projects / mesa.git / blob
? search:


46ba005945d99bc4394d75fac6ce5d9c6be6e422
[mesa.git] / src / compiler / spirv / spirv_to_nir.c
1 /*
2 * Copyright © 2015 Intel Corporation
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 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
33
34 struct spec_constant_value {
35 bool is_double;
36 union {
37 uint32_t data32;
38 uint64_t data64;
39 };
40 };
41
42 void
43 _vtn_warn(const char *file, int line, const char *msg, ...)
44 {
45 char *formatted;
46 va_list args;
47
48 va_start(args, msg);
49 formatted = ralloc_vasprintf(NULL, msg, args);
50 va_end(args);
51
52 fprintf(stderr, "%s:%d WARNING: %s\n", file, line, formatted);
53
54 ralloc_free(formatted);
55 }
56
57 static struct vtn_ssa_value *
58 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
59 {
60 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
61 val->type = type;
62
63 if (glsl_type_is_vector_or_scalar(type)) {
64 unsigned num_components = glsl_get_vector_elements(val->type);
65 unsigned bit_size = glsl_get_bit_size(val->type);
66 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
67 } else {
68 unsigned elems = glsl_get_length(val->type);
69 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
70 if (glsl_type_is_matrix(type)) {
71 const struct glsl_type *elem_type =
72 glsl_vector_type(glsl_get_base_type(type),
73 glsl_get_vector_elements(type));
74
75 for (unsigned i = 0; i < elems; i++)
76 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
77 } else if (glsl_type_is_array(type)) {
78 const struct glsl_type *elem_type = glsl_get_array_element(type);
79 for (unsigned i = 0; i < elems; i++)
80 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
81 } else {
82 for (unsigned i = 0; i < elems; i++) {
83 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
84 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
85 }
86 }
87 }
88
89 return val;
90 }
91
92 static struct vtn_ssa_value *
93 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
94 const struct glsl_type *type)
95 {
96 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
97
98 if (entry)
99 return entry->data;
100
101 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
102 val->type = type;
103
104 switch (glsl_get_base_type(type)) {
105 case GLSL_TYPE_INT:
106 case GLSL_TYPE_UINT:
107 case GLSL_TYPE_INT64:
108 case GLSL_TYPE_UINT64:
109 case GLSL_TYPE_BOOL:
110 case GLSL_TYPE_FLOAT:
111 case GLSL_TYPE_DOUBLE: {
112 int bit_size = glsl_get_bit_size(type);
113 if (glsl_type_is_vector_or_scalar(type)) {
114 unsigned num_components = glsl_get_vector_elements(val->type);
115 nir_load_const_instr *load =
116 nir_load_const_instr_create(b->shader, num_components, bit_size);
117
118 load->value = constant->values[0];
119
120 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
121 val->def = &load->def;
122 } else {
123 assert(glsl_type_is_matrix(type));
124 unsigned rows = glsl_get_vector_elements(val->type);
125 unsigned columns = glsl_get_matrix_columns(val->type);
126 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
127
128 for (unsigned i = 0; i < columns; i++) {
129 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
130 col_val->type = glsl_get_column_type(val->type);
131 nir_load_const_instr *load =
132 nir_load_const_instr_create(b->shader, rows, bit_size);
133
134 load->value = constant->values[i];
135
136 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
137 col_val->def = &load->def;
138
139 val->elems[i] = col_val;
140 }
141 }
142 break;
143 }
144
145 case GLSL_TYPE_ARRAY: {
146 unsigned elems = glsl_get_length(val->type);
147 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
148 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
149 for (unsigned i = 0; i < elems; i++)
150 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
151 elem_type);
152 break;
153 }
154
155 case GLSL_TYPE_STRUCT: {
156 unsigned elems = glsl_get_length(val->type);
157 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
158 for (unsigned i = 0; i < elems; i++) {
159 const struct glsl_type *elem_type =
160 glsl_get_struct_field(val->type, i);
161 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
162 elem_type);
163 }
164 break;
165 }
166
167 default:
168 unreachable("bad constant type");
169 }
170
171 return val;
172 }
173
174 struct vtn_ssa_value *
175 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
176 {
177 struct vtn_value *val = vtn_untyped_value(b, value_id);
178 switch (val->value_type) {
179 case vtn_value_type_undef:
180 return vtn_undef_ssa_value(b, val->type->type);
181
182 case vtn_value_type_constant:
183 return vtn_const_ssa_value(b, val->constant, val->const_type);
184
185 case vtn_value_type_ssa:
186 return val->ssa;
187
188 case vtn_value_type_pointer:
189 assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
190 struct vtn_ssa_value *ssa =
191 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
192 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
193 return ssa;
194
195 default:
196 unreachable("Invalid type for an SSA value");
197 }
198 }
199
200 static char *
201 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
202 unsigned word_count, unsigned *words_used)
203 {
204 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
205 if (words_used) {
206 /* Ammount of space taken by the string (including the null) */
207 unsigned len = strlen(dup) + 1;
208 *words_used = DIV_ROUND_UP(len, sizeof(*words));
209 }
210 return dup;
211 }
212
213 const uint32_t *
214 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
215 const uint32_t *end, vtn_instruction_handler handler)
216 {
217 b->file = NULL;
218 b->line = -1;
219 b->col = -1;
220
221 const uint32_t *w = start;
222 while (w < end) {
223 SpvOp opcode = w[0] & SpvOpCodeMask;
224 unsigned count = w[0] >> SpvWordCountShift;
225 assert(count >= 1 && w + count <= end);
226
227 switch (opcode) {
228 case SpvOpNop:
229 break; /* Do nothing */
230
231 case SpvOpLine:
232 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
233 b->line = w[2];
234 b->col = w[3];
235 break;
236
237 case SpvOpNoLine:
238 b->file = NULL;
239 b->line = -1;
240 b->col = -1;
241 break;
242
243 default:
244 if (!handler(b, opcode, w, count))
245 return w;
246 break;
247 }
248
249 w += count;
250 }
251 assert(w == end);
252 return w;
253 }
254
255 static void
256 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
257 const uint32_t *w, unsigned count)
258 {
259 switch (opcode) {
260 case SpvOpExtInstImport: {
261 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
262 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
263 val->ext_handler = vtn_handle_glsl450_instruction;
264 } else {
265 unreachable("Unsupported extension");
266 }
267 break;
268 }
269
270 case SpvOpExtInst: {
271 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
272 bool handled = val->ext_handler(b, w[4], w, count);
273 (void)handled;
274 assert(handled);
275 break;
276 }
277
278 default:
279 unreachable("Unhandled opcode");
280 }
281 }
282
283 static void
284 _foreach_decoration_helper(struct vtn_builder *b,
285 struct vtn_value *base_value,
286 int parent_member,
287 struct vtn_value *value,
288 vtn_decoration_foreach_cb cb, void *data)
289 {
290 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
291 int member;
292 if (dec->scope == VTN_DEC_DECORATION) {
293 member = parent_member;
294 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
295 assert(parent_member == -1);
296 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
297 } else {
298 /* Not a decoration */
299 continue;
300 }
301
302 if (dec->group) {
303 assert(dec->group->value_type == vtn_value_type_decoration_group);
304 _foreach_decoration_helper(b, base_value, member, dec->group,
305 cb, data);
306 } else {
307 cb(b, base_value, member, dec, data);
308 }
309 }
310 }
311
312 /** Iterates (recursively if needed) over all of the decorations on a value
313 *
314 * This function iterates over all of the decorations applied to a given
315 * value. If it encounters a decoration group, it recurses into the group
316 * and iterates over all of those decorations as well.
317 */
318 void
319 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
320 vtn_decoration_foreach_cb cb, void *data)
321 {
322 _foreach_decoration_helper(b, value, -1, value, cb, data);
323 }
324
325 void
326 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
327 vtn_execution_mode_foreach_cb cb, void *data)
328 {
329 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
330 if (dec->scope != VTN_DEC_EXECUTION_MODE)
331 continue;
332
333 assert(dec->group == NULL);
334 cb(b, value, dec, data);
335 }
336 }
337
338 static void
339 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
340 const uint32_t *w, unsigned count)
341 {
342 const uint32_t *w_end = w + count;
343 const uint32_t target = w[1];
344 w += 2;
345
346 switch (opcode) {
347 case SpvOpDecorationGroup:
348 vtn_push_value(b, target, vtn_value_type_decoration_group);
349 break;
350
351 case SpvOpDecorate:
352 case SpvOpMemberDecorate:
353 case SpvOpExecutionMode: {
354 struct vtn_value *val = &b->values[target];
355
356 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
357 switch (opcode) {
358 case SpvOpDecorate:
359 dec->scope = VTN_DEC_DECORATION;
360 break;
361 case SpvOpMemberDecorate:
362 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
363 break;
364 case SpvOpExecutionMode:
365 dec->scope = VTN_DEC_EXECUTION_MODE;
366 break;
367 default:
368 unreachable("Invalid decoration opcode");
369 }
370 dec->decoration = *(w++);
371 dec->literals = w;
372
373 /* Link into the list */
374 dec->next = val->decoration;
375 val->decoration = dec;
376 break;
377 }
378
379 case SpvOpGroupMemberDecorate:
380 case SpvOpGroupDecorate: {
381 struct vtn_value *group =
382 vtn_value(b, target, vtn_value_type_decoration_group);
383
384 for (; w < w_end; w++) {
385 struct vtn_value *val = vtn_untyped_value(b, *w);
386 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
387
388 dec->group = group;
389 if (opcode == SpvOpGroupDecorate) {
390 dec->scope = VTN_DEC_DECORATION;
391 } else {
392 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
393 }
394
395 /* Link into the list */
396 dec->next = val->decoration;
397 val->decoration = dec;
398 }
399 break;
400 }
401
402 default:
403 unreachable("Unhandled opcode");
404 }
405 }
406
407 struct member_decoration_ctx {
408 unsigned num_fields;
409 struct glsl_struct_field *fields;
410 struct vtn_type *type;
411 };
412
413 /* does a shallow copy of a vtn_type */
414
415 static struct vtn_type *
416 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
417 {
418 struct vtn_type *dest = ralloc(b, struct vtn_type);
419 *dest = *src;
420
421 switch (src->base_type) {
422 case vtn_base_type_void:
423 case vtn_base_type_scalar:
424 case vtn_base_type_vector:
425 case vtn_base_type_matrix:
426 case vtn_base_type_array:
427 case vtn_base_type_pointer:
428 case vtn_base_type_image:
429 case vtn_base_type_sampler:
430 /* Nothing more to do */
431 break;
432
433 case vtn_base_type_struct:
434 dest->members = ralloc_array(b, struct vtn_type *, src->length);
435 memcpy(dest->members, src->members,
436 src->length * sizeof(src->members[0]));
437
438 dest->offsets = ralloc_array(b, unsigned, src->length);
439 memcpy(dest->offsets, src->offsets,
440 src->length * sizeof(src->offsets[0]));
441 break;
442
443 case vtn_base_type_function:
444 dest->params = ralloc_array(b, struct vtn_type *, src->length);
445 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
446 break;
447 }
448
449 return dest;
450 }
451
452 static struct vtn_type *
453 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
454 {
455 type->members[member] = vtn_type_copy(b, type->members[member]);
456 type = type->members[member];
457
458 /* We may have an array of matrices.... Oh, joy! */
459 while (glsl_type_is_array(type->type)) {
460 type->array_element = vtn_type_copy(b, type->array_element);
461 type = type->array_element;
462 }
463
464 assert(glsl_type_is_matrix(type->type));
465
466 return type;
467 }
468
469 static void
470 struct_member_decoration_cb(struct vtn_builder *b,
471 struct vtn_value *val, int member,
472 const struct vtn_decoration *dec, void *void_ctx)
473 {
474 struct member_decoration_ctx *ctx = void_ctx;
475
476 if (member < 0)
477 return;
478
479 assert(member < ctx->num_fields);
480
481 switch (dec->decoration) {
482 case SpvDecorationNonWritable:
483 case SpvDecorationNonReadable:
484 case SpvDecorationRelaxedPrecision:
485 case SpvDecorationVolatile:
486 case SpvDecorationCoherent:
487 case SpvDecorationUniform:
488 break; /* FIXME: Do nothing with this for now. */
489 case SpvDecorationNoPerspective:
490 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
491 break;
492 case SpvDecorationFlat:
493 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
494 break;
495 case SpvDecorationCentroid:
496 ctx->fields[member].centroid = true;
497 break;
498 case SpvDecorationSample:
499 ctx->fields[member].sample = true;
500 break;
501 case SpvDecorationStream:
502 /* Vulkan only allows one GS stream */
503 assert(dec->literals[0] == 0);
504 break;
505 case SpvDecorationLocation:
506 ctx->fields[member].location = dec->literals[0];
507 break;
508 case SpvDecorationComponent:
509 break; /* FIXME: What should we do with these? */
510 case SpvDecorationBuiltIn:
511 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
512 ctx->type->members[member]->is_builtin = true;
513 ctx->type->members[member]->builtin = dec->literals[0];
514 ctx->type->builtin_block = true;
515 break;
516 case SpvDecorationOffset:
517 ctx->type->offsets[member] = dec->literals[0];
518 break;
519 case SpvDecorationMatrixStride:
520 /* Handled as a second pass */
521 break;
522 case SpvDecorationColMajor:
523 break; /* Nothing to do here. Column-major is the default. */
524 case SpvDecorationRowMajor:
525 mutable_matrix_member(b, ctx->type, member)->row_major = true;
526 break;
527
528 case SpvDecorationPatch:
529 break;
530
531 case SpvDecorationSpecId:
532 case SpvDecorationBlock:
533 case SpvDecorationBufferBlock:
534 case SpvDecorationArrayStride:
535 case SpvDecorationGLSLShared:
536 case SpvDecorationGLSLPacked:
537 case SpvDecorationInvariant:
538 case SpvDecorationRestrict:
539 case SpvDecorationAliased:
540 case SpvDecorationConstant:
541 case SpvDecorationIndex:
542 case SpvDecorationBinding:
543 case SpvDecorationDescriptorSet:
544 case SpvDecorationLinkageAttributes:
545 case SpvDecorationNoContraction:
546 case SpvDecorationInputAttachmentIndex:
547 vtn_warn("Decoration not allowed on struct members: %s",
548 spirv_decoration_to_string(dec->decoration));
549 break;
550
551 case SpvDecorationXfbBuffer:
552 case SpvDecorationXfbStride:
553 vtn_warn("Vulkan does not have transform feedback");
554 break;
555
556 case SpvDecorationCPacked:
557 case SpvDecorationSaturatedConversion:
558 case SpvDecorationFuncParamAttr:
559 case SpvDecorationFPRoundingMode:
560 case SpvDecorationFPFastMathMode:
561 case SpvDecorationAlignment:
562 vtn_warn("Decoration only allowed for CL-style kernels: %s",
563 spirv_decoration_to_string(dec->decoration));
564 break;
565
566 default:
567 unreachable("Unhandled decoration");
568 }
569 }
570
571 /* Matrix strides are handled as a separate pass because we need to know
572 * whether the matrix is row-major or not first.
573 */
574 static void
575 struct_member_matrix_stride_cb(struct vtn_builder *b,
576 struct vtn_value *val, int member,
577 const struct vtn_decoration *dec,
578 void *void_ctx)
579 {
580 if (dec->decoration != SpvDecorationMatrixStride)
581 return;
582 assert(member >= 0);
583
584 struct member_decoration_ctx *ctx = void_ctx;
585
586 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
587 if (mat_type->row_major) {
588 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
589 mat_type->stride = mat_type->array_element->stride;
590 mat_type->array_element->stride = dec->literals[0];
591 } else {
592 assert(mat_type->array_element->stride > 0);
593 mat_type->stride = dec->literals[0];
594 }
595 }
596
597 static void
598 type_decoration_cb(struct vtn_builder *b,
599 struct vtn_value *val, int member,
600 const struct vtn_decoration *dec, void *ctx)
601 {
602 struct vtn_type *type = val->type;
603
604 if (member != -1)
605 return;
606
607 switch (dec->decoration) {
608 case SpvDecorationArrayStride:
609 assert(type->base_type == vtn_base_type_matrix ||
610 type->base_type == vtn_base_type_array ||
611 type->base_type == vtn_base_type_pointer);
612 type->stride = dec->literals[0];
613 break;
614 case SpvDecorationBlock:
615 assert(type->base_type == vtn_base_type_struct);
616 type->block = true;
617 break;
618 case SpvDecorationBufferBlock:
619 assert(type->base_type == vtn_base_type_struct);
620 type->buffer_block = true;
621 break;
622 case SpvDecorationGLSLShared:
623 case SpvDecorationGLSLPacked:
624 /* Ignore these, since we get explicit offsets anyways */
625 break;
626
627 case SpvDecorationRowMajor:
628 case SpvDecorationColMajor:
629 case SpvDecorationMatrixStride:
630 case SpvDecorationBuiltIn:
631 case SpvDecorationNoPerspective:
632 case SpvDecorationFlat:
633 case SpvDecorationPatch:
634 case SpvDecorationCentroid:
635 case SpvDecorationSample:
636 case SpvDecorationVolatile:
637 case SpvDecorationCoherent:
638 case SpvDecorationNonWritable:
639 case SpvDecorationNonReadable:
640 case SpvDecorationUniform:
641 case SpvDecorationStream:
642 case SpvDecorationLocation:
643 case SpvDecorationComponent:
644 case SpvDecorationOffset:
645 case SpvDecorationXfbBuffer:
646 case SpvDecorationXfbStride:
647 vtn_warn("Decoration only allowed for struct members: %s",
648 spirv_decoration_to_string(dec->decoration));
649 break;
650
651 case SpvDecorationRelaxedPrecision:
652 case SpvDecorationSpecId:
653 case SpvDecorationInvariant:
654 case SpvDecorationRestrict:
655 case SpvDecorationAliased:
656 case SpvDecorationConstant:
657 case SpvDecorationIndex:
658 case SpvDecorationBinding:
659 case SpvDecorationDescriptorSet:
660 case SpvDecorationLinkageAttributes:
661 case SpvDecorationNoContraction:
662 case SpvDecorationInputAttachmentIndex:
663 vtn_warn("Decoration not allowed on types: %s",
664 spirv_decoration_to_string(dec->decoration));
665 break;
666
667 case SpvDecorationCPacked:
668 case SpvDecorationSaturatedConversion:
669 case SpvDecorationFuncParamAttr:
670 case SpvDecorationFPRoundingMode:
671 case SpvDecorationFPFastMathMode:
672 case SpvDecorationAlignment:
673 vtn_warn("Decoration only allowed for CL-style kernels: %s",
674 spirv_decoration_to_string(dec->decoration));
675 break;
676
677 default:
678 unreachable("Unhandled decoration");
679 }
680 }
681
682 static unsigned
683 translate_image_format(SpvImageFormat format)
684 {
685 switch (format) {
686 case SpvImageFormatUnknown: return 0; /* GL_NONE */
687 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
688 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
689 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
690 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
691 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
692 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
693 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
694 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
695 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
696 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
697 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
698 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
699 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
700 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
701 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
702 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
703 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
704 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
705 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
706 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
707 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
708 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
709 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
710 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
711 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
712 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
713 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
714 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
715 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
716 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
717 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
718 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
719 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
720 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
721 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
722 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
723 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
724 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
725 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
726 default:
727 unreachable("Invalid image format");
728 return 0;
729 }
730 }
731
732 static void
733 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
734 const uint32_t *w, unsigned count)
735 {
736 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
737
738 val->type = rzalloc(b, struct vtn_type);
739 val->type->val = val;
740
741 switch (opcode) {
742 case SpvOpTypeVoid:
743 val->type->base_type = vtn_base_type_void;
744 val->type->type = glsl_void_type();
745 break;
746 case SpvOpTypeBool:
747 val->type->base_type = vtn_base_type_scalar;
748 val->type->type = glsl_bool_type();
749 break;
750 case SpvOpTypeInt: {
751 int bit_size = w[2];
752 const bool signedness = w[3];
753 val->type->base_type = vtn_base_type_scalar;
754 if (bit_size == 64)
755 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
756 else
757 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
758 break;
759 }
760 case SpvOpTypeFloat: {
761 int bit_size = w[2];
762 val->type->base_type = vtn_base_type_scalar;
763 val->type->type = bit_size == 64 ? glsl_double_type() : glsl_float_type();
764 break;
765 }
766
767 case SpvOpTypeVector: {
768 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
769 unsigned elems = w[3];
770
771 assert(glsl_type_is_scalar(base->type));
772 val->type->base_type = vtn_base_type_vector;
773 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
774 val->type->stride = glsl_get_bit_size(base->type) / 8;
775 val->type->array_element = base;
776 break;
777 }
778
779 case SpvOpTypeMatrix: {
780 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
781 unsigned columns = w[3];
782
783 assert(glsl_type_is_vector(base->type));
784 val->type->base_type = vtn_base_type_matrix;
785 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
786 glsl_get_vector_elements(base->type),
787 columns);
788 assert(!glsl_type_is_error(val->type->type));
789 val->type->length = columns;
790 val->type->array_element = base;
791 val->type->row_major = false;
792 val->type->stride = 0;
793 break;
794 }
795
796 case SpvOpTypeRuntimeArray:
797 case SpvOpTypeArray: {
798 struct vtn_type *array_element =
799 vtn_value(b, w[2], vtn_value_type_type)->type;
800
801 if (opcode == SpvOpTypeRuntimeArray) {
802 /* A length of 0 is used to denote unsized arrays */
803 val->type->length = 0;
804 } else {
805 val->type->length =
806 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
807 }
808
809 val->type->base_type = vtn_base_type_array;
810 val->type->type = glsl_array_type(array_element->type, val->type->length);
811 val->type->array_element = array_element;
812 val->type->stride = 0;
813 break;
814 }
815
816 case SpvOpTypeStruct: {
817 unsigned num_fields = count - 2;
818 val->type->base_type = vtn_base_type_struct;
819 val->type->length = num_fields;
820 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
821 val->type->offsets = ralloc_array(b, unsigned, num_fields);
822
823 NIR_VLA(struct glsl_struct_field, fields, count);
824 for (unsigned i = 0; i < num_fields; i++) {
825 val->type->members[i] =
826 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
827 fields[i] = (struct glsl_struct_field) {
828 .type = val->type->members[i]->type,
829 .name = ralloc_asprintf(b, "field%d", i),
830 .location = -1,
831 };
832 }
833
834 struct member_decoration_ctx ctx = {
835 .num_fields = num_fields,
836 .fields = fields,
837 .type = val->type
838 };
839
840 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
841 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
842
843 const char *name = val->name ? val->name : "struct";
844
845 val->type->type = glsl_struct_type(fields, num_fields, name);
846 break;
847 }
848
849 case SpvOpTypeFunction: {
850 val->type->base_type = vtn_base_type_function;
851 val->type->type = NULL;
852
853 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
854
855 const unsigned num_params = count - 3;
856 val->type->length = num_params;
857 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
858 for (unsigned i = 0; i < count - 3; i++) {
859 val->type->params[i] =
860 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
861 }
862 break;
863 }
864
865 case SpvOpTypePointer: {
866 SpvStorageClass storage_class = w[2];
867 struct vtn_type *deref_type =
868 vtn_value(b, w[3], vtn_value_type_type)->type;
869
870 val->type->base_type = vtn_base_type_pointer;
871 val->type->storage_class = storage_class;
872 val->type->deref = deref_type;
873
874 if (storage_class == SpvStorageClassUniform ||
875 storage_class == SpvStorageClassStorageBuffer) {
876 /* These can actually be stored to nir_variables and used as SSA
877 * values so they need a real glsl_type.
878 */
879 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
880 }
881 break;
882 }
883
884 case SpvOpTypeImage: {
885 val->type->base_type = vtn_base_type_image;
886
887 const struct glsl_type *sampled_type =
888 vtn_value(b, w[2], vtn_value_type_type)->type->type;
889
890 assert(glsl_type_is_vector_or_scalar(sampled_type));
891
892 enum glsl_sampler_dim dim;
893 switch ((SpvDim)w[3]) {
894 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
895 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
896 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
897 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
898 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
899 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
900 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
901 default:
902 unreachable("Invalid SPIR-V Sampler dimension");
903 }
904
905 bool is_shadow = w[4];
906 bool is_array = w[5];
907 bool multisampled = w[6];
908 unsigned sampled = w[7];
909 SpvImageFormat format = w[8];
910
911 if (count > 9)
912 val->type->access_qualifier = w[9];
913 else
914 val->type->access_qualifier = SpvAccessQualifierReadWrite;
915
916 if (multisampled) {
917 if (dim == GLSL_SAMPLER_DIM_2D)
918 dim = GLSL_SAMPLER_DIM_MS;
919 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
920 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
921 else
922 unreachable("Unsupported multisampled image type");
923 }
924
925 val->type->image_format = translate_image_format(format);
926
927 if (sampled == 1) {
928 val->type->sampled = true;
929 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
930 glsl_get_base_type(sampled_type));
931 } else if (sampled == 2) {
932 assert(!is_shadow);
933 val->type->sampled = false;
934 val->type->type = glsl_image_type(dim, is_array,
935 glsl_get_base_type(sampled_type));
936 } else {
937 unreachable("We need to know if the image will be sampled");
938 }
939 break;
940 }
941
942 case SpvOpTypeSampledImage:
943 val->type = vtn_value(b, w[2], vtn_value_type_type)->type;
944 break;
945
946 case SpvOpTypeSampler:
947 /* The actual sampler type here doesn't really matter. It gets
948 * thrown away the moment you combine it with an image. What really
949 * matters is that it's a sampler type as opposed to an integer type
950 * so the backend knows what to do.
951 */
952 val->type->base_type = vtn_base_type_sampler;
953 val->type->type = glsl_bare_sampler_type();
954 break;
955
956 case SpvOpTypeOpaque:
957 case SpvOpTypeEvent:
958 case SpvOpTypeDeviceEvent:
959 case SpvOpTypeReserveId:
960 case SpvOpTypeQueue:
961 case SpvOpTypePipe:
962 default:
963 unreachable("Unhandled opcode");
964 }
965
966 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
967 }
968
969 static nir_constant *
970 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
971 {
972 nir_constant *c = rzalloc(b, nir_constant);
973
974 /* For pointers and other typeless things, we have to return something but
975 * it doesn't matter what.
976 */
977 if (!type)
978 return c;
979
980 switch (glsl_get_base_type(type)) {
981 case GLSL_TYPE_INT:
982 case GLSL_TYPE_UINT:
983 case GLSL_TYPE_INT64:
984 case GLSL_TYPE_UINT64:
985 case GLSL_TYPE_BOOL:
986 case GLSL_TYPE_FLOAT:
987 case GLSL_TYPE_DOUBLE:
988 /* Nothing to do here. It's already initialized to zero */
989 break;
990
991 case GLSL_TYPE_ARRAY:
992 assert(glsl_get_length(type) > 0);
993 c->num_elements = glsl_get_length(type);
994 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
995
996 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
997 for (unsigned i = 1; i < c->num_elements; i++)
998 c->elements[i] = c->elements[0];
999 break;
1000
1001 case GLSL_TYPE_STRUCT:
1002 c->num_elements = glsl_get_length(type);
1003 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1004
1005 for (unsigned i = 0; i < c->num_elements; i++) {
1006 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1007 }
1008 break;
1009
1010 default:
1011 unreachable("Invalid type for null constant");
1012 }
1013
1014 return c;
1015 }
1016
1017 static void
1018 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1019 int member, const struct vtn_decoration *dec,
1020 void *data)
1021 {
1022 assert(member == -1);
1023 if (dec->decoration != SpvDecorationSpecId)
1024 return;
1025
1026 struct spec_constant_value *const_value = data;
1027
1028 for (unsigned i = 0; i < b->num_specializations; i++) {
1029 if (b->specializations[i].id == dec->literals[0]) {
1030 if (const_value->is_double)
1031 const_value->data64 = b->specializations[i].data64;
1032 else
1033 const_value->data32 = b->specializations[i].data32;
1034 return;
1035 }
1036 }
1037 }
1038
1039 static uint32_t
1040 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1041 uint32_t const_value)
1042 {
1043 struct spec_constant_value data;
1044 data.is_double = false;
1045 data.data32 = const_value;
1046 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1047 return data.data32;
1048 }
1049
1050 static uint64_t
1051 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1052 uint64_t const_value)
1053 {
1054 struct spec_constant_value data;
1055 data.is_double = true;
1056 data.data64 = const_value;
1057 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1058 return data.data64;
1059 }
1060
1061 static void
1062 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1063 struct vtn_value *val,
1064 int member,
1065 const struct vtn_decoration *dec,
1066 void *data)
1067 {
1068 assert(member == -1);
1069 if (dec->decoration != SpvDecorationBuiltIn ||
1070 dec->literals[0] != SpvBuiltInWorkgroupSize)
1071 return;
1072
1073 assert(val->const_type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1074
1075 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1076 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1077 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1078 }
1079
1080 static void
1081 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1082 const uint32_t *w, unsigned count)
1083 {
1084 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1085 val->const_type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
1086 val->constant = rzalloc(b, nir_constant);
1087 switch (opcode) {
1088 case SpvOpConstantTrue:
1089 assert(val->const_type == glsl_bool_type());
1090 val->constant->values[0].u32[0] = NIR_TRUE;
1091 break;
1092 case SpvOpConstantFalse:
1093 assert(val->const_type == glsl_bool_type());
1094 val->constant->values[0].u32[0] = NIR_FALSE;
1095 break;
1096
1097 case SpvOpSpecConstantTrue:
1098 case SpvOpSpecConstantFalse: {
1099 assert(val->const_type == glsl_bool_type());
1100 uint32_t int_val =
1101 get_specialization(b, val, (opcode == SpvOpSpecConstantTrue));
1102 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1103 break;
1104 }
1105
1106 case SpvOpConstant: {
1107 assert(glsl_type_is_scalar(val->const_type));
1108 int bit_size = glsl_get_bit_size(val->const_type);
1109 if (bit_size == 64) {
1110 val->constant->values->u32[0] = w[3];
1111 val->constant->values->u32[1] = w[4];
1112 } else {
1113 assert(bit_size == 32);
1114 val->constant->values->u32[0] = w[3];
1115 }
1116 break;
1117 }
1118 case SpvOpSpecConstant: {
1119 assert(glsl_type_is_scalar(val->const_type));
1120 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1121 int bit_size = glsl_get_bit_size(val->const_type);
1122 if (bit_size == 64)
1123 val->constant->values[0].u64[0] =
1124 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1125 else
1126 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1127 break;
1128 }
1129 case SpvOpSpecConstantComposite:
1130 case SpvOpConstantComposite: {
1131 unsigned elem_count = count - 3;
1132 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1133 for (unsigned i = 0; i < elem_count; i++)
1134 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1135
1136 switch (glsl_get_base_type(val->const_type)) {
1137 case GLSL_TYPE_UINT:
1138 case GLSL_TYPE_INT:
1139 case GLSL_TYPE_UINT64:
1140 case GLSL_TYPE_INT64:
1141 case GLSL_TYPE_FLOAT:
1142 case GLSL_TYPE_BOOL:
1143 case GLSL_TYPE_DOUBLE: {
1144 int bit_size = glsl_get_bit_size(val->const_type);
1145 if (glsl_type_is_matrix(val->const_type)) {
1146 assert(glsl_get_matrix_columns(val->const_type) == elem_count);
1147 for (unsigned i = 0; i < elem_count; i++)
1148 val->constant->values[i] = elems[i]->values[0];
1149 } else {
1150 assert(glsl_type_is_vector(val->const_type));
1151 assert(glsl_get_vector_elements(val->const_type) == elem_count);
1152 for (unsigned i = 0; i < elem_count; i++) {
1153 if (bit_size == 64) {
1154 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1155 } else {
1156 assert(bit_size == 32);
1157 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1158 }
1159 }
1160 }
1161 ralloc_free(elems);
1162 break;
1163 }
1164 case GLSL_TYPE_STRUCT:
1165 case GLSL_TYPE_ARRAY:
1166 ralloc_steal(val->constant, elems);
1167 val->constant->num_elements = elem_count;
1168 val->constant->elements = elems;
1169 break;
1170
1171 default:
1172 unreachable("Unsupported type for constants");
1173 }
1174 break;
1175 }
1176
1177 case SpvOpSpecConstantOp: {
1178 SpvOp opcode = get_specialization(b, val, w[3]);
1179 switch (opcode) {
1180 case SpvOpVectorShuffle: {
1181 struct vtn_value *v0 = &b->values[w[4]];
1182 struct vtn_value *v1 = &b->values[w[5]];
1183
1184 assert(v0->value_type == vtn_value_type_constant ||
1185 v0->value_type == vtn_value_type_undef);
1186 assert(v1->value_type == vtn_value_type_constant ||
1187 v1->value_type == vtn_value_type_undef);
1188
1189 unsigned len0 = v0->value_type == vtn_value_type_constant ?
1190 glsl_get_vector_elements(v0->const_type) :
1191 glsl_get_vector_elements(v0->type->type);
1192 unsigned len1 = v1->value_type == vtn_value_type_constant ?
1193 glsl_get_vector_elements(v1->const_type) :
1194 glsl_get_vector_elements(v1->type->type);
1195
1196 assert(len0 + len1 < 16);
1197
1198 unsigned bit_size = glsl_get_bit_size(val->const_type);
1199 unsigned bit_size0 = v0->value_type == vtn_value_type_constant ?
1200 glsl_get_bit_size(v0->const_type) :
1201 glsl_get_bit_size(v0->type->type);
1202 unsigned bit_size1 = v1->value_type == vtn_value_type_constant ?
1203 glsl_get_bit_size(v1->const_type) :
1204 glsl_get_bit_size(v1->type->type);
1205
1206 assert(bit_size == bit_size0 && bit_size == bit_size1);
1207 (void)bit_size0; (void)bit_size1;
1208
1209 if (bit_size == 64) {
1210 uint64_t u64[8];
1211 if (v0->value_type == vtn_value_type_constant) {
1212 for (unsigned i = 0; i < len0; i++)
1213 u64[i] = v0->constant->values[0].u64[i];
1214 }
1215 if (v1->value_type == vtn_value_type_constant) {
1216 for (unsigned i = 0; i < len1; i++)
1217 u64[len0 + i] = v1->constant->values[0].u64[i];
1218 }
1219
1220 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1221 uint32_t comp = w[i + 6];
1222 /* If component is not used, set the value to a known constant
1223 * to detect if it is wrongly used.
1224 */
1225 if (comp == (uint32_t)-1)
1226 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1227 else
1228 val->constant->values[0].u64[j] = u64[comp];
1229 }
1230 } else {
1231 uint32_t u32[8];
1232 if (v0->value_type == vtn_value_type_constant) {
1233 for (unsigned i = 0; i < len0; i++)
1234 u32[i] = v0->constant->values[0].u32[i];
1235 }
1236 if (v1->value_type == vtn_value_type_constant) {
1237 for (unsigned i = 0; i < len1; i++)
1238 u32[len0 + i] = v1->constant->values[0].u32[i];
1239 }
1240
1241 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1242 uint32_t comp = w[i + 6];
1243 /* If component is not used, set the value to a known constant
1244 * to detect if it is wrongly used.
1245 */
1246 if (comp == (uint32_t)-1)
1247 val->constant->values[0].u32[j] = 0xdeadbeef;
1248 else
1249 val->constant->values[0].u32[j] = u32[comp];
1250 }
1251 }
1252 break;
1253 }
1254
1255 case SpvOpCompositeExtract:
1256 case SpvOpCompositeInsert: {
1257 struct vtn_value *comp;
1258 unsigned deref_start;
1259 struct nir_constant **c;
1260 if (opcode == SpvOpCompositeExtract) {
1261 comp = vtn_value(b, w[4], vtn_value_type_constant);
1262 deref_start = 5;
1263 c = &comp->constant;
1264 } else {
1265 comp = vtn_value(b, w[5], vtn_value_type_constant);
1266 deref_start = 6;
1267 val->constant = nir_constant_clone(comp->constant,
1268 (nir_variable *)b);
1269 c = &val->constant;
1270 }
1271
1272 int elem = -1;
1273 int col = 0;
1274 const struct glsl_type *type = comp->const_type;
1275 for (unsigned i = deref_start; i < count; i++) {
1276 switch (glsl_get_base_type(type)) {
1277 case GLSL_TYPE_UINT:
1278 case GLSL_TYPE_INT:
1279 case GLSL_TYPE_UINT64:
1280 case GLSL_TYPE_INT64:
1281 case GLSL_TYPE_FLOAT:
1282 case GLSL_TYPE_DOUBLE:
1283 case GLSL_TYPE_BOOL:
1284 /* If we hit this granularity, we're picking off an element */
1285 if (glsl_type_is_matrix(type)) {
1286 assert(col == 0 && elem == -1);
1287 col = w[i];
1288 elem = 0;
1289 type = glsl_get_column_type(type);
1290 } else {
1291 assert(elem <= 0 && glsl_type_is_vector(type));
1292 elem = w[i];
1293 type = glsl_scalar_type(glsl_get_base_type(type));
1294 }
1295 continue;
1296
1297 case GLSL_TYPE_ARRAY:
1298 c = &(*c)->elements[w[i]];
1299 type = glsl_get_array_element(type);
1300 continue;
1301
1302 case GLSL_TYPE_STRUCT:
1303 c = &(*c)->elements[w[i]];
1304 type = glsl_get_struct_field(type, w[i]);
1305 continue;
1306
1307 default:
1308 unreachable("Invalid constant type");
1309 }
1310 }
1311
1312 if (opcode == SpvOpCompositeExtract) {
1313 if (elem == -1) {
1314 val->constant = *c;
1315 } else {
1316 unsigned num_components = glsl_get_vector_elements(type);
1317 unsigned bit_size = glsl_get_bit_size(type);
1318 for (unsigned i = 0; i < num_components; i++)
1319 if (bit_size == 64) {
1320 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1321 } else {
1322 assert(bit_size == 32);
1323 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1324 }
1325 }
1326 } else {
1327 struct vtn_value *insert =
1328 vtn_value(b, w[4], vtn_value_type_constant);
1329 assert(insert->const_type == type);
1330 if (elem == -1) {
1331 *c = insert->constant;
1332 } else {
1333 unsigned num_components = glsl_get_vector_elements(type);
1334 unsigned bit_size = glsl_get_bit_size(type);
1335 for (unsigned i = 0; i < num_components; i++)
1336 if (bit_size == 64) {
1337 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1338 } else {
1339 assert(bit_size == 32);
1340 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1341 }
1342 }
1343 }
1344 break;
1345 }
1346
1347 default: {
1348 bool swap;
1349 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->const_type);
1350 nir_alu_type src_alu_type = dst_alu_type;
1351 nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap, src_alu_type, dst_alu_type);
1352
1353 unsigned num_components = glsl_get_vector_elements(val->const_type);
1354 unsigned bit_size =
1355 glsl_get_bit_size(val->const_type);
1356
1357 nir_const_value src[4];
1358 assert(count <= 7);
1359 for (unsigned i = 0; i < count - 4; i++) {
1360 nir_constant *c =
1361 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1362
1363 unsigned j = swap ? 1 - i : i;
1364 assert(bit_size == 32);
1365 src[j] = c->values[0];
1366 }
1367
1368 val->constant->values[0] =
1369 nir_eval_const_opcode(op, num_components, bit_size, src);
1370 break;
1371 } /* default */
1372 }
1373 break;
1374 }
1375
1376 case SpvOpConstantNull:
1377 val->constant = vtn_null_constant(b, val->const_type);
1378 break;
1379
1380 case SpvOpConstantSampler:
1381 unreachable("OpConstantSampler requires Kernel Capability");
1382 break;
1383
1384 default:
1385 unreachable("Unhandled opcode");
1386 }
1387
1388 /* Now that we have the value, update the workgroup size if needed */
1389 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1390 }
1391
1392 static void
1393 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1394 const uint32_t *w, unsigned count)
1395 {
1396 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1397 struct vtn_function *vtn_callee =
1398 vtn_value(b, w[3], vtn_value_type_function)->func;
1399 struct nir_function *callee = vtn_callee->impl->function;
1400
1401 vtn_callee->referenced = true;
1402
1403 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1404 for (unsigned i = 0; i < call->num_params; i++) {
1405 unsigned arg_id = w[4 + i];
1406 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1407 if (arg->value_type == vtn_value_type_pointer &&
1408 arg->pointer->ptr_type->type == NULL) {
1409 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1410 call->params[i] = nir_deref_var_clone(d, call);
1411 } else {
1412 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1413
1414 /* Make a temporary to store the argument in */
1415 nir_variable *tmp =
1416 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1417 call->params[i] = nir_deref_var_create(call, tmp);
1418
1419 vtn_local_store(b, arg_ssa, call->params[i]);
1420 }
1421 }
1422
1423 nir_variable *out_tmp = NULL;
1424 assert(res_type->type == callee->return_type);
1425 if (!glsl_type_is_void(callee->return_type)) {
1426 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1427 "out_tmp");
1428 call->return_deref = nir_deref_var_create(call, out_tmp);
1429 }
1430
1431 nir_builder_instr_insert(&b->nb, &call->instr);
1432
1433 if (glsl_type_is_void(callee->return_type)) {
1434 vtn_push_value(b, w[2], vtn_value_type_undef);
1435 } else {
1436 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1437 }
1438 }
1439
1440 struct vtn_ssa_value *
1441 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1442 {
1443 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1444 val->type = type;
1445
1446 if (!glsl_type_is_vector_or_scalar(type)) {
1447 unsigned elems = glsl_get_length(type);
1448 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1449 for (unsigned i = 0; i < elems; i++) {
1450 const struct glsl_type *child_type;
1451
1452 switch (glsl_get_base_type(type)) {
1453 case GLSL_TYPE_INT:
1454 case GLSL_TYPE_UINT:
1455 case GLSL_TYPE_INT64:
1456 case GLSL_TYPE_UINT64:
1457 case GLSL_TYPE_BOOL:
1458 case GLSL_TYPE_FLOAT:
1459 case GLSL_TYPE_DOUBLE:
1460 child_type = glsl_get_column_type(type);
1461 break;
1462 case GLSL_TYPE_ARRAY:
1463 child_type = glsl_get_array_element(type);
1464 break;
1465 case GLSL_TYPE_STRUCT:
1466 child_type = glsl_get_struct_field(type, i);
1467 break;
1468 default:
1469 unreachable("unkown base type");
1470 }
1471
1472 val->elems[i] = vtn_create_ssa_value(b, child_type);
1473 }
1474 }
1475
1476 return val;
1477 }
1478
1479 static nir_tex_src
1480 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1481 {
1482 nir_tex_src src;
1483 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1484 src.src_type = type;
1485 return src;
1486 }
1487
1488 static void
1489 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1490 const uint32_t *w, unsigned count)
1491 {
1492 if (opcode == SpvOpSampledImage) {
1493 struct vtn_value *val =
1494 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1495 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1496 val->sampled_image->type =
1497 vtn_value(b, w[1], vtn_value_type_type)->type;
1498 val->sampled_image->image =
1499 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1500 val->sampled_image->sampler =
1501 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1502 return;
1503 } else if (opcode == SpvOpImage) {
1504 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1505 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1506 if (src_val->value_type == vtn_value_type_sampled_image) {
1507 val->pointer = src_val->sampled_image->image;
1508 } else {
1509 assert(src_val->value_type == vtn_value_type_pointer);
1510 val->pointer = src_val->pointer;
1511 }
1512 return;
1513 }
1514
1515 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1516 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1517
1518 struct vtn_sampled_image sampled;
1519 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1520 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1521 sampled = *sampled_val->sampled_image;
1522 } else {
1523 assert(sampled_val->value_type == vtn_value_type_pointer);
1524 sampled.type = sampled_val->pointer->type;
1525 sampled.image = NULL;
1526 sampled.sampler = sampled_val->pointer;
1527 }
1528
1529 const struct glsl_type *image_type = sampled.type->type;
1530 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1531 const bool is_array = glsl_sampler_type_is_array(image_type);
1532 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1533
1534 /* Figure out the base texture operation */
1535 nir_texop texop;
1536 switch (opcode) {
1537 case SpvOpImageSampleImplicitLod:
1538 case SpvOpImageSampleDrefImplicitLod:
1539 case SpvOpImageSampleProjImplicitLod:
1540 case SpvOpImageSampleProjDrefImplicitLod:
1541 texop = nir_texop_tex;
1542 break;
1543
1544 case SpvOpImageSampleExplicitLod:
1545 case SpvOpImageSampleDrefExplicitLod:
1546 case SpvOpImageSampleProjExplicitLod:
1547 case SpvOpImageSampleProjDrefExplicitLod:
1548 texop = nir_texop_txl;
1549 break;
1550
1551 case SpvOpImageFetch:
1552 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1553 texop = nir_texop_txf_ms;
1554 } else {
1555 texop = nir_texop_txf;
1556 }
1557 break;
1558
1559 case SpvOpImageGather:
1560 case SpvOpImageDrefGather:
1561 texop = nir_texop_tg4;
1562 break;
1563
1564 case SpvOpImageQuerySizeLod:
1565 case SpvOpImageQuerySize:
1566 texop = nir_texop_txs;
1567 break;
1568
1569 case SpvOpImageQueryLod:
1570 texop = nir_texop_lod;
1571 break;
1572
1573 case SpvOpImageQueryLevels:
1574 texop = nir_texop_query_levels;
1575 break;
1576
1577 case SpvOpImageQuerySamples:
1578 texop = nir_texop_texture_samples;
1579 break;
1580
1581 default:
1582 unreachable("Unhandled opcode");
1583 }
1584
1585 nir_tex_src srcs[8]; /* 8 should be enough */
1586 nir_tex_src *p = srcs;
1587
1588 unsigned idx = 4;
1589
1590 struct nir_ssa_def *coord;
1591 unsigned coord_components;
1592 switch (opcode) {
1593 case SpvOpImageSampleImplicitLod:
1594 case SpvOpImageSampleExplicitLod:
1595 case SpvOpImageSampleDrefImplicitLod:
1596 case SpvOpImageSampleDrefExplicitLod:
1597 case SpvOpImageSampleProjImplicitLod:
1598 case SpvOpImageSampleProjExplicitLod:
1599 case SpvOpImageSampleProjDrefImplicitLod:
1600 case SpvOpImageSampleProjDrefExplicitLod:
1601 case SpvOpImageFetch:
1602 case SpvOpImageGather:
1603 case SpvOpImageDrefGather:
1604 case SpvOpImageQueryLod: {
1605 /* All these types have the coordinate as their first real argument */
1606 switch (sampler_dim) {
1607 case GLSL_SAMPLER_DIM_1D:
1608 case GLSL_SAMPLER_DIM_BUF:
1609 coord_components = 1;
1610 break;
1611 case GLSL_SAMPLER_DIM_2D:
1612 case GLSL_SAMPLER_DIM_RECT:
1613 case GLSL_SAMPLER_DIM_MS:
1614 coord_components = 2;
1615 break;
1616 case GLSL_SAMPLER_DIM_3D:
1617 case GLSL_SAMPLER_DIM_CUBE:
1618 coord_components = 3;
1619 break;
1620 default:
1621 unreachable("Invalid sampler type");
1622 }
1623
1624 if (is_array && texop != nir_texop_lod)
1625 coord_components++;
1626
1627 coord = vtn_ssa_value(b, w[idx++])->def;
1628 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1629 (1 << coord_components) - 1));
1630 p->src_type = nir_tex_src_coord;
1631 p++;
1632 break;
1633 }
1634
1635 default:
1636 coord = NULL;
1637 coord_components = 0;
1638 break;
1639 }
1640
1641 switch (opcode) {
1642 case SpvOpImageSampleProjImplicitLod:
1643 case SpvOpImageSampleProjExplicitLod:
1644 case SpvOpImageSampleProjDrefImplicitLod:
1645 case SpvOpImageSampleProjDrefExplicitLod:
1646 /* These have the projector as the last coordinate component */
1647 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1648 p->src_type = nir_tex_src_projector;
1649 p++;
1650 break;
1651
1652 default:
1653 break;
1654 }
1655
1656 unsigned gather_component = 0;
1657 switch (opcode) {
1658 case SpvOpImageSampleDrefImplicitLod:
1659 case SpvOpImageSampleDrefExplicitLod:
1660 case SpvOpImageSampleProjDrefImplicitLod:
1661 case SpvOpImageSampleProjDrefExplicitLod:
1662 case SpvOpImageDrefGather:
1663 /* These all have an explicit depth value as their next source */
1664 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
1665 break;
1666
1667 case SpvOpImageGather:
1668 /* This has a component as its next source */
1669 gather_component =
1670 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
1671 break;
1672
1673 default:
1674 break;
1675 }
1676
1677 /* For OpImageQuerySizeLod, we always have an LOD */
1678 if (opcode == SpvOpImageQuerySizeLod)
1679 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1680
1681 /* Now we need to handle some number of optional arguments */
1682 const struct vtn_ssa_value *gather_offsets = NULL;
1683 if (idx < count) {
1684 uint32_t operands = w[idx++];
1685
1686 if (operands & SpvImageOperandsBiasMask) {
1687 assert(texop == nir_texop_tex);
1688 texop = nir_texop_txb;
1689 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1690 }
1691
1692 if (operands & SpvImageOperandsLodMask) {
1693 assert(texop == nir_texop_txl || texop == nir_texop_txf ||
1694 texop == nir_texop_txs);
1695 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1696 }
1697
1698 if (operands & SpvImageOperandsGradMask) {
1699 assert(texop == nir_texop_txl);
1700 texop = nir_texop_txd;
1701 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
1702 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
1703 }
1704
1705 if (operands & SpvImageOperandsOffsetMask ||
1706 operands & SpvImageOperandsConstOffsetMask)
1707 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
1708
1709 if (operands & SpvImageOperandsConstOffsetsMask) {
1710 gather_offsets = vtn_ssa_value(b, w[idx++]);
1711 (*p++) = (nir_tex_src){};
1712 }
1713
1714 if (operands & SpvImageOperandsSampleMask) {
1715 assert(texop == nir_texop_txf_ms);
1716 texop = nir_texop_txf_ms;
1717 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
1718 }
1719 }
1720 /* We should have now consumed exactly all of the arguments */
1721 assert(idx == count);
1722
1723 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
1724 instr->op = texop;
1725
1726 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
1727
1728 instr->coord_components = coord_components;
1729 instr->sampler_dim = sampler_dim;
1730 instr->is_array = is_array;
1731 instr->is_shadow = is_shadow;
1732 instr->is_new_style_shadow =
1733 is_shadow && glsl_get_components(ret_type->type) == 1;
1734 instr->component = gather_component;
1735
1736 switch (glsl_get_sampler_result_type(image_type)) {
1737 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
1738 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
1739 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
1740 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
1741 default:
1742 unreachable("Invalid base type for sampler result");
1743 }
1744
1745 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
1746 nir_deref_var *texture;
1747 if (sampled.image) {
1748 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
1749 texture = image;
1750 } else {
1751 texture = sampler;
1752 }
1753
1754 instr->texture = nir_deref_var_clone(texture, instr);
1755
1756 switch (instr->op) {
1757 case nir_texop_tex:
1758 case nir_texop_txb:
1759 case nir_texop_txl:
1760 case nir_texop_txd:
1761 case nir_texop_tg4:
1762 /* These operations require a sampler */
1763 instr->sampler = nir_deref_var_clone(sampler, instr);
1764 break;
1765 case nir_texop_txf:
1766 case nir_texop_txf_ms:
1767 case nir_texop_txs:
1768 case nir_texop_lod:
1769 case nir_texop_query_levels:
1770 case nir_texop_texture_samples:
1771 case nir_texop_samples_identical:
1772 /* These don't */
1773 instr->sampler = NULL;
1774 break;
1775 case nir_texop_txf_ms_mcs:
1776 unreachable("unexpected nir_texop_txf_ms_mcs");
1777 }
1778
1779 nir_ssa_dest_init(&instr->instr, &instr->dest,
1780 nir_tex_instr_dest_size(instr), 32, NULL);
1781
1782 assert(glsl_get_vector_elements(ret_type->type) ==
1783 nir_tex_instr_dest_size(instr));
1784
1785 nir_ssa_def *def;
1786 nir_instr *instruction;
1787 if (gather_offsets) {
1788 assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
1789 assert(glsl_get_length(gather_offsets->type) == 4);
1790 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
1791
1792 /* Copy the current instruction 4x */
1793 for (uint32_t i = 1; i < 4; i++) {
1794 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
1795 instrs[i]->op = instr->op;
1796 instrs[i]->coord_components = instr->coord_components;
1797 instrs[i]->sampler_dim = instr->sampler_dim;
1798 instrs[i]->is_array = instr->is_array;
1799 instrs[i]->is_shadow = instr->is_shadow;
1800 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
1801 instrs[i]->component = instr->component;
1802 instrs[i]->dest_type = instr->dest_type;
1803 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
1804 instrs[i]->sampler = NULL;
1805
1806 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
1807
1808 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
1809 nir_tex_instr_dest_size(instr), 32, NULL);
1810 }
1811
1812 /* Fill in the last argument with the offset from the passed in offsets
1813 * and insert the instruction into the stream.
1814 */
1815 for (uint32_t i = 0; i < 4; i++) {
1816 nir_tex_src src;
1817 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
1818 src.src_type = nir_tex_src_offset;
1819 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
1820 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
1821 }
1822
1823 /* Combine the results of the 4 instructions by taking their .w
1824 * components
1825 */
1826 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
1827 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
1828 vec4->dest.write_mask = 0xf;
1829 for (uint32_t i = 0; i < 4; i++) {
1830 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
1831 vec4->src[i].swizzle[0] = 3;
1832 }
1833 def = &vec4->dest.dest.ssa;
1834 instruction = &vec4->instr;
1835 } else {
1836 def = &instr->dest.ssa;
1837 instruction = &instr->instr;
1838 }
1839
1840 val->ssa = vtn_create_ssa_value(b, ret_type->type);
1841 val->ssa->def = def;
1842
1843 nir_builder_instr_insert(&b->nb, instruction);
1844 }
1845
1846 static void
1847 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
1848 const uint32_t *w, nir_src *src)
1849 {
1850 switch (opcode) {
1851 case SpvOpAtomicIIncrement:
1852 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
1853 break;
1854
1855 case SpvOpAtomicIDecrement:
1856 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
1857 break;
1858
1859 case SpvOpAtomicISub:
1860 src[0] =
1861 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
1862 break;
1863
1864 case SpvOpAtomicCompareExchange:
1865 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
1866 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
1867 break;
1868
1869 case SpvOpAtomicExchange:
1870 case SpvOpAtomicIAdd:
1871 case SpvOpAtomicSMin:
1872 case SpvOpAtomicUMin:
1873 case SpvOpAtomicSMax:
1874 case SpvOpAtomicUMax:
1875 case SpvOpAtomicAnd:
1876 case SpvOpAtomicOr:
1877 case SpvOpAtomicXor:
1878 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
1879 break;
1880
1881 default:
1882 unreachable("Invalid SPIR-V atomic");
1883 }
1884 }
1885
1886 static nir_ssa_def *
1887 get_image_coord(struct vtn_builder *b, uint32_t value)
1888 {
1889 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
1890
1891 /* The image_load_store intrinsics assume a 4-dim coordinate */
1892 unsigned dim = glsl_get_vector_elements(coord->type);
1893 unsigned swizzle[4];
1894 for (unsigned i = 0; i < 4; i++)
1895 swizzle[i] = MIN2(i, dim - 1);
1896
1897 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
1898 }
1899
1900 static void
1901 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
1902 const uint32_t *w, unsigned count)
1903 {
1904 /* Just get this one out of the way */
1905 if (opcode == SpvOpImageTexelPointer) {
1906 struct vtn_value *val =
1907 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
1908 val->image = ralloc(b, struct vtn_image_pointer);
1909
1910 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1911 val->image->coord = get_image_coord(b, w[4]);
1912 val->image->sample = vtn_ssa_value(b, w[5])->def;
1913 return;
1914 }
1915
1916 struct vtn_image_pointer image;
1917
1918 switch (opcode) {
1919 case SpvOpAtomicExchange:
1920 case SpvOpAtomicCompareExchange:
1921 case SpvOpAtomicCompareExchangeWeak:
1922 case SpvOpAtomicIIncrement:
1923 case SpvOpAtomicIDecrement:
1924 case SpvOpAtomicIAdd:
1925 case SpvOpAtomicISub:
1926 case SpvOpAtomicLoad:
1927 case SpvOpAtomicSMin:
1928 case SpvOpAtomicUMin:
1929 case SpvOpAtomicSMax:
1930 case SpvOpAtomicUMax:
1931 case SpvOpAtomicAnd:
1932 case SpvOpAtomicOr:
1933 case SpvOpAtomicXor:
1934 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
1935 break;
1936
1937 case SpvOpAtomicStore:
1938 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
1939 break;
1940
1941 case SpvOpImageQuerySize:
1942 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1943 image.coord = NULL;
1944 image.sample = NULL;
1945 break;
1946
1947 case SpvOpImageRead:
1948 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1949 image.coord = get_image_coord(b, w[4]);
1950
1951 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
1952 assert(w[5] == SpvImageOperandsSampleMask);
1953 image.sample = vtn_ssa_value(b, w[6])->def;
1954 } else {
1955 image.sample = nir_ssa_undef(&b->nb, 1, 32);
1956 }
1957 break;
1958
1959 case SpvOpImageWrite:
1960 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
1961 image.coord = get_image_coord(b, w[2]);
1962
1963 /* texel = w[3] */
1964
1965 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
1966 assert(w[4] == SpvImageOperandsSampleMask);
1967 image.sample = vtn_ssa_value(b, w[5])->def;
1968 } else {
1969 image.sample = nir_ssa_undef(&b->nb, 1, 32);
1970 }
1971 break;
1972
1973 default:
1974 unreachable("Invalid image opcode");
1975 }
1976
1977 nir_intrinsic_op op;
1978 switch (opcode) {
1979 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1980 OP(ImageQuerySize, size)
1981 OP(ImageRead, load)
1982 OP(ImageWrite, store)
1983 OP(AtomicLoad, load)
1984 OP(AtomicStore, store)
1985 OP(AtomicExchange, atomic_exchange)
1986 OP(AtomicCompareExchange, atomic_comp_swap)
1987 OP(AtomicIIncrement, atomic_add)
1988 OP(AtomicIDecrement, atomic_add)
1989 OP(AtomicIAdd, atomic_add)
1990 OP(AtomicISub, atomic_add)
1991 OP(AtomicSMin, atomic_min)
1992 OP(AtomicUMin, atomic_min)
1993 OP(AtomicSMax, atomic_max)
1994 OP(AtomicUMax, atomic_max)
1995 OP(AtomicAnd, atomic_and)
1996 OP(AtomicOr, atomic_or)
1997 OP(AtomicXor, atomic_xor)
1998 #undef OP
1999 default:
2000 unreachable("Invalid image opcode");
2001 }
2002
2003 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2004
2005 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2006 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2007
2008 /* ImageQuerySize doesn't take any extra parameters */
2009 if (opcode != SpvOpImageQuerySize) {
2010 /* The image coordinate is always 4 components but we may not have that
2011 * many. Swizzle to compensate.
2012 */
2013 unsigned swiz[4];
2014 for (unsigned i = 0; i < 4; i++)
2015 swiz[i] = i < image.coord->num_components ? i : 0;
2016 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2017 swiz, 4, false));
2018 intrin->src[1] = nir_src_for_ssa(image.sample);
2019 }
2020
2021 switch (opcode) {
2022 case SpvOpAtomicLoad:
2023 case SpvOpImageQuerySize:
2024 case SpvOpImageRead:
2025 break;
2026 case SpvOpAtomicStore:
2027 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2028 break;
2029 case SpvOpImageWrite:
2030 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2031 break;
2032
2033 case SpvOpAtomicCompareExchange:
2034 case SpvOpAtomicIIncrement:
2035 case SpvOpAtomicIDecrement:
2036 case SpvOpAtomicExchange:
2037 case SpvOpAtomicIAdd:
2038 case SpvOpAtomicISub:
2039 case SpvOpAtomicSMin:
2040 case SpvOpAtomicUMin:
2041 case SpvOpAtomicSMax:
2042 case SpvOpAtomicUMax:
2043 case SpvOpAtomicAnd:
2044 case SpvOpAtomicOr:
2045 case SpvOpAtomicXor:
2046 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2047 break;
2048
2049 default:
2050 unreachable("Invalid image opcode");
2051 }
2052
2053 if (opcode != SpvOpImageWrite) {
2054 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2055 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2056
2057 unsigned dest_components =
2058 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2059 if (intrin->intrinsic == nir_intrinsic_image_size) {
2060 dest_components = intrin->num_components =
2061 glsl_get_vector_elements(type->type);
2062 }
2063
2064 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2065 dest_components, 32, NULL);
2066
2067 nir_builder_instr_insert(&b->nb, &intrin->instr);
2068
2069 val->ssa = vtn_create_ssa_value(b, type->type);
2070 val->ssa->def = &intrin->dest.ssa;
2071 } else {
2072 nir_builder_instr_insert(&b->nb, &intrin->instr);
2073 }
2074 }
2075
2076 static nir_intrinsic_op
2077 get_ssbo_nir_atomic_op(SpvOp opcode)
2078 {
2079 switch (opcode) {
2080 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2081 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2082 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2083 OP(AtomicExchange, atomic_exchange)
2084 OP(AtomicCompareExchange, atomic_comp_swap)
2085 OP(AtomicIIncrement, atomic_add)
2086 OP(AtomicIDecrement, atomic_add)
2087 OP(AtomicIAdd, atomic_add)
2088 OP(AtomicISub, atomic_add)
2089 OP(AtomicSMin, atomic_imin)
2090 OP(AtomicUMin, atomic_umin)
2091 OP(AtomicSMax, atomic_imax)
2092 OP(AtomicUMax, atomic_umax)
2093 OP(AtomicAnd, atomic_and)
2094 OP(AtomicOr, atomic_or)
2095 OP(AtomicXor, atomic_xor)
2096 #undef OP
2097 default:
2098 unreachable("Invalid SSBO atomic");
2099 }
2100 }
2101
2102 static nir_intrinsic_op
2103 get_shared_nir_atomic_op(SpvOp opcode)
2104 {
2105 switch (opcode) {
2106 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2107 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2108 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2109 OP(AtomicExchange, atomic_exchange)
2110 OP(AtomicCompareExchange, atomic_comp_swap)
2111 OP(AtomicIIncrement, atomic_add)
2112 OP(AtomicIDecrement, atomic_add)
2113 OP(AtomicIAdd, atomic_add)
2114 OP(AtomicISub, atomic_add)
2115 OP(AtomicSMin, atomic_imin)
2116 OP(AtomicUMin, atomic_umin)
2117 OP(AtomicSMax, atomic_imax)
2118 OP(AtomicUMax, atomic_umax)
2119 OP(AtomicAnd, atomic_and)
2120 OP(AtomicOr, atomic_or)
2121 OP(AtomicXor, atomic_xor)
2122 #undef OP
2123 default:
2124 unreachable("Invalid shared atomic");
2125 }
2126 }
2127
2128 static void
2129 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2130 const uint32_t *w, unsigned count)
2131 {
2132 struct vtn_pointer *ptr;
2133 nir_intrinsic_instr *atomic;
2134
2135 switch (opcode) {
2136 case SpvOpAtomicLoad:
2137 case SpvOpAtomicExchange:
2138 case SpvOpAtomicCompareExchange:
2139 case SpvOpAtomicCompareExchangeWeak:
2140 case SpvOpAtomicIIncrement:
2141 case SpvOpAtomicIDecrement:
2142 case SpvOpAtomicIAdd:
2143 case SpvOpAtomicISub:
2144 case SpvOpAtomicSMin:
2145 case SpvOpAtomicUMin:
2146 case SpvOpAtomicSMax:
2147 case SpvOpAtomicUMax:
2148 case SpvOpAtomicAnd:
2149 case SpvOpAtomicOr:
2150 case SpvOpAtomicXor:
2151 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2152 break;
2153
2154 case SpvOpAtomicStore:
2155 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2156 break;
2157
2158 default:
2159 unreachable("Invalid SPIR-V atomic");
2160 }
2161
2162 /*
2163 SpvScope scope = w[4];
2164 SpvMemorySemanticsMask semantics = w[5];
2165 */
2166
2167 if (ptr->mode == vtn_variable_mode_workgroup) {
2168 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2169 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2170 nir_intrinsic_op op = get_shared_nir_atomic_op(opcode);
2171 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2172 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2173
2174 switch (opcode) {
2175 case SpvOpAtomicLoad:
2176 atomic->num_components = glsl_get_vector_elements(deref_type);
2177 break;
2178
2179 case SpvOpAtomicStore:
2180 atomic->num_components = glsl_get_vector_elements(deref_type);
2181 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2182 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2183 break;
2184
2185 case SpvOpAtomicExchange:
2186 case SpvOpAtomicCompareExchange:
2187 case SpvOpAtomicCompareExchangeWeak:
2188 case SpvOpAtomicIIncrement:
2189 case SpvOpAtomicIDecrement:
2190 case SpvOpAtomicIAdd:
2191 case SpvOpAtomicISub:
2192 case SpvOpAtomicSMin:
2193 case SpvOpAtomicUMin:
2194 case SpvOpAtomicSMax:
2195 case SpvOpAtomicUMax:
2196 case SpvOpAtomicAnd:
2197 case SpvOpAtomicOr:
2198 case SpvOpAtomicXor:
2199 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2200 break;
2201
2202 default:
2203 unreachable("Invalid SPIR-V atomic");
2204
2205 }
2206 } else {
2207 assert(ptr->mode == vtn_variable_mode_ssbo);
2208 nir_ssa_def *offset, *index;
2209 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2210
2211 nir_intrinsic_op op = get_ssbo_nir_atomic_op(opcode);
2212
2213 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2214
2215 switch (opcode) {
2216 case SpvOpAtomicLoad:
2217 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2218 atomic->src[0] = nir_src_for_ssa(index);
2219 atomic->src[1] = nir_src_for_ssa(offset);
2220 break;
2221
2222 case SpvOpAtomicStore:
2223 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2224 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2225 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2226 atomic->src[1] = nir_src_for_ssa(index);
2227 atomic->src[2] = nir_src_for_ssa(offset);
2228 break;
2229
2230 case SpvOpAtomicExchange:
2231 case SpvOpAtomicCompareExchange:
2232 case SpvOpAtomicCompareExchangeWeak:
2233 case SpvOpAtomicIIncrement:
2234 case SpvOpAtomicIDecrement:
2235 case SpvOpAtomicIAdd:
2236 case SpvOpAtomicISub:
2237 case SpvOpAtomicSMin:
2238 case SpvOpAtomicUMin:
2239 case SpvOpAtomicSMax:
2240 case SpvOpAtomicUMax:
2241 case SpvOpAtomicAnd:
2242 case SpvOpAtomicOr:
2243 case SpvOpAtomicXor:
2244 atomic->src[0] = nir_src_for_ssa(index);
2245 atomic->src[1] = nir_src_for_ssa(offset);
2246 fill_common_atomic_sources(b, opcode, w, &atomic->src[2]);
2247 break;
2248
2249 default:
2250 unreachable("Invalid SPIR-V atomic");
2251 }
2252 }
2253
2254 if (opcode != SpvOpAtomicStore) {
2255 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2256
2257 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2258 glsl_get_vector_elements(type->type),
2259 glsl_get_bit_size(type->type), NULL);
2260
2261 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2262 val->ssa = rzalloc(b, struct vtn_ssa_value);
2263 val->ssa->def = &atomic->dest.ssa;
2264 val->ssa->type = type->type;
2265 }
2266
2267 nir_builder_instr_insert(&b->nb, &atomic->instr);
2268 }
2269
2270 static nir_alu_instr *
2271 create_vec(nir_shader *shader, unsigned num_components, unsigned bit_size)
2272 {
2273 nir_op op;
2274 switch (num_components) {
2275 case 1: op = nir_op_fmov; break;
2276 case 2: op = nir_op_vec2; break;
2277 case 3: op = nir_op_vec3; break;
2278 case 4: op = nir_op_vec4; break;
2279 default: unreachable("bad vector size");
2280 }
2281
2282 nir_alu_instr *vec = nir_alu_instr_create(shader, op);
2283 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2284 bit_size, NULL);
2285 vec->dest.write_mask = (1 << num_components) - 1;
2286
2287 return vec;
2288 }
2289
2290 struct vtn_ssa_value *
2291 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2292 {
2293 if (src->transposed)
2294 return src->transposed;
2295
2296 struct vtn_ssa_value *dest =
2297 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2298
2299 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2300 nir_alu_instr *vec = create_vec(b->shader,
2301 glsl_get_matrix_columns(src->type),
2302 glsl_get_bit_size(src->type));
2303 if (glsl_type_is_vector_or_scalar(src->type)) {
2304 vec->src[0].src = nir_src_for_ssa(src->def);
2305 vec->src[0].swizzle[0] = i;
2306 } else {
2307 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2308 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2309 vec->src[j].swizzle[0] = i;
2310 }
2311 }
2312 nir_builder_instr_insert(&b->nb, &vec->instr);
2313 dest->elems[i]->def = &vec->dest.dest.ssa;
2314 }
2315
2316 dest->transposed = src;
2317
2318 return dest;
2319 }
2320
2321 nir_ssa_def *
2322 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2323 {
2324 unsigned swiz[4] = { index };
2325 return nir_swizzle(&b->nb, src, swiz, 1, true);
2326 }
2327
2328 nir_ssa_def *
2329 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2330 unsigned index)
2331 {
2332 nir_alu_instr *vec = create_vec(b->shader, src->num_components,
2333 src->bit_size);
2334
2335 for (unsigned i = 0; i < src->num_components; i++) {
2336 if (i == index) {
2337 vec->src[i].src = nir_src_for_ssa(insert);
2338 } else {
2339 vec->src[i].src = nir_src_for_ssa(src);
2340 vec->src[i].swizzle[0] = i;
2341 }
2342 }
2343
2344 nir_builder_instr_insert(&b->nb, &vec->instr);
2345
2346 return &vec->dest.dest.ssa;
2347 }
2348
2349 nir_ssa_def *
2350 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2351 nir_ssa_def *index)
2352 {
2353 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2354 for (unsigned i = 1; i < src->num_components; i++)
2355 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2356 vtn_vector_extract(b, src, i), dest);
2357
2358 return dest;
2359 }
2360
2361 nir_ssa_def *
2362 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2363 nir_ssa_def *insert, nir_ssa_def *index)
2364 {
2365 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2366 for (unsigned i = 1; i < src->num_components; i++)
2367 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2368 vtn_vector_insert(b, src, insert, i), dest);
2369
2370 return dest;
2371 }
2372
2373 static nir_ssa_def *
2374 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2375 nir_ssa_def *src0, nir_ssa_def *src1,
2376 const uint32_t *indices)
2377 {
2378 nir_alu_instr *vec = create_vec(b->shader, num_components, src0->bit_size);
2379
2380 for (unsigned i = 0; i < num_components; i++) {
2381 uint32_t index = indices[i];
2382 if (index == 0xffffffff) {
2383 vec->src[i].src =
2384 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2385 } else if (index < src0->num_components) {
2386 vec->src[i].src = nir_src_for_ssa(src0);
2387 vec->src[i].swizzle[0] = index;
2388 } else {
2389 vec->src[i].src = nir_src_for_ssa(src1);
2390 vec->src[i].swizzle[0] = index - src0->num_components;
2391 }
2392 }
2393
2394 nir_builder_instr_insert(&b->nb, &vec->instr);
2395
2396 return &vec->dest.dest.ssa;
2397 }
2398
2399 /*
2400 * Concatentates a number of vectors/scalars together to produce a vector
2401 */
2402 static nir_ssa_def *
2403 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2404 unsigned num_srcs, nir_ssa_def **srcs)
2405 {
2406 nir_alu_instr *vec = create_vec(b->shader, num_components,
2407 srcs[0]->bit_size);
2408
2409 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2410 *
2411 * "When constructing a vector, there must be at least two Constituent
2412 * operands."
2413 */
2414 assert(num_srcs >= 2);
2415
2416 unsigned dest_idx = 0;
2417 for (unsigned i = 0; i < num_srcs; i++) {
2418 nir_ssa_def *src = srcs[i];
2419 assert(dest_idx + src->num_components <= num_components);
2420 for (unsigned j = 0; j < src->num_components; j++) {
2421 vec->src[dest_idx].src = nir_src_for_ssa(src);
2422 vec->src[dest_idx].swizzle[0] = j;
2423 dest_idx++;
2424 }
2425 }
2426
2427 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2428 *
2429 * "When constructing a vector, the total number of components in all
2430 * the operands must equal the number of components in Result Type."
2431 */
2432 assert(dest_idx == num_components);
2433
2434 nir_builder_instr_insert(&b->nb, &vec->instr);
2435
2436 return &vec->dest.dest.ssa;
2437 }
2438
2439 static struct vtn_ssa_value *
2440 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2441 {
2442 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2443 dest->type = src->type;
2444
2445 if (glsl_type_is_vector_or_scalar(src->type)) {
2446 dest->def = src->def;
2447 } else {
2448 unsigned elems = glsl_get_length(src->type);
2449
2450 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2451 for (unsigned i = 0; i < elems; i++)
2452 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2453 }
2454
2455 return dest;
2456 }
2457
2458 static struct vtn_ssa_value *
2459 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2460 struct vtn_ssa_value *insert, const uint32_t *indices,
2461 unsigned num_indices)
2462 {
2463 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2464
2465 struct vtn_ssa_value *cur = dest;
2466 unsigned i;
2467 for (i = 0; i < num_indices - 1; i++) {
2468 cur = cur->elems[indices[i]];
2469 }
2470
2471 if (glsl_type_is_vector_or_scalar(cur->type)) {
2472 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2473 * the component granularity. In that case, the last index will be
2474 * the index to insert the scalar into the vector.
2475 */
2476
2477 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2478 } else {
2479 cur->elems[indices[i]] = insert;
2480 }
2481
2482 return dest;
2483 }
2484
2485 static struct vtn_ssa_value *
2486 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2487 const uint32_t *indices, unsigned num_indices)
2488 {
2489 struct vtn_ssa_value *cur = src;
2490 for (unsigned i = 0; i < num_indices; i++) {
2491 if (glsl_type_is_vector_or_scalar(cur->type)) {
2492 assert(i == num_indices - 1);
2493 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2494 * the component granularity. The last index will be the index of the
2495 * vector to extract.
2496 */
2497
2498 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2499 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2500 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2501 return ret;
2502 } else {
2503 cur = cur->elems[indices[i]];
2504 }
2505 }
2506
2507 return cur;
2508 }
2509
2510 static void
2511 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2512 const uint32_t *w, unsigned count)
2513 {
2514 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2515 const struct glsl_type *type =
2516 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2517 val->ssa = vtn_create_ssa_value(b, type);
2518
2519 switch (opcode) {
2520 case SpvOpVectorExtractDynamic:
2521 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2522 vtn_ssa_value(b, w[4])->def);
2523 break;
2524
2525 case SpvOpVectorInsertDynamic:
2526 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2527 vtn_ssa_value(b, w[4])->def,
2528 vtn_ssa_value(b, w[5])->def);
2529 break;
2530
2531 case SpvOpVectorShuffle:
2532 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2533 vtn_ssa_value(b, w[3])->def,
2534 vtn_ssa_value(b, w[4])->def,
2535 w + 5);
2536 break;
2537
2538 case SpvOpCompositeConstruct: {
2539 unsigned elems = count - 3;
2540 if (glsl_type_is_vector_or_scalar(type)) {
2541 nir_ssa_def *srcs[4];
2542 for (unsigned i = 0; i < elems; i++)
2543 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2544 val->ssa->def =
2545 vtn_vector_construct(b, glsl_get_vector_elements(type),
2546 elems, srcs);
2547 } else {
2548 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2549 for (unsigned i = 0; i < elems; i++)
2550 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2551 }
2552 break;
2553 }
2554 case SpvOpCompositeExtract:
2555 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2556 w + 4, count - 4);
2557 break;
2558
2559 case SpvOpCompositeInsert:
2560 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2561 vtn_ssa_value(b, w[3]),
2562 w + 5, count - 5);
2563 break;
2564
2565 case SpvOpCopyObject:
2566 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2567 break;
2568
2569 default:
2570 unreachable("unknown composite operation");
2571 }
2572 }
2573
2574 static void
2575 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
2576 const uint32_t *w, unsigned count)
2577 {
2578 nir_intrinsic_op intrinsic_op;
2579 switch (opcode) {
2580 case SpvOpEmitVertex:
2581 case SpvOpEmitStreamVertex:
2582 intrinsic_op = nir_intrinsic_emit_vertex;
2583 break;
2584 case SpvOpEndPrimitive:
2585 case SpvOpEndStreamPrimitive:
2586 intrinsic_op = nir_intrinsic_end_primitive;
2587 break;
2588 case SpvOpMemoryBarrier:
2589 intrinsic_op = nir_intrinsic_memory_barrier;
2590 break;
2591 case SpvOpControlBarrier:
2592 intrinsic_op = nir_intrinsic_barrier;
2593 break;
2594 default:
2595 unreachable("unknown barrier instruction");
2596 }
2597
2598 nir_intrinsic_instr *intrin =
2599 nir_intrinsic_instr_create(b->shader, intrinsic_op);
2600
2601 if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
2602 nir_intrinsic_set_stream_id(intrin, w[1]);
2603
2604 nir_builder_instr_insert(&b->nb, &intrin->instr);
2605 }
2606
2607 static unsigned
2608 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode)
2609 {
2610 switch (mode) {
2611 case SpvExecutionModeInputPoints:
2612 case SpvExecutionModeOutputPoints:
2613 return 0; /* GL_POINTS */
2614 case SpvExecutionModeInputLines:
2615 return 1; /* GL_LINES */
2616 case SpvExecutionModeInputLinesAdjacency:
2617 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2618 case SpvExecutionModeTriangles:
2619 return 4; /* GL_TRIANGLES */
2620 case SpvExecutionModeInputTrianglesAdjacency:
2621 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2622 case SpvExecutionModeQuads:
2623 return 7; /* GL_QUADS */
2624 case SpvExecutionModeIsolines:
2625 return 0x8E7A; /* GL_ISOLINES */
2626 case SpvExecutionModeOutputLineStrip:
2627 return 3; /* GL_LINE_STRIP */
2628 case SpvExecutionModeOutputTriangleStrip:
2629 return 5; /* GL_TRIANGLE_STRIP */
2630 default:
2631 unreachable("Invalid primitive type");
2632 return 4;
2633 }
2634 }
2635
2636 static unsigned
2637 vertices_in_from_spv_execution_mode(SpvExecutionMode mode)
2638 {
2639 switch (mode) {
2640 case SpvExecutionModeInputPoints:
2641 return 1;
2642 case SpvExecutionModeInputLines:
2643 return 2;
2644 case SpvExecutionModeInputLinesAdjacency:
2645 return 4;
2646 case SpvExecutionModeTriangles:
2647 return 3;
2648 case SpvExecutionModeInputTrianglesAdjacency:
2649 return 6;
2650 default:
2651 unreachable("Invalid GS input mode");
2652 return 0;
2653 }
2654 }
2655
2656 static gl_shader_stage
2657 stage_for_execution_model(SpvExecutionModel model)
2658 {
2659 switch (model) {
2660 case SpvExecutionModelVertex:
2661 return MESA_SHADER_VERTEX;
2662 case SpvExecutionModelTessellationControl:
2663 return MESA_SHADER_TESS_CTRL;
2664 case SpvExecutionModelTessellationEvaluation:
2665 return MESA_SHADER_TESS_EVAL;
2666 case SpvExecutionModelGeometry:
2667 return MESA_SHADER_GEOMETRY;
2668 case SpvExecutionModelFragment:
2669 return MESA_SHADER_FRAGMENT;
2670 case SpvExecutionModelGLCompute:
2671 return MESA_SHADER_COMPUTE;
2672 default:
2673 unreachable("Unsupported execution model");
2674 }
2675 }
2676
2677 #define spv_check_supported(name, cap) do { \
2678 if (!(b->options && b->options->caps.name)) \
2679 vtn_warn("Unsupported SPIR-V capability: %s", \
2680 spirv_capability_to_string(cap)); \
2681 } while(0)
2682
2683 static bool
2684 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
2685 const uint32_t *w, unsigned count)
2686 {
2687 switch (opcode) {
2688 case SpvOpSource:
2689 case SpvOpSourceExtension:
2690 case SpvOpSourceContinued:
2691 case SpvOpExtension:
2692 /* Unhandled, but these are for debug so that's ok. */
2693 break;
2694
2695 case SpvOpCapability: {
2696 SpvCapability cap = w[1];
2697 switch (cap) {
2698 case SpvCapabilityMatrix:
2699 case SpvCapabilityShader:
2700 case SpvCapabilityGeometry:
2701 case SpvCapabilityGeometryPointSize:
2702 case SpvCapabilityUniformBufferArrayDynamicIndexing:
2703 case SpvCapabilitySampledImageArrayDynamicIndexing:
2704 case SpvCapabilityStorageBufferArrayDynamicIndexing:
2705 case SpvCapabilityStorageImageArrayDynamicIndexing:
2706 case SpvCapabilityImageRect:
2707 case SpvCapabilitySampledRect:
2708 case SpvCapabilitySampled1D:
2709 case SpvCapabilityImage1D:
2710 case SpvCapabilitySampledCubeArray:
2711 case SpvCapabilityImageCubeArray:
2712 case SpvCapabilitySampledBuffer:
2713 case SpvCapabilityImageBuffer:
2714 case SpvCapabilityImageQuery:
2715 case SpvCapabilityDerivativeControl:
2716 case SpvCapabilityInterpolationFunction:
2717 case SpvCapabilityMultiViewport:
2718 case SpvCapabilitySampleRateShading:
2719 case SpvCapabilityClipDistance:
2720 case SpvCapabilityCullDistance:
2721 case SpvCapabilityInputAttachment:
2722 case SpvCapabilityImageGatherExtended:
2723 case SpvCapabilityStorageImageExtendedFormats:
2724 break;
2725
2726 case SpvCapabilityGeometryStreams:
2727 case SpvCapabilityLinkage:
2728 case SpvCapabilityVector16:
2729 case SpvCapabilityFloat16Buffer:
2730 case SpvCapabilityFloat16:
2731 case SpvCapabilityInt64Atomics:
2732 case SpvCapabilityAtomicStorage:
2733 case SpvCapabilityInt16:
2734 case SpvCapabilityStorageImageMultisample:
2735 case SpvCapabilityInt8:
2736 case SpvCapabilitySparseResidency:
2737 case SpvCapabilityMinLod:
2738 case SpvCapabilityTransformFeedback:
2739 vtn_warn("Unsupported SPIR-V capability: %s",
2740 spirv_capability_to_string(cap));
2741 break;
2742
2743 case SpvCapabilityFloat64:
2744 spv_check_supported(float64, cap);
2745 break;
2746 case SpvCapabilityInt64:
2747 spv_check_supported(int64, cap);
2748 break;
2749
2750 case SpvCapabilityAddresses:
2751 case SpvCapabilityKernel:
2752 case SpvCapabilityImageBasic:
2753 case SpvCapabilityImageReadWrite:
2754 case SpvCapabilityImageMipmap:
2755 case SpvCapabilityPipes:
2756 case SpvCapabilityGroups:
2757 case SpvCapabilityDeviceEnqueue:
2758 case SpvCapabilityLiteralSampler:
2759 case SpvCapabilityGenericPointer:
2760 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2761 spirv_capability_to_string(cap));
2762 break;
2763
2764 case SpvCapabilityImageMSArray:
2765 spv_check_supported(image_ms_array, cap);
2766 break;
2767
2768 case SpvCapabilityTessellation:
2769 case SpvCapabilityTessellationPointSize:
2770 spv_check_supported(tessellation, cap);
2771 break;
2772
2773 case SpvCapabilityDrawParameters:
2774 spv_check_supported(draw_parameters, cap);
2775 break;
2776
2777 case SpvCapabilityStorageImageReadWithoutFormat:
2778 spv_check_supported(image_read_without_format, cap);
2779 break;
2780
2781 case SpvCapabilityStorageImageWriteWithoutFormat:
2782 spv_check_supported(image_write_without_format, cap);
2783 break;
2784
2785 case SpvCapabilityMultiView:
2786 spv_check_supported(multiview, cap);
2787 break;
2788
2789 case SpvCapabilityVariablePointersStorageBuffer:
2790 case SpvCapabilityVariablePointers:
2791 spv_check_supported(variable_pointers, cap);
2792 break;
2793
2794 default:
2795 unreachable("Unhandled capability");
2796 }
2797 break;
2798 }
2799
2800 case SpvOpExtInstImport:
2801 vtn_handle_extension(b, opcode, w, count);
2802 break;
2803
2804 case SpvOpMemoryModel:
2805 assert(w[1] == SpvAddressingModelLogical);
2806 assert(w[2] == SpvMemoryModelSimple ||
2807 w[2] == SpvMemoryModelGLSL450);
2808 break;
2809
2810 case SpvOpEntryPoint: {
2811 struct vtn_value *entry_point = &b->values[w[2]];
2812 /* Let this be a name label regardless */
2813 unsigned name_words;
2814 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
2815
2816 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
2817 stage_for_execution_model(w[1]) != b->entry_point_stage)
2818 break;
2819
2820 assert(b->entry_point == NULL);
2821 b->entry_point = entry_point;
2822 break;
2823 }
2824
2825 case SpvOpString:
2826 vtn_push_value(b, w[1], vtn_value_type_string)->str =
2827 vtn_string_literal(b, &w[2], count - 2, NULL);
2828 break;
2829
2830 case SpvOpName:
2831 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
2832 break;
2833
2834 case SpvOpMemberName:
2835 /* TODO */
2836 break;
2837
2838 case SpvOpExecutionMode:
2839 case SpvOpDecorationGroup:
2840 case SpvOpDecorate:
2841 case SpvOpMemberDecorate:
2842 case SpvOpGroupDecorate:
2843 case SpvOpGroupMemberDecorate:
2844 vtn_handle_decoration(b, opcode, w, count);
2845 break;
2846
2847 default:
2848 return false; /* End of preamble */
2849 }
2850
2851 return true;
2852 }
2853
2854 static void
2855 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
2856 const struct vtn_decoration *mode, void *data)
2857 {
2858 assert(b->entry_point == entry_point);
2859
2860 switch(mode->exec_mode) {
2861 case SpvExecutionModeOriginUpperLeft:
2862 case SpvExecutionModeOriginLowerLeft:
2863 b->origin_upper_left =
2864 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
2865 break;
2866
2867 case SpvExecutionModeEarlyFragmentTests:
2868 assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2869 b->shader->info.fs.early_fragment_tests = true;
2870 break;
2871
2872 case SpvExecutionModeInvocations:
2873 assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
2874 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
2875 break;
2876
2877 case SpvExecutionModeDepthReplacing:
2878 assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2879 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
2880 break;
2881 case SpvExecutionModeDepthGreater:
2882 assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2883 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
2884 break;
2885 case SpvExecutionModeDepthLess:
2886 assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2887 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
2888 break;
2889 case SpvExecutionModeDepthUnchanged:
2890 assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2891 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2892 break;
2893
2894 case SpvExecutionModeLocalSize:
2895 assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
2896 b->shader->info.cs.local_size[0] = mode->literals[0];
2897 b->shader->info.cs.local_size[1] = mode->literals[1];
2898 b->shader->info.cs.local_size[2] = mode->literals[2];
2899 break;
2900 case SpvExecutionModeLocalSizeHint:
2901 break; /* Nothing to do with this */
2902
2903 case SpvExecutionModeOutputVertices:
2904 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2905 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
2906 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
2907 } else {
2908 assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
2909 b->shader->info.gs.vertices_out = mode->literals[0];
2910 }
2911 break;
2912
2913 case SpvExecutionModeInputPoints:
2914 case SpvExecutionModeInputLines:
2915 case SpvExecutionModeInputLinesAdjacency:
2916 case SpvExecutionModeTriangles:
2917 case SpvExecutionModeInputTrianglesAdjacency:
2918 case SpvExecutionModeQuads:
2919 case SpvExecutionModeIsolines:
2920 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2921 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
2922 b->shader->info.tess.primitive_mode =
2923 gl_primitive_from_spv_execution_mode(mode->exec_mode);
2924 } else {
2925 assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
2926 b->shader->info.gs.vertices_in =
2927 vertices_in_from_spv_execution_mode(mode->exec_mode);
2928 }
2929 break;
2930
2931 case SpvExecutionModeOutputPoints:
2932 case SpvExecutionModeOutputLineStrip:
2933 case SpvExecutionModeOutputTriangleStrip:
2934 assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
2935 b->shader->info.gs.output_primitive =
2936 gl_primitive_from_spv_execution_mode(mode->exec_mode);
2937 break;
2938
2939 case SpvExecutionModeSpacingEqual:
2940 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2941 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2942 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
2943 break;
2944 case SpvExecutionModeSpacingFractionalEven:
2945 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2946 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2947 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
2948 break;
2949 case SpvExecutionModeSpacingFractionalOdd:
2950 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2951 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2952 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
2953 break;
2954 case SpvExecutionModeVertexOrderCw:
2955 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2956 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2957 b->shader->info.tess.ccw = false;
2958 break;
2959 case SpvExecutionModeVertexOrderCcw:
2960 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2961 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2962 b->shader->info.tess.ccw = true;
2963 break;
2964 case SpvExecutionModePointMode:
2965 assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
2966 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
2967 b->shader->info.tess.point_mode = true;
2968 break;
2969
2970 case SpvExecutionModePixelCenterInteger:
2971 b->pixel_center_integer = true;
2972 break;
2973
2974 case SpvExecutionModeXfb:
2975 unreachable("Unhandled execution mode");
2976 break;
2977
2978 case SpvExecutionModeVecTypeHint:
2979 case SpvExecutionModeContractionOff:
2980 break; /* OpenCL */
2981
2982 default:
2983 unreachable("Unhandled execution mode");
2984 }
2985 }
2986
2987 static bool
2988 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
2989 const uint32_t *w, unsigned count)
2990 {
2991 switch (opcode) {
2992 case SpvOpSource:
2993 case SpvOpSourceContinued:
2994 case SpvOpSourceExtension:
2995 case SpvOpExtension:
2996 case SpvOpCapability:
2997 case SpvOpExtInstImport:
2998 case SpvOpMemoryModel:
2999 case SpvOpEntryPoint:
3000 case SpvOpExecutionMode:
3001 case SpvOpString:
3002 case SpvOpName:
3003 case SpvOpMemberName:
3004 case SpvOpDecorationGroup:
3005 case SpvOpDecorate:
3006 case SpvOpMemberDecorate:
3007 case SpvOpGroupDecorate:
3008 case SpvOpGroupMemberDecorate:
3009 unreachable("Invalid opcode types and variables section");
3010 break;
3011
3012 case SpvOpTypeVoid:
3013 case SpvOpTypeBool:
3014 case SpvOpTypeInt:
3015 case SpvOpTypeFloat:
3016 case SpvOpTypeVector:
3017 case SpvOpTypeMatrix:
3018 case SpvOpTypeImage:
3019 case SpvOpTypeSampler:
3020 case SpvOpTypeSampledImage:
3021 case SpvOpTypeArray:
3022 case SpvOpTypeRuntimeArray:
3023 case SpvOpTypeStruct:
3024 case SpvOpTypeOpaque:
3025 case SpvOpTypePointer:
3026 case SpvOpTypeFunction:
3027 case SpvOpTypeEvent:
3028 case SpvOpTypeDeviceEvent:
3029 case SpvOpTypeReserveId:
3030 case SpvOpTypeQueue:
3031 case SpvOpTypePipe:
3032 vtn_handle_type(b, opcode, w, count);
3033 break;
3034
3035 case SpvOpConstantTrue:
3036 case SpvOpConstantFalse:
3037 case SpvOpConstant:
3038 case SpvOpConstantComposite:
3039 case SpvOpConstantSampler:
3040 case SpvOpConstantNull:
3041 case SpvOpSpecConstantTrue:
3042 case SpvOpSpecConstantFalse:
3043 case SpvOpSpecConstant:
3044 case SpvOpSpecConstantComposite:
3045 case SpvOpSpecConstantOp:
3046 vtn_handle_constant(b, opcode, w, count);
3047 break;
3048
3049 case SpvOpUndef:
3050 case SpvOpVariable:
3051 vtn_handle_variables(b, opcode, w, count);
3052 break;
3053
3054 default:
3055 return false; /* End of preamble */
3056 }
3057
3058 return true;
3059 }
3060
3061 static bool
3062 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3063 const uint32_t *w, unsigned count)
3064 {
3065 switch (opcode) {
3066 case SpvOpLabel:
3067 break;
3068
3069 case SpvOpLoopMerge:
3070 case SpvOpSelectionMerge:
3071 /* This is handled by cfg pre-pass and walk_blocks */
3072 break;
3073
3074 case SpvOpUndef: {
3075 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3076 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3077 break;
3078 }
3079
3080 case SpvOpExtInst:
3081 vtn_handle_extension(b, opcode, w, count);
3082 break;
3083
3084 case SpvOpVariable:
3085 case SpvOpLoad:
3086 case SpvOpStore:
3087 case SpvOpCopyMemory:
3088 case SpvOpCopyMemorySized:
3089 case SpvOpAccessChain:
3090 case SpvOpPtrAccessChain:
3091 case SpvOpInBoundsAccessChain:
3092 case SpvOpArrayLength:
3093 vtn_handle_variables(b, opcode, w, count);
3094 break;
3095
3096 case SpvOpFunctionCall:
3097 vtn_handle_function_call(b, opcode, w, count);
3098 break;
3099
3100 case SpvOpSampledImage:
3101 case SpvOpImage:
3102 case SpvOpImageSampleImplicitLod:
3103 case SpvOpImageSampleExplicitLod:
3104 case SpvOpImageSampleDrefImplicitLod:
3105 case SpvOpImageSampleDrefExplicitLod:
3106 case SpvOpImageSampleProjImplicitLod:
3107 case SpvOpImageSampleProjExplicitLod:
3108 case SpvOpImageSampleProjDrefImplicitLod:
3109 case SpvOpImageSampleProjDrefExplicitLod:
3110 case SpvOpImageFetch:
3111 case SpvOpImageGather:
3112 case SpvOpImageDrefGather:
3113 case SpvOpImageQuerySizeLod:
3114 case SpvOpImageQueryLod:
3115 case SpvOpImageQueryLevels:
3116 case SpvOpImageQuerySamples:
3117 vtn_handle_texture(b, opcode, w, count);
3118 break;
3119
3120 case SpvOpImageRead:
3121 case SpvOpImageWrite:
3122 case SpvOpImageTexelPointer:
3123 vtn_handle_image(b, opcode, w, count);
3124 break;
3125
3126 case SpvOpImageQuerySize: {
3127 struct vtn_pointer *image =
3128 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3129 if (image->mode == vtn_variable_mode_image) {
3130 vtn_handle_image(b, opcode, w, count);
3131 } else {
3132 assert(image->mode == vtn_variable_mode_sampler);
3133 vtn_handle_texture(b, opcode, w, count);
3134 }
3135 break;
3136 }
3137
3138 case SpvOpAtomicLoad:
3139 case SpvOpAtomicExchange:
3140 case SpvOpAtomicCompareExchange:
3141 case SpvOpAtomicCompareExchangeWeak:
3142 case SpvOpAtomicIIncrement:
3143 case SpvOpAtomicIDecrement:
3144 case SpvOpAtomicIAdd:
3145 case SpvOpAtomicISub:
3146 case SpvOpAtomicSMin:
3147 case SpvOpAtomicUMin:
3148 case SpvOpAtomicSMax:
3149 case SpvOpAtomicUMax:
3150 case SpvOpAtomicAnd:
3151 case SpvOpAtomicOr:
3152 case SpvOpAtomicXor: {
3153 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3154 if (pointer->value_type == vtn_value_type_image_pointer) {
3155 vtn_handle_image(b, opcode, w, count);
3156 } else {
3157 assert(pointer->value_type == vtn_value_type_pointer);
3158 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3159 }
3160 break;
3161 }
3162
3163 case SpvOpAtomicStore: {
3164 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3165 if (pointer->value_type == vtn_value_type_image_pointer) {
3166 vtn_handle_image(b, opcode, w, count);
3167 } else {
3168 assert(pointer->value_type == vtn_value_type_pointer);
3169 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3170 }
3171 break;
3172 }
3173
3174 case SpvOpSelect: {
3175 /* Handle OpSelect up-front here because it needs to be able to handle
3176 * pointers and not just regular vectors and scalars.
3177 */
3178 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3179 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3180 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3181 vtn_ssa_value(b, w[4])->def,
3182 vtn_ssa_value(b, w[5])->def);
3183 vtn_push_ssa(b, w[2], res_type, ssa);
3184 break;
3185 }
3186
3187 case SpvOpSNegate:
3188 case SpvOpFNegate:
3189 case SpvOpNot:
3190 case SpvOpAny:
3191 case SpvOpAll:
3192 case SpvOpConvertFToU:
3193 case SpvOpConvertFToS:
3194 case SpvOpConvertSToF:
3195 case SpvOpConvertUToF:
3196 case SpvOpUConvert:
3197 case SpvOpSConvert:
3198 case SpvOpFConvert:
3199 case SpvOpQuantizeToF16:
3200 case SpvOpConvertPtrToU:
3201 case SpvOpConvertUToPtr:
3202 case SpvOpPtrCastToGeneric:
3203 case SpvOpGenericCastToPtr:
3204 case SpvOpBitcast:
3205 case SpvOpIsNan:
3206 case SpvOpIsInf:
3207 case SpvOpIsFinite:
3208 case SpvOpIsNormal:
3209 case SpvOpSignBitSet:
3210 case SpvOpLessOrGreater:
3211 case SpvOpOrdered:
3212 case SpvOpUnordered:
3213 case SpvOpIAdd:
3214 case SpvOpFAdd:
3215 case SpvOpISub:
3216 case SpvOpFSub:
3217 case SpvOpIMul:
3218 case SpvOpFMul:
3219 case SpvOpUDiv:
3220 case SpvOpSDiv:
3221 case SpvOpFDiv:
3222 case SpvOpUMod:
3223 case SpvOpSRem:
3224 case SpvOpSMod:
3225 case SpvOpFRem:
3226 case SpvOpFMod:
3227 case SpvOpVectorTimesScalar:
3228 case SpvOpDot:
3229 case SpvOpIAddCarry:
3230 case SpvOpISubBorrow:
3231 case SpvOpUMulExtended:
3232 case SpvOpSMulExtended:
3233 case SpvOpShiftRightLogical:
3234 case SpvOpShiftRightArithmetic:
3235 case SpvOpShiftLeftLogical:
3236 case SpvOpLogicalEqual:
3237 case SpvOpLogicalNotEqual:
3238 case SpvOpLogicalOr:
3239 case SpvOpLogicalAnd:
3240 case SpvOpLogicalNot:
3241 case SpvOpBitwiseOr:
3242 case SpvOpBitwiseXor:
3243 case SpvOpBitwiseAnd:
3244 case SpvOpIEqual:
3245 case SpvOpFOrdEqual:
3246 case SpvOpFUnordEqual:
3247 case SpvOpINotEqual:
3248 case SpvOpFOrdNotEqual:
3249 case SpvOpFUnordNotEqual:
3250 case SpvOpULessThan:
3251 case SpvOpSLessThan:
3252 case SpvOpFOrdLessThan:
3253 case SpvOpFUnordLessThan:
3254 case SpvOpUGreaterThan:
3255 case SpvOpSGreaterThan:
3256 case SpvOpFOrdGreaterThan:
3257 case SpvOpFUnordGreaterThan:
3258 case SpvOpULessThanEqual:
3259 case SpvOpSLessThanEqual:
3260 case SpvOpFOrdLessThanEqual:
3261 case SpvOpFUnordLessThanEqual:
3262 case SpvOpUGreaterThanEqual:
3263 case SpvOpSGreaterThanEqual:
3264 case SpvOpFOrdGreaterThanEqual:
3265 case SpvOpFUnordGreaterThanEqual:
3266 case SpvOpDPdx:
3267 case SpvOpDPdy:
3268 case SpvOpFwidth:
3269 case SpvOpDPdxFine:
3270 case SpvOpDPdyFine:
3271 case SpvOpFwidthFine:
3272 case SpvOpDPdxCoarse:
3273 case SpvOpDPdyCoarse:
3274 case SpvOpFwidthCoarse:
3275 case SpvOpBitFieldInsert:
3276 case SpvOpBitFieldSExtract:
3277 case SpvOpBitFieldUExtract:
3278 case SpvOpBitReverse:
3279 case SpvOpBitCount:
3280 case SpvOpTranspose:
3281 case SpvOpOuterProduct:
3282 case SpvOpMatrixTimesScalar:
3283 case SpvOpVectorTimesMatrix:
3284 case SpvOpMatrixTimesVector:
3285 case SpvOpMatrixTimesMatrix:
3286 vtn_handle_alu(b, opcode, w, count);
3287 break;
3288
3289 case SpvOpVectorExtractDynamic:
3290 case SpvOpVectorInsertDynamic:
3291 case SpvOpVectorShuffle:
3292 case SpvOpCompositeConstruct:
3293 case SpvOpCompositeExtract:
3294 case SpvOpCompositeInsert:
3295 case SpvOpCopyObject:
3296 vtn_handle_composite(b, opcode, w, count);
3297 break;
3298
3299 case SpvOpEmitVertex:
3300 case SpvOpEndPrimitive:
3301 case SpvOpEmitStreamVertex:
3302 case SpvOpEndStreamPrimitive:
3303 case SpvOpControlBarrier:
3304 case SpvOpMemoryBarrier:
3305 vtn_handle_barrier(b, opcode, w, count);
3306 break;
3307
3308 default:
3309 unreachable("Unhandled opcode");
3310 }
3311
3312 return true;
3313 }
3314
3315 nir_function *
3316 spirv_to_nir(const uint32_t *words, size_t word_count,
3317 struct nir_spirv_specialization *spec, unsigned num_spec,
3318 gl_shader_stage stage, const char *entry_point_name,
3319 const struct spirv_to_nir_options *options,
3320 const nir_shader_compiler_options *nir_options)
3321 {
3322 /* Initialize the stn_builder object */
3323 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3324 exec_list_make_empty(&b->functions);
3325 b->entry_point_stage = stage;
3326 b->entry_point_name = entry_point_name;
3327 b->options = options;
3328
3329 const uint32_t *word_end = words + word_count;
3330
3331 /* Handle the SPIR-V header (first 4 dwords) */
3332 assert(word_count > 5);
3333
3334 assert(words[0] == SpvMagicNumber);
3335 assert(words[1] >= 0x10000);
3336 /* words[2] == generator magic */
3337 unsigned value_id_bound = words[3];
3338 assert(words[4] == 0);
3339
3340 words+= 5;
3341
3342 b->value_id_bound = value_id_bound;
3343 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3344
3345 /* Handle all the preamble instructions */
3346 words = vtn_foreach_instruction(b, words, word_end,
3347 vtn_handle_preamble_instruction);
3348
3349 if (b->entry_point == NULL) {
3350 assert(!"Entry point not found");
3351 ralloc_free(b);
3352 return NULL;
3353 }
3354
3355 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3356
3357 /* Set shader info defaults */
3358 b->shader->info.gs.invocations = 1;
3359
3360 /* Parse execution modes */
3361 vtn_foreach_execution_mode(b, b->entry_point,
3362 vtn_handle_execution_mode, NULL);
3363
3364 b->specializations = spec;
3365 b->num_specializations = num_spec;
3366
3367 /* Handle all variable, type, and constant instructions */
3368 words = vtn_foreach_instruction(b, words, word_end,
3369 vtn_handle_variable_or_type_instruction);
3370
3371 vtn_build_cfg(b, words, word_end);
3372
3373 assert(b->entry_point->value_type == vtn_value_type_function);
3374 b->entry_point->func->referenced = true;
3375
3376 bool progress;
3377 do {
3378 progress = false;
3379 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3380 if (func->referenced && !func->emitted) {
3381 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3382 _mesa_key_pointer_equal);
3383
3384 vtn_function_emit(b, func, vtn_handle_body_instruction);
3385 progress = true;
3386 }
3387 }
3388 } while (progress);
3389
3390 assert(b->entry_point->value_type == vtn_value_type_function);
3391 nir_function *entry_point = b->entry_point->func->impl->function;
3392 assert(entry_point);
3393
3394 /* Unparent the shader from the vtn_builder before we delete the builder */
3395 ralloc_steal(NULL, b->shader);
3396
3397 ralloc_free(b);
3398
3399 return entry_point;
3400 }