2 * Copyright © 2015 Intel Corporation
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:
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
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
24 * Jason Ekstrand (jason@jlekstrand.net)
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"
37 vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
38 size_t spirv_offset, const char *message)
40 if (b->options->debug.func) {
41 b->options->debug.func(b->options->debug.private_data,
42 level, spirv_offset, message);
46 if (level >= NIR_SPIRV_DEBUG_LEVEL_WARNING)
47 fprintf(stderr, "%s\n", message);
52 vtn_logf(struct vtn_builder *b, enum nir_spirv_debug_level level,
53 size_t spirv_offset, const char *fmt, ...)
59 msg = ralloc_vasprintf(NULL, fmt, args);
62 vtn_log(b, level, spirv_offset, msg);
68 vtn_log_err(struct vtn_builder *b,
69 enum nir_spirv_debug_level level, const char *prefix,
70 const char *file, unsigned line,
71 const char *fmt, va_list args)
75 msg = ralloc_strdup(NULL, prefix);
78 ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
81 ralloc_asprintf_append(&msg, " ");
83 ralloc_vasprintf_append(&msg, fmt, args);
85 ralloc_asprintf_append(&msg, "\n %zu bytes into the SPIR-V binary",
89 ralloc_asprintf_append(&msg,
90 "\n in SPIR-V source file %s, line %d, col %d",
91 b->file, b->line, b->col);
94 vtn_log(b, level, b->spirv_offset, msg);
100 vtn_dump_shader(struct vtn_builder *b, const char *path, const char *prefix)
105 int len = snprintf(filename, sizeof(filename), "%s/%s-%d.spirv",
106 path, prefix, idx++);
107 if (len < 0 || len >= sizeof(filename))
110 FILE *f = fopen(filename, "w");
114 fwrite(b->spirv, sizeof(*b->spirv), b->spirv_word_count, f);
117 vtn_info("SPIR-V shader dumped to %s", filename);
121 _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
122 const char *fmt, ...)
127 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_WARNING, "SPIR-V WARNING:\n",
128 file, line, fmt, args);
133 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
134 const char *fmt, ...)
139 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V parsing FAILED:\n",
140 file, line, fmt, args);
143 const char *dump_path = getenv("MESA_SPIRV_FAIL_DUMP_PATH");
145 vtn_dump_shader(b, dump_path, "fail");
147 longjmp(b->fail_jump, 1);
150 struct spec_constant_value {
158 static struct vtn_ssa_value *
159 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
161 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
164 if (glsl_type_is_vector_or_scalar(type)) {
165 unsigned num_components = glsl_get_vector_elements(val->type);
166 unsigned bit_size = glsl_get_bit_size(val->type);
167 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
169 unsigned elems = glsl_get_length(val->type);
170 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
171 if (glsl_type_is_matrix(type)) {
172 const struct glsl_type *elem_type =
173 glsl_vector_type(glsl_get_base_type(type),
174 glsl_get_vector_elements(type));
176 for (unsigned i = 0; i < elems; i++)
177 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
178 } else if (glsl_type_is_array(type)) {
179 const struct glsl_type *elem_type = glsl_get_array_element(type);
180 for (unsigned i = 0; i < elems; i++)
181 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
183 for (unsigned i = 0; i < elems; i++) {
184 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
185 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
193 static struct vtn_ssa_value *
194 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
195 const struct glsl_type *type)
197 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
202 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
205 switch (glsl_get_base_type(type)) {
208 case GLSL_TYPE_INT16:
209 case GLSL_TYPE_UINT16:
210 case GLSL_TYPE_INT64:
211 case GLSL_TYPE_UINT64:
213 case GLSL_TYPE_FLOAT:
214 case GLSL_TYPE_FLOAT16:
215 case GLSL_TYPE_DOUBLE: {
216 int bit_size = glsl_get_bit_size(type);
217 if (glsl_type_is_vector_or_scalar(type)) {
218 unsigned num_components = glsl_get_vector_elements(val->type);
219 nir_load_const_instr *load =
220 nir_load_const_instr_create(b->shader, num_components, bit_size);
222 load->value = constant->values[0];
224 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
225 val->def = &load->def;
227 assert(glsl_type_is_matrix(type));
228 unsigned rows = glsl_get_vector_elements(val->type);
229 unsigned columns = glsl_get_matrix_columns(val->type);
230 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
232 for (unsigned i = 0; i < columns; i++) {
233 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
234 col_val->type = glsl_get_column_type(val->type);
235 nir_load_const_instr *load =
236 nir_load_const_instr_create(b->shader, rows, bit_size);
238 load->value = constant->values[i];
240 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
241 col_val->def = &load->def;
243 val->elems[i] = col_val;
249 case GLSL_TYPE_ARRAY: {
250 unsigned elems = glsl_get_length(val->type);
251 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
252 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
253 for (unsigned i = 0; i < elems; i++)
254 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
259 case GLSL_TYPE_STRUCT: {
260 unsigned elems = glsl_get_length(val->type);
261 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
262 for (unsigned i = 0; i < elems; i++) {
263 const struct glsl_type *elem_type =
264 glsl_get_struct_field(val->type, i);
265 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
272 vtn_fail("bad constant type");
278 struct vtn_ssa_value *
279 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
281 struct vtn_value *val = vtn_untyped_value(b, value_id);
282 switch (val->value_type) {
283 case vtn_value_type_undef:
284 return vtn_undef_ssa_value(b, val->type->type);
286 case vtn_value_type_constant:
287 return vtn_const_ssa_value(b, val->constant, val->type->type);
289 case vtn_value_type_ssa:
292 case vtn_value_type_pointer:
293 vtn_assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
294 struct vtn_ssa_value *ssa =
295 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
296 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
300 vtn_fail("Invalid type for an SSA value");
305 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
306 unsigned word_count, unsigned *words_used)
308 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
310 /* Ammount of space taken by the string (including the null) */
311 unsigned len = strlen(dup) + 1;
312 *words_used = DIV_ROUND_UP(len, sizeof(*words));
318 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
319 const uint32_t *end, vtn_instruction_handler handler)
325 const uint32_t *w = start;
327 SpvOp opcode = w[0] & SpvOpCodeMask;
328 unsigned count = w[0] >> SpvWordCountShift;
329 vtn_assert(count >= 1 && w + count <= end);
331 b->spirv_offset = (uint8_t *)w - (uint8_t *)b->spirv;
335 break; /* Do nothing */
338 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
350 if (!handler(b, opcode, w, count))
368 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
369 const uint32_t *w, unsigned count)
372 case SpvOpExtInstImport: {
373 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
374 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
375 val->ext_handler = vtn_handle_glsl450_instruction;
377 vtn_fail("Unsupported extension");
383 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
384 bool handled = val->ext_handler(b, w[4], w, count);
390 vtn_fail("Unhandled opcode");
395 _foreach_decoration_helper(struct vtn_builder *b,
396 struct vtn_value *base_value,
398 struct vtn_value *value,
399 vtn_decoration_foreach_cb cb, void *data)
401 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
403 if (dec->scope == VTN_DEC_DECORATION) {
404 member = parent_member;
405 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
406 vtn_fail_if(value->value_type != vtn_value_type_type ||
407 value->type->base_type != vtn_base_type_struct,
408 "OpMemberDecorate and OpGroupMemberDecorate are only "
409 "allowed on OpTypeStruct");
410 /* This means we haven't recursed yet */
411 assert(value == base_value);
413 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
415 vtn_fail_if(member >= base_value->type->length,
416 "OpMemberDecorate specifies member %d but the "
417 "OpTypeStruct has only %u members",
418 member, base_value->type->length);
420 /* Not a decoration */
421 assert(dec->scope == VTN_DEC_EXECUTION_MODE);
426 assert(dec->group->value_type == vtn_value_type_decoration_group);
427 _foreach_decoration_helper(b, base_value, member, dec->group,
430 cb(b, base_value, member, dec, data);
435 /** Iterates (recursively if needed) over all of the decorations on a value
437 * This function iterates over all of the decorations applied to a given
438 * value. If it encounters a decoration group, it recurses into the group
439 * and iterates over all of those decorations as well.
442 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
443 vtn_decoration_foreach_cb cb, void *data)
445 _foreach_decoration_helper(b, value, -1, value, cb, data);
449 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
450 vtn_execution_mode_foreach_cb cb, void *data)
452 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
453 if (dec->scope != VTN_DEC_EXECUTION_MODE)
456 assert(dec->group == NULL);
457 cb(b, value, dec, data);
462 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
463 const uint32_t *w, unsigned count)
465 const uint32_t *w_end = w + count;
466 const uint32_t target = w[1];
470 case SpvOpDecorationGroup:
471 vtn_push_value(b, target, vtn_value_type_decoration_group);
475 case SpvOpMemberDecorate:
476 case SpvOpExecutionMode: {
477 struct vtn_value *val = vtn_untyped_value(b, target);
479 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
482 dec->scope = VTN_DEC_DECORATION;
484 case SpvOpMemberDecorate:
485 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
486 vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
487 "Member argument of OpMemberDecorate too large");
489 case SpvOpExecutionMode:
490 dec->scope = VTN_DEC_EXECUTION_MODE;
493 unreachable("Invalid decoration opcode");
495 dec->decoration = *(w++);
498 /* Link into the list */
499 dec->next = val->decoration;
500 val->decoration = dec;
504 case SpvOpGroupMemberDecorate:
505 case SpvOpGroupDecorate: {
506 struct vtn_value *group =
507 vtn_value(b, target, vtn_value_type_decoration_group);
509 for (; w < w_end; w++) {
510 struct vtn_value *val = vtn_untyped_value(b, *w);
511 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
514 if (opcode == SpvOpGroupDecorate) {
515 dec->scope = VTN_DEC_DECORATION;
517 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
518 vtn_fail_if(dec->scope < 0, /* Check for overflow */
519 "Member argument of OpGroupMemberDecorate too large");
522 /* Link into the list */
523 dec->next = val->decoration;
524 val->decoration = dec;
530 unreachable("Unhandled opcode");
534 struct member_decoration_ctx {
536 struct glsl_struct_field *fields;
537 struct vtn_type *type;
540 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
541 * OpStore, or OpCopyMemory between them without breaking anything.
542 * Technically, the SPIR-V rules require the exact same type ID but this lets
543 * us internally be a bit looser.
546 vtn_types_compatible(struct vtn_builder *b,
547 struct vtn_type *t1, struct vtn_type *t2)
549 if (t1->id == t2->id)
552 if (t1->base_type != t2->base_type)
555 switch (t1->base_type) {
556 case vtn_base_type_void:
557 case vtn_base_type_scalar:
558 case vtn_base_type_vector:
559 case vtn_base_type_matrix:
560 case vtn_base_type_image:
561 case vtn_base_type_sampler:
562 case vtn_base_type_sampled_image:
563 return t1->type == t2->type;
565 case vtn_base_type_array:
566 return t1->length == t2->length &&
567 vtn_types_compatible(b, t1->array_element, t2->array_element);
569 case vtn_base_type_pointer:
570 return vtn_types_compatible(b, t1->deref, t2->deref);
572 case vtn_base_type_struct:
573 if (t1->length != t2->length)
576 for (unsigned i = 0; i < t1->length; i++) {
577 if (!vtn_types_compatible(b, t1->members[i], t2->members[i]))
582 case vtn_base_type_function:
583 /* This case shouldn't get hit since you can't copy around function
584 * types. Just require them to be identical.
589 vtn_fail("Invalid base type");
592 /* does a shallow copy of a vtn_type */
594 static struct vtn_type *
595 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
597 struct vtn_type *dest = ralloc(b, struct vtn_type);
600 switch (src->base_type) {
601 case vtn_base_type_void:
602 case vtn_base_type_scalar:
603 case vtn_base_type_vector:
604 case vtn_base_type_matrix:
605 case vtn_base_type_array:
606 case vtn_base_type_pointer:
607 case vtn_base_type_image:
608 case vtn_base_type_sampler:
609 case vtn_base_type_sampled_image:
610 /* Nothing more to do */
613 case vtn_base_type_struct:
614 dest->members = ralloc_array(b, struct vtn_type *, src->length);
615 memcpy(dest->members, src->members,
616 src->length * sizeof(src->members[0]));
618 dest->offsets = ralloc_array(b, unsigned, src->length);
619 memcpy(dest->offsets, src->offsets,
620 src->length * sizeof(src->offsets[0]));
623 case vtn_base_type_function:
624 dest->params = ralloc_array(b, struct vtn_type *, src->length);
625 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
632 static struct vtn_type *
633 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
635 type->members[member] = vtn_type_copy(b, type->members[member]);
636 type = type->members[member];
638 /* We may have an array of matrices.... Oh, joy! */
639 while (glsl_type_is_array(type->type)) {
640 type->array_element = vtn_type_copy(b, type->array_element);
641 type = type->array_element;
644 vtn_assert(glsl_type_is_matrix(type->type));
650 struct_member_decoration_cb(struct vtn_builder *b,
651 struct vtn_value *val, int member,
652 const struct vtn_decoration *dec, void *void_ctx)
654 struct member_decoration_ctx *ctx = void_ctx;
659 assert(member < ctx->num_fields);
661 switch (dec->decoration) {
662 case SpvDecorationNonWritable:
663 case SpvDecorationNonReadable:
664 case SpvDecorationRelaxedPrecision:
665 case SpvDecorationVolatile:
666 case SpvDecorationCoherent:
667 case SpvDecorationUniform:
668 break; /* FIXME: Do nothing with this for now. */
669 case SpvDecorationNoPerspective:
670 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
672 case SpvDecorationFlat:
673 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
675 case SpvDecorationCentroid:
676 ctx->fields[member].centroid = true;
678 case SpvDecorationSample:
679 ctx->fields[member].sample = true;
681 case SpvDecorationStream:
682 /* Vulkan only allows one GS stream */
683 vtn_assert(dec->literals[0] == 0);
685 case SpvDecorationLocation:
686 ctx->fields[member].location = dec->literals[0];
688 case SpvDecorationComponent:
689 break; /* FIXME: What should we do with these? */
690 case SpvDecorationBuiltIn:
691 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
692 ctx->type->members[member]->is_builtin = true;
693 ctx->type->members[member]->builtin = dec->literals[0];
694 ctx->type->builtin_block = true;
696 case SpvDecorationOffset:
697 ctx->type->offsets[member] = dec->literals[0];
699 case SpvDecorationMatrixStride:
700 /* Handled as a second pass */
702 case SpvDecorationColMajor:
703 break; /* Nothing to do here. Column-major is the default. */
704 case SpvDecorationRowMajor:
705 mutable_matrix_member(b, ctx->type, member)->row_major = true;
708 case SpvDecorationPatch:
711 case SpvDecorationSpecId:
712 case SpvDecorationBlock:
713 case SpvDecorationBufferBlock:
714 case SpvDecorationArrayStride:
715 case SpvDecorationGLSLShared:
716 case SpvDecorationGLSLPacked:
717 case SpvDecorationInvariant:
718 case SpvDecorationRestrict:
719 case SpvDecorationAliased:
720 case SpvDecorationConstant:
721 case SpvDecorationIndex:
722 case SpvDecorationBinding:
723 case SpvDecorationDescriptorSet:
724 case SpvDecorationLinkageAttributes:
725 case SpvDecorationNoContraction:
726 case SpvDecorationInputAttachmentIndex:
727 vtn_warn("Decoration not allowed on struct members: %s",
728 spirv_decoration_to_string(dec->decoration));
731 case SpvDecorationXfbBuffer:
732 case SpvDecorationXfbStride:
733 vtn_warn("Vulkan does not have transform feedback");
736 case SpvDecorationCPacked:
737 case SpvDecorationSaturatedConversion:
738 case SpvDecorationFuncParamAttr:
739 case SpvDecorationFPRoundingMode:
740 case SpvDecorationFPFastMathMode:
741 case SpvDecorationAlignment:
742 vtn_warn("Decoration only allowed for CL-style kernels: %s",
743 spirv_decoration_to_string(dec->decoration));
747 vtn_fail("Unhandled decoration");
751 /* Matrix strides are handled as a separate pass because we need to know
752 * whether the matrix is row-major or not first.
755 struct_member_matrix_stride_cb(struct vtn_builder *b,
756 struct vtn_value *val, int member,
757 const struct vtn_decoration *dec,
760 if (dec->decoration != SpvDecorationMatrixStride)
763 vtn_fail_if(member < 0,
764 "The MatrixStride decoration is only allowed on members "
767 struct member_decoration_ctx *ctx = void_ctx;
769 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
770 if (mat_type->row_major) {
771 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
772 mat_type->stride = mat_type->array_element->stride;
773 mat_type->array_element->stride = dec->literals[0];
775 vtn_assert(mat_type->array_element->stride > 0);
776 mat_type->stride = dec->literals[0];
781 type_decoration_cb(struct vtn_builder *b,
782 struct vtn_value *val, int member,
783 const struct vtn_decoration *dec, void *ctx)
785 struct vtn_type *type = val->type;
788 /* This should have been handled by OpTypeStruct */
789 assert(val->type->base_type == vtn_base_type_struct);
790 assert(member >= 0 && member < val->type->length);
794 switch (dec->decoration) {
795 case SpvDecorationArrayStride:
796 vtn_assert(type->base_type == vtn_base_type_matrix ||
797 type->base_type == vtn_base_type_array ||
798 type->base_type == vtn_base_type_pointer);
799 type->stride = dec->literals[0];
801 case SpvDecorationBlock:
802 vtn_assert(type->base_type == vtn_base_type_struct);
805 case SpvDecorationBufferBlock:
806 vtn_assert(type->base_type == vtn_base_type_struct);
807 type->buffer_block = true;
809 case SpvDecorationGLSLShared:
810 case SpvDecorationGLSLPacked:
811 /* Ignore these, since we get explicit offsets anyways */
814 case SpvDecorationRowMajor:
815 case SpvDecorationColMajor:
816 case SpvDecorationMatrixStride:
817 case SpvDecorationBuiltIn:
818 case SpvDecorationNoPerspective:
819 case SpvDecorationFlat:
820 case SpvDecorationPatch:
821 case SpvDecorationCentroid:
822 case SpvDecorationSample:
823 case SpvDecorationVolatile:
824 case SpvDecorationCoherent:
825 case SpvDecorationNonWritable:
826 case SpvDecorationNonReadable:
827 case SpvDecorationUniform:
828 case SpvDecorationStream:
829 case SpvDecorationLocation:
830 case SpvDecorationComponent:
831 case SpvDecorationOffset:
832 case SpvDecorationXfbBuffer:
833 case SpvDecorationXfbStride:
834 vtn_warn("Decoration only allowed for struct members: %s",
835 spirv_decoration_to_string(dec->decoration));
838 case SpvDecorationRelaxedPrecision:
839 case SpvDecorationSpecId:
840 case SpvDecorationInvariant:
841 case SpvDecorationRestrict:
842 case SpvDecorationAliased:
843 case SpvDecorationConstant:
844 case SpvDecorationIndex:
845 case SpvDecorationBinding:
846 case SpvDecorationDescriptorSet:
847 case SpvDecorationLinkageAttributes:
848 case SpvDecorationNoContraction:
849 case SpvDecorationInputAttachmentIndex:
850 vtn_warn("Decoration not allowed on types: %s",
851 spirv_decoration_to_string(dec->decoration));
854 case SpvDecorationCPacked:
855 case SpvDecorationSaturatedConversion:
856 case SpvDecorationFuncParamAttr:
857 case SpvDecorationFPRoundingMode:
858 case SpvDecorationFPFastMathMode:
859 case SpvDecorationAlignment:
860 vtn_warn("Decoration only allowed for CL-style kernels: %s",
861 spirv_decoration_to_string(dec->decoration));
865 vtn_fail("Unhandled decoration");
870 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
873 case SpvImageFormatUnknown: return 0; /* GL_NONE */
874 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
875 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
876 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
877 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
878 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
879 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
880 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
881 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
882 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
883 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
884 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
885 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
886 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
887 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
888 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
889 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
890 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
891 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
892 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
893 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
894 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
895 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
896 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
897 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
898 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
899 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
900 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
901 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
902 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
903 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
904 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
905 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
906 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
907 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
908 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
909 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
910 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
911 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
912 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
914 vtn_fail("Invalid image format");
918 static struct vtn_type *
919 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
920 uint32_t *size_out, uint32_t *align_out)
922 switch (type->base_type) {
923 case vtn_base_type_scalar: {
924 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
925 *size_out = comp_size;
926 *align_out = comp_size;
930 case vtn_base_type_vector: {
931 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
932 assert(type->length > 0 && type->length <= 4);
933 unsigned align_comps = type->length == 3 ? 4 : type->length;
934 *size_out = comp_size * type->length,
935 *align_out = comp_size * align_comps;
939 case vtn_base_type_matrix:
940 case vtn_base_type_array: {
941 /* We're going to add an array stride */
942 type = vtn_type_copy(b, type);
943 uint32_t elem_size, elem_align;
944 type->array_element = vtn_type_layout_std430(b, type->array_element,
945 &elem_size, &elem_align);
946 type->stride = vtn_align_u32(elem_size, elem_align);
947 *size_out = type->stride * type->length;
948 *align_out = elem_align;
952 case vtn_base_type_struct: {
953 /* We're going to add member offsets */
954 type = vtn_type_copy(b, type);
957 for (unsigned i = 0; i < type->length; i++) {
958 uint32_t mem_size, mem_align;
959 type->members[i] = vtn_type_layout_std430(b, type->members[i],
960 &mem_size, &mem_align);
961 offset = vtn_align_u32(offset, mem_align);
962 type->offsets[i] = offset;
964 align = MAX2(align, mem_align);
972 unreachable("Invalid SPIR-V type for std430");
977 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
978 const uint32_t *w, unsigned count)
980 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
982 val->type = rzalloc(b, struct vtn_type);
983 val->type->id = w[1];
987 val->type->base_type = vtn_base_type_void;
988 val->type->type = glsl_void_type();
991 val->type->base_type = vtn_base_type_scalar;
992 val->type->type = glsl_bool_type();
993 val->type->length = 1;
997 const bool signedness = w[3];
998 val->type->base_type = vtn_base_type_scalar;
1001 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
1004 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
1007 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
1010 vtn_fail("Invalid int bit size");
1012 val->type->length = 1;
1016 case SpvOpTypeFloat: {
1017 int bit_size = w[2];
1018 val->type->base_type = vtn_base_type_scalar;
1021 val->type->type = glsl_float16_t_type();
1024 val->type->type = glsl_float_type();
1027 val->type->type = glsl_double_type();
1030 vtn_fail("Invalid float bit size");
1032 val->type->length = 1;
1036 case SpvOpTypeVector: {
1037 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1038 unsigned elems = w[3];
1040 vtn_fail_if(base->base_type != vtn_base_type_scalar,
1041 "Base type for OpTypeVector must be a scalar");
1042 vtn_fail_if(elems < 2 || elems > 4,
1043 "Invalid component count for OpTypeVector");
1045 val->type->base_type = vtn_base_type_vector;
1046 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
1047 val->type->length = elems;
1048 val->type->stride = glsl_get_bit_size(base->type) / 8;
1049 val->type->array_element = base;
1053 case SpvOpTypeMatrix: {
1054 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1055 unsigned columns = w[3];
1057 vtn_fail_if(base->base_type != vtn_base_type_vector,
1058 "Base type for OpTypeMatrix must be a vector");
1059 vtn_fail_if(columns < 2 || columns > 4,
1060 "Invalid column count for OpTypeMatrix");
1062 val->type->base_type = vtn_base_type_matrix;
1063 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
1064 glsl_get_vector_elements(base->type),
1066 vtn_fail_if(glsl_type_is_error(val->type->type),
1067 "Unsupported base type for OpTypeMatrix");
1068 assert(!glsl_type_is_error(val->type->type));
1069 val->type->length = columns;
1070 val->type->array_element = base;
1071 val->type->row_major = false;
1072 val->type->stride = 0;
1076 case SpvOpTypeRuntimeArray:
1077 case SpvOpTypeArray: {
1078 struct vtn_type *array_element =
1079 vtn_value(b, w[2], vtn_value_type_type)->type;
1081 if (opcode == SpvOpTypeRuntimeArray) {
1082 /* A length of 0 is used to denote unsized arrays */
1083 val->type->length = 0;
1086 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
1089 val->type->base_type = vtn_base_type_array;
1090 val->type->type = glsl_array_type(array_element->type, val->type->length);
1091 val->type->array_element = array_element;
1092 val->type->stride = 0;
1096 case SpvOpTypeStruct: {
1097 unsigned num_fields = count - 2;
1098 val->type->base_type = vtn_base_type_struct;
1099 val->type->length = num_fields;
1100 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
1101 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1103 NIR_VLA(struct glsl_struct_field, fields, count);
1104 for (unsigned i = 0; i < num_fields; i++) {
1105 val->type->members[i] =
1106 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1107 fields[i] = (struct glsl_struct_field) {
1108 .type = val->type->members[i]->type,
1109 .name = ralloc_asprintf(b, "field%d", i),
1114 struct member_decoration_ctx ctx = {
1115 .num_fields = num_fields,
1120 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1121 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1123 const char *name = val->name ? val->name : "struct";
1125 val->type->type = glsl_struct_type(fields, num_fields, name);
1129 case SpvOpTypeFunction: {
1130 val->type->base_type = vtn_base_type_function;
1131 val->type->type = NULL;
1133 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1135 const unsigned num_params = count - 3;
1136 val->type->length = num_params;
1137 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1138 for (unsigned i = 0; i < count - 3; i++) {
1139 val->type->params[i] =
1140 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1145 case SpvOpTypePointer: {
1146 SpvStorageClass storage_class = w[2];
1147 struct vtn_type *deref_type =
1148 vtn_value(b, w[3], vtn_value_type_type)->type;
1150 val->type->base_type = vtn_base_type_pointer;
1151 val->type->storage_class = storage_class;
1152 val->type->deref = deref_type;
1154 if (storage_class == SpvStorageClassUniform ||
1155 storage_class == SpvStorageClassStorageBuffer) {
1156 /* These can actually be stored to nir_variables and used as SSA
1157 * values so they need a real glsl_type.
1159 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1162 if (storage_class == SpvStorageClassWorkgroup &&
1163 b->options->lower_workgroup_access_to_offsets) {
1164 uint32_t size, align;
1165 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1167 val->type->length = size;
1168 val->type->align = align;
1169 /* These can actually be stored to nir_variables and used as SSA
1170 * values so they need a real glsl_type.
1172 val->type->type = glsl_uint_type();
1177 case SpvOpTypeImage: {
1178 val->type->base_type = vtn_base_type_image;
1180 const struct vtn_type *sampled_type =
1181 vtn_value(b, w[2], vtn_value_type_type)->type;
1183 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1184 glsl_get_bit_size(sampled_type->type) != 32,
1185 "Sampled type of OpTypeImage must be a 32-bit scalar");
1187 enum glsl_sampler_dim dim;
1188 switch ((SpvDim)w[3]) {
1189 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1190 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1191 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1192 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1193 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1194 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1195 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1197 vtn_fail("Invalid SPIR-V image dimensionality");
1200 bool is_shadow = w[4];
1201 bool is_array = w[5];
1202 bool multisampled = w[6];
1203 unsigned sampled = w[7];
1204 SpvImageFormat format = w[8];
1207 val->type->access_qualifier = w[9];
1209 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1212 if (dim == GLSL_SAMPLER_DIM_2D)
1213 dim = GLSL_SAMPLER_DIM_MS;
1214 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1215 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1217 vtn_fail("Unsupported multisampled image type");
1220 val->type->image_format = translate_image_format(b, format);
1222 enum glsl_base_type sampled_base_type =
1223 glsl_get_base_type(sampled_type->type);
1225 val->type->sampled = true;
1226 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1228 } else if (sampled == 2) {
1229 vtn_assert(!is_shadow);
1230 val->type->sampled = false;
1231 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1233 vtn_fail("We need to know if the image will be sampled");
1238 case SpvOpTypeSampledImage:
1239 val->type->base_type = vtn_base_type_sampled_image;
1240 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1241 val->type->type = val->type->image->type;
1244 case SpvOpTypeSampler:
1245 /* The actual sampler type here doesn't really matter. It gets
1246 * thrown away the moment you combine it with an image. What really
1247 * matters is that it's a sampler type as opposed to an integer type
1248 * so the backend knows what to do.
1250 val->type->base_type = vtn_base_type_sampler;
1251 val->type->type = glsl_bare_sampler_type();
1254 case SpvOpTypeOpaque:
1255 case SpvOpTypeEvent:
1256 case SpvOpTypeDeviceEvent:
1257 case SpvOpTypeReserveId:
1258 case SpvOpTypeQueue:
1261 vtn_fail("Unhandled opcode");
1264 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1267 static nir_constant *
1268 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1270 nir_constant *c = rzalloc(b, nir_constant);
1272 /* For pointers and other typeless things, we have to return something but
1273 * it doesn't matter what.
1278 switch (glsl_get_base_type(type)) {
1280 case GLSL_TYPE_UINT:
1281 case GLSL_TYPE_INT16:
1282 case GLSL_TYPE_UINT16:
1283 case GLSL_TYPE_INT64:
1284 case GLSL_TYPE_UINT64:
1285 case GLSL_TYPE_BOOL:
1286 case GLSL_TYPE_FLOAT:
1287 case GLSL_TYPE_FLOAT16:
1288 case GLSL_TYPE_DOUBLE:
1289 /* Nothing to do here. It's already initialized to zero */
1292 case GLSL_TYPE_ARRAY:
1293 vtn_assert(glsl_get_length(type) > 0);
1294 c->num_elements = glsl_get_length(type);
1295 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1297 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1298 for (unsigned i = 1; i < c->num_elements; i++)
1299 c->elements[i] = c->elements[0];
1302 case GLSL_TYPE_STRUCT:
1303 c->num_elements = glsl_get_length(type);
1304 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1306 for (unsigned i = 0; i < c->num_elements; i++) {
1307 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1312 vtn_fail("Invalid type for null constant");
1319 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1320 int member, const struct vtn_decoration *dec,
1323 vtn_assert(member == -1);
1324 if (dec->decoration != SpvDecorationSpecId)
1327 struct spec_constant_value *const_value = data;
1329 for (unsigned i = 0; i < b->num_specializations; i++) {
1330 if (b->specializations[i].id == dec->literals[0]) {
1331 if (const_value->is_double)
1332 const_value->data64 = b->specializations[i].data64;
1334 const_value->data32 = b->specializations[i].data32;
1341 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1342 uint32_t const_value)
1344 struct spec_constant_value data;
1345 data.is_double = false;
1346 data.data32 = const_value;
1347 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1352 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1353 uint64_t const_value)
1355 struct spec_constant_value data;
1356 data.is_double = true;
1357 data.data64 = const_value;
1358 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1363 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1364 struct vtn_value *val,
1366 const struct vtn_decoration *dec,
1369 vtn_assert(member == -1);
1370 if (dec->decoration != SpvDecorationBuiltIn ||
1371 dec->literals[0] != SpvBuiltInWorkgroupSize)
1374 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1376 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1377 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1378 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1382 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1383 const uint32_t *w, unsigned count)
1385 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1386 val->constant = rzalloc(b, nir_constant);
1388 case SpvOpConstantTrue:
1389 case SpvOpConstantFalse:
1390 case SpvOpSpecConstantTrue:
1391 case SpvOpSpecConstantFalse: {
1392 vtn_fail_if(val->type->type != glsl_bool_type(),
1393 "Result type of %s must be OpTypeBool",
1394 spirv_op_to_string(opcode));
1396 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1397 opcode == SpvOpSpecConstantTrue);
1399 if (opcode == SpvOpSpecConstantTrue ||
1400 opcode == SpvOpSpecConstantFalse)
1401 int_val = get_specialization(b, val, int_val);
1403 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1407 case SpvOpConstant: {
1408 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1409 "Result type of %s must be a scalar",
1410 spirv_op_to_string(opcode));
1411 int bit_size = glsl_get_bit_size(val->type->type);
1414 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1417 val->constant->values->u32[0] = w[3];
1420 val->constant->values->u16[0] = w[3];
1423 vtn_fail("Unsupported SpvOpConstant bit size");
1428 case SpvOpSpecConstant: {
1429 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1430 "Result type of %s must be a scalar",
1431 spirv_op_to_string(opcode));
1432 int bit_size = glsl_get_bit_size(val->type->type);
1435 val->constant->values[0].u64[0] =
1436 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1439 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1442 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1445 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1450 case SpvOpSpecConstantComposite:
1451 case SpvOpConstantComposite: {
1452 unsigned elem_count = count - 3;
1453 vtn_fail_if(elem_count != val->type->length,
1454 "%s has %u constituents, expected %u",
1455 spirv_op_to_string(opcode), elem_count, val->type->length);
1457 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1458 for (unsigned i = 0; i < elem_count; i++)
1459 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1461 switch (val->type->base_type) {
1462 case vtn_base_type_vector: {
1463 assert(glsl_type_is_vector(val->type->type));
1464 int bit_size = glsl_get_bit_size(val->type->type);
1465 for (unsigned i = 0; i < elem_count; i++) {
1468 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1471 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1474 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1477 vtn_fail("Invalid SpvOpConstantComposite bit size");
1483 case vtn_base_type_matrix:
1484 assert(glsl_type_is_matrix(val->type->type));
1485 for (unsigned i = 0; i < elem_count; i++)
1486 val->constant->values[i] = elems[i]->values[0];
1489 case vtn_base_type_struct:
1490 case vtn_base_type_array:
1491 ralloc_steal(val->constant, elems);
1492 val->constant->num_elements = elem_count;
1493 val->constant->elements = elems;
1497 vtn_fail("Result type of %s must be a composite type",
1498 spirv_op_to_string(opcode));
1503 case SpvOpSpecConstantOp: {
1504 SpvOp opcode = get_specialization(b, val, w[3]);
1506 case SpvOpVectorShuffle: {
1507 struct vtn_value *v0 = &b->values[w[4]];
1508 struct vtn_value *v1 = &b->values[w[5]];
1510 vtn_assert(v0->value_type == vtn_value_type_constant ||
1511 v0->value_type == vtn_value_type_undef);
1512 vtn_assert(v1->value_type == vtn_value_type_constant ||
1513 v1->value_type == vtn_value_type_undef);
1515 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1516 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1518 vtn_assert(len0 + len1 < 16);
1520 unsigned bit_size = glsl_get_bit_size(val->type->type);
1521 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1522 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1524 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1525 (void)bit_size0; (void)bit_size1;
1527 if (bit_size == 64) {
1529 if (v0->value_type == vtn_value_type_constant) {
1530 for (unsigned i = 0; i < len0; i++)
1531 u64[i] = v0->constant->values[0].u64[i];
1533 if (v1->value_type == vtn_value_type_constant) {
1534 for (unsigned i = 0; i < len1; i++)
1535 u64[len0 + i] = v1->constant->values[0].u64[i];
1538 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1539 uint32_t comp = w[i + 6];
1540 /* If component is not used, set the value to a known constant
1541 * to detect if it is wrongly used.
1543 if (comp == (uint32_t)-1)
1544 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1546 val->constant->values[0].u64[j] = u64[comp];
1549 /* This is for both 32-bit and 16-bit values */
1551 if (v0->value_type == vtn_value_type_constant) {
1552 for (unsigned i = 0; i < len0; i++)
1553 u32[i] = v0->constant->values[0].u32[i];
1555 if (v1->value_type == vtn_value_type_constant) {
1556 for (unsigned i = 0; i < len1; i++)
1557 u32[len0 + i] = v1->constant->values[0].u32[i];
1560 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1561 uint32_t comp = w[i + 6];
1562 /* If component is not used, set the value to a known constant
1563 * to detect if it is wrongly used.
1565 if (comp == (uint32_t)-1)
1566 val->constant->values[0].u32[j] = 0xdeadbeef;
1568 val->constant->values[0].u32[j] = u32[comp];
1574 case SpvOpCompositeExtract:
1575 case SpvOpCompositeInsert: {
1576 struct vtn_value *comp;
1577 unsigned deref_start;
1578 struct nir_constant **c;
1579 if (opcode == SpvOpCompositeExtract) {
1580 comp = vtn_value(b, w[4], vtn_value_type_constant);
1582 c = &comp->constant;
1584 comp = vtn_value(b, w[5], vtn_value_type_constant);
1586 val->constant = nir_constant_clone(comp->constant,
1593 const struct vtn_type *type = comp->type;
1594 for (unsigned i = deref_start; i < count; i++) {
1595 vtn_fail_if(w[i] > type->length,
1596 "%uth index of %s is %u but the type has only "
1597 "%u elements", i - deref_start,
1598 spirv_op_to_string(opcode), w[i], type->length);
1600 switch (type->base_type) {
1601 case vtn_base_type_vector:
1603 type = type->array_element;
1606 case vtn_base_type_matrix:
1607 assert(col == 0 && elem == -1);
1610 type = type->array_element;
1613 case vtn_base_type_array:
1614 c = &(*c)->elements[w[i]];
1615 type = type->array_element;
1618 case vtn_base_type_struct:
1619 c = &(*c)->elements[w[i]];
1620 type = type->members[w[i]];
1624 vtn_fail("%s must only index into composite types",
1625 spirv_op_to_string(opcode));
1629 if (opcode == SpvOpCompositeExtract) {
1633 unsigned num_components = type->length;
1634 unsigned bit_size = glsl_get_bit_size(type->type);
1635 for (unsigned i = 0; i < num_components; i++)
1638 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1641 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1644 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1647 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1651 struct vtn_value *insert =
1652 vtn_value(b, w[4], vtn_value_type_constant);
1653 vtn_assert(insert->type == type);
1655 *c = insert->constant;
1657 unsigned num_components = type->length;
1658 unsigned bit_size = glsl_get_bit_size(type->type);
1659 for (unsigned i = 0; i < num_components; i++)
1662 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1665 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1668 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1671 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1680 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1681 nir_alu_type src_alu_type = dst_alu_type;
1682 unsigned num_components = glsl_get_vector_elements(val->type->type);
1685 vtn_assert(count <= 7);
1690 /* We have a source in a conversion */
1692 nir_get_nir_type_for_glsl_type(
1693 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1694 /* We use the bitsize of the conversion source to evaluate the opcode later */
1695 bit_size = glsl_get_bit_size(
1696 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1699 bit_size = glsl_get_bit_size(val->type->type);
1702 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1705 nir_const_value src[4];
1707 for (unsigned i = 0; i < count - 4; i++) {
1709 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1711 unsigned j = swap ? 1 - i : i;
1712 src[j] = c->values[0];
1715 val->constant->values[0] =
1716 nir_eval_const_opcode(op, num_components, bit_size, src);
1723 case SpvOpConstantNull:
1724 val->constant = vtn_null_constant(b, val->type->type);
1727 case SpvOpConstantSampler:
1728 vtn_fail("OpConstantSampler requires Kernel Capability");
1732 vtn_fail("Unhandled opcode");
1735 /* Now that we have the value, update the workgroup size if needed */
1736 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1740 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1741 const uint32_t *w, unsigned count)
1743 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1744 struct vtn_function *vtn_callee =
1745 vtn_value(b, w[3], vtn_value_type_function)->func;
1746 struct nir_function *callee = vtn_callee->impl->function;
1748 vtn_callee->referenced = true;
1750 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1751 for (unsigned i = 0; i < call->num_params; i++) {
1752 unsigned arg_id = w[4 + i];
1753 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1754 if (arg->value_type == vtn_value_type_pointer &&
1755 arg->pointer->ptr_type->type == NULL) {
1756 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1757 call->params[i] = nir_deref_var_clone(d, call);
1759 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1761 /* Make a temporary to store the argument in */
1763 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1764 call->params[i] = nir_deref_var_create(call, tmp);
1766 vtn_local_store(b, arg_ssa, call->params[i]);
1770 nir_variable *out_tmp = NULL;
1771 vtn_assert(res_type->type == callee->return_type);
1772 if (!glsl_type_is_void(callee->return_type)) {
1773 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1775 call->return_deref = nir_deref_var_create(call, out_tmp);
1778 nir_builder_instr_insert(&b->nb, &call->instr);
1780 if (glsl_type_is_void(callee->return_type)) {
1781 vtn_push_value(b, w[2], vtn_value_type_undef);
1783 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1787 struct vtn_ssa_value *
1788 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1790 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1793 if (!glsl_type_is_vector_or_scalar(type)) {
1794 unsigned elems = glsl_get_length(type);
1795 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1796 for (unsigned i = 0; i < elems; i++) {
1797 const struct glsl_type *child_type;
1799 switch (glsl_get_base_type(type)) {
1801 case GLSL_TYPE_UINT:
1802 case GLSL_TYPE_INT16:
1803 case GLSL_TYPE_UINT16:
1804 case GLSL_TYPE_INT64:
1805 case GLSL_TYPE_UINT64:
1806 case GLSL_TYPE_BOOL:
1807 case GLSL_TYPE_FLOAT:
1808 case GLSL_TYPE_FLOAT16:
1809 case GLSL_TYPE_DOUBLE:
1810 child_type = glsl_get_column_type(type);
1812 case GLSL_TYPE_ARRAY:
1813 child_type = glsl_get_array_element(type);
1815 case GLSL_TYPE_STRUCT:
1816 child_type = glsl_get_struct_field(type, i);
1819 vtn_fail("unkown base type");
1822 val->elems[i] = vtn_create_ssa_value(b, child_type);
1830 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1833 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1834 src.src_type = type;
1839 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1840 const uint32_t *w, unsigned count)
1842 if (opcode == SpvOpSampledImage) {
1843 struct vtn_value *val =
1844 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1845 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1846 val->sampled_image->type =
1847 vtn_value(b, w[1], vtn_value_type_type)->type;
1848 val->sampled_image->image =
1849 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1850 val->sampled_image->sampler =
1851 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1853 } else if (opcode == SpvOpImage) {
1854 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1855 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1856 if (src_val->value_type == vtn_value_type_sampled_image) {
1857 val->pointer = src_val->sampled_image->image;
1859 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1860 val->pointer = src_val->pointer;
1865 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1866 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1868 struct vtn_sampled_image sampled;
1869 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1870 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1871 sampled = *sampled_val->sampled_image;
1873 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1874 sampled.type = sampled_val->pointer->type;
1875 sampled.image = NULL;
1876 sampled.sampler = sampled_val->pointer;
1879 const struct glsl_type *image_type = sampled.type->type;
1880 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1881 const bool is_array = glsl_sampler_type_is_array(image_type);
1882 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1884 /* Figure out the base texture operation */
1887 case SpvOpImageSampleImplicitLod:
1888 case SpvOpImageSampleDrefImplicitLod:
1889 case SpvOpImageSampleProjImplicitLod:
1890 case SpvOpImageSampleProjDrefImplicitLod:
1891 texop = nir_texop_tex;
1894 case SpvOpImageSampleExplicitLod:
1895 case SpvOpImageSampleDrefExplicitLod:
1896 case SpvOpImageSampleProjExplicitLod:
1897 case SpvOpImageSampleProjDrefExplicitLod:
1898 texop = nir_texop_txl;
1901 case SpvOpImageFetch:
1902 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1903 texop = nir_texop_txf_ms;
1905 texop = nir_texop_txf;
1909 case SpvOpImageGather:
1910 case SpvOpImageDrefGather:
1911 texop = nir_texop_tg4;
1914 case SpvOpImageQuerySizeLod:
1915 case SpvOpImageQuerySize:
1916 texop = nir_texop_txs;
1919 case SpvOpImageQueryLod:
1920 texop = nir_texop_lod;
1923 case SpvOpImageQueryLevels:
1924 texop = nir_texop_query_levels;
1927 case SpvOpImageQuerySamples:
1928 texop = nir_texop_texture_samples;
1932 vtn_fail("Unhandled opcode");
1935 nir_tex_src srcs[8]; /* 8 should be enough */
1936 nir_tex_src *p = srcs;
1940 struct nir_ssa_def *coord;
1941 unsigned coord_components;
1943 case SpvOpImageSampleImplicitLod:
1944 case SpvOpImageSampleExplicitLod:
1945 case SpvOpImageSampleDrefImplicitLod:
1946 case SpvOpImageSampleDrefExplicitLod:
1947 case SpvOpImageSampleProjImplicitLod:
1948 case SpvOpImageSampleProjExplicitLod:
1949 case SpvOpImageSampleProjDrefImplicitLod:
1950 case SpvOpImageSampleProjDrefExplicitLod:
1951 case SpvOpImageFetch:
1952 case SpvOpImageGather:
1953 case SpvOpImageDrefGather:
1954 case SpvOpImageQueryLod: {
1955 /* All these types have the coordinate as their first real argument */
1956 switch (sampler_dim) {
1957 case GLSL_SAMPLER_DIM_1D:
1958 case GLSL_SAMPLER_DIM_BUF:
1959 coord_components = 1;
1961 case GLSL_SAMPLER_DIM_2D:
1962 case GLSL_SAMPLER_DIM_RECT:
1963 case GLSL_SAMPLER_DIM_MS:
1964 coord_components = 2;
1966 case GLSL_SAMPLER_DIM_3D:
1967 case GLSL_SAMPLER_DIM_CUBE:
1968 coord_components = 3;
1971 vtn_fail("Invalid sampler type");
1974 if (is_array && texop != nir_texop_lod)
1977 coord = vtn_ssa_value(b, w[idx++])->def;
1978 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1979 (1 << coord_components) - 1));
1980 p->src_type = nir_tex_src_coord;
1987 coord_components = 0;
1992 case SpvOpImageSampleProjImplicitLod:
1993 case SpvOpImageSampleProjExplicitLod:
1994 case SpvOpImageSampleProjDrefImplicitLod:
1995 case SpvOpImageSampleProjDrefExplicitLod:
1996 /* These have the projector as the last coordinate component */
1997 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1998 p->src_type = nir_tex_src_projector;
2006 unsigned gather_component = 0;
2008 case SpvOpImageSampleDrefImplicitLod:
2009 case SpvOpImageSampleDrefExplicitLod:
2010 case SpvOpImageSampleProjDrefImplicitLod:
2011 case SpvOpImageSampleProjDrefExplicitLod:
2012 case SpvOpImageDrefGather:
2013 /* These all have an explicit depth value as their next source */
2014 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
2017 case SpvOpImageGather:
2018 /* This has a component as its next source */
2020 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
2027 /* For OpImageQuerySizeLod, we always have an LOD */
2028 if (opcode == SpvOpImageQuerySizeLod)
2029 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2031 /* Now we need to handle some number of optional arguments */
2032 const struct vtn_ssa_value *gather_offsets = NULL;
2034 uint32_t operands = w[idx++];
2036 if (operands & SpvImageOperandsBiasMask) {
2037 vtn_assert(texop == nir_texop_tex);
2038 texop = nir_texop_txb;
2039 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
2042 if (operands & SpvImageOperandsLodMask) {
2043 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
2044 texop == nir_texop_txs);
2045 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2048 if (operands & SpvImageOperandsGradMask) {
2049 vtn_assert(texop == nir_texop_txl);
2050 texop = nir_texop_txd;
2051 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
2052 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
2055 if (operands & SpvImageOperandsOffsetMask ||
2056 operands & SpvImageOperandsConstOffsetMask)
2057 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
2059 if (operands & SpvImageOperandsConstOffsetsMask) {
2060 gather_offsets = vtn_ssa_value(b, w[idx++]);
2061 (*p++) = (nir_tex_src){};
2064 if (operands & SpvImageOperandsSampleMask) {
2065 vtn_assert(texop == nir_texop_txf_ms);
2066 texop = nir_texop_txf_ms;
2067 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
2070 /* We should have now consumed exactly all of the arguments */
2071 vtn_assert(idx == count);
2073 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
2076 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
2078 instr->coord_components = coord_components;
2079 instr->sampler_dim = sampler_dim;
2080 instr->is_array = is_array;
2081 instr->is_shadow = is_shadow;
2082 instr->is_new_style_shadow =
2083 is_shadow && glsl_get_components(ret_type->type) == 1;
2084 instr->component = gather_component;
2086 switch (glsl_get_sampler_result_type(image_type)) {
2087 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
2088 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
2089 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
2090 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
2092 vtn_fail("Invalid base type for sampler result");
2095 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
2096 nir_deref_var *texture;
2097 if (sampled.image) {
2098 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
2104 instr->texture = nir_deref_var_clone(texture, instr);
2106 switch (instr->op) {
2112 /* These operations require a sampler */
2113 instr->sampler = nir_deref_var_clone(sampler, instr);
2116 case nir_texop_txf_ms:
2119 case nir_texop_query_levels:
2120 case nir_texop_texture_samples:
2121 case nir_texop_samples_identical:
2123 instr->sampler = NULL;
2125 case nir_texop_txf_ms_mcs:
2126 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2129 nir_ssa_dest_init(&instr->instr, &instr->dest,
2130 nir_tex_instr_dest_size(instr), 32, NULL);
2132 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2133 nir_tex_instr_dest_size(instr));
2136 nir_instr *instruction;
2137 if (gather_offsets) {
2138 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2139 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2140 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2142 /* Copy the current instruction 4x */
2143 for (uint32_t i = 1; i < 4; i++) {
2144 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2145 instrs[i]->op = instr->op;
2146 instrs[i]->coord_components = instr->coord_components;
2147 instrs[i]->sampler_dim = instr->sampler_dim;
2148 instrs[i]->is_array = instr->is_array;
2149 instrs[i]->is_shadow = instr->is_shadow;
2150 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2151 instrs[i]->component = instr->component;
2152 instrs[i]->dest_type = instr->dest_type;
2153 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
2154 instrs[i]->sampler = NULL;
2156 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2158 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2159 nir_tex_instr_dest_size(instr), 32, NULL);
2162 /* Fill in the last argument with the offset from the passed in offsets
2163 * and insert the instruction into the stream.
2165 for (uint32_t i = 0; i < 4; i++) {
2167 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2168 src.src_type = nir_tex_src_offset;
2169 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2170 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2173 /* Combine the results of the 4 instructions by taking their .w
2176 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2177 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2178 vec4->dest.write_mask = 0xf;
2179 for (uint32_t i = 0; i < 4; i++) {
2180 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2181 vec4->src[i].swizzle[0] = 3;
2183 def = &vec4->dest.dest.ssa;
2184 instruction = &vec4->instr;
2186 def = &instr->dest.ssa;
2187 instruction = &instr->instr;
2190 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2191 val->ssa->def = def;
2193 nir_builder_instr_insert(&b->nb, instruction);
2197 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2198 const uint32_t *w, nir_src *src)
2201 case SpvOpAtomicIIncrement:
2202 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2205 case SpvOpAtomicIDecrement:
2206 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2209 case SpvOpAtomicISub:
2211 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2214 case SpvOpAtomicCompareExchange:
2215 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2216 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2219 case SpvOpAtomicExchange:
2220 case SpvOpAtomicIAdd:
2221 case SpvOpAtomicSMin:
2222 case SpvOpAtomicUMin:
2223 case SpvOpAtomicSMax:
2224 case SpvOpAtomicUMax:
2225 case SpvOpAtomicAnd:
2227 case SpvOpAtomicXor:
2228 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2232 vtn_fail("Invalid SPIR-V atomic");
2236 static nir_ssa_def *
2237 get_image_coord(struct vtn_builder *b, uint32_t value)
2239 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2241 /* The image_load_store intrinsics assume a 4-dim coordinate */
2242 unsigned dim = glsl_get_vector_elements(coord->type);
2243 unsigned swizzle[4];
2244 for (unsigned i = 0; i < 4; i++)
2245 swizzle[i] = MIN2(i, dim - 1);
2247 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2251 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2252 const uint32_t *w, unsigned count)
2254 /* Just get this one out of the way */
2255 if (opcode == SpvOpImageTexelPointer) {
2256 struct vtn_value *val =
2257 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2258 val->image = ralloc(b, struct vtn_image_pointer);
2260 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2261 val->image->coord = get_image_coord(b, w[4]);
2262 val->image->sample = vtn_ssa_value(b, w[5])->def;
2266 struct vtn_image_pointer image;
2269 case SpvOpAtomicExchange:
2270 case SpvOpAtomicCompareExchange:
2271 case SpvOpAtomicCompareExchangeWeak:
2272 case SpvOpAtomicIIncrement:
2273 case SpvOpAtomicIDecrement:
2274 case SpvOpAtomicIAdd:
2275 case SpvOpAtomicISub:
2276 case SpvOpAtomicLoad:
2277 case SpvOpAtomicSMin:
2278 case SpvOpAtomicUMin:
2279 case SpvOpAtomicSMax:
2280 case SpvOpAtomicUMax:
2281 case SpvOpAtomicAnd:
2283 case SpvOpAtomicXor:
2284 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2287 case SpvOpAtomicStore:
2288 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2291 case SpvOpImageQuerySize:
2292 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2294 image.sample = NULL;
2297 case SpvOpImageRead:
2298 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2299 image.coord = get_image_coord(b, w[4]);
2301 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2302 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2303 image.sample = vtn_ssa_value(b, w[6])->def;
2305 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2309 case SpvOpImageWrite:
2310 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2311 image.coord = get_image_coord(b, w[2]);
2315 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2316 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2317 image.sample = vtn_ssa_value(b, w[5])->def;
2319 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2324 vtn_fail("Invalid image opcode");
2327 nir_intrinsic_op op;
2329 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2330 OP(ImageQuerySize, size)
2332 OP(ImageWrite, store)
2333 OP(AtomicLoad, load)
2334 OP(AtomicStore, store)
2335 OP(AtomicExchange, atomic_exchange)
2336 OP(AtomicCompareExchange, atomic_comp_swap)
2337 OP(AtomicIIncrement, atomic_add)
2338 OP(AtomicIDecrement, atomic_add)
2339 OP(AtomicIAdd, atomic_add)
2340 OP(AtomicISub, atomic_add)
2341 OP(AtomicSMin, atomic_min)
2342 OP(AtomicUMin, atomic_min)
2343 OP(AtomicSMax, atomic_max)
2344 OP(AtomicUMax, atomic_max)
2345 OP(AtomicAnd, atomic_and)
2346 OP(AtomicOr, atomic_or)
2347 OP(AtomicXor, atomic_xor)
2350 vtn_fail("Invalid image opcode");
2353 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2355 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2356 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2358 /* ImageQuerySize doesn't take any extra parameters */
2359 if (opcode != SpvOpImageQuerySize) {
2360 /* The image coordinate is always 4 components but we may not have that
2361 * many. Swizzle to compensate.
2364 for (unsigned i = 0; i < 4; i++)
2365 swiz[i] = i < image.coord->num_components ? i : 0;
2366 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2368 intrin->src[1] = nir_src_for_ssa(image.sample);
2372 case SpvOpAtomicLoad:
2373 case SpvOpImageQuerySize:
2374 case SpvOpImageRead:
2376 case SpvOpAtomicStore:
2377 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2379 case SpvOpImageWrite:
2380 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2383 case SpvOpAtomicCompareExchange:
2384 case SpvOpAtomicIIncrement:
2385 case SpvOpAtomicIDecrement:
2386 case SpvOpAtomicExchange:
2387 case SpvOpAtomicIAdd:
2388 case SpvOpAtomicISub:
2389 case SpvOpAtomicSMin:
2390 case SpvOpAtomicUMin:
2391 case SpvOpAtomicSMax:
2392 case SpvOpAtomicUMax:
2393 case SpvOpAtomicAnd:
2395 case SpvOpAtomicXor:
2396 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2400 vtn_fail("Invalid image opcode");
2403 if (opcode != SpvOpImageWrite) {
2404 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2405 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2407 unsigned dest_components =
2408 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2409 if (intrin->intrinsic == nir_intrinsic_image_size) {
2410 dest_components = intrin->num_components =
2411 glsl_get_vector_elements(type->type);
2414 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2415 dest_components, 32, NULL);
2417 nir_builder_instr_insert(&b->nb, &intrin->instr);
2419 val->ssa = vtn_create_ssa_value(b, type->type);
2420 val->ssa->def = &intrin->dest.ssa;
2422 nir_builder_instr_insert(&b->nb, &intrin->instr);
2426 static nir_intrinsic_op
2427 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2430 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2431 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2432 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2433 OP(AtomicExchange, atomic_exchange)
2434 OP(AtomicCompareExchange, atomic_comp_swap)
2435 OP(AtomicIIncrement, atomic_add)
2436 OP(AtomicIDecrement, atomic_add)
2437 OP(AtomicIAdd, atomic_add)
2438 OP(AtomicISub, atomic_add)
2439 OP(AtomicSMin, atomic_imin)
2440 OP(AtomicUMin, atomic_umin)
2441 OP(AtomicSMax, atomic_imax)
2442 OP(AtomicUMax, atomic_umax)
2443 OP(AtomicAnd, atomic_and)
2444 OP(AtomicOr, atomic_or)
2445 OP(AtomicXor, atomic_xor)
2448 vtn_fail("Invalid SSBO atomic");
2452 static nir_intrinsic_op
2453 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2456 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2457 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2458 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2459 OP(AtomicExchange, atomic_exchange)
2460 OP(AtomicCompareExchange, atomic_comp_swap)
2461 OP(AtomicIIncrement, atomic_add)
2462 OP(AtomicIDecrement, atomic_add)
2463 OP(AtomicIAdd, atomic_add)
2464 OP(AtomicISub, atomic_add)
2465 OP(AtomicSMin, atomic_imin)
2466 OP(AtomicUMin, atomic_umin)
2467 OP(AtomicSMax, atomic_imax)
2468 OP(AtomicUMax, atomic_umax)
2469 OP(AtomicAnd, atomic_and)
2470 OP(AtomicOr, atomic_or)
2471 OP(AtomicXor, atomic_xor)
2474 vtn_fail("Invalid shared atomic");
2478 static nir_intrinsic_op
2479 get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2482 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2483 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2484 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2485 OP(AtomicExchange, atomic_exchange)
2486 OP(AtomicCompareExchange, atomic_comp_swap)
2487 OP(AtomicIIncrement, atomic_add)
2488 OP(AtomicIDecrement, atomic_add)
2489 OP(AtomicIAdd, atomic_add)
2490 OP(AtomicISub, atomic_add)
2491 OP(AtomicSMin, atomic_imin)
2492 OP(AtomicUMin, atomic_umin)
2493 OP(AtomicSMax, atomic_imax)
2494 OP(AtomicUMax, atomic_umax)
2495 OP(AtomicAnd, atomic_and)
2496 OP(AtomicOr, atomic_or)
2497 OP(AtomicXor, atomic_xor)
2500 vtn_fail("Invalid shared atomic");
2505 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2506 const uint32_t *w, unsigned count)
2508 struct vtn_pointer *ptr;
2509 nir_intrinsic_instr *atomic;
2512 case SpvOpAtomicLoad:
2513 case SpvOpAtomicExchange:
2514 case SpvOpAtomicCompareExchange:
2515 case SpvOpAtomicCompareExchangeWeak:
2516 case SpvOpAtomicIIncrement:
2517 case SpvOpAtomicIDecrement:
2518 case SpvOpAtomicIAdd:
2519 case SpvOpAtomicISub:
2520 case SpvOpAtomicSMin:
2521 case SpvOpAtomicUMin:
2522 case SpvOpAtomicSMax:
2523 case SpvOpAtomicUMax:
2524 case SpvOpAtomicAnd:
2526 case SpvOpAtomicXor:
2527 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2530 case SpvOpAtomicStore:
2531 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2535 vtn_fail("Invalid SPIR-V atomic");
2539 SpvScope scope = w[4];
2540 SpvMemorySemanticsMask semantics = w[5];
2543 if (ptr->mode == vtn_variable_mode_workgroup &&
2544 !b->options->lower_workgroup_access_to_offsets) {
2545 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2546 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2547 nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
2548 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2549 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2552 case SpvOpAtomicLoad:
2553 atomic->num_components = glsl_get_vector_elements(deref_type);
2556 case SpvOpAtomicStore:
2557 atomic->num_components = glsl_get_vector_elements(deref_type);
2558 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2559 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2562 case SpvOpAtomicExchange:
2563 case SpvOpAtomicCompareExchange:
2564 case SpvOpAtomicCompareExchangeWeak:
2565 case SpvOpAtomicIIncrement:
2566 case SpvOpAtomicIDecrement:
2567 case SpvOpAtomicIAdd:
2568 case SpvOpAtomicISub:
2569 case SpvOpAtomicSMin:
2570 case SpvOpAtomicUMin:
2571 case SpvOpAtomicSMax:
2572 case SpvOpAtomicUMax:
2573 case SpvOpAtomicAnd:
2575 case SpvOpAtomicXor:
2576 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2580 vtn_fail("Invalid SPIR-V atomic");
2584 nir_ssa_def *offset, *index;
2585 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2587 nir_intrinsic_op op;
2588 if (ptr->mode == vtn_variable_mode_ssbo) {
2589 op = get_ssbo_nir_atomic_op(b, opcode);
2591 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2592 b->options->lower_workgroup_access_to_offsets);
2593 op = get_shared_nir_atomic_op(b, opcode);
2596 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2600 case SpvOpAtomicLoad:
2601 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2602 if (ptr->mode == vtn_variable_mode_ssbo)
2603 atomic->src[src++] = nir_src_for_ssa(index);
2604 atomic->src[src++] = nir_src_for_ssa(offset);
2607 case SpvOpAtomicStore:
2608 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2609 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2610 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2611 if (ptr->mode == vtn_variable_mode_ssbo)
2612 atomic->src[src++] = nir_src_for_ssa(index);
2613 atomic->src[src++] = nir_src_for_ssa(offset);
2616 case SpvOpAtomicExchange:
2617 case SpvOpAtomicCompareExchange:
2618 case SpvOpAtomicCompareExchangeWeak:
2619 case SpvOpAtomicIIncrement:
2620 case SpvOpAtomicIDecrement:
2621 case SpvOpAtomicIAdd:
2622 case SpvOpAtomicISub:
2623 case SpvOpAtomicSMin:
2624 case SpvOpAtomicUMin:
2625 case SpvOpAtomicSMax:
2626 case SpvOpAtomicUMax:
2627 case SpvOpAtomicAnd:
2629 case SpvOpAtomicXor:
2630 if (ptr->mode == vtn_variable_mode_ssbo)
2631 atomic->src[src++] = nir_src_for_ssa(index);
2632 atomic->src[src++] = nir_src_for_ssa(offset);
2633 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2637 vtn_fail("Invalid SPIR-V atomic");
2641 if (opcode != SpvOpAtomicStore) {
2642 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2644 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2645 glsl_get_vector_elements(type->type),
2646 glsl_get_bit_size(type->type), NULL);
2648 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2649 val->ssa = rzalloc(b, struct vtn_ssa_value);
2650 val->ssa->def = &atomic->dest.ssa;
2651 val->ssa->type = type->type;
2654 nir_builder_instr_insert(&b->nb, &atomic->instr);
2657 static nir_alu_instr *
2658 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2661 switch (num_components) {
2662 case 1: op = nir_op_fmov; break;
2663 case 2: op = nir_op_vec2; break;
2664 case 3: op = nir_op_vec3; break;
2665 case 4: op = nir_op_vec4; break;
2666 default: vtn_fail("bad vector size");
2669 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2670 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2672 vec->dest.write_mask = (1 << num_components) - 1;
2677 struct vtn_ssa_value *
2678 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2680 if (src->transposed)
2681 return src->transposed;
2683 struct vtn_ssa_value *dest =
2684 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2686 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2687 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2688 glsl_get_bit_size(src->type));
2689 if (glsl_type_is_vector_or_scalar(src->type)) {
2690 vec->src[0].src = nir_src_for_ssa(src->def);
2691 vec->src[0].swizzle[0] = i;
2693 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2694 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2695 vec->src[j].swizzle[0] = i;
2698 nir_builder_instr_insert(&b->nb, &vec->instr);
2699 dest->elems[i]->def = &vec->dest.dest.ssa;
2702 dest->transposed = src;
2708 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2710 unsigned swiz[4] = { index };
2711 return nir_swizzle(&b->nb, src, swiz, 1, true);
2715 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2718 nir_alu_instr *vec = create_vec(b, src->num_components,
2721 for (unsigned i = 0; i < src->num_components; i++) {
2723 vec->src[i].src = nir_src_for_ssa(insert);
2725 vec->src[i].src = nir_src_for_ssa(src);
2726 vec->src[i].swizzle[0] = i;
2730 nir_builder_instr_insert(&b->nb, &vec->instr);
2732 return &vec->dest.dest.ssa;
2736 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2739 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2740 for (unsigned i = 1; i < src->num_components; i++)
2741 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2742 vtn_vector_extract(b, src, i), dest);
2748 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2749 nir_ssa_def *insert, nir_ssa_def *index)
2751 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2752 for (unsigned i = 1; i < src->num_components; i++)
2753 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2754 vtn_vector_insert(b, src, insert, i), dest);
2759 static nir_ssa_def *
2760 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2761 nir_ssa_def *src0, nir_ssa_def *src1,
2762 const uint32_t *indices)
2764 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2766 for (unsigned i = 0; i < num_components; i++) {
2767 uint32_t index = indices[i];
2768 if (index == 0xffffffff) {
2770 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2771 } else if (index < src0->num_components) {
2772 vec->src[i].src = nir_src_for_ssa(src0);
2773 vec->src[i].swizzle[0] = index;
2775 vec->src[i].src = nir_src_for_ssa(src1);
2776 vec->src[i].swizzle[0] = index - src0->num_components;
2780 nir_builder_instr_insert(&b->nb, &vec->instr);
2782 return &vec->dest.dest.ssa;
2786 * Concatentates a number of vectors/scalars together to produce a vector
2788 static nir_ssa_def *
2789 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2790 unsigned num_srcs, nir_ssa_def **srcs)
2792 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2794 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2796 * "When constructing a vector, there must be at least two Constituent
2799 vtn_assert(num_srcs >= 2);
2801 unsigned dest_idx = 0;
2802 for (unsigned i = 0; i < num_srcs; i++) {
2803 nir_ssa_def *src = srcs[i];
2804 vtn_assert(dest_idx + src->num_components <= num_components);
2805 for (unsigned j = 0; j < src->num_components; j++) {
2806 vec->src[dest_idx].src = nir_src_for_ssa(src);
2807 vec->src[dest_idx].swizzle[0] = j;
2812 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2814 * "When constructing a vector, the total number of components in all
2815 * the operands must equal the number of components in Result Type."
2817 vtn_assert(dest_idx == num_components);
2819 nir_builder_instr_insert(&b->nb, &vec->instr);
2821 return &vec->dest.dest.ssa;
2824 static struct vtn_ssa_value *
2825 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2827 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2828 dest->type = src->type;
2830 if (glsl_type_is_vector_or_scalar(src->type)) {
2831 dest->def = src->def;
2833 unsigned elems = glsl_get_length(src->type);
2835 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2836 for (unsigned i = 0; i < elems; i++)
2837 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2843 static struct vtn_ssa_value *
2844 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2845 struct vtn_ssa_value *insert, const uint32_t *indices,
2846 unsigned num_indices)
2848 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2850 struct vtn_ssa_value *cur = dest;
2852 for (i = 0; i < num_indices - 1; i++) {
2853 cur = cur->elems[indices[i]];
2856 if (glsl_type_is_vector_or_scalar(cur->type)) {
2857 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2858 * the component granularity. In that case, the last index will be
2859 * the index to insert the scalar into the vector.
2862 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2864 cur->elems[indices[i]] = insert;
2870 static struct vtn_ssa_value *
2871 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2872 const uint32_t *indices, unsigned num_indices)
2874 struct vtn_ssa_value *cur = src;
2875 for (unsigned i = 0; i < num_indices; i++) {
2876 if (glsl_type_is_vector_or_scalar(cur->type)) {
2877 vtn_assert(i == num_indices - 1);
2878 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2879 * the component granularity. The last index will be the index of the
2880 * vector to extract.
2883 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2884 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2885 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2888 cur = cur->elems[indices[i]];
2896 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2897 const uint32_t *w, unsigned count)
2899 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2900 const struct glsl_type *type =
2901 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2902 val->ssa = vtn_create_ssa_value(b, type);
2905 case SpvOpVectorExtractDynamic:
2906 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2907 vtn_ssa_value(b, w[4])->def);
2910 case SpvOpVectorInsertDynamic:
2911 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2912 vtn_ssa_value(b, w[4])->def,
2913 vtn_ssa_value(b, w[5])->def);
2916 case SpvOpVectorShuffle:
2917 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2918 vtn_ssa_value(b, w[3])->def,
2919 vtn_ssa_value(b, w[4])->def,
2923 case SpvOpCompositeConstruct: {
2924 unsigned elems = count - 3;
2925 if (glsl_type_is_vector_or_scalar(type)) {
2926 nir_ssa_def *srcs[4];
2927 for (unsigned i = 0; i < elems; i++)
2928 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2930 vtn_vector_construct(b, glsl_get_vector_elements(type),
2933 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2934 for (unsigned i = 0; i < elems; i++)
2935 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2939 case SpvOpCompositeExtract:
2940 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2944 case SpvOpCompositeInsert:
2945 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2946 vtn_ssa_value(b, w[3]),
2950 case SpvOpCopyObject:
2951 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2955 vtn_fail("unknown composite operation");
2960 vtn_emit_barrier(struct vtn_builder *b, nir_intrinsic_op op)
2962 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2963 nir_builder_instr_insert(&b->nb, &intrin->instr);
2967 vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
2968 SpvMemorySemanticsMask semantics)
2970 static const SpvMemorySemanticsMask all_memory_semantics =
2971 SpvMemorySemanticsUniformMemoryMask |
2972 SpvMemorySemanticsWorkgroupMemoryMask |
2973 SpvMemorySemanticsAtomicCounterMemoryMask |
2974 SpvMemorySemanticsImageMemoryMask;
2976 /* If we're not actually doing a memory barrier, bail */
2977 if (!(semantics & all_memory_semantics))
2980 /* GL and Vulkan don't have these */
2981 vtn_assert(scope != SpvScopeCrossDevice);
2983 if (scope == SpvScopeSubgroup)
2984 return; /* Nothing to do here */
2986 if (scope == SpvScopeWorkgroup) {
2987 vtn_emit_barrier(b, nir_intrinsic_group_memory_barrier);
2991 /* There's only two scopes thing left */
2992 vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice);
2994 if ((semantics & all_memory_semantics) == all_memory_semantics) {
2995 vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
2999 /* Issue a bunch of more specific barriers */
3000 uint32_t bits = semantics;
3002 SpvMemorySemanticsMask semantic = 1 << u_bit_scan(&bits);
3004 case SpvMemorySemanticsUniformMemoryMask:
3005 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer);
3007 case SpvMemorySemanticsWorkgroupMemoryMask:
3008 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared);
3010 case SpvMemorySemanticsAtomicCounterMemoryMask:
3011 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter);
3013 case SpvMemorySemanticsImageMemoryMask:
3014 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image);
3023 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
3024 const uint32_t *w, unsigned count)
3027 case SpvOpEmitVertex:
3028 case SpvOpEmitStreamVertex:
3029 case SpvOpEndPrimitive:
3030 case SpvOpEndStreamPrimitive: {
3031 nir_intrinsic_op intrinsic_op;
3033 case SpvOpEmitVertex:
3034 case SpvOpEmitStreamVertex:
3035 intrinsic_op = nir_intrinsic_emit_vertex;
3037 case SpvOpEndPrimitive:
3038 case SpvOpEndStreamPrimitive:
3039 intrinsic_op = nir_intrinsic_end_primitive;
3042 unreachable("Invalid opcode");
3045 nir_intrinsic_instr *intrin =
3046 nir_intrinsic_instr_create(b->shader, intrinsic_op);
3049 case SpvOpEmitStreamVertex:
3050 case SpvOpEndStreamPrimitive:
3051 nir_intrinsic_set_stream_id(intrin, w[1]);
3057 nir_builder_instr_insert(&b->nb, &intrin->instr);
3061 case SpvOpMemoryBarrier: {
3062 SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
3063 SpvMemorySemanticsMask semantics =
3064 vtn_constant_value(b, w[2])->values[0].u32[0];
3065 vtn_emit_memory_barrier(b, scope, semantics);
3069 case SpvOpControlBarrier: {
3070 SpvScope execution_scope =
3071 vtn_constant_value(b, w[1])->values[0].u32[0];
3072 if (execution_scope == SpvScopeWorkgroup)
3073 vtn_emit_barrier(b, nir_intrinsic_barrier);
3075 SpvScope memory_scope =
3076 vtn_constant_value(b, w[2])->values[0].u32[0];
3077 SpvMemorySemanticsMask memory_semantics =
3078 vtn_constant_value(b, w[3])->values[0].u32[0];
3079 vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
3084 unreachable("unknown barrier instruction");
3089 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
3090 SpvExecutionMode mode)
3093 case SpvExecutionModeInputPoints:
3094 case SpvExecutionModeOutputPoints:
3095 return 0; /* GL_POINTS */
3096 case SpvExecutionModeInputLines:
3097 return 1; /* GL_LINES */
3098 case SpvExecutionModeInputLinesAdjacency:
3099 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3100 case SpvExecutionModeTriangles:
3101 return 4; /* GL_TRIANGLES */
3102 case SpvExecutionModeInputTrianglesAdjacency:
3103 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3104 case SpvExecutionModeQuads:
3105 return 7; /* GL_QUADS */
3106 case SpvExecutionModeIsolines:
3107 return 0x8E7A; /* GL_ISOLINES */
3108 case SpvExecutionModeOutputLineStrip:
3109 return 3; /* GL_LINE_STRIP */
3110 case SpvExecutionModeOutputTriangleStrip:
3111 return 5; /* GL_TRIANGLE_STRIP */
3113 vtn_fail("Invalid primitive type");
3118 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
3119 SpvExecutionMode mode)
3122 case SpvExecutionModeInputPoints:
3124 case SpvExecutionModeInputLines:
3126 case SpvExecutionModeInputLinesAdjacency:
3128 case SpvExecutionModeTriangles:
3130 case SpvExecutionModeInputTrianglesAdjacency:
3133 vtn_fail("Invalid GS input mode");
3137 static gl_shader_stage
3138 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
3141 case SpvExecutionModelVertex:
3142 return MESA_SHADER_VERTEX;
3143 case SpvExecutionModelTessellationControl:
3144 return MESA_SHADER_TESS_CTRL;
3145 case SpvExecutionModelTessellationEvaluation:
3146 return MESA_SHADER_TESS_EVAL;
3147 case SpvExecutionModelGeometry:
3148 return MESA_SHADER_GEOMETRY;
3149 case SpvExecutionModelFragment:
3150 return MESA_SHADER_FRAGMENT;
3151 case SpvExecutionModelGLCompute:
3152 return MESA_SHADER_COMPUTE;
3154 vtn_fail("Unsupported execution model");
3158 #define spv_check_supported(name, cap) do { \
3159 if (!(b->options && b->options->caps.name)) \
3160 vtn_warn("Unsupported SPIR-V capability: %s", \
3161 spirv_capability_to_string(cap)); \
3165 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
3166 const uint32_t *w, unsigned count)
3173 case SpvSourceLanguageUnknown: lang = "unknown"; break;
3174 case SpvSourceLanguageESSL: lang = "ESSL"; break;
3175 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
3176 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
3177 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
3178 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
3181 uint32_t version = w[2];
3184 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
3186 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
3190 case SpvOpSourceExtension:
3191 case SpvOpSourceContinued:
3192 case SpvOpExtension:
3193 /* Unhandled, but these are for debug so that's ok. */
3196 case SpvOpCapability: {
3197 SpvCapability cap = w[1];
3199 case SpvCapabilityMatrix:
3200 case SpvCapabilityShader:
3201 case SpvCapabilityGeometry:
3202 case SpvCapabilityGeometryPointSize:
3203 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3204 case SpvCapabilitySampledImageArrayDynamicIndexing:
3205 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3206 case SpvCapabilityStorageImageArrayDynamicIndexing:
3207 case SpvCapabilityImageRect:
3208 case SpvCapabilitySampledRect:
3209 case SpvCapabilitySampled1D:
3210 case SpvCapabilityImage1D:
3211 case SpvCapabilitySampledCubeArray:
3212 case SpvCapabilityImageCubeArray:
3213 case SpvCapabilitySampledBuffer:
3214 case SpvCapabilityImageBuffer:
3215 case SpvCapabilityImageQuery:
3216 case SpvCapabilityDerivativeControl:
3217 case SpvCapabilityInterpolationFunction:
3218 case SpvCapabilityMultiViewport:
3219 case SpvCapabilitySampleRateShading:
3220 case SpvCapabilityClipDistance:
3221 case SpvCapabilityCullDistance:
3222 case SpvCapabilityInputAttachment:
3223 case SpvCapabilityImageGatherExtended:
3224 case SpvCapabilityStorageImageExtendedFormats:
3227 case SpvCapabilityGeometryStreams:
3228 case SpvCapabilityLinkage:
3229 case SpvCapabilityVector16:
3230 case SpvCapabilityFloat16Buffer:
3231 case SpvCapabilityFloat16:
3232 case SpvCapabilityInt64Atomics:
3233 case SpvCapabilityAtomicStorage:
3234 case SpvCapabilityInt16:
3235 case SpvCapabilityStorageImageMultisample:
3236 case SpvCapabilityInt8:
3237 case SpvCapabilitySparseResidency:
3238 case SpvCapabilityMinLod:
3239 case SpvCapabilityTransformFeedback:
3240 vtn_warn("Unsupported SPIR-V capability: %s",
3241 spirv_capability_to_string(cap));
3244 case SpvCapabilityFloat64:
3245 spv_check_supported(float64, cap);
3247 case SpvCapabilityInt64:
3248 spv_check_supported(int64, cap);
3251 case SpvCapabilityAddresses:
3252 case SpvCapabilityKernel:
3253 case SpvCapabilityImageBasic:
3254 case SpvCapabilityImageReadWrite:
3255 case SpvCapabilityImageMipmap:
3256 case SpvCapabilityPipes:
3257 case SpvCapabilityGroups:
3258 case SpvCapabilityDeviceEnqueue:
3259 case SpvCapabilityLiteralSampler:
3260 case SpvCapabilityGenericPointer:
3261 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3262 spirv_capability_to_string(cap));
3265 case SpvCapabilityImageMSArray:
3266 spv_check_supported(image_ms_array, cap);
3269 case SpvCapabilityTessellation:
3270 case SpvCapabilityTessellationPointSize:
3271 spv_check_supported(tessellation, cap);
3274 case SpvCapabilityDrawParameters:
3275 spv_check_supported(draw_parameters, cap);
3278 case SpvCapabilityStorageImageReadWithoutFormat:
3279 spv_check_supported(image_read_without_format, cap);
3282 case SpvCapabilityStorageImageWriteWithoutFormat:
3283 spv_check_supported(image_write_without_format, cap);
3286 case SpvCapabilityDeviceGroup:
3287 spv_check_supported(device_group, cap);
3290 case SpvCapabilityMultiView:
3291 spv_check_supported(multiview, cap);
3294 case SpvCapabilityVariablePointersStorageBuffer:
3295 case SpvCapabilityVariablePointers:
3296 spv_check_supported(variable_pointers, cap);
3299 case SpvCapabilityStorageUniformBufferBlock16:
3300 case SpvCapabilityStorageUniform16:
3301 case SpvCapabilityStoragePushConstant16:
3302 case SpvCapabilityStorageInputOutput16:
3303 spv_check_supported(storage_16bit, cap);
3306 case SpvCapabilityShaderViewportIndexLayerEXT:
3307 spv_check_supported(shader_viewport_index_layer, cap);
3311 vtn_fail("Unhandled capability");
3316 case SpvOpExtInstImport:
3317 vtn_handle_extension(b, opcode, w, count);
3320 case SpvOpMemoryModel:
3321 vtn_assert(w[1] == SpvAddressingModelLogical);
3322 vtn_assert(w[2] == SpvMemoryModelSimple ||
3323 w[2] == SpvMemoryModelGLSL450);
3326 case SpvOpEntryPoint: {
3327 struct vtn_value *entry_point = &b->values[w[2]];
3328 /* Let this be a name label regardless */
3329 unsigned name_words;
3330 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3332 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3333 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3336 vtn_assert(b->entry_point == NULL);
3337 b->entry_point = entry_point;
3342 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3343 vtn_string_literal(b, &w[2], count - 2, NULL);
3347 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3350 case SpvOpMemberName:
3354 case SpvOpExecutionMode:
3355 case SpvOpDecorationGroup:
3357 case SpvOpMemberDecorate:
3358 case SpvOpGroupDecorate:
3359 case SpvOpGroupMemberDecorate:
3360 vtn_handle_decoration(b, opcode, w, count);
3364 return false; /* End of preamble */
3371 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3372 const struct vtn_decoration *mode, void *data)
3374 vtn_assert(b->entry_point == entry_point);
3376 switch(mode->exec_mode) {
3377 case SpvExecutionModeOriginUpperLeft:
3378 case SpvExecutionModeOriginLowerLeft:
3379 b->origin_upper_left =
3380 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3383 case SpvExecutionModeEarlyFragmentTests:
3384 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3385 b->shader->info.fs.early_fragment_tests = true;
3388 case SpvExecutionModeInvocations:
3389 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3390 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3393 case SpvExecutionModeDepthReplacing:
3394 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3395 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3397 case SpvExecutionModeDepthGreater:
3398 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3399 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3401 case SpvExecutionModeDepthLess:
3402 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3403 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3405 case SpvExecutionModeDepthUnchanged:
3406 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3407 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3410 case SpvExecutionModeLocalSize:
3411 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3412 b->shader->info.cs.local_size[0] = mode->literals[0];
3413 b->shader->info.cs.local_size[1] = mode->literals[1];
3414 b->shader->info.cs.local_size[2] = mode->literals[2];
3416 case SpvExecutionModeLocalSizeHint:
3417 break; /* Nothing to do with this */
3419 case SpvExecutionModeOutputVertices:
3420 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3421 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3422 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3424 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3425 b->shader->info.gs.vertices_out = mode->literals[0];
3429 case SpvExecutionModeInputPoints:
3430 case SpvExecutionModeInputLines:
3431 case SpvExecutionModeInputLinesAdjacency:
3432 case SpvExecutionModeTriangles:
3433 case SpvExecutionModeInputTrianglesAdjacency:
3434 case SpvExecutionModeQuads:
3435 case SpvExecutionModeIsolines:
3436 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3437 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3438 b->shader->info.tess.primitive_mode =
3439 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3441 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3442 b->shader->info.gs.vertices_in =
3443 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3447 case SpvExecutionModeOutputPoints:
3448 case SpvExecutionModeOutputLineStrip:
3449 case SpvExecutionModeOutputTriangleStrip:
3450 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3451 b->shader->info.gs.output_primitive =
3452 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3455 case SpvExecutionModeSpacingEqual:
3456 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3457 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3458 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3460 case SpvExecutionModeSpacingFractionalEven:
3461 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3462 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3463 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3465 case SpvExecutionModeSpacingFractionalOdd:
3466 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3467 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3468 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3470 case SpvExecutionModeVertexOrderCw:
3471 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3472 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3473 b->shader->info.tess.ccw = false;
3475 case SpvExecutionModeVertexOrderCcw:
3476 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3477 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3478 b->shader->info.tess.ccw = true;
3480 case SpvExecutionModePointMode:
3481 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3482 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3483 b->shader->info.tess.point_mode = true;
3486 case SpvExecutionModePixelCenterInteger:
3487 b->pixel_center_integer = true;
3490 case SpvExecutionModeXfb:
3491 vtn_fail("Unhandled execution mode");
3494 case SpvExecutionModeVecTypeHint:
3495 case SpvExecutionModeContractionOff:
3499 vtn_fail("Unhandled execution mode");
3504 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3505 const uint32_t *w, unsigned count)
3507 vtn_set_instruction_result_type(b, opcode, w, count);
3511 case SpvOpSourceContinued:
3512 case SpvOpSourceExtension:
3513 case SpvOpExtension:
3514 case SpvOpCapability:
3515 case SpvOpExtInstImport:
3516 case SpvOpMemoryModel:
3517 case SpvOpEntryPoint:
3518 case SpvOpExecutionMode:
3521 case SpvOpMemberName:
3522 case SpvOpDecorationGroup:
3524 case SpvOpMemberDecorate:
3525 case SpvOpGroupDecorate:
3526 case SpvOpGroupMemberDecorate:
3527 vtn_fail("Invalid opcode types and variables section");
3533 case SpvOpTypeFloat:
3534 case SpvOpTypeVector:
3535 case SpvOpTypeMatrix:
3536 case SpvOpTypeImage:
3537 case SpvOpTypeSampler:
3538 case SpvOpTypeSampledImage:
3539 case SpvOpTypeArray:
3540 case SpvOpTypeRuntimeArray:
3541 case SpvOpTypeStruct:
3542 case SpvOpTypeOpaque:
3543 case SpvOpTypePointer:
3544 case SpvOpTypeFunction:
3545 case SpvOpTypeEvent:
3546 case SpvOpTypeDeviceEvent:
3547 case SpvOpTypeReserveId:
3548 case SpvOpTypeQueue:
3550 vtn_handle_type(b, opcode, w, count);
3553 case SpvOpConstantTrue:
3554 case SpvOpConstantFalse:
3556 case SpvOpConstantComposite:
3557 case SpvOpConstantSampler:
3558 case SpvOpConstantNull:
3559 case SpvOpSpecConstantTrue:
3560 case SpvOpSpecConstantFalse:
3561 case SpvOpSpecConstant:
3562 case SpvOpSpecConstantComposite:
3563 case SpvOpSpecConstantOp:
3564 vtn_handle_constant(b, opcode, w, count);
3569 vtn_handle_variables(b, opcode, w, count);
3573 return false; /* End of preamble */
3580 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3581 const uint32_t *w, unsigned count)
3587 case SpvOpLoopMerge:
3588 case SpvOpSelectionMerge:
3589 /* This is handled by cfg pre-pass and walk_blocks */
3593 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3594 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3599 vtn_handle_extension(b, opcode, w, count);
3605 case SpvOpCopyMemory:
3606 case SpvOpCopyMemorySized:
3607 case SpvOpAccessChain:
3608 case SpvOpPtrAccessChain:
3609 case SpvOpInBoundsAccessChain:
3610 case SpvOpArrayLength:
3611 vtn_handle_variables(b, opcode, w, count);
3614 case SpvOpFunctionCall:
3615 vtn_handle_function_call(b, opcode, w, count);
3618 case SpvOpSampledImage:
3620 case SpvOpImageSampleImplicitLod:
3621 case SpvOpImageSampleExplicitLod:
3622 case SpvOpImageSampleDrefImplicitLod:
3623 case SpvOpImageSampleDrefExplicitLod:
3624 case SpvOpImageSampleProjImplicitLod:
3625 case SpvOpImageSampleProjExplicitLod:
3626 case SpvOpImageSampleProjDrefImplicitLod:
3627 case SpvOpImageSampleProjDrefExplicitLod:
3628 case SpvOpImageFetch:
3629 case SpvOpImageGather:
3630 case SpvOpImageDrefGather:
3631 case SpvOpImageQuerySizeLod:
3632 case SpvOpImageQueryLod:
3633 case SpvOpImageQueryLevels:
3634 case SpvOpImageQuerySamples:
3635 vtn_handle_texture(b, opcode, w, count);
3638 case SpvOpImageRead:
3639 case SpvOpImageWrite:
3640 case SpvOpImageTexelPointer:
3641 vtn_handle_image(b, opcode, w, count);
3644 case SpvOpImageQuerySize: {
3645 struct vtn_pointer *image =
3646 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3647 if (image->mode == vtn_variable_mode_image) {
3648 vtn_handle_image(b, opcode, w, count);
3650 vtn_assert(image->mode == vtn_variable_mode_sampler);
3651 vtn_handle_texture(b, opcode, w, count);
3656 case SpvOpAtomicLoad:
3657 case SpvOpAtomicExchange:
3658 case SpvOpAtomicCompareExchange:
3659 case SpvOpAtomicCompareExchangeWeak:
3660 case SpvOpAtomicIIncrement:
3661 case SpvOpAtomicIDecrement:
3662 case SpvOpAtomicIAdd:
3663 case SpvOpAtomicISub:
3664 case SpvOpAtomicSMin:
3665 case SpvOpAtomicUMin:
3666 case SpvOpAtomicSMax:
3667 case SpvOpAtomicUMax:
3668 case SpvOpAtomicAnd:
3670 case SpvOpAtomicXor: {
3671 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3672 if (pointer->value_type == vtn_value_type_image_pointer) {
3673 vtn_handle_image(b, opcode, w, count);
3675 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3676 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3681 case SpvOpAtomicStore: {
3682 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3683 if (pointer->value_type == vtn_value_type_image_pointer) {
3684 vtn_handle_image(b, opcode, w, count);
3686 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3687 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3693 /* Handle OpSelect up-front here because it needs to be able to handle
3694 * pointers and not just regular vectors and scalars.
3696 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3697 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3698 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3699 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3701 const struct glsl_type *sel_type;
3702 switch (res_val->type->base_type) {
3703 case vtn_base_type_scalar:
3704 sel_type = glsl_bool_type();
3706 case vtn_base_type_vector:
3707 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3709 case vtn_base_type_pointer:
3710 /* We need to have actual storage for pointer types */
3711 vtn_fail_if(res_val->type->type == NULL,
3712 "Invalid pointer result type for OpSelect");
3713 sel_type = glsl_bool_type();
3716 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3719 if (unlikely(sel_val->type->type != sel_type)) {
3720 if (sel_val->type->type == glsl_bool_type()) {
3721 /* This case is illegal but some older versions of GLSLang produce
3722 * it. The GLSLang issue was fixed on March 30, 2017:
3724 * https://github.com/KhronosGroup/glslang/issues/809
3726 * Unfortunately, there are applications in the wild which are
3727 * shipping with this bug so it isn't nice to fail on them so we
3728 * throw a warning instead. It's not actually a problem for us as
3729 * nir_builder will just splat the condition out which is most
3730 * likely what the client wanted anyway.
3732 vtn_warn("Condition type of OpSelect must have the same number "
3733 "of components as Result Type");
3735 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3736 "of Boolean type. It must have the same number of "
3737 "components as Result Type");
3741 vtn_fail_if(obj1_val->type != res_val->type ||
3742 obj2_val->type != res_val->type,
3743 "Object types must match the result type in OpSelect");
3745 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3746 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3747 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3748 vtn_ssa_value(b, w[4])->def,
3749 vtn_ssa_value(b, w[5])->def);
3750 vtn_push_ssa(b, w[2], res_type, ssa);
3759 case SpvOpConvertFToU:
3760 case SpvOpConvertFToS:
3761 case SpvOpConvertSToF:
3762 case SpvOpConvertUToF:
3766 case SpvOpQuantizeToF16:
3767 case SpvOpConvertPtrToU:
3768 case SpvOpConvertUToPtr:
3769 case SpvOpPtrCastToGeneric:
3770 case SpvOpGenericCastToPtr:
3776 case SpvOpSignBitSet:
3777 case SpvOpLessOrGreater:
3779 case SpvOpUnordered:
3794 case SpvOpVectorTimesScalar:
3796 case SpvOpIAddCarry:
3797 case SpvOpISubBorrow:
3798 case SpvOpUMulExtended:
3799 case SpvOpSMulExtended:
3800 case SpvOpShiftRightLogical:
3801 case SpvOpShiftRightArithmetic:
3802 case SpvOpShiftLeftLogical:
3803 case SpvOpLogicalEqual:
3804 case SpvOpLogicalNotEqual:
3805 case SpvOpLogicalOr:
3806 case SpvOpLogicalAnd:
3807 case SpvOpLogicalNot:
3808 case SpvOpBitwiseOr:
3809 case SpvOpBitwiseXor:
3810 case SpvOpBitwiseAnd:
3812 case SpvOpFOrdEqual:
3813 case SpvOpFUnordEqual:
3814 case SpvOpINotEqual:
3815 case SpvOpFOrdNotEqual:
3816 case SpvOpFUnordNotEqual:
3817 case SpvOpULessThan:
3818 case SpvOpSLessThan:
3819 case SpvOpFOrdLessThan:
3820 case SpvOpFUnordLessThan:
3821 case SpvOpUGreaterThan:
3822 case SpvOpSGreaterThan:
3823 case SpvOpFOrdGreaterThan:
3824 case SpvOpFUnordGreaterThan:
3825 case SpvOpULessThanEqual:
3826 case SpvOpSLessThanEqual:
3827 case SpvOpFOrdLessThanEqual:
3828 case SpvOpFUnordLessThanEqual:
3829 case SpvOpUGreaterThanEqual:
3830 case SpvOpSGreaterThanEqual:
3831 case SpvOpFOrdGreaterThanEqual:
3832 case SpvOpFUnordGreaterThanEqual:
3838 case SpvOpFwidthFine:
3839 case SpvOpDPdxCoarse:
3840 case SpvOpDPdyCoarse:
3841 case SpvOpFwidthCoarse:
3842 case SpvOpBitFieldInsert:
3843 case SpvOpBitFieldSExtract:
3844 case SpvOpBitFieldUExtract:
3845 case SpvOpBitReverse:
3847 case SpvOpTranspose:
3848 case SpvOpOuterProduct:
3849 case SpvOpMatrixTimesScalar:
3850 case SpvOpVectorTimesMatrix:
3851 case SpvOpMatrixTimesVector:
3852 case SpvOpMatrixTimesMatrix:
3853 vtn_handle_alu(b, opcode, w, count);
3856 case SpvOpVectorExtractDynamic:
3857 case SpvOpVectorInsertDynamic:
3858 case SpvOpVectorShuffle:
3859 case SpvOpCompositeConstruct:
3860 case SpvOpCompositeExtract:
3861 case SpvOpCompositeInsert:
3862 case SpvOpCopyObject:
3863 vtn_handle_composite(b, opcode, w, count);
3866 case SpvOpEmitVertex:
3867 case SpvOpEndPrimitive:
3868 case SpvOpEmitStreamVertex:
3869 case SpvOpEndStreamPrimitive:
3870 case SpvOpControlBarrier:
3871 case SpvOpMemoryBarrier:
3872 vtn_handle_barrier(b, opcode, w, count);
3876 vtn_fail("Unhandled opcode");
3883 spirv_to_nir(const uint32_t *words, size_t word_count,
3884 struct nir_spirv_specialization *spec, unsigned num_spec,
3885 gl_shader_stage stage, const char *entry_point_name,
3886 const struct spirv_to_nir_options *options,
3887 const nir_shader_compiler_options *nir_options)
3889 /* Initialize the stn_builder object */
3890 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3892 b->spirv_word_count = word_count;
3896 exec_list_make_empty(&b->functions);
3897 b->entry_point_stage = stage;
3898 b->entry_point_name = entry_point_name;
3899 b->options = options;
3901 /* See also _vtn_fail() */
3902 if (setjmp(b->fail_jump)) {
3907 const uint32_t *word_end = words + word_count;
3909 /* Handle the SPIR-V header (first 4 dwords) */
3910 vtn_assert(word_count > 5);
3912 vtn_assert(words[0] == SpvMagicNumber);
3913 vtn_assert(words[1] >= 0x10000);
3914 /* words[2] == generator magic */
3915 unsigned value_id_bound = words[3];
3916 vtn_assert(words[4] == 0);
3920 b->value_id_bound = value_id_bound;
3921 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3923 /* Handle all the preamble instructions */
3924 words = vtn_foreach_instruction(b, words, word_end,
3925 vtn_handle_preamble_instruction);
3927 if (b->entry_point == NULL) {
3928 vtn_fail("Entry point not found");
3933 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3935 /* Set shader info defaults */
3936 b->shader->info.gs.invocations = 1;
3938 /* Parse execution modes */
3939 vtn_foreach_execution_mode(b, b->entry_point,
3940 vtn_handle_execution_mode, NULL);
3942 b->specializations = spec;
3943 b->num_specializations = num_spec;
3945 /* Handle all variable, type, and constant instructions */
3946 words = vtn_foreach_instruction(b, words, word_end,
3947 vtn_handle_variable_or_type_instruction);
3949 /* Set types on all vtn_values */
3950 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
3952 vtn_build_cfg(b, words, word_end);
3954 assert(b->entry_point->value_type == vtn_value_type_function);
3955 b->entry_point->func->referenced = true;
3960 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3961 if (func->referenced && !func->emitted) {
3962 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3963 _mesa_key_pointer_equal);
3965 vtn_function_emit(b, func, vtn_handle_body_instruction);
3971 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
3972 nir_function *entry_point = b->entry_point->func->impl->function;
3973 vtn_assert(entry_point);
3975 /* Unparent the shader from the vtn_builder before we delete the builder */
3976 ralloc_steal(NULL, b->shader);