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