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 /* does a shallow copy of a vtn_type */
542 static struct vtn_type *
543 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
545 struct vtn_type *dest = ralloc(b, struct vtn_type);
548 switch (src->base_type) {
549 case vtn_base_type_void:
550 case vtn_base_type_scalar:
551 case vtn_base_type_vector:
552 case vtn_base_type_matrix:
553 case vtn_base_type_array:
554 case vtn_base_type_pointer:
555 case vtn_base_type_image:
556 case vtn_base_type_sampler:
557 case vtn_base_type_sampled_image:
558 /* Nothing more to do */
561 case vtn_base_type_struct:
562 dest->members = ralloc_array(b, struct vtn_type *, src->length);
563 memcpy(dest->members, src->members,
564 src->length * sizeof(src->members[0]));
566 dest->offsets = ralloc_array(b, unsigned, src->length);
567 memcpy(dest->offsets, src->offsets,
568 src->length * sizeof(src->offsets[0]));
571 case vtn_base_type_function:
572 dest->params = ralloc_array(b, struct vtn_type *, src->length);
573 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
580 static struct vtn_type *
581 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
583 type->members[member] = vtn_type_copy(b, type->members[member]);
584 type = type->members[member];
586 /* We may have an array of matrices.... Oh, joy! */
587 while (glsl_type_is_array(type->type)) {
588 type->array_element = vtn_type_copy(b, type->array_element);
589 type = type->array_element;
592 vtn_assert(glsl_type_is_matrix(type->type));
598 struct_member_decoration_cb(struct vtn_builder *b,
599 struct vtn_value *val, int member,
600 const struct vtn_decoration *dec, void *void_ctx)
602 struct member_decoration_ctx *ctx = void_ctx;
607 assert(member < ctx->num_fields);
609 switch (dec->decoration) {
610 case SpvDecorationNonWritable:
611 case SpvDecorationNonReadable:
612 case SpvDecorationRelaxedPrecision:
613 case SpvDecorationVolatile:
614 case SpvDecorationCoherent:
615 case SpvDecorationUniform:
616 break; /* FIXME: Do nothing with this for now. */
617 case SpvDecorationNoPerspective:
618 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
620 case SpvDecorationFlat:
621 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
623 case SpvDecorationCentroid:
624 ctx->fields[member].centroid = true;
626 case SpvDecorationSample:
627 ctx->fields[member].sample = true;
629 case SpvDecorationStream:
630 /* Vulkan only allows one GS stream */
631 vtn_assert(dec->literals[0] == 0);
633 case SpvDecorationLocation:
634 ctx->fields[member].location = dec->literals[0];
636 case SpvDecorationComponent:
637 break; /* FIXME: What should we do with these? */
638 case SpvDecorationBuiltIn:
639 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
640 ctx->type->members[member]->is_builtin = true;
641 ctx->type->members[member]->builtin = dec->literals[0];
642 ctx->type->builtin_block = true;
644 case SpvDecorationOffset:
645 ctx->type->offsets[member] = dec->literals[0];
647 case SpvDecorationMatrixStride:
648 /* Handled as a second pass */
650 case SpvDecorationColMajor:
651 break; /* Nothing to do here. Column-major is the default. */
652 case SpvDecorationRowMajor:
653 mutable_matrix_member(b, ctx->type, member)->row_major = true;
656 case SpvDecorationPatch:
659 case SpvDecorationSpecId:
660 case SpvDecorationBlock:
661 case SpvDecorationBufferBlock:
662 case SpvDecorationArrayStride:
663 case SpvDecorationGLSLShared:
664 case SpvDecorationGLSLPacked:
665 case SpvDecorationInvariant:
666 case SpvDecorationRestrict:
667 case SpvDecorationAliased:
668 case SpvDecorationConstant:
669 case SpvDecorationIndex:
670 case SpvDecorationBinding:
671 case SpvDecorationDescriptorSet:
672 case SpvDecorationLinkageAttributes:
673 case SpvDecorationNoContraction:
674 case SpvDecorationInputAttachmentIndex:
675 vtn_warn("Decoration not allowed on struct members: %s",
676 spirv_decoration_to_string(dec->decoration));
679 case SpvDecorationXfbBuffer:
680 case SpvDecorationXfbStride:
681 vtn_warn("Vulkan does not have transform feedback");
684 case SpvDecorationCPacked:
685 case SpvDecorationSaturatedConversion:
686 case SpvDecorationFuncParamAttr:
687 case SpvDecorationFPRoundingMode:
688 case SpvDecorationFPFastMathMode:
689 case SpvDecorationAlignment:
690 vtn_warn("Decoration only allowed for CL-style kernels: %s",
691 spirv_decoration_to_string(dec->decoration));
695 vtn_fail("Unhandled decoration");
699 /* Matrix strides are handled as a separate pass because we need to know
700 * whether the matrix is row-major or not first.
703 struct_member_matrix_stride_cb(struct vtn_builder *b,
704 struct vtn_value *val, int member,
705 const struct vtn_decoration *dec,
708 if (dec->decoration != SpvDecorationMatrixStride)
711 vtn_fail_if(member < 0,
712 "The MatrixStride decoration is only allowed on members "
715 struct member_decoration_ctx *ctx = void_ctx;
717 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
718 if (mat_type->row_major) {
719 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
720 mat_type->stride = mat_type->array_element->stride;
721 mat_type->array_element->stride = dec->literals[0];
723 vtn_assert(mat_type->array_element->stride > 0);
724 mat_type->stride = dec->literals[0];
729 type_decoration_cb(struct vtn_builder *b,
730 struct vtn_value *val, int member,
731 const struct vtn_decoration *dec, void *ctx)
733 struct vtn_type *type = val->type;
736 /* This should have been handled by OpTypeStruct */
737 assert(val->type->base_type == vtn_base_type_struct);
738 assert(member >= 0 && member < val->type->length);
742 switch (dec->decoration) {
743 case SpvDecorationArrayStride:
744 vtn_assert(type->base_type == vtn_base_type_matrix ||
745 type->base_type == vtn_base_type_array ||
746 type->base_type == vtn_base_type_pointer);
747 type->stride = dec->literals[0];
749 case SpvDecorationBlock:
750 vtn_assert(type->base_type == vtn_base_type_struct);
753 case SpvDecorationBufferBlock:
754 vtn_assert(type->base_type == vtn_base_type_struct);
755 type->buffer_block = true;
757 case SpvDecorationGLSLShared:
758 case SpvDecorationGLSLPacked:
759 /* Ignore these, since we get explicit offsets anyways */
762 case SpvDecorationRowMajor:
763 case SpvDecorationColMajor:
764 case SpvDecorationMatrixStride:
765 case SpvDecorationBuiltIn:
766 case SpvDecorationNoPerspective:
767 case SpvDecorationFlat:
768 case SpvDecorationPatch:
769 case SpvDecorationCentroid:
770 case SpvDecorationSample:
771 case SpvDecorationVolatile:
772 case SpvDecorationCoherent:
773 case SpvDecorationNonWritable:
774 case SpvDecorationNonReadable:
775 case SpvDecorationUniform:
776 case SpvDecorationStream:
777 case SpvDecorationLocation:
778 case SpvDecorationComponent:
779 case SpvDecorationOffset:
780 case SpvDecorationXfbBuffer:
781 case SpvDecorationXfbStride:
782 vtn_warn("Decoration only allowed for struct members: %s",
783 spirv_decoration_to_string(dec->decoration));
786 case SpvDecorationRelaxedPrecision:
787 case SpvDecorationSpecId:
788 case SpvDecorationInvariant:
789 case SpvDecorationRestrict:
790 case SpvDecorationAliased:
791 case SpvDecorationConstant:
792 case SpvDecorationIndex:
793 case SpvDecorationBinding:
794 case SpvDecorationDescriptorSet:
795 case SpvDecorationLinkageAttributes:
796 case SpvDecorationNoContraction:
797 case SpvDecorationInputAttachmentIndex:
798 vtn_warn("Decoration not allowed on types: %s",
799 spirv_decoration_to_string(dec->decoration));
802 case SpvDecorationCPacked:
803 case SpvDecorationSaturatedConversion:
804 case SpvDecorationFuncParamAttr:
805 case SpvDecorationFPRoundingMode:
806 case SpvDecorationFPFastMathMode:
807 case SpvDecorationAlignment:
808 vtn_warn("Decoration only allowed for CL-style kernels: %s",
809 spirv_decoration_to_string(dec->decoration));
813 vtn_fail("Unhandled decoration");
818 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
821 case SpvImageFormatUnknown: return 0; /* GL_NONE */
822 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
823 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
824 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
825 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
826 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
827 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
828 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
829 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
830 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
831 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
832 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
833 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
834 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
835 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
836 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
837 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
838 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
839 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
840 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
841 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
842 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
843 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
844 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
845 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
846 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
847 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
848 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
849 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
850 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
851 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
852 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
853 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
854 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
855 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
856 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
857 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
858 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
859 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
860 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
862 vtn_fail("Invalid image format");
866 static struct vtn_type *
867 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
868 uint32_t *size_out, uint32_t *align_out)
870 switch (type->base_type) {
871 case vtn_base_type_scalar: {
872 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
873 *size_out = comp_size;
874 *align_out = comp_size;
878 case vtn_base_type_vector: {
879 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
880 assert(type->length > 0 && type->length <= 4);
881 unsigned align_comps = type->length == 3 ? 4 : type->length;
882 *size_out = comp_size * type->length,
883 *align_out = comp_size * align_comps;
887 case vtn_base_type_matrix:
888 case vtn_base_type_array: {
889 /* We're going to add an array stride */
890 type = vtn_type_copy(b, type);
891 uint32_t elem_size, elem_align;
892 type->array_element = vtn_type_layout_std430(b, type->array_element,
893 &elem_size, &elem_align);
894 type->stride = vtn_align_u32(elem_size, elem_align);
895 *size_out = type->stride * type->length;
896 *align_out = elem_align;
900 case vtn_base_type_struct: {
901 /* We're going to add member offsets */
902 type = vtn_type_copy(b, type);
905 for (unsigned i = 0; i < type->length; i++) {
906 uint32_t mem_size, mem_align;
907 type->members[i] = vtn_type_layout_std430(b, type->members[i],
908 &mem_size, &mem_align);
909 offset = vtn_align_u32(offset, mem_align);
910 type->offsets[i] = offset;
912 align = MAX2(align, mem_align);
920 unreachable("Invalid SPIR-V type for std430");
925 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
926 const uint32_t *w, unsigned count)
928 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
930 val->type = rzalloc(b, struct vtn_type);
931 val->type->id = w[1];
935 val->type->base_type = vtn_base_type_void;
936 val->type->type = glsl_void_type();
939 val->type->base_type = vtn_base_type_scalar;
940 val->type->type = glsl_bool_type();
941 val->type->length = 1;
945 const bool signedness = w[3];
946 val->type->base_type = vtn_base_type_scalar;
949 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
952 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
955 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
958 vtn_fail("Invalid int bit size");
960 val->type->length = 1;
964 case SpvOpTypeFloat: {
966 val->type->base_type = vtn_base_type_scalar;
969 val->type->type = glsl_float16_t_type();
972 val->type->type = glsl_float_type();
975 val->type->type = glsl_double_type();
978 vtn_fail("Invalid float bit size");
980 val->type->length = 1;
984 case SpvOpTypeVector: {
985 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
986 unsigned elems = w[3];
988 vtn_fail_if(base->base_type != vtn_base_type_scalar,
989 "Base type for OpTypeVector must be a scalar");
990 vtn_fail_if(elems < 2 || elems > 4,
991 "Invalid component count for OpTypeVector");
993 val->type->base_type = vtn_base_type_vector;
994 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
995 val->type->length = elems;
996 val->type->stride = glsl_get_bit_size(base->type) / 8;
997 val->type->array_element = base;
1001 case SpvOpTypeMatrix: {
1002 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1003 unsigned columns = w[3];
1005 vtn_fail_if(base->base_type != vtn_base_type_vector,
1006 "Base type for OpTypeMatrix must be a vector");
1007 vtn_fail_if(columns < 2 || columns > 4,
1008 "Invalid column count for OpTypeMatrix");
1010 val->type->base_type = vtn_base_type_matrix;
1011 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
1012 glsl_get_vector_elements(base->type),
1014 vtn_fail_if(glsl_type_is_error(val->type->type),
1015 "Unsupported base type for OpTypeMatrix");
1016 assert(!glsl_type_is_error(val->type->type));
1017 val->type->length = columns;
1018 val->type->array_element = base;
1019 val->type->row_major = false;
1020 val->type->stride = 0;
1024 case SpvOpTypeRuntimeArray:
1025 case SpvOpTypeArray: {
1026 struct vtn_type *array_element =
1027 vtn_value(b, w[2], vtn_value_type_type)->type;
1029 if (opcode == SpvOpTypeRuntimeArray) {
1030 /* A length of 0 is used to denote unsized arrays */
1031 val->type->length = 0;
1034 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
1037 val->type->base_type = vtn_base_type_array;
1038 val->type->type = glsl_array_type(array_element->type, val->type->length);
1039 val->type->array_element = array_element;
1040 val->type->stride = 0;
1044 case SpvOpTypeStruct: {
1045 unsigned num_fields = count - 2;
1046 val->type->base_type = vtn_base_type_struct;
1047 val->type->length = num_fields;
1048 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
1049 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1051 NIR_VLA(struct glsl_struct_field, fields, count);
1052 for (unsigned i = 0; i < num_fields; i++) {
1053 val->type->members[i] =
1054 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1055 fields[i] = (struct glsl_struct_field) {
1056 .type = val->type->members[i]->type,
1057 .name = ralloc_asprintf(b, "field%d", i),
1062 struct member_decoration_ctx ctx = {
1063 .num_fields = num_fields,
1068 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1069 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1071 const char *name = val->name ? val->name : "struct";
1073 val->type->type = glsl_struct_type(fields, num_fields, name);
1077 case SpvOpTypeFunction: {
1078 val->type->base_type = vtn_base_type_function;
1079 val->type->type = NULL;
1081 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1083 const unsigned num_params = count - 3;
1084 val->type->length = num_params;
1085 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1086 for (unsigned i = 0; i < count - 3; i++) {
1087 val->type->params[i] =
1088 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1093 case SpvOpTypePointer: {
1094 SpvStorageClass storage_class = w[2];
1095 struct vtn_type *deref_type =
1096 vtn_value(b, w[3], vtn_value_type_type)->type;
1098 val->type->base_type = vtn_base_type_pointer;
1099 val->type->storage_class = storage_class;
1100 val->type->deref = deref_type;
1102 if (storage_class == SpvStorageClassUniform ||
1103 storage_class == SpvStorageClassStorageBuffer) {
1104 /* These can actually be stored to nir_variables and used as SSA
1105 * values so they need a real glsl_type.
1107 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1110 if (storage_class == SpvStorageClassWorkgroup &&
1111 b->options->lower_workgroup_access_to_offsets) {
1112 uint32_t size, align;
1113 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1115 val->type->length = size;
1116 val->type->align = align;
1117 /* These can actually be stored to nir_variables and used as SSA
1118 * values so they need a real glsl_type.
1120 val->type->type = glsl_uint_type();
1125 case SpvOpTypeImage: {
1126 val->type->base_type = vtn_base_type_image;
1128 const struct vtn_type *sampled_type =
1129 vtn_value(b, w[2], vtn_value_type_type)->type;
1131 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1132 glsl_get_bit_size(sampled_type->type) != 32,
1133 "Sampled type of OpTypeImage must be a 32-bit scalar");
1135 enum glsl_sampler_dim dim;
1136 switch ((SpvDim)w[3]) {
1137 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1138 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1139 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1140 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1141 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1142 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1143 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1145 vtn_fail("Invalid SPIR-V image dimensionality");
1148 bool is_shadow = w[4];
1149 bool is_array = w[5];
1150 bool multisampled = w[6];
1151 unsigned sampled = w[7];
1152 SpvImageFormat format = w[8];
1155 val->type->access_qualifier = w[9];
1157 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1160 if (dim == GLSL_SAMPLER_DIM_2D)
1161 dim = GLSL_SAMPLER_DIM_MS;
1162 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1163 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1165 vtn_fail("Unsupported multisampled image type");
1168 val->type->image_format = translate_image_format(b, format);
1170 enum glsl_base_type sampled_base_type =
1171 glsl_get_base_type(sampled_type->type);
1173 val->type->sampled = true;
1174 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1176 } else if (sampled == 2) {
1177 vtn_assert(!is_shadow);
1178 val->type->sampled = false;
1179 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1181 vtn_fail("We need to know if the image will be sampled");
1186 case SpvOpTypeSampledImage:
1187 val->type->base_type = vtn_base_type_sampled_image;
1188 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1189 val->type->type = val->type->image->type;
1192 case SpvOpTypeSampler:
1193 /* The actual sampler type here doesn't really matter. It gets
1194 * thrown away the moment you combine it with an image. What really
1195 * matters is that it's a sampler type as opposed to an integer type
1196 * so the backend knows what to do.
1198 val->type->base_type = vtn_base_type_sampler;
1199 val->type->type = glsl_bare_sampler_type();
1202 case SpvOpTypeOpaque:
1203 case SpvOpTypeEvent:
1204 case SpvOpTypeDeviceEvent:
1205 case SpvOpTypeReserveId:
1206 case SpvOpTypeQueue:
1209 vtn_fail("Unhandled opcode");
1212 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1215 static nir_constant *
1216 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1218 nir_constant *c = rzalloc(b, nir_constant);
1220 /* For pointers and other typeless things, we have to return something but
1221 * it doesn't matter what.
1226 switch (glsl_get_base_type(type)) {
1228 case GLSL_TYPE_UINT:
1229 case GLSL_TYPE_INT16:
1230 case GLSL_TYPE_UINT16:
1231 case GLSL_TYPE_INT64:
1232 case GLSL_TYPE_UINT64:
1233 case GLSL_TYPE_BOOL:
1234 case GLSL_TYPE_FLOAT:
1235 case GLSL_TYPE_FLOAT16:
1236 case GLSL_TYPE_DOUBLE:
1237 /* Nothing to do here. It's already initialized to zero */
1240 case GLSL_TYPE_ARRAY:
1241 vtn_assert(glsl_get_length(type) > 0);
1242 c->num_elements = glsl_get_length(type);
1243 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1245 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1246 for (unsigned i = 1; i < c->num_elements; i++)
1247 c->elements[i] = c->elements[0];
1250 case GLSL_TYPE_STRUCT:
1251 c->num_elements = glsl_get_length(type);
1252 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1254 for (unsigned i = 0; i < c->num_elements; i++) {
1255 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1260 vtn_fail("Invalid type for null constant");
1267 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1268 int member, const struct vtn_decoration *dec,
1271 vtn_assert(member == -1);
1272 if (dec->decoration != SpvDecorationSpecId)
1275 struct spec_constant_value *const_value = data;
1277 for (unsigned i = 0; i < b->num_specializations; i++) {
1278 if (b->specializations[i].id == dec->literals[0]) {
1279 if (const_value->is_double)
1280 const_value->data64 = b->specializations[i].data64;
1282 const_value->data32 = b->specializations[i].data32;
1289 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1290 uint32_t const_value)
1292 struct spec_constant_value data;
1293 data.is_double = false;
1294 data.data32 = const_value;
1295 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1300 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1301 uint64_t const_value)
1303 struct spec_constant_value data;
1304 data.is_double = true;
1305 data.data64 = const_value;
1306 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1311 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1312 struct vtn_value *val,
1314 const struct vtn_decoration *dec,
1317 vtn_assert(member == -1);
1318 if (dec->decoration != SpvDecorationBuiltIn ||
1319 dec->literals[0] != SpvBuiltInWorkgroupSize)
1322 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1324 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1325 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1326 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1330 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1331 const uint32_t *w, unsigned count)
1333 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1334 val->constant = rzalloc(b, nir_constant);
1336 case SpvOpConstantTrue:
1337 case SpvOpConstantFalse:
1338 case SpvOpSpecConstantTrue:
1339 case SpvOpSpecConstantFalse: {
1340 vtn_fail_if(val->type->type != glsl_bool_type(),
1341 "Result type of %s must be OpTypeBool",
1342 spirv_op_to_string(opcode));
1344 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1345 opcode == SpvOpSpecConstantTrue);
1347 if (opcode == SpvOpSpecConstantTrue ||
1348 opcode == SpvOpSpecConstantFalse)
1349 int_val = get_specialization(b, val, int_val);
1351 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1355 case SpvOpConstant: {
1356 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1357 "Result type of %s must be a scalar",
1358 spirv_op_to_string(opcode));
1359 int bit_size = glsl_get_bit_size(val->type->type);
1362 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1365 val->constant->values->u32[0] = w[3];
1368 val->constant->values->u16[0] = w[3];
1371 vtn_fail("Unsupported SpvOpConstant bit size");
1376 case SpvOpSpecConstant: {
1377 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1378 "Result type of %s must be a scalar",
1379 spirv_op_to_string(opcode));
1380 int bit_size = glsl_get_bit_size(val->type->type);
1383 val->constant->values[0].u64[0] =
1384 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1387 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1390 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1393 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1398 case SpvOpSpecConstantComposite:
1399 case SpvOpConstantComposite: {
1400 unsigned elem_count = count - 3;
1401 vtn_fail_if(elem_count != val->type->length,
1402 "%s has %u constituents, expected %u",
1403 spirv_op_to_string(opcode), elem_count, val->type->length);
1405 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1406 for (unsigned i = 0; i < elem_count; i++)
1407 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1409 switch (val->type->base_type) {
1410 case vtn_base_type_vector: {
1411 assert(glsl_type_is_vector(val->type->type));
1412 int bit_size = glsl_get_bit_size(val->type->type);
1413 for (unsigned i = 0; i < elem_count; i++) {
1416 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1419 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1422 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1425 vtn_fail("Invalid SpvOpConstantComposite bit size");
1431 case vtn_base_type_matrix:
1432 assert(glsl_type_is_matrix(val->type->type));
1433 for (unsigned i = 0; i < elem_count; i++)
1434 val->constant->values[i] = elems[i]->values[0];
1437 case vtn_base_type_struct:
1438 case vtn_base_type_array:
1439 ralloc_steal(val->constant, elems);
1440 val->constant->num_elements = elem_count;
1441 val->constant->elements = elems;
1445 vtn_fail("Result type of %s must be a composite type",
1446 spirv_op_to_string(opcode));
1451 case SpvOpSpecConstantOp: {
1452 SpvOp opcode = get_specialization(b, val, w[3]);
1454 case SpvOpVectorShuffle: {
1455 struct vtn_value *v0 = &b->values[w[4]];
1456 struct vtn_value *v1 = &b->values[w[5]];
1458 vtn_assert(v0->value_type == vtn_value_type_constant ||
1459 v0->value_type == vtn_value_type_undef);
1460 vtn_assert(v1->value_type == vtn_value_type_constant ||
1461 v1->value_type == vtn_value_type_undef);
1463 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1464 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1466 vtn_assert(len0 + len1 < 16);
1468 unsigned bit_size = glsl_get_bit_size(val->type->type);
1469 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1470 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1472 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1473 (void)bit_size0; (void)bit_size1;
1475 if (bit_size == 64) {
1477 if (v0->value_type == vtn_value_type_constant) {
1478 for (unsigned i = 0; i < len0; i++)
1479 u64[i] = v0->constant->values[0].u64[i];
1481 if (v1->value_type == vtn_value_type_constant) {
1482 for (unsigned i = 0; i < len1; i++)
1483 u64[len0 + i] = v1->constant->values[0].u64[i];
1486 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1487 uint32_t comp = w[i + 6];
1488 /* If component is not used, set the value to a known constant
1489 * to detect if it is wrongly used.
1491 if (comp == (uint32_t)-1)
1492 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1494 val->constant->values[0].u64[j] = u64[comp];
1497 /* This is for both 32-bit and 16-bit values */
1499 if (v0->value_type == vtn_value_type_constant) {
1500 for (unsigned i = 0; i < len0; i++)
1501 u32[i] = v0->constant->values[0].u32[i];
1503 if (v1->value_type == vtn_value_type_constant) {
1504 for (unsigned i = 0; i < len1; i++)
1505 u32[len0 + i] = v1->constant->values[0].u32[i];
1508 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1509 uint32_t comp = w[i + 6];
1510 /* If component is not used, set the value to a known constant
1511 * to detect if it is wrongly used.
1513 if (comp == (uint32_t)-1)
1514 val->constant->values[0].u32[j] = 0xdeadbeef;
1516 val->constant->values[0].u32[j] = u32[comp];
1522 case SpvOpCompositeExtract:
1523 case SpvOpCompositeInsert: {
1524 struct vtn_value *comp;
1525 unsigned deref_start;
1526 struct nir_constant **c;
1527 if (opcode == SpvOpCompositeExtract) {
1528 comp = vtn_value(b, w[4], vtn_value_type_constant);
1530 c = &comp->constant;
1532 comp = vtn_value(b, w[5], vtn_value_type_constant);
1534 val->constant = nir_constant_clone(comp->constant,
1541 const struct vtn_type *type = comp->type;
1542 for (unsigned i = deref_start; i < count; i++) {
1543 vtn_fail_if(w[i] > type->length,
1544 "%uth index of %s is %u but the type has only "
1545 "%u elements", i - deref_start,
1546 spirv_op_to_string(opcode), w[i], type->length);
1548 switch (type->base_type) {
1549 case vtn_base_type_vector:
1551 type = type->array_element;
1554 case vtn_base_type_matrix:
1555 assert(col == 0 && elem == -1);
1558 type = type->array_element;
1561 case vtn_base_type_array:
1562 c = &(*c)->elements[w[i]];
1563 type = type->array_element;
1566 case vtn_base_type_struct:
1567 c = &(*c)->elements[w[i]];
1568 type = type->members[w[i]];
1572 vtn_fail("%s must only index into composite types",
1573 spirv_op_to_string(opcode));
1577 if (opcode == SpvOpCompositeExtract) {
1581 unsigned num_components = type->length;
1582 unsigned bit_size = glsl_get_bit_size(type->type);
1583 for (unsigned i = 0; i < num_components; i++)
1586 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1589 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1592 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1595 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1599 struct vtn_value *insert =
1600 vtn_value(b, w[4], vtn_value_type_constant);
1601 vtn_assert(insert->type == type);
1603 *c = insert->constant;
1605 unsigned num_components = type->length;
1606 unsigned bit_size = glsl_get_bit_size(type->type);
1607 for (unsigned i = 0; i < num_components; i++)
1610 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1613 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1616 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1619 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1628 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1629 nir_alu_type src_alu_type = dst_alu_type;
1630 unsigned num_components = glsl_get_vector_elements(val->type->type);
1633 vtn_assert(count <= 7);
1638 /* We have a source in a conversion */
1640 nir_get_nir_type_for_glsl_type(
1641 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1642 /* We use the bitsize of the conversion source to evaluate the opcode later */
1643 bit_size = glsl_get_bit_size(
1644 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1647 bit_size = glsl_get_bit_size(val->type->type);
1650 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1653 nir_const_value src[4];
1655 for (unsigned i = 0; i < count - 4; i++) {
1657 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1659 unsigned j = swap ? 1 - i : i;
1660 src[j] = c->values[0];
1663 val->constant->values[0] =
1664 nir_eval_const_opcode(op, num_components, bit_size, src);
1671 case SpvOpConstantNull:
1672 val->constant = vtn_null_constant(b, val->type->type);
1675 case SpvOpConstantSampler:
1676 vtn_fail("OpConstantSampler requires Kernel Capability");
1680 vtn_fail("Unhandled opcode");
1683 /* Now that we have the value, update the workgroup size if needed */
1684 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1688 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1689 const uint32_t *w, unsigned count)
1691 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1692 struct vtn_function *vtn_callee =
1693 vtn_value(b, w[3], vtn_value_type_function)->func;
1694 struct nir_function *callee = vtn_callee->impl->function;
1696 vtn_callee->referenced = true;
1698 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1699 for (unsigned i = 0; i < call->num_params; i++) {
1700 unsigned arg_id = w[4 + i];
1701 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1702 if (arg->value_type == vtn_value_type_pointer &&
1703 arg->pointer->ptr_type->type == NULL) {
1704 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1705 call->params[i] = nir_deref_var_clone(d, call);
1707 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1709 /* Make a temporary to store the argument in */
1711 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1712 call->params[i] = nir_deref_var_create(call, tmp);
1714 vtn_local_store(b, arg_ssa, call->params[i]);
1718 nir_variable *out_tmp = NULL;
1719 vtn_assert(res_type->type == callee->return_type);
1720 if (!glsl_type_is_void(callee->return_type)) {
1721 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1723 call->return_deref = nir_deref_var_create(call, out_tmp);
1726 nir_builder_instr_insert(&b->nb, &call->instr);
1728 if (glsl_type_is_void(callee->return_type)) {
1729 vtn_push_value(b, w[2], vtn_value_type_undef);
1731 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1735 struct vtn_ssa_value *
1736 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1738 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1741 if (!glsl_type_is_vector_or_scalar(type)) {
1742 unsigned elems = glsl_get_length(type);
1743 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1744 for (unsigned i = 0; i < elems; i++) {
1745 const struct glsl_type *child_type;
1747 switch (glsl_get_base_type(type)) {
1749 case GLSL_TYPE_UINT:
1750 case GLSL_TYPE_INT16:
1751 case GLSL_TYPE_UINT16:
1752 case GLSL_TYPE_INT64:
1753 case GLSL_TYPE_UINT64:
1754 case GLSL_TYPE_BOOL:
1755 case GLSL_TYPE_FLOAT:
1756 case GLSL_TYPE_FLOAT16:
1757 case GLSL_TYPE_DOUBLE:
1758 child_type = glsl_get_column_type(type);
1760 case GLSL_TYPE_ARRAY:
1761 child_type = glsl_get_array_element(type);
1763 case GLSL_TYPE_STRUCT:
1764 child_type = glsl_get_struct_field(type, i);
1767 vtn_fail("unkown base type");
1770 val->elems[i] = vtn_create_ssa_value(b, child_type);
1778 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1781 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1782 src.src_type = type;
1787 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1788 const uint32_t *w, unsigned count)
1790 if (opcode == SpvOpSampledImage) {
1791 struct vtn_value *val =
1792 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1793 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1794 val->sampled_image->type =
1795 vtn_value(b, w[1], vtn_value_type_type)->type;
1796 val->sampled_image->image =
1797 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1798 val->sampled_image->sampler =
1799 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1801 } else if (opcode == SpvOpImage) {
1802 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1803 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1804 if (src_val->value_type == vtn_value_type_sampled_image) {
1805 val->pointer = src_val->sampled_image->image;
1807 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1808 val->pointer = src_val->pointer;
1813 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1814 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1816 struct vtn_sampled_image sampled;
1817 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1818 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1819 sampled = *sampled_val->sampled_image;
1821 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1822 sampled.type = sampled_val->pointer->type;
1823 sampled.image = NULL;
1824 sampled.sampler = sampled_val->pointer;
1827 const struct glsl_type *image_type = sampled.type->type;
1828 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1829 const bool is_array = glsl_sampler_type_is_array(image_type);
1830 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1832 /* Figure out the base texture operation */
1835 case SpvOpImageSampleImplicitLod:
1836 case SpvOpImageSampleDrefImplicitLod:
1837 case SpvOpImageSampleProjImplicitLod:
1838 case SpvOpImageSampleProjDrefImplicitLod:
1839 texop = nir_texop_tex;
1842 case SpvOpImageSampleExplicitLod:
1843 case SpvOpImageSampleDrefExplicitLod:
1844 case SpvOpImageSampleProjExplicitLod:
1845 case SpvOpImageSampleProjDrefExplicitLod:
1846 texop = nir_texop_txl;
1849 case SpvOpImageFetch:
1850 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1851 texop = nir_texop_txf_ms;
1853 texop = nir_texop_txf;
1857 case SpvOpImageGather:
1858 case SpvOpImageDrefGather:
1859 texop = nir_texop_tg4;
1862 case SpvOpImageQuerySizeLod:
1863 case SpvOpImageQuerySize:
1864 texop = nir_texop_txs;
1867 case SpvOpImageQueryLod:
1868 texop = nir_texop_lod;
1871 case SpvOpImageQueryLevels:
1872 texop = nir_texop_query_levels;
1875 case SpvOpImageQuerySamples:
1876 texop = nir_texop_texture_samples;
1880 vtn_fail("Unhandled opcode");
1883 nir_tex_src srcs[8]; /* 8 should be enough */
1884 nir_tex_src *p = srcs;
1888 struct nir_ssa_def *coord;
1889 unsigned coord_components;
1891 case SpvOpImageSampleImplicitLod:
1892 case SpvOpImageSampleExplicitLod:
1893 case SpvOpImageSampleDrefImplicitLod:
1894 case SpvOpImageSampleDrefExplicitLod:
1895 case SpvOpImageSampleProjImplicitLod:
1896 case SpvOpImageSampleProjExplicitLod:
1897 case SpvOpImageSampleProjDrefImplicitLod:
1898 case SpvOpImageSampleProjDrefExplicitLod:
1899 case SpvOpImageFetch:
1900 case SpvOpImageGather:
1901 case SpvOpImageDrefGather:
1902 case SpvOpImageQueryLod: {
1903 /* All these types have the coordinate as their first real argument */
1904 switch (sampler_dim) {
1905 case GLSL_SAMPLER_DIM_1D:
1906 case GLSL_SAMPLER_DIM_BUF:
1907 coord_components = 1;
1909 case GLSL_SAMPLER_DIM_2D:
1910 case GLSL_SAMPLER_DIM_RECT:
1911 case GLSL_SAMPLER_DIM_MS:
1912 coord_components = 2;
1914 case GLSL_SAMPLER_DIM_3D:
1915 case GLSL_SAMPLER_DIM_CUBE:
1916 coord_components = 3;
1919 vtn_fail("Invalid sampler type");
1922 if (is_array && texop != nir_texop_lod)
1925 coord = vtn_ssa_value(b, w[idx++])->def;
1926 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1927 (1 << coord_components) - 1));
1928 p->src_type = nir_tex_src_coord;
1935 coord_components = 0;
1940 case SpvOpImageSampleProjImplicitLod:
1941 case SpvOpImageSampleProjExplicitLod:
1942 case SpvOpImageSampleProjDrefImplicitLod:
1943 case SpvOpImageSampleProjDrefExplicitLod:
1944 /* These have the projector as the last coordinate component */
1945 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1946 p->src_type = nir_tex_src_projector;
1954 unsigned gather_component = 0;
1956 case SpvOpImageSampleDrefImplicitLod:
1957 case SpvOpImageSampleDrefExplicitLod:
1958 case SpvOpImageSampleProjDrefImplicitLod:
1959 case SpvOpImageSampleProjDrefExplicitLod:
1960 case SpvOpImageDrefGather:
1961 /* These all have an explicit depth value as their next source */
1962 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
1965 case SpvOpImageGather:
1966 /* This has a component as its next source */
1968 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
1975 /* For OpImageQuerySizeLod, we always have an LOD */
1976 if (opcode == SpvOpImageQuerySizeLod)
1977 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1979 /* Now we need to handle some number of optional arguments */
1980 const struct vtn_ssa_value *gather_offsets = NULL;
1982 uint32_t operands = w[idx++];
1984 if (operands & SpvImageOperandsBiasMask) {
1985 vtn_assert(texop == nir_texop_tex);
1986 texop = nir_texop_txb;
1987 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1990 if (operands & SpvImageOperandsLodMask) {
1991 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
1992 texop == nir_texop_txs);
1993 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1996 if (operands & SpvImageOperandsGradMask) {
1997 vtn_assert(texop == nir_texop_txl);
1998 texop = nir_texop_txd;
1999 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
2000 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
2003 if (operands & SpvImageOperandsOffsetMask ||
2004 operands & SpvImageOperandsConstOffsetMask)
2005 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
2007 if (operands & SpvImageOperandsConstOffsetsMask) {
2008 gather_offsets = vtn_ssa_value(b, w[idx++]);
2009 (*p++) = (nir_tex_src){};
2012 if (operands & SpvImageOperandsSampleMask) {
2013 vtn_assert(texop == nir_texop_txf_ms);
2014 texop = nir_texop_txf_ms;
2015 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
2018 /* We should have now consumed exactly all of the arguments */
2019 vtn_assert(idx == count);
2021 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
2024 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
2026 instr->coord_components = coord_components;
2027 instr->sampler_dim = sampler_dim;
2028 instr->is_array = is_array;
2029 instr->is_shadow = is_shadow;
2030 instr->is_new_style_shadow =
2031 is_shadow && glsl_get_components(ret_type->type) == 1;
2032 instr->component = gather_component;
2034 switch (glsl_get_sampler_result_type(image_type)) {
2035 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
2036 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
2037 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
2038 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
2040 vtn_fail("Invalid base type for sampler result");
2043 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
2044 nir_deref_var *texture;
2045 if (sampled.image) {
2046 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
2052 instr->texture = nir_deref_var_clone(texture, instr);
2054 switch (instr->op) {
2060 /* These operations require a sampler */
2061 instr->sampler = nir_deref_var_clone(sampler, instr);
2064 case nir_texop_txf_ms:
2067 case nir_texop_query_levels:
2068 case nir_texop_texture_samples:
2069 case nir_texop_samples_identical:
2071 instr->sampler = NULL;
2073 case nir_texop_txf_ms_mcs:
2074 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2077 nir_ssa_dest_init(&instr->instr, &instr->dest,
2078 nir_tex_instr_dest_size(instr), 32, NULL);
2080 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2081 nir_tex_instr_dest_size(instr));
2084 nir_instr *instruction;
2085 if (gather_offsets) {
2086 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2087 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2088 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2090 /* Copy the current instruction 4x */
2091 for (uint32_t i = 1; i < 4; i++) {
2092 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2093 instrs[i]->op = instr->op;
2094 instrs[i]->coord_components = instr->coord_components;
2095 instrs[i]->sampler_dim = instr->sampler_dim;
2096 instrs[i]->is_array = instr->is_array;
2097 instrs[i]->is_shadow = instr->is_shadow;
2098 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2099 instrs[i]->component = instr->component;
2100 instrs[i]->dest_type = instr->dest_type;
2101 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
2102 instrs[i]->sampler = NULL;
2104 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2106 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2107 nir_tex_instr_dest_size(instr), 32, NULL);
2110 /* Fill in the last argument with the offset from the passed in offsets
2111 * and insert the instruction into the stream.
2113 for (uint32_t i = 0; i < 4; i++) {
2115 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2116 src.src_type = nir_tex_src_offset;
2117 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2118 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2121 /* Combine the results of the 4 instructions by taking their .w
2124 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2125 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2126 vec4->dest.write_mask = 0xf;
2127 for (uint32_t i = 0; i < 4; i++) {
2128 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2129 vec4->src[i].swizzle[0] = 3;
2131 def = &vec4->dest.dest.ssa;
2132 instruction = &vec4->instr;
2134 def = &instr->dest.ssa;
2135 instruction = &instr->instr;
2138 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2139 val->ssa->def = def;
2141 nir_builder_instr_insert(&b->nb, instruction);
2145 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2146 const uint32_t *w, nir_src *src)
2149 case SpvOpAtomicIIncrement:
2150 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2153 case SpvOpAtomicIDecrement:
2154 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2157 case SpvOpAtomicISub:
2159 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2162 case SpvOpAtomicCompareExchange:
2163 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2164 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2167 case SpvOpAtomicExchange:
2168 case SpvOpAtomicIAdd:
2169 case SpvOpAtomicSMin:
2170 case SpvOpAtomicUMin:
2171 case SpvOpAtomicSMax:
2172 case SpvOpAtomicUMax:
2173 case SpvOpAtomicAnd:
2175 case SpvOpAtomicXor:
2176 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2180 vtn_fail("Invalid SPIR-V atomic");
2184 static nir_ssa_def *
2185 get_image_coord(struct vtn_builder *b, uint32_t value)
2187 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2189 /* The image_load_store intrinsics assume a 4-dim coordinate */
2190 unsigned dim = glsl_get_vector_elements(coord->type);
2191 unsigned swizzle[4];
2192 for (unsigned i = 0; i < 4; i++)
2193 swizzle[i] = MIN2(i, dim - 1);
2195 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2199 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2200 const uint32_t *w, unsigned count)
2202 /* Just get this one out of the way */
2203 if (opcode == SpvOpImageTexelPointer) {
2204 struct vtn_value *val =
2205 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2206 val->image = ralloc(b, struct vtn_image_pointer);
2208 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2209 val->image->coord = get_image_coord(b, w[4]);
2210 val->image->sample = vtn_ssa_value(b, w[5])->def;
2214 struct vtn_image_pointer image;
2217 case SpvOpAtomicExchange:
2218 case SpvOpAtomicCompareExchange:
2219 case SpvOpAtomicCompareExchangeWeak:
2220 case SpvOpAtomicIIncrement:
2221 case SpvOpAtomicIDecrement:
2222 case SpvOpAtomicIAdd:
2223 case SpvOpAtomicISub:
2224 case SpvOpAtomicLoad:
2225 case SpvOpAtomicSMin:
2226 case SpvOpAtomicUMin:
2227 case SpvOpAtomicSMax:
2228 case SpvOpAtomicUMax:
2229 case SpvOpAtomicAnd:
2231 case SpvOpAtomicXor:
2232 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2235 case SpvOpAtomicStore:
2236 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2239 case SpvOpImageQuerySize:
2240 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2242 image.sample = NULL;
2245 case SpvOpImageRead:
2246 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2247 image.coord = get_image_coord(b, w[4]);
2249 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2250 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2251 image.sample = vtn_ssa_value(b, w[6])->def;
2253 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2257 case SpvOpImageWrite:
2258 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2259 image.coord = get_image_coord(b, w[2]);
2263 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2264 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2265 image.sample = vtn_ssa_value(b, w[5])->def;
2267 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2272 vtn_fail("Invalid image opcode");
2275 nir_intrinsic_op op;
2277 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2278 OP(ImageQuerySize, size)
2280 OP(ImageWrite, store)
2281 OP(AtomicLoad, load)
2282 OP(AtomicStore, store)
2283 OP(AtomicExchange, atomic_exchange)
2284 OP(AtomicCompareExchange, atomic_comp_swap)
2285 OP(AtomicIIncrement, atomic_add)
2286 OP(AtomicIDecrement, atomic_add)
2287 OP(AtomicIAdd, atomic_add)
2288 OP(AtomicISub, atomic_add)
2289 OP(AtomicSMin, atomic_min)
2290 OP(AtomicUMin, atomic_min)
2291 OP(AtomicSMax, atomic_max)
2292 OP(AtomicUMax, atomic_max)
2293 OP(AtomicAnd, atomic_and)
2294 OP(AtomicOr, atomic_or)
2295 OP(AtomicXor, atomic_xor)
2298 vtn_fail("Invalid image opcode");
2301 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2303 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2304 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2306 /* ImageQuerySize doesn't take any extra parameters */
2307 if (opcode != SpvOpImageQuerySize) {
2308 /* The image coordinate is always 4 components but we may not have that
2309 * many. Swizzle to compensate.
2312 for (unsigned i = 0; i < 4; i++)
2313 swiz[i] = i < image.coord->num_components ? i : 0;
2314 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2316 intrin->src[1] = nir_src_for_ssa(image.sample);
2320 case SpvOpAtomicLoad:
2321 case SpvOpImageQuerySize:
2322 case SpvOpImageRead:
2324 case SpvOpAtomicStore:
2325 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2327 case SpvOpImageWrite:
2328 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2331 case SpvOpAtomicCompareExchange:
2332 case SpvOpAtomicIIncrement:
2333 case SpvOpAtomicIDecrement:
2334 case SpvOpAtomicExchange:
2335 case SpvOpAtomicIAdd:
2336 case SpvOpAtomicISub:
2337 case SpvOpAtomicSMin:
2338 case SpvOpAtomicUMin:
2339 case SpvOpAtomicSMax:
2340 case SpvOpAtomicUMax:
2341 case SpvOpAtomicAnd:
2343 case SpvOpAtomicXor:
2344 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2348 vtn_fail("Invalid image opcode");
2351 if (opcode != SpvOpImageWrite) {
2352 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2353 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2355 unsigned dest_components =
2356 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2357 if (intrin->intrinsic == nir_intrinsic_image_size) {
2358 dest_components = intrin->num_components =
2359 glsl_get_vector_elements(type->type);
2362 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2363 dest_components, 32, NULL);
2365 nir_builder_instr_insert(&b->nb, &intrin->instr);
2367 val->ssa = vtn_create_ssa_value(b, type->type);
2368 val->ssa->def = &intrin->dest.ssa;
2370 nir_builder_instr_insert(&b->nb, &intrin->instr);
2374 static nir_intrinsic_op
2375 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2378 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2379 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2380 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2381 OP(AtomicExchange, atomic_exchange)
2382 OP(AtomicCompareExchange, atomic_comp_swap)
2383 OP(AtomicIIncrement, atomic_add)
2384 OP(AtomicIDecrement, atomic_add)
2385 OP(AtomicIAdd, atomic_add)
2386 OP(AtomicISub, atomic_add)
2387 OP(AtomicSMin, atomic_imin)
2388 OP(AtomicUMin, atomic_umin)
2389 OP(AtomicSMax, atomic_imax)
2390 OP(AtomicUMax, atomic_umax)
2391 OP(AtomicAnd, atomic_and)
2392 OP(AtomicOr, atomic_or)
2393 OP(AtomicXor, atomic_xor)
2396 vtn_fail("Invalid SSBO atomic");
2400 static nir_intrinsic_op
2401 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2404 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2405 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2406 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2407 OP(AtomicExchange, atomic_exchange)
2408 OP(AtomicCompareExchange, atomic_comp_swap)
2409 OP(AtomicIIncrement, atomic_add)
2410 OP(AtomicIDecrement, atomic_add)
2411 OP(AtomicIAdd, atomic_add)
2412 OP(AtomicISub, atomic_add)
2413 OP(AtomicSMin, atomic_imin)
2414 OP(AtomicUMin, atomic_umin)
2415 OP(AtomicSMax, atomic_imax)
2416 OP(AtomicUMax, atomic_umax)
2417 OP(AtomicAnd, atomic_and)
2418 OP(AtomicOr, atomic_or)
2419 OP(AtomicXor, atomic_xor)
2422 vtn_fail("Invalid shared atomic");
2426 static nir_intrinsic_op
2427 get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2430 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2431 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2432 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##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 shared atomic");
2453 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2454 const uint32_t *w, unsigned count)
2456 struct vtn_pointer *ptr;
2457 nir_intrinsic_instr *atomic;
2460 case SpvOpAtomicLoad:
2461 case SpvOpAtomicExchange:
2462 case SpvOpAtomicCompareExchange:
2463 case SpvOpAtomicCompareExchangeWeak:
2464 case SpvOpAtomicIIncrement:
2465 case SpvOpAtomicIDecrement:
2466 case SpvOpAtomicIAdd:
2467 case SpvOpAtomicISub:
2468 case SpvOpAtomicSMin:
2469 case SpvOpAtomicUMin:
2470 case SpvOpAtomicSMax:
2471 case SpvOpAtomicUMax:
2472 case SpvOpAtomicAnd:
2474 case SpvOpAtomicXor:
2475 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2478 case SpvOpAtomicStore:
2479 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2483 vtn_fail("Invalid SPIR-V atomic");
2487 SpvScope scope = w[4];
2488 SpvMemorySemanticsMask semantics = w[5];
2491 if (ptr->mode == vtn_variable_mode_workgroup &&
2492 !b->options->lower_workgroup_access_to_offsets) {
2493 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2494 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2495 nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
2496 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2497 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2500 case SpvOpAtomicLoad:
2501 atomic->num_components = glsl_get_vector_elements(deref_type);
2504 case SpvOpAtomicStore:
2505 atomic->num_components = glsl_get_vector_elements(deref_type);
2506 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2507 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2510 case SpvOpAtomicExchange:
2511 case SpvOpAtomicCompareExchange:
2512 case SpvOpAtomicCompareExchangeWeak:
2513 case SpvOpAtomicIIncrement:
2514 case SpvOpAtomicIDecrement:
2515 case SpvOpAtomicIAdd:
2516 case SpvOpAtomicISub:
2517 case SpvOpAtomicSMin:
2518 case SpvOpAtomicUMin:
2519 case SpvOpAtomicSMax:
2520 case SpvOpAtomicUMax:
2521 case SpvOpAtomicAnd:
2523 case SpvOpAtomicXor:
2524 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2528 vtn_fail("Invalid SPIR-V atomic");
2532 nir_ssa_def *offset, *index;
2533 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2535 nir_intrinsic_op op;
2536 if (ptr->mode == vtn_variable_mode_ssbo) {
2537 op = get_ssbo_nir_atomic_op(b, opcode);
2539 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2540 b->options->lower_workgroup_access_to_offsets);
2541 op = get_shared_nir_atomic_op(b, opcode);
2544 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2548 case SpvOpAtomicLoad:
2549 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2550 if (ptr->mode == vtn_variable_mode_ssbo)
2551 atomic->src[src++] = nir_src_for_ssa(index);
2552 atomic->src[src++] = nir_src_for_ssa(offset);
2555 case SpvOpAtomicStore:
2556 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2557 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2558 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2559 if (ptr->mode == vtn_variable_mode_ssbo)
2560 atomic->src[src++] = nir_src_for_ssa(index);
2561 atomic->src[src++] = nir_src_for_ssa(offset);
2564 case SpvOpAtomicExchange:
2565 case SpvOpAtomicCompareExchange:
2566 case SpvOpAtomicCompareExchangeWeak:
2567 case SpvOpAtomicIIncrement:
2568 case SpvOpAtomicIDecrement:
2569 case SpvOpAtomicIAdd:
2570 case SpvOpAtomicISub:
2571 case SpvOpAtomicSMin:
2572 case SpvOpAtomicUMin:
2573 case SpvOpAtomicSMax:
2574 case SpvOpAtomicUMax:
2575 case SpvOpAtomicAnd:
2577 case SpvOpAtomicXor:
2578 if (ptr->mode == vtn_variable_mode_ssbo)
2579 atomic->src[src++] = nir_src_for_ssa(index);
2580 atomic->src[src++] = nir_src_for_ssa(offset);
2581 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2585 vtn_fail("Invalid SPIR-V atomic");
2589 if (opcode != SpvOpAtomicStore) {
2590 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2592 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2593 glsl_get_vector_elements(type->type),
2594 glsl_get_bit_size(type->type), NULL);
2596 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2597 val->ssa = rzalloc(b, struct vtn_ssa_value);
2598 val->ssa->def = &atomic->dest.ssa;
2599 val->ssa->type = type->type;
2602 nir_builder_instr_insert(&b->nb, &atomic->instr);
2605 static nir_alu_instr *
2606 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2609 switch (num_components) {
2610 case 1: op = nir_op_fmov; break;
2611 case 2: op = nir_op_vec2; break;
2612 case 3: op = nir_op_vec3; break;
2613 case 4: op = nir_op_vec4; break;
2614 default: vtn_fail("bad vector size");
2617 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2618 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2620 vec->dest.write_mask = (1 << num_components) - 1;
2625 struct vtn_ssa_value *
2626 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2628 if (src->transposed)
2629 return src->transposed;
2631 struct vtn_ssa_value *dest =
2632 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2634 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2635 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2636 glsl_get_bit_size(src->type));
2637 if (glsl_type_is_vector_or_scalar(src->type)) {
2638 vec->src[0].src = nir_src_for_ssa(src->def);
2639 vec->src[0].swizzle[0] = i;
2641 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2642 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2643 vec->src[j].swizzle[0] = i;
2646 nir_builder_instr_insert(&b->nb, &vec->instr);
2647 dest->elems[i]->def = &vec->dest.dest.ssa;
2650 dest->transposed = src;
2656 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2658 unsigned swiz[4] = { index };
2659 return nir_swizzle(&b->nb, src, swiz, 1, true);
2663 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2666 nir_alu_instr *vec = create_vec(b, src->num_components,
2669 for (unsigned i = 0; i < src->num_components; i++) {
2671 vec->src[i].src = nir_src_for_ssa(insert);
2673 vec->src[i].src = nir_src_for_ssa(src);
2674 vec->src[i].swizzle[0] = i;
2678 nir_builder_instr_insert(&b->nb, &vec->instr);
2680 return &vec->dest.dest.ssa;
2684 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2687 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2688 for (unsigned i = 1; i < src->num_components; i++)
2689 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2690 vtn_vector_extract(b, src, i), dest);
2696 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2697 nir_ssa_def *insert, nir_ssa_def *index)
2699 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2700 for (unsigned i = 1; i < src->num_components; i++)
2701 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2702 vtn_vector_insert(b, src, insert, i), dest);
2707 static nir_ssa_def *
2708 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2709 nir_ssa_def *src0, nir_ssa_def *src1,
2710 const uint32_t *indices)
2712 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2714 for (unsigned i = 0; i < num_components; i++) {
2715 uint32_t index = indices[i];
2716 if (index == 0xffffffff) {
2718 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2719 } else if (index < src0->num_components) {
2720 vec->src[i].src = nir_src_for_ssa(src0);
2721 vec->src[i].swizzle[0] = index;
2723 vec->src[i].src = nir_src_for_ssa(src1);
2724 vec->src[i].swizzle[0] = index - src0->num_components;
2728 nir_builder_instr_insert(&b->nb, &vec->instr);
2730 return &vec->dest.dest.ssa;
2734 * Concatentates a number of vectors/scalars together to produce a vector
2736 static nir_ssa_def *
2737 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2738 unsigned num_srcs, nir_ssa_def **srcs)
2740 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2742 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2744 * "When constructing a vector, there must be at least two Constituent
2747 vtn_assert(num_srcs >= 2);
2749 unsigned dest_idx = 0;
2750 for (unsigned i = 0; i < num_srcs; i++) {
2751 nir_ssa_def *src = srcs[i];
2752 vtn_assert(dest_idx + src->num_components <= num_components);
2753 for (unsigned j = 0; j < src->num_components; j++) {
2754 vec->src[dest_idx].src = nir_src_for_ssa(src);
2755 vec->src[dest_idx].swizzle[0] = j;
2760 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2762 * "When constructing a vector, the total number of components in all
2763 * the operands must equal the number of components in Result Type."
2765 vtn_assert(dest_idx == num_components);
2767 nir_builder_instr_insert(&b->nb, &vec->instr);
2769 return &vec->dest.dest.ssa;
2772 static struct vtn_ssa_value *
2773 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2775 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2776 dest->type = src->type;
2778 if (glsl_type_is_vector_or_scalar(src->type)) {
2779 dest->def = src->def;
2781 unsigned elems = glsl_get_length(src->type);
2783 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2784 for (unsigned i = 0; i < elems; i++)
2785 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2791 static struct vtn_ssa_value *
2792 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2793 struct vtn_ssa_value *insert, const uint32_t *indices,
2794 unsigned num_indices)
2796 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2798 struct vtn_ssa_value *cur = dest;
2800 for (i = 0; i < num_indices - 1; i++) {
2801 cur = cur->elems[indices[i]];
2804 if (glsl_type_is_vector_or_scalar(cur->type)) {
2805 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2806 * the component granularity. In that case, the last index will be
2807 * the index to insert the scalar into the vector.
2810 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2812 cur->elems[indices[i]] = insert;
2818 static struct vtn_ssa_value *
2819 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2820 const uint32_t *indices, unsigned num_indices)
2822 struct vtn_ssa_value *cur = src;
2823 for (unsigned i = 0; i < num_indices; i++) {
2824 if (glsl_type_is_vector_or_scalar(cur->type)) {
2825 vtn_assert(i == num_indices - 1);
2826 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2827 * the component granularity. The last index will be the index of the
2828 * vector to extract.
2831 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2832 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2833 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2836 cur = cur->elems[indices[i]];
2844 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2845 const uint32_t *w, unsigned count)
2847 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2848 const struct glsl_type *type =
2849 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2850 val->ssa = vtn_create_ssa_value(b, type);
2853 case SpvOpVectorExtractDynamic:
2854 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2855 vtn_ssa_value(b, w[4])->def);
2858 case SpvOpVectorInsertDynamic:
2859 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2860 vtn_ssa_value(b, w[4])->def,
2861 vtn_ssa_value(b, w[5])->def);
2864 case SpvOpVectorShuffle:
2865 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2866 vtn_ssa_value(b, w[3])->def,
2867 vtn_ssa_value(b, w[4])->def,
2871 case SpvOpCompositeConstruct: {
2872 unsigned elems = count - 3;
2873 if (glsl_type_is_vector_or_scalar(type)) {
2874 nir_ssa_def *srcs[4];
2875 for (unsigned i = 0; i < elems; i++)
2876 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2878 vtn_vector_construct(b, glsl_get_vector_elements(type),
2881 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2882 for (unsigned i = 0; i < elems; i++)
2883 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2887 case SpvOpCompositeExtract:
2888 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2892 case SpvOpCompositeInsert:
2893 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2894 vtn_ssa_value(b, w[3]),
2898 case SpvOpCopyObject:
2899 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2903 vtn_fail("unknown composite operation");
2908 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
2909 const uint32_t *w, unsigned count)
2911 nir_intrinsic_op intrinsic_op;
2913 case SpvOpEmitVertex:
2914 case SpvOpEmitStreamVertex:
2915 intrinsic_op = nir_intrinsic_emit_vertex;
2917 case SpvOpEndPrimitive:
2918 case SpvOpEndStreamPrimitive:
2919 intrinsic_op = nir_intrinsic_end_primitive;
2921 case SpvOpMemoryBarrier:
2922 intrinsic_op = nir_intrinsic_memory_barrier;
2924 case SpvOpControlBarrier:
2925 intrinsic_op = nir_intrinsic_barrier;
2928 vtn_fail("unknown barrier instruction");
2931 nir_intrinsic_instr *intrin =
2932 nir_intrinsic_instr_create(b->shader, intrinsic_op);
2934 if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
2935 nir_intrinsic_set_stream_id(intrin, w[1]);
2937 nir_builder_instr_insert(&b->nb, &intrin->instr);
2941 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
2942 SpvExecutionMode mode)
2945 case SpvExecutionModeInputPoints:
2946 case SpvExecutionModeOutputPoints:
2947 return 0; /* GL_POINTS */
2948 case SpvExecutionModeInputLines:
2949 return 1; /* GL_LINES */
2950 case SpvExecutionModeInputLinesAdjacency:
2951 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2952 case SpvExecutionModeTriangles:
2953 return 4; /* GL_TRIANGLES */
2954 case SpvExecutionModeInputTrianglesAdjacency:
2955 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2956 case SpvExecutionModeQuads:
2957 return 7; /* GL_QUADS */
2958 case SpvExecutionModeIsolines:
2959 return 0x8E7A; /* GL_ISOLINES */
2960 case SpvExecutionModeOutputLineStrip:
2961 return 3; /* GL_LINE_STRIP */
2962 case SpvExecutionModeOutputTriangleStrip:
2963 return 5; /* GL_TRIANGLE_STRIP */
2965 vtn_fail("Invalid primitive type");
2970 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
2971 SpvExecutionMode mode)
2974 case SpvExecutionModeInputPoints:
2976 case SpvExecutionModeInputLines:
2978 case SpvExecutionModeInputLinesAdjacency:
2980 case SpvExecutionModeTriangles:
2982 case SpvExecutionModeInputTrianglesAdjacency:
2985 vtn_fail("Invalid GS input mode");
2989 static gl_shader_stage
2990 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
2993 case SpvExecutionModelVertex:
2994 return MESA_SHADER_VERTEX;
2995 case SpvExecutionModelTessellationControl:
2996 return MESA_SHADER_TESS_CTRL;
2997 case SpvExecutionModelTessellationEvaluation:
2998 return MESA_SHADER_TESS_EVAL;
2999 case SpvExecutionModelGeometry:
3000 return MESA_SHADER_GEOMETRY;
3001 case SpvExecutionModelFragment:
3002 return MESA_SHADER_FRAGMENT;
3003 case SpvExecutionModelGLCompute:
3004 return MESA_SHADER_COMPUTE;
3006 vtn_fail("Unsupported execution model");
3010 #define spv_check_supported(name, cap) do { \
3011 if (!(b->options && b->options->caps.name)) \
3012 vtn_warn("Unsupported SPIR-V capability: %s", \
3013 spirv_capability_to_string(cap)); \
3017 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
3018 const uint32_t *w, unsigned count)
3025 case SpvSourceLanguageUnknown: lang = "unknown"; break;
3026 case SpvSourceLanguageESSL: lang = "ESSL"; break;
3027 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
3028 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
3029 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
3030 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
3033 uint32_t version = w[2];
3036 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
3038 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
3042 case SpvOpSourceExtension:
3043 case SpvOpSourceContinued:
3044 case SpvOpExtension:
3045 /* Unhandled, but these are for debug so that's ok. */
3048 case SpvOpCapability: {
3049 SpvCapability cap = w[1];
3051 case SpvCapabilityMatrix:
3052 case SpvCapabilityShader:
3053 case SpvCapabilityGeometry:
3054 case SpvCapabilityGeometryPointSize:
3055 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3056 case SpvCapabilitySampledImageArrayDynamicIndexing:
3057 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3058 case SpvCapabilityStorageImageArrayDynamicIndexing:
3059 case SpvCapabilityImageRect:
3060 case SpvCapabilitySampledRect:
3061 case SpvCapabilitySampled1D:
3062 case SpvCapabilityImage1D:
3063 case SpvCapabilitySampledCubeArray:
3064 case SpvCapabilityImageCubeArray:
3065 case SpvCapabilitySampledBuffer:
3066 case SpvCapabilityImageBuffer:
3067 case SpvCapabilityImageQuery:
3068 case SpvCapabilityDerivativeControl:
3069 case SpvCapabilityInterpolationFunction:
3070 case SpvCapabilityMultiViewport:
3071 case SpvCapabilitySampleRateShading:
3072 case SpvCapabilityClipDistance:
3073 case SpvCapabilityCullDistance:
3074 case SpvCapabilityInputAttachment:
3075 case SpvCapabilityImageGatherExtended:
3076 case SpvCapabilityStorageImageExtendedFormats:
3079 case SpvCapabilityGeometryStreams:
3080 case SpvCapabilityLinkage:
3081 case SpvCapabilityVector16:
3082 case SpvCapabilityFloat16Buffer:
3083 case SpvCapabilityFloat16:
3084 case SpvCapabilityInt64Atomics:
3085 case SpvCapabilityAtomicStorage:
3086 case SpvCapabilityInt16:
3087 case SpvCapabilityStorageImageMultisample:
3088 case SpvCapabilityInt8:
3089 case SpvCapabilitySparseResidency:
3090 case SpvCapabilityMinLod:
3091 case SpvCapabilityTransformFeedback:
3092 vtn_warn("Unsupported SPIR-V capability: %s",
3093 spirv_capability_to_string(cap));
3096 case SpvCapabilityFloat64:
3097 spv_check_supported(float64, cap);
3099 case SpvCapabilityInt64:
3100 spv_check_supported(int64, cap);
3103 case SpvCapabilityAddresses:
3104 case SpvCapabilityKernel:
3105 case SpvCapabilityImageBasic:
3106 case SpvCapabilityImageReadWrite:
3107 case SpvCapabilityImageMipmap:
3108 case SpvCapabilityPipes:
3109 case SpvCapabilityGroups:
3110 case SpvCapabilityDeviceEnqueue:
3111 case SpvCapabilityLiteralSampler:
3112 case SpvCapabilityGenericPointer:
3113 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3114 spirv_capability_to_string(cap));
3117 case SpvCapabilityImageMSArray:
3118 spv_check_supported(image_ms_array, cap);
3121 case SpvCapabilityTessellation:
3122 case SpvCapabilityTessellationPointSize:
3123 spv_check_supported(tessellation, cap);
3126 case SpvCapabilityDrawParameters:
3127 spv_check_supported(draw_parameters, cap);
3130 case SpvCapabilityStorageImageReadWithoutFormat:
3131 spv_check_supported(image_read_without_format, cap);
3134 case SpvCapabilityStorageImageWriteWithoutFormat:
3135 spv_check_supported(image_write_without_format, cap);
3138 case SpvCapabilityMultiView:
3139 spv_check_supported(multiview, cap);
3142 case SpvCapabilityVariablePointersStorageBuffer:
3143 case SpvCapabilityVariablePointers:
3144 spv_check_supported(variable_pointers, cap);
3147 case SpvCapabilityStorageUniformBufferBlock16:
3148 case SpvCapabilityStorageUniform16:
3149 case SpvCapabilityStoragePushConstant16:
3150 case SpvCapabilityStorageInputOutput16:
3151 spv_check_supported(storage_16bit, cap);
3155 vtn_fail("Unhandled capability");
3160 case SpvOpExtInstImport:
3161 vtn_handle_extension(b, opcode, w, count);
3164 case SpvOpMemoryModel:
3165 vtn_assert(w[1] == SpvAddressingModelLogical);
3166 vtn_assert(w[2] == SpvMemoryModelSimple ||
3167 w[2] == SpvMemoryModelGLSL450);
3170 case SpvOpEntryPoint: {
3171 struct vtn_value *entry_point = &b->values[w[2]];
3172 /* Let this be a name label regardless */
3173 unsigned name_words;
3174 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3176 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3177 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3180 vtn_assert(b->entry_point == NULL);
3181 b->entry_point = entry_point;
3186 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3187 vtn_string_literal(b, &w[2], count - 2, NULL);
3191 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3194 case SpvOpMemberName:
3198 case SpvOpExecutionMode:
3199 case SpvOpDecorationGroup:
3201 case SpvOpMemberDecorate:
3202 case SpvOpGroupDecorate:
3203 case SpvOpGroupMemberDecorate:
3204 vtn_handle_decoration(b, opcode, w, count);
3208 return false; /* End of preamble */
3215 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3216 const struct vtn_decoration *mode, void *data)
3218 vtn_assert(b->entry_point == entry_point);
3220 switch(mode->exec_mode) {
3221 case SpvExecutionModeOriginUpperLeft:
3222 case SpvExecutionModeOriginLowerLeft:
3223 b->origin_upper_left =
3224 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3227 case SpvExecutionModeEarlyFragmentTests:
3228 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3229 b->shader->info.fs.early_fragment_tests = true;
3232 case SpvExecutionModeInvocations:
3233 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3234 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3237 case SpvExecutionModeDepthReplacing:
3238 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3239 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3241 case SpvExecutionModeDepthGreater:
3242 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3243 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3245 case SpvExecutionModeDepthLess:
3246 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3247 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3249 case SpvExecutionModeDepthUnchanged:
3250 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3251 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3254 case SpvExecutionModeLocalSize:
3255 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3256 b->shader->info.cs.local_size[0] = mode->literals[0];
3257 b->shader->info.cs.local_size[1] = mode->literals[1];
3258 b->shader->info.cs.local_size[2] = mode->literals[2];
3260 case SpvExecutionModeLocalSizeHint:
3261 break; /* Nothing to do with this */
3263 case SpvExecutionModeOutputVertices:
3264 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3265 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3266 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3268 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3269 b->shader->info.gs.vertices_out = mode->literals[0];
3273 case SpvExecutionModeInputPoints:
3274 case SpvExecutionModeInputLines:
3275 case SpvExecutionModeInputLinesAdjacency:
3276 case SpvExecutionModeTriangles:
3277 case SpvExecutionModeInputTrianglesAdjacency:
3278 case SpvExecutionModeQuads:
3279 case SpvExecutionModeIsolines:
3280 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3281 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3282 b->shader->info.tess.primitive_mode =
3283 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3285 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3286 b->shader->info.gs.vertices_in =
3287 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3291 case SpvExecutionModeOutputPoints:
3292 case SpvExecutionModeOutputLineStrip:
3293 case SpvExecutionModeOutputTriangleStrip:
3294 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3295 b->shader->info.gs.output_primitive =
3296 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3299 case SpvExecutionModeSpacingEqual:
3300 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3301 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3302 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3304 case SpvExecutionModeSpacingFractionalEven:
3305 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3306 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3307 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3309 case SpvExecutionModeSpacingFractionalOdd:
3310 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3311 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3312 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3314 case SpvExecutionModeVertexOrderCw:
3315 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3316 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3317 b->shader->info.tess.ccw = false;
3319 case SpvExecutionModeVertexOrderCcw:
3320 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3321 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3322 b->shader->info.tess.ccw = true;
3324 case SpvExecutionModePointMode:
3325 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3326 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3327 b->shader->info.tess.point_mode = true;
3330 case SpvExecutionModePixelCenterInteger:
3331 b->pixel_center_integer = true;
3334 case SpvExecutionModeXfb:
3335 vtn_fail("Unhandled execution mode");
3338 case SpvExecutionModeVecTypeHint:
3339 case SpvExecutionModeContractionOff:
3343 vtn_fail("Unhandled execution mode");
3348 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3349 const uint32_t *w, unsigned count)
3351 vtn_set_instruction_result_type(b, opcode, w, count);
3355 case SpvOpSourceContinued:
3356 case SpvOpSourceExtension:
3357 case SpvOpExtension:
3358 case SpvOpCapability:
3359 case SpvOpExtInstImport:
3360 case SpvOpMemoryModel:
3361 case SpvOpEntryPoint:
3362 case SpvOpExecutionMode:
3365 case SpvOpMemberName:
3366 case SpvOpDecorationGroup:
3368 case SpvOpMemberDecorate:
3369 case SpvOpGroupDecorate:
3370 case SpvOpGroupMemberDecorate:
3371 vtn_fail("Invalid opcode types and variables section");
3377 case SpvOpTypeFloat:
3378 case SpvOpTypeVector:
3379 case SpvOpTypeMatrix:
3380 case SpvOpTypeImage:
3381 case SpvOpTypeSampler:
3382 case SpvOpTypeSampledImage:
3383 case SpvOpTypeArray:
3384 case SpvOpTypeRuntimeArray:
3385 case SpvOpTypeStruct:
3386 case SpvOpTypeOpaque:
3387 case SpvOpTypePointer:
3388 case SpvOpTypeFunction:
3389 case SpvOpTypeEvent:
3390 case SpvOpTypeDeviceEvent:
3391 case SpvOpTypeReserveId:
3392 case SpvOpTypeQueue:
3394 vtn_handle_type(b, opcode, w, count);
3397 case SpvOpConstantTrue:
3398 case SpvOpConstantFalse:
3400 case SpvOpConstantComposite:
3401 case SpvOpConstantSampler:
3402 case SpvOpConstantNull:
3403 case SpvOpSpecConstantTrue:
3404 case SpvOpSpecConstantFalse:
3405 case SpvOpSpecConstant:
3406 case SpvOpSpecConstantComposite:
3407 case SpvOpSpecConstantOp:
3408 vtn_handle_constant(b, opcode, w, count);
3413 vtn_handle_variables(b, opcode, w, count);
3417 return false; /* End of preamble */
3424 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3425 const uint32_t *w, unsigned count)
3431 case SpvOpLoopMerge:
3432 case SpvOpSelectionMerge:
3433 /* This is handled by cfg pre-pass and walk_blocks */
3437 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3438 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3443 vtn_handle_extension(b, opcode, w, count);
3449 case SpvOpCopyMemory:
3450 case SpvOpCopyMemorySized:
3451 case SpvOpAccessChain:
3452 case SpvOpPtrAccessChain:
3453 case SpvOpInBoundsAccessChain:
3454 case SpvOpArrayLength:
3455 vtn_handle_variables(b, opcode, w, count);
3458 case SpvOpFunctionCall:
3459 vtn_handle_function_call(b, opcode, w, count);
3462 case SpvOpSampledImage:
3464 case SpvOpImageSampleImplicitLod:
3465 case SpvOpImageSampleExplicitLod:
3466 case SpvOpImageSampleDrefImplicitLod:
3467 case SpvOpImageSampleDrefExplicitLod:
3468 case SpvOpImageSampleProjImplicitLod:
3469 case SpvOpImageSampleProjExplicitLod:
3470 case SpvOpImageSampleProjDrefImplicitLod:
3471 case SpvOpImageSampleProjDrefExplicitLod:
3472 case SpvOpImageFetch:
3473 case SpvOpImageGather:
3474 case SpvOpImageDrefGather:
3475 case SpvOpImageQuerySizeLod:
3476 case SpvOpImageQueryLod:
3477 case SpvOpImageQueryLevels:
3478 case SpvOpImageQuerySamples:
3479 vtn_handle_texture(b, opcode, w, count);
3482 case SpvOpImageRead:
3483 case SpvOpImageWrite:
3484 case SpvOpImageTexelPointer:
3485 vtn_handle_image(b, opcode, w, count);
3488 case SpvOpImageQuerySize: {
3489 struct vtn_pointer *image =
3490 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3491 if (image->mode == vtn_variable_mode_image) {
3492 vtn_handle_image(b, opcode, w, count);
3494 vtn_assert(image->mode == vtn_variable_mode_sampler);
3495 vtn_handle_texture(b, opcode, w, count);
3500 case SpvOpAtomicLoad:
3501 case SpvOpAtomicExchange:
3502 case SpvOpAtomicCompareExchange:
3503 case SpvOpAtomicCompareExchangeWeak:
3504 case SpvOpAtomicIIncrement:
3505 case SpvOpAtomicIDecrement:
3506 case SpvOpAtomicIAdd:
3507 case SpvOpAtomicISub:
3508 case SpvOpAtomicSMin:
3509 case SpvOpAtomicUMin:
3510 case SpvOpAtomicSMax:
3511 case SpvOpAtomicUMax:
3512 case SpvOpAtomicAnd:
3514 case SpvOpAtomicXor: {
3515 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3516 if (pointer->value_type == vtn_value_type_image_pointer) {
3517 vtn_handle_image(b, opcode, w, count);
3519 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3520 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3525 case SpvOpAtomicStore: {
3526 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3527 if (pointer->value_type == vtn_value_type_image_pointer) {
3528 vtn_handle_image(b, opcode, w, count);
3530 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3531 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3537 /* Handle OpSelect up-front here because it needs to be able to handle
3538 * pointers and not just regular vectors and scalars.
3540 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3541 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3542 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3543 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3545 const struct glsl_type *sel_type;
3546 switch (res_val->type->base_type) {
3547 case vtn_base_type_scalar:
3548 sel_type = glsl_bool_type();
3550 case vtn_base_type_vector:
3551 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3553 case vtn_base_type_pointer:
3554 /* We need to have actual storage for pointer types */
3555 vtn_fail_if(res_val->type->type == NULL,
3556 "Invalid pointer result type for OpSelect");
3557 sel_type = glsl_bool_type();
3560 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3563 if (unlikely(sel_val->type->type != sel_type)) {
3564 if (sel_val->type->type == glsl_bool_type()) {
3565 /* This case is illegal but some older versions of GLSLang produce
3566 * it. The GLSLang issue was fixed on March 30, 2017:
3568 * https://github.com/KhronosGroup/glslang/issues/809
3570 * Unfortunately, there are applications in the wild which are
3571 * shipping with this bug so it isn't nice to fail on them so we
3572 * throw a warning instead. It's not actually a problem for us as
3573 * nir_builder will just splat the condition out which is most
3574 * likely what the client wanted anyway.
3576 vtn_warn("Condition type of OpSelect must have the same number "
3577 "of components as Result Type");
3579 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3580 "of Boolean type. It must have the same number of "
3581 "components as Result Type");
3585 vtn_fail_if(obj1_val->type != res_val->type ||
3586 obj2_val->type != res_val->type,
3587 "Object types must match the result type in OpSelect");
3589 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3590 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3591 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3592 vtn_ssa_value(b, w[4])->def,
3593 vtn_ssa_value(b, w[5])->def);
3594 vtn_push_ssa(b, w[2], res_type, ssa);
3603 case SpvOpConvertFToU:
3604 case SpvOpConvertFToS:
3605 case SpvOpConvertSToF:
3606 case SpvOpConvertUToF:
3610 case SpvOpQuantizeToF16:
3611 case SpvOpConvertPtrToU:
3612 case SpvOpConvertUToPtr:
3613 case SpvOpPtrCastToGeneric:
3614 case SpvOpGenericCastToPtr:
3620 case SpvOpSignBitSet:
3621 case SpvOpLessOrGreater:
3623 case SpvOpUnordered:
3638 case SpvOpVectorTimesScalar:
3640 case SpvOpIAddCarry:
3641 case SpvOpISubBorrow:
3642 case SpvOpUMulExtended:
3643 case SpvOpSMulExtended:
3644 case SpvOpShiftRightLogical:
3645 case SpvOpShiftRightArithmetic:
3646 case SpvOpShiftLeftLogical:
3647 case SpvOpLogicalEqual:
3648 case SpvOpLogicalNotEqual:
3649 case SpvOpLogicalOr:
3650 case SpvOpLogicalAnd:
3651 case SpvOpLogicalNot:
3652 case SpvOpBitwiseOr:
3653 case SpvOpBitwiseXor:
3654 case SpvOpBitwiseAnd:
3656 case SpvOpFOrdEqual:
3657 case SpvOpFUnordEqual:
3658 case SpvOpINotEqual:
3659 case SpvOpFOrdNotEqual:
3660 case SpvOpFUnordNotEqual:
3661 case SpvOpULessThan:
3662 case SpvOpSLessThan:
3663 case SpvOpFOrdLessThan:
3664 case SpvOpFUnordLessThan:
3665 case SpvOpUGreaterThan:
3666 case SpvOpSGreaterThan:
3667 case SpvOpFOrdGreaterThan:
3668 case SpvOpFUnordGreaterThan:
3669 case SpvOpULessThanEqual:
3670 case SpvOpSLessThanEqual:
3671 case SpvOpFOrdLessThanEqual:
3672 case SpvOpFUnordLessThanEqual:
3673 case SpvOpUGreaterThanEqual:
3674 case SpvOpSGreaterThanEqual:
3675 case SpvOpFOrdGreaterThanEqual:
3676 case SpvOpFUnordGreaterThanEqual:
3682 case SpvOpFwidthFine:
3683 case SpvOpDPdxCoarse:
3684 case SpvOpDPdyCoarse:
3685 case SpvOpFwidthCoarse:
3686 case SpvOpBitFieldInsert:
3687 case SpvOpBitFieldSExtract:
3688 case SpvOpBitFieldUExtract:
3689 case SpvOpBitReverse:
3691 case SpvOpTranspose:
3692 case SpvOpOuterProduct:
3693 case SpvOpMatrixTimesScalar:
3694 case SpvOpVectorTimesMatrix:
3695 case SpvOpMatrixTimesVector:
3696 case SpvOpMatrixTimesMatrix:
3697 vtn_handle_alu(b, opcode, w, count);
3700 case SpvOpVectorExtractDynamic:
3701 case SpvOpVectorInsertDynamic:
3702 case SpvOpVectorShuffle:
3703 case SpvOpCompositeConstruct:
3704 case SpvOpCompositeExtract:
3705 case SpvOpCompositeInsert:
3706 case SpvOpCopyObject:
3707 vtn_handle_composite(b, opcode, w, count);
3710 case SpvOpEmitVertex:
3711 case SpvOpEndPrimitive:
3712 case SpvOpEmitStreamVertex:
3713 case SpvOpEndStreamPrimitive:
3714 case SpvOpControlBarrier:
3715 case SpvOpMemoryBarrier:
3716 vtn_handle_barrier(b, opcode, w, count);
3720 vtn_fail("Unhandled opcode");
3727 spirv_to_nir(const uint32_t *words, size_t word_count,
3728 struct nir_spirv_specialization *spec, unsigned num_spec,
3729 gl_shader_stage stage, const char *entry_point_name,
3730 const struct spirv_to_nir_options *options,
3731 const nir_shader_compiler_options *nir_options)
3733 /* Initialize the stn_builder object */
3734 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3736 b->spirv_word_count = word_count;
3740 exec_list_make_empty(&b->functions);
3741 b->entry_point_stage = stage;
3742 b->entry_point_name = entry_point_name;
3743 b->options = options;
3745 /* See also _vtn_fail() */
3746 if (setjmp(b->fail_jump)) {
3751 const uint32_t *word_end = words + word_count;
3753 /* Handle the SPIR-V header (first 4 dwords) */
3754 vtn_assert(word_count > 5);
3756 vtn_assert(words[0] == SpvMagicNumber);
3757 vtn_assert(words[1] >= 0x10000);
3758 /* words[2] == generator magic */
3759 unsigned value_id_bound = words[3];
3760 vtn_assert(words[4] == 0);
3764 b->value_id_bound = value_id_bound;
3765 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3767 /* Handle all the preamble instructions */
3768 words = vtn_foreach_instruction(b, words, word_end,
3769 vtn_handle_preamble_instruction);
3771 if (b->entry_point == NULL) {
3772 vtn_fail("Entry point not found");
3777 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3779 /* Set shader info defaults */
3780 b->shader->info.gs.invocations = 1;
3782 /* Parse execution modes */
3783 vtn_foreach_execution_mode(b, b->entry_point,
3784 vtn_handle_execution_mode, NULL);
3786 b->specializations = spec;
3787 b->num_specializations = num_spec;
3789 /* Handle all variable, type, and constant instructions */
3790 words = vtn_foreach_instruction(b, words, word_end,
3791 vtn_handle_variable_or_type_instruction);
3793 /* Set types on all vtn_values */
3794 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
3796 vtn_build_cfg(b, words, word_end);
3798 assert(b->entry_point->value_type == vtn_value_type_function);
3799 b->entry_point->func->referenced = true;
3804 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3805 if (func->referenced && !func->emitted) {
3806 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3807 _mesa_key_pointer_equal);
3809 vtn_function_emit(b, func, vtn_handle_body_instruction);
3815 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
3816 nir_function *entry_point = b->entry_point->func->impl->function;
3817 vtn_assert(entry_point);
3819 /* Unparent the shader from the vtn_builder before we delete the builder */
3820 ralloc_steal(NULL, b->shader);