xref: /aosp_15_r20/external/mesa3d/src/compiler/nir/nir_conversion_builder.h (revision 6104692788411f58d303aa86923a9ff6ecaded22)
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          *low = nir_imm_intN_t(b, 0, src_bit_size);
273          if (src_base_type == nir_type_uint || src_bit_size > dest_bit_size)
274             *high = nir_imm_intN_t(b, uhigh, src_bit_size);
275       } else {
276          *low = nir_imm_floatN_t(b, 0.0f, src_bit_size);
277          *high = nir_imm_floatN_t(b, uhigh, src_bit_size);
278       }
279       break;
280    }
281    case nir_type_float: {
282       double flow, fhigh;
283       switch (dest_bit_size) {
284       case 16:
285          flow = -65504.0f;
286          fhigh = 65504.0f;
287          break;
288       case 32:
289          flow = -FLT_MAX;
290          fhigh = FLT_MAX;
291          break;
292       case 64:
293          flow = -DBL_MAX;
294          fhigh = DBL_MAX;
295          break;
296       default:
297          unreachable("Unhandled bit size");
298       }
299 
300       switch (src_base_type) {
301       case nir_type_int: {
302          int64_t src_ilow, src_ihigh;
303          if (src_bit_size == 64) {
304             src_ilow = INT64_MIN;
305             src_ihigh = INT64_MAX;
306          } else {
307             src_ilow = -(1ll << (src_bit_size - 1));
308             src_ihigh = (1ll << (src_bit_size - 1)) - 1;
309          }
310          if (src_ilow < flow)
311             *low = nir_imm_intN_t(b, flow, src_bit_size);
312          if (src_ihigh > fhigh)
313             *high = nir_imm_intN_t(b, fhigh, src_bit_size);
314          break;
315       }
316       case nir_type_uint: {
317          uint64_t src_uhigh = src_bit_size == 64 ? ~0ull : (1ull << src_bit_size) - 1;
318          if (src_uhigh > fhigh)
319             *high = nir_imm_intN_t(b, fhigh, src_bit_size);
320          break;
321       }
322       case nir_type_float:
323          *low = nir_imm_floatN_t(b, flow, src_bit_size);
324          *high = nir_imm_floatN_t(b, fhigh, src_bit_size);
325          break;
326       default:
327          unreachable("Clamping from unknown type");
328       }
329       break;
330    }
331    default:
332       unreachable("clamping to unknown type");
333       break;
334    }
335 }
336 
337 /**
338  * Clamp the value into the widest representatble range of the
339  * destination type with cmp + bcsel.
340  *
341  * val/val_type: The variables used for bcsel
342  * src/src_type: The variables used for comparison
343  * dest_type: The type which determines the range used for comparison
344  */
345 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)346 nir_clamp_to_type_range(nir_builder *b,
347                         nir_def *val, nir_alu_type val_type,
348                         nir_def *src, nir_alu_type src_type,
349                         nir_alu_type dest_type)
350 {
351    assert(nir_alu_type_get_type_size(src_type) == 0 ||
352           nir_alu_type_get_type_size(src_type) == src->bit_size);
353    src_type |= src->bit_size;
354    if (nir_alu_type_range_contains_type_range(dest_type, src_type))
355       return val;
356 
357    /* limits of the destination type, expressed in the source type */
358    nir_def *low = NULL, *high = NULL;
359    nir_get_clamp_limits(b, src_type, dest_type, &low, &high);
360 
361    nir_def *low_cond = NULL, *high_cond = NULL;
362    switch (nir_alu_type_get_base_type(src_type)) {
363    case nir_type_int:
364       low_cond = low ? nir_ilt(b, src, low) : NULL;
365       high_cond = high ? nir_ilt(b, high, src) : NULL;
366       break;
367    case nir_type_uint:
368       low_cond = low ? nir_ult(b, src, low) : NULL;
369       high_cond = high ? nir_ult(b, high, src) : NULL;
370       break;
371    case nir_type_float:
372       low_cond = low ? nir_fge(b, low, src) : NULL;
373       high_cond = high ? nir_fge(b, src, high) : NULL;
374       break;
375    default:
376       unreachable("clamping from unknown type");
377    }
378 
379    nir_def *val_low = low, *val_high = high;
380    if (val_type != src_type) {
381       nir_get_clamp_limits(b, val_type, dest_type, &val_low, &val_high);
382    }
383 
384    nir_def *res = val;
385    if (low_cond && val_low)
386       res = nir_bcsel(b, low_cond, val_low, res);
387    if (high_cond && val_high)
388       res = nir_bcsel(b, high_cond, val_high, res);
389 
390    return res;
391 }
392 
393 static inline nir_rounding_mode
nir_simplify_conversion_rounding(nir_alu_type src_type,nir_alu_type dest_type,nir_rounding_mode rounding)394 nir_simplify_conversion_rounding(nir_alu_type src_type,
395                                  nir_alu_type dest_type,
396                                  nir_rounding_mode rounding)
397 {
398    nir_alu_type src_base_type = nir_alu_type_get_base_type(src_type);
399    nir_alu_type dest_base_type = nir_alu_type_get_base_type(dest_type);
400    unsigned src_bit_size = nir_alu_type_get_type_size(src_type);
401    unsigned dest_bit_size = nir_alu_type_get_type_size(dest_type);
402    assert(src_bit_size > 0 && dest_bit_size > 0);
403 
404    if (rounding == nir_rounding_mode_undef)
405       return rounding;
406 
407    /* Pure integer conversion doesn't have any rounding */
408    if (src_base_type != nir_type_float &&
409        dest_base_type != nir_type_float)
410       return nir_rounding_mode_undef;
411 
412    /* Float down-casts don't round */
413    if (src_base_type == nir_type_float &&
414        dest_base_type == nir_type_float &&
415        dest_bit_size >= src_bit_size)
416       return nir_rounding_mode_undef;
417 
418    /* Regular float to int conversions are RTZ */
419    if (src_base_type == nir_type_float &&
420        dest_base_type != nir_type_float &&
421        rounding == nir_rounding_mode_rtz)
422       return nir_rounding_mode_undef;
423 
424    /* The CL spec requires regular conversions to float to be RTNE */
425    if (dest_base_type == nir_type_float &&
426        rounding == nir_rounding_mode_rtne)
427       return nir_rounding_mode_undef;
428 
429    /* Couldn't simplify */
430    return rounding;
431 }
432 
433 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)434 nir_convert_with_rounding(nir_builder *b,
435                           nir_def *src, nir_alu_type src_type,
436                           nir_alu_type dest_type,
437                           nir_rounding_mode round,
438                           bool clamp)
439 {
440    /* Some stuff wants sized types */
441    assert(nir_alu_type_get_type_size(src_type) == 0 ||
442           nir_alu_type_get_type_size(src_type) == src->bit_size);
443    src_type |= src->bit_size;
444 
445    /* Split types from bit sizes */
446    nir_alu_type src_base_type = nir_alu_type_get_base_type(src_type);
447    nir_alu_type dest_base_type = nir_alu_type_get_base_type(dest_type);
448    unsigned dest_bit_size = nir_alu_type_get_type_size(dest_type);
449 
450    /* Try to simplify the conversion if we can */
451    clamp = clamp &&
452            !nir_alu_type_range_contains_type_range(dest_type, src_type);
453    round = nir_simplify_conversion_rounding(src_type, dest_type, round);
454 
455    /* For float -> int/uint conversions, we might not be able to represent
456     * the destination range in the source float accurately. For these cases,
457     * do the comparison in float range, but the bcsel in the destination range.
458     */
459    bool clamp_after_conversion = clamp &&
460                                  src_base_type == nir_type_float &&
461                                  dest_base_type != nir_type_float;
462 
463    /*
464     * If we don't care about rounding and clamping, we can just use NIR's
465     * built-in ops. There is also a special case for SPIR-V in shaders, where
466     * f32/f64 -> f16 conversions can have one of two rounding modes applied,
467     * which NIR has built-in opcodes for.
468     *
469     * For the rest, we have our own implementation of rounding and clamping.
470     */
471    bool trivial_convert;
472    if (!clamp && round == nir_rounding_mode_undef) {
473       trivial_convert = true;
474    } else if (!clamp && src_type == nir_type_float32 &&
475               dest_type == nir_type_float16 &&
476               (round == nir_rounding_mode_rtne ||
477                round == nir_rounding_mode_rtz)) {
478       trivial_convert = true;
479    } else {
480       trivial_convert = false;
481    }
482 
483    if (trivial_convert)
484       return nir_type_convert(b, src, src_type, dest_type, round);
485 
486    nir_def *dest = src;
487 
488    /* clamp the result into range */
489    if (clamp && !clamp_after_conversion)
490       dest = nir_clamp_to_type_range(b, src, src_type, src, src_type, dest_type);
491 
492    /* round with selected rounding mode */
493    if (!trivial_convert && round != nir_rounding_mode_undef) {
494       if (src_base_type == nir_type_float) {
495          if (dest_base_type == nir_type_float) {
496             dest = nir_round_float_to_float(b, dest, dest_bit_size, round);
497          } else {
498             dest = nir_round_float_to_int(b, dest, round);
499          }
500       } else {
501          dest = nir_round_int_to_float(b, dest, src_type, dest_bit_size, round);
502       }
503 
504       round = nir_rounding_mode_undef;
505    }
506 
507    /* now we can convert the value */
508    nir_op op = nir_type_conversion_op(src_type, dest_type, round);
509    dest = nir_build_alu(b, op, dest, NULL, NULL, NULL);
510 
511    if (clamp_after_conversion)
512       dest = nir_clamp_to_type_range(b, dest, dest_type, src, src_type, dest_type);
513 
514    return dest;
515 }
516 
517 #ifdef __cplusplus
518 }
519 #endif
520 
521 #endif /* NIR_CONVERSION_BUILDER_H */
522