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