_size);
278 } else {
279 *low = nir_imm_floatN_t(b, 0.0f, src_bit_size);
280 *high = nir_imm_floatN_t(b, uhigh, src_bit_size);
281 }
282 break;
283 }
284 case nir_type_float: {
285 double flow, fhigh;
286 switch (dest_bit_size) {
287 case 16:
288 flow = -65504.0f;
289 fhigh = 65504.0f;
290 break;
291 case 32:
292 flow = -FLT_MAX;
293 fhigh = FLT_MAX;
294 break;
295 case 64:
296 flow = -DBL_MAX;
297 fhigh = DBL_MAX;
298 break;
299 default:
300 unreachable("Unhandled bit size");
301 }
302
303 switch (src_base_type) {
304 case nir_type_int: {
305 int64_t src_ilow, src_ihigh;
306 if (src_bit_size == 64) {
307 src_ilow = INT64_MIN;
308 src_ihigh = INT64_MAX;
309 } else {
310 src_ilow = -(1ll << (src_bit_size - 1));
311 src_ihigh = (1ll << (src_bit_size - 1)) - 1;
312 }
313 if (src_ilow < flow)
314 *low = nir_imm_intN_t(b, flow, src_bit_size);
315 if (src_ihigh > fhigh)
316 *high = nir_imm_intN_t(b, fhigh, src_bit_size);
317 break;
318 }
319 case nir_type_uint: {
320 uint64_t src_uhigh = src_bit_size == 64 ? ~0ull : (1ull << src_bit_size) - 1;
321 if (src_uhigh > fhigh)
322 *high = nir_imm_intN_t(b, fhigh, src_bit_size);
323 break;
324 }
325 case nir_type_float:
326 *low = nir_imm_floatN_t(b, flow, src_bit_size);
327 *high = nir_imm_floatN_t(b, fhigh, src_bit_size);
328 break;
329 default:
330 unreachable("Clamping from unknown type");
331 }
332 break;
333 }
334 default:
335 unreachable("clamping to unknown type");
336 break;
337 }
338 }
339
340 /**
341 * Clamp the value into the widest representatble range of the
342 * destination type with cmp + bcsel.
343 *
344 * val/val_type: The variables used for bcsel
345 * src/src_type: The variables used for comparison
346 * dest_type: The type which determines the range used for comparison
347 */
348 static inline nir_def *
nir_clamp_to_type_range(nir_builder * b,nir_def * val,nir_alu_type val_type,nir_def * src,nir_alu_type src_type,nir_alu_type dest_type)349 nir_clamp_to_type_range(nir_builder *b,
350 nir_def *val, nir_alu_type val_type,
351 nir_def *src, nir_alu_type src_type,
352 nir_alu_type dest_type)
353 {
354 assert(nir_alu_type_get_type_size(src_type) == 0 ||
355 nir_alu_type_get_type_size(src_type) == src->bit_size);
356 src_type |= src->bit_size;
357 if (nir_alu_type_range_contains_type_range(dest_type, src_type))
358 return val;
359
360 /* limits of the destination type, expressed in the source type */
361 nir_def *low = NULL, *high = NULL;
362 nir_get_clamp_limits(b, src_type, dest_type, &low, &high);
363
364 nir_def *low_cond = NULL, *high_cond = NULL;
365 switch (nir_alu_type_get_base_type(src_type)) {
366 case nir_type_int:
367 low_cond = low ? nir_ilt(b, src, low) : NULL;
368 high_cond = high ? nir_ilt(b, high, src) : NULL;
369 break;
370 case nir_type_uint:
371 low_cond = low ? nir_ult(b, src, low) : NULL;
372 high_cond = high ? nir_ult(b, high, src) : NULL;
373 break;
374 case nir_type_float:
375 low_cond = low ? nir_fge(b, low, src) : NULL;
376 high_cond = high ? nir_fge(b, src, high) : NULL;
377 break;
378 default:
379 unreachable("clamping from unknown type");
380 }
381
382 nir_def *val_low = low, *val_high = high;
383 if (val_type != src_type) {
384 nir_get_clamp_limits(b, val_type, dest_type, &val_low, &val_high);
385 }
386
387 nir_def *res = val;
388 if (low_cond && val_low)
389 res = nir_bcsel(b, low_cond, val_low, res);
390 if (high_cond && val_high)
391 res = nir_bcsel(b, high_cond, val_high, res);
392
393 return res;
394 }
395
396 static inline nir_rounding_mode
nir_simplify_conversion_rounding(nir_alu_type src_type,nir_alu_type dest_type,nir_rounding_mode rounding)397 nir_simplify_conversion_rounding(nir_alu_type src_type,
398 nir_alu_type dest_type,
399 nir_rounding_mode rounding)
400 {
401 nir_alu_type src_base_type = nir_alu_type_get_base_type(src_type);
402 nir_alu_type dest_base_type = nir_alu_type_get_base_type(dest_type);
403 unsigned src_bit_size = nir_alu_type_get_type_size(src_type);
404 unsigned dest_bit_size = nir_alu_type_get_type_size(dest_type);
405 assert(src_bit_size > 0 && dest_bit_size > 0);
406
407 if (rounding == nir_rounding_mode_undef)
408 return rounding;
409
410 /* Pure integer conversion doesn't have any rounding */
411 if (src_base_type != nir_type_float &&
412 dest_base_type != nir_type_float)
413 return nir_rounding_mode_undef;
414
415 /* Float down-casts don't round */
416 if (src_base_type == nir_type_float &&
417 dest_base_type == nir_type_float &&
418 dest_bit_size >= src_bit_size)
419 return nir_rounding_mode_undef;
420
421 /* Regular float to int conversions are RTZ */
422 if (src_base_type == nir_type_float &&
423 dest_base_type != nir_type_float &&
424 rounding == nir_rounding_mode_rtz)
425 return nir_rounding_mode_undef;
426
427 /* The CL spec requires regular conversions to float to be RTNE */
428 if (dest_base_type == nir_type_float &&
429 rounding == nir_rounding_mode_rtne)
430 return nir_rounding_mode_undef;
431
432 /* Couldn't simplify */
433 return rounding;
434 }
435
436 static inline nir_def *
nir_convert_with_rounding(nir_builder * b,nir_def * src,nir_alu_type src_type,nir_alu_type dest_type,nir_rounding_mode round,bool clamp)437 nir_convert_with_rounding(nir_builder *b,
438 nir_def *src, nir_alu_type src_type,
439 nir_alu_type dest_type,
440 nir_rounding_mode round,
441 bool clamp)
442 {
443 /* Some stuff wants sized types */
444 assert(nir_alu_type_get_type_size(src_type) == 0 ||
445 nir_alu_type_get_type_size(src_type) == src->bit_size);
446 src_type |= src->bit_size;
447
448 /* Split types from bit sizes */
449 nir_alu_type src_base_type = nir_alu_type_get_base_type(src_type);
450 nir_alu_type dest_base_type = nir_alu_type_get_base_type(dest_type);
451 unsigned dest_bit_size = nir_alu_type_get_type_size(dest_type);
452
453 /* Try to simplify the conversion if we can */
454 clamp = clamp &&
455 !nir_alu_type_range_contains_type_range(dest_type, src_type);
456 round = nir_simplify_conversion_rounding(src_type, dest_type, round);
457
458 /* For float -> int/uint conversions, we might not be able to represent
459 * the destination range in the source float accurately. For these cases,
460 * do the comparison in float range, but the bcsel in the destination range.
461 */
462 bool clamp_after_conversion = clamp &&
463 src_base_type == nir_type_float &&
464 dest_base_type != nir_type_float;
465
466 /*
467 * If we don't care about rounding and clamping, we can just use NIR's
468 * built-in ops. There is also a special case for SPIR-V in shaders, where
469 * f32/f64 -> f16 conversions can have one of two rounding modes applied,
470 * which NIR has built-in opcodes for.
471 *
472 * For the rest, we have our own implementation of rounding and clamping.
473 */
474 bool trivial_convert;
475 if (!clamp && round == nir_rounding_mode_undef) {
476 trivial_convert = true;
477 } else if (!clamp && src_type == nir_type_float32 &&
478 dest_type == nir_type_float16 &&
479 (round == nir_rounding_mode_rtne ||
480 round == nir_rounding_mode_rtz)) {
481 trivial_convert = true;
482 } else {
483 trivial_convert = false;
484 }
485
486 if (trivial_convert)
487 return nir_type_convert(b, src, src_type, dest_type, round);
488
489 nir_def *dest = src;
490
491 /* clamp the result into range */
492 if (clamp && !clamp_after_conversion)
493 dest = nir_clamp_to_type_range(b, src, src_type, src, src_type, dest_type);
494
495 /* round with selected rounding mode */
496 if (!trivial_convert && round != nir_rounding_mode_undef) {
497 if (src_base_type == nir_type_float) {
498 if (dest_base_type == nir_type_float) {
499 dest = nir_round_float_to_float(b, dest, dest_bit_size, round);
500 } else {
501 dest = nir_round_float_to_int(b, dest, round);
502 }
503 } else {
504 dest = nir_round_int_to_float(b, dest, src_type, dest_bit_size, round);
505 }
506
507 round = nir_rounding_mode_undef;
508 }
509
510 /* now we can convert the value */
511 nir_op op = nir_type_conversion_op(src_type, dest_type, round);
512 dest = nir_build_alu(b, op, dest, NULL, NULL, NULL);
513
514 if (clamp_after_conversion)
515 dest = nir_clamp_to_type_range(b, dest, dest_type, src, src_type, dest_type);
516
517 return dest;
518 }
519
520 #ifdef __cplusplus
521 }
522 #endif
523
524 #endif /* NIR_CONVERSION_BUILDER_H */
525