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"
35 vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
36 size_t spirv_offset, const char *message)
38 if (b->options->debug.func) {
39 b->options->debug.func(b->options->debug.private_data,
40 level, spirv_offset, message);
44 if (level >= NIR_SPIRV_DEBUG_LEVEL_WARNING)
45 fprintf(stderr, "%s\n", message);
50 vtn_logf(struct vtn_builder *b, enum nir_spirv_debug_level level,
51 size_t spirv_offset, const char *fmt, ...)
57 msg = ralloc_vasprintf(NULL, fmt, args);
60 vtn_log(b, level, spirv_offset, msg);
66 vtn_log_err(struct vtn_builder *b,
67 enum nir_spirv_debug_level level, const char *prefix,
68 const char *file, unsigned line,
69 const char *fmt, va_list args)
73 msg = ralloc_strdup(NULL, prefix);
76 ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
79 ralloc_asprintf_append(&msg, " ");
81 ralloc_vasprintf_append(&msg, fmt, args);
83 ralloc_asprintf_append(&msg, "\n %zu bytes into the SPIR-V binary",
87 ralloc_asprintf_append(&msg,
88 "\n in SPIR-V source file %s, line %d, col %d",
89 b->file, b->line, b->col);
92 vtn_log(b, level, b->spirv_offset, msg);
98 _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
104 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_WARNING, "SPIR-V WARNING:\n",
105 file, line, fmt, args);
110 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
111 const char *fmt, ...)
116 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V parsing FAILED:\n",
117 file, line, fmt, args);
120 longjmp(b->fail_jump, 1);
123 struct spec_constant_value {
131 static struct vtn_ssa_value *
132 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
134 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
137 if (glsl_type_is_vector_or_scalar(type)) {
138 unsigned num_components = glsl_get_vector_elements(val->type);
139 unsigned bit_size = glsl_get_bit_size(val->type);
140 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
142 unsigned elems = glsl_get_length(val->type);
143 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
144 if (glsl_type_is_matrix(type)) {
145 const struct glsl_type *elem_type =
146 glsl_vector_type(glsl_get_base_type(type),
147 glsl_get_vector_elements(type));
149 for (unsigned i = 0; i < elems; i++)
150 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
151 } else if (glsl_type_is_array(type)) {
152 const struct glsl_type *elem_type = glsl_get_array_element(type);
153 for (unsigned i = 0; i < elems; i++)
154 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
156 for (unsigned i = 0; i < elems; i++) {
157 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
158 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
166 static struct vtn_ssa_value *
167 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
168 const struct glsl_type *type)
170 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
175 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
178 switch (glsl_get_base_type(type)) {
181 case GLSL_TYPE_INT16:
182 case GLSL_TYPE_UINT16:
183 case GLSL_TYPE_INT64:
184 case GLSL_TYPE_UINT64:
186 case GLSL_TYPE_FLOAT:
187 case GLSL_TYPE_FLOAT16:
188 case GLSL_TYPE_DOUBLE: {
189 int bit_size = glsl_get_bit_size(type);
190 if (glsl_type_is_vector_or_scalar(type)) {
191 unsigned num_components = glsl_get_vector_elements(val->type);
192 nir_load_const_instr *load =
193 nir_load_const_instr_create(b->shader, num_components, bit_size);
195 load->value = constant->values[0];
197 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
198 val->def = &load->def;
200 assert(glsl_type_is_matrix(type));
201 unsigned rows = glsl_get_vector_elements(val->type);
202 unsigned columns = glsl_get_matrix_columns(val->type);
203 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
205 for (unsigned i = 0; i < columns; i++) {
206 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
207 col_val->type = glsl_get_column_type(val->type);
208 nir_load_const_instr *load =
209 nir_load_const_instr_create(b->shader, rows, bit_size);
211 load->value = constant->values[i];
213 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
214 col_val->def = &load->def;
216 val->elems[i] = col_val;
222 case GLSL_TYPE_ARRAY: {
223 unsigned elems = glsl_get_length(val->type);
224 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
225 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
226 for (unsigned i = 0; i < elems; i++)
227 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
232 case GLSL_TYPE_STRUCT: {
233 unsigned elems = glsl_get_length(val->type);
234 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
235 for (unsigned i = 0; i < elems; i++) {
236 const struct glsl_type *elem_type =
237 glsl_get_struct_field(val->type, i);
238 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
245 vtn_fail("bad constant type");
251 struct vtn_ssa_value *
252 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
254 struct vtn_value *val = vtn_untyped_value(b, value_id);
255 switch (val->value_type) {
256 case vtn_value_type_undef:
257 return vtn_undef_ssa_value(b, val->type->type);
259 case vtn_value_type_constant:
260 return vtn_const_ssa_value(b, val->constant, val->type->type);
262 case vtn_value_type_ssa:
265 case vtn_value_type_pointer:
266 vtn_assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
267 struct vtn_ssa_value *ssa =
268 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
269 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
273 vtn_fail("Invalid type for an SSA value");
278 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
279 unsigned word_count, unsigned *words_used)
281 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
283 /* Ammount of space taken by the string (including the null) */
284 unsigned len = strlen(dup) + 1;
285 *words_used = DIV_ROUND_UP(len, sizeof(*words));
291 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
292 const uint32_t *end, vtn_instruction_handler handler)
298 const uint32_t *w = start;
300 SpvOp opcode = w[0] & SpvOpCodeMask;
301 unsigned count = w[0] >> SpvWordCountShift;
302 vtn_assert(count >= 1 && w + count <= end);
304 b->spirv_offset = (uint8_t *)w - (uint8_t *)b->spirv;
308 break; /* Do nothing */
311 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
323 if (!handler(b, opcode, w, count))
341 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
342 const uint32_t *w, unsigned count)
345 case SpvOpExtInstImport: {
346 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
347 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
348 val->ext_handler = vtn_handle_glsl450_instruction;
350 vtn_fail("Unsupported extension");
356 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
357 bool handled = val->ext_handler(b, w[4], w, count);
363 vtn_fail("Unhandled opcode");
368 _foreach_decoration_helper(struct vtn_builder *b,
369 struct vtn_value *base_value,
371 struct vtn_value *value,
372 vtn_decoration_foreach_cb cb, void *data)
374 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
376 if (dec->scope == VTN_DEC_DECORATION) {
377 member = parent_member;
378 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
379 vtn_assert(parent_member == -1);
380 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
382 /* Not a decoration */
387 vtn_assert(dec->group->value_type == vtn_value_type_decoration_group);
388 _foreach_decoration_helper(b, base_value, member, dec->group,
391 cb(b, base_value, member, dec, data);
396 /** Iterates (recursively if needed) over all of the decorations on a value
398 * This function iterates over all of the decorations applied to a given
399 * value. If it encounters a decoration group, it recurses into the group
400 * and iterates over all of those decorations as well.
403 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
404 vtn_decoration_foreach_cb cb, void *data)
406 _foreach_decoration_helper(b, value, -1, value, cb, data);
410 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
411 vtn_execution_mode_foreach_cb cb, void *data)
413 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
414 if (dec->scope != VTN_DEC_EXECUTION_MODE)
417 vtn_assert(dec->group == NULL);
418 cb(b, value, dec, data);
423 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
424 const uint32_t *w, unsigned count)
426 const uint32_t *w_end = w + count;
427 const uint32_t target = w[1];
431 case SpvOpDecorationGroup:
432 vtn_push_value(b, target, vtn_value_type_decoration_group);
436 case SpvOpMemberDecorate:
437 case SpvOpExecutionMode: {
438 struct vtn_value *val = &b->values[target];
440 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
443 dec->scope = VTN_DEC_DECORATION;
445 case SpvOpMemberDecorate:
446 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
448 case SpvOpExecutionMode:
449 dec->scope = VTN_DEC_EXECUTION_MODE;
452 vtn_fail("Invalid decoration opcode");
454 dec->decoration = *(w++);
457 /* Link into the list */
458 dec->next = val->decoration;
459 val->decoration = dec;
463 case SpvOpGroupMemberDecorate:
464 case SpvOpGroupDecorate: {
465 struct vtn_value *group =
466 vtn_value(b, target, vtn_value_type_decoration_group);
468 for (; w < w_end; w++) {
469 struct vtn_value *val = vtn_untyped_value(b, *w);
470 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
473 if (opcode == SpvOpGroupDecorate) {
474 dec->scope = VTN_DEC_DECORATION;
476 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
479 /* Link into the list */
480 dec->next = val->decoration;
481 val->decoration = dec;
487 vtn_fail("Unhandled opcode");
491 struct member_decoration_ctx {
493 struct glsl_struct_field *fields;
494 struct vtn_type *type;
497 /* does a shallow copy of a vtn_type */
499 static struct vtn_type *
500 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
502 struct vtn_type *dest = ralloc(b, struct vtn_type);
505 switch (src->base_type) {
506 case vtn_base_type_void:
507 case vtn_base_type_scalar:
508 case vtn_base_type_vector:
509 case vtn_base_type_matrix:
510 case vtn_base_type_array:
511 case vtn_base_type_pointer:
512 case vtn_base_type_image:
513 case vtn_base_type_sampler:
514 case vtn_base_type_sampled_image:
515 /* Nothing more to do */
518 case vtn_base_type_struct:
519 dest->members = ralloc_array(b, struct vtn_type *, src->length);
520 memcpy(dest->members, src->members,
521 src->length * sizeof(src->members[0]));
523 dest->offsets = ralloc_array(b, unsigned, src->length);
524 memcpy(dest->offsets, src->offsets,
525 src->length * sizeof(src->offsets[0]));
528 case vtn_base_type_function:
529 dest->params = ralloc_array(b, struct vtn_type *, src->length);
530 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
537 static struct vtn_type *
538 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
540 type->members[member] = vtn_type_copy(b, type->members[member]);
541 type = type->members[member];
543 /* We may have an array of matrices.... Oh, joy! */
544 while (glsl_type_is_array(type->type)) {
545 type->array_element = vtn_type_copy(b, type->array_element);
546 type = type->array_element;
549 vtn_assert(glsl_type_is_matrix(type->type));
555 struct_member_decoration_cb(struct vtn_builder *b,
556 struct vtn_value *val, int member,
557 const struct vtn_decoration *dec, void *void_ctx)
559 struct member_decoration_ctx *ctx = void_ctx;
564 vtn_assert(member < ctx->num_fields);
566 switch (dec->decoration) {
567 case SpvDecorationNonWritable:
568 case SpvDecorationNonReadable:
569 case SpvDecorationRelaxedPrecision:
570 case SpvDecorationVolatile:
571 case SpvDecorationCoherent:
572 case SpvDecorationUniform:
573 break; /* FIXME: Do nothing with this for now. */
574 case SpvDecorationNoPerspective:
575 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
577 case SpvDecorationFlat:
578 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
580 case SpvDecorationCentroid:
581 ctx->fields[member].centroid = true;
583 case SpvDecorationSample:
584 ctx->fields[member].sample = true;
586 case SpvDecorationStream:
587 /* Vulkan only allows one GS stream */
588 vtn_assert(dec->literals[0] == 0);
590 case SpvDecorationLocation:
591 ctx->fields[member].location = dec->literals[0];
593 case SpvDecorationComponent:
594 break; /* FIXME: What should we do with these? */
595 case SpvDecorationBuiltIn:
596 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
597 ctx->type->members[member]->is_builtin = true;
598 ctx->type->members[member]->builtin = dec->literals[0];
599 ctx->type->builtin_block = true;
601 case SpvDecorationOffset:
602 ctx->type->offsets[member] = dec->literals[0];
604 case SpvDecorationMatrixStride:
605 /* Handled as a second pass */
607 case SpvDecorationColMajor:
608 break; /* Nothing to do here. Column-major is the default. */
609 case SpvDecorationRowMajor:
610 mutable_matrix_member(b, ctx->type, member)->row_major = true;
613 case SpvDecorationPatch:
616 case SpvDecorationSpecId:
617 case SpvDecorationBlock:
618 case SpvDecorationBufferBlock:
619 case SpvDecorationArrayStride:
620 case SpvDecorationGLSLShared:
621 case SpvDecorationGLSLPacked:
622 case SpvDecorationInvariant:
623 case SpvDecorationRestrict:
624 case SpvDecorationAliased:
625 case SpvDecorationConstant:
626 case SpvDecorationIndex:
627 case SpvDecorationBinding:
628 case SpvDecorationDescriptorSet:
629 case SpvDecorationLinkageAttributes:
630 case SpvDecorationNoContraction:
631 case SpvDecorationInputAttachmentIndex:
632 vtn_warn("Decoration not allowed on struct members: %s",
633 spirv_decoration_to_string(dec->decoration));
636 case SpvDecorationXfbBuffer:
637 case SpvDecorationXfbStride:
638 vtn_warn("Vulkan does not have transform feedback");
641 case SpvDecorationCPacked:
642 case SpvDecorationSaturatedConversion:
643 case SpvDecorationFuncParamAttr:
644 case SpvDecorationFPRoundingMode:
645 case SpvDecorationFPFastMathMode:
646 case SpvDecorationAlignment:
647 vtn_warn("Decoration only allowed for CL-style kernels: %s",
648 spirv_decoration_to_string(dec->decoration));
652 vtn_fail("Unhandled decoration");
656 /* Matrix strides are handled as a separate pass because we need to know
657 * whether the matrix is row-major or not first.
660 struct_member_matrix_stride_cb(struct vtn_builder *b,
661 struct vtn_value *val, int member,
662 const struct vtn_decoration *dec,
665 if (dec->decoration != SpvDecorationMatrixStride)
667 vtn_assert(member >= 0);
669 struct member_decoration_ctx *ctx = void_ctx;
671 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
672 if (mat_type->row_major) {
673 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
674 mat_type->stride = mat_type->array_element->stride;
675 mat_type->array_element->stride = dec->literals[0];
677 vtn_assert(mat_type->array_element->stride > 0);
678 mat_type->stride = dec->literals[0];
683 type_decoration_cb(struct vtn_builder *b,
684 struct vtn_value *val, int member,
685 const struct vtn_decoration *dec, void *ctx)
687 struct vtn_type *type = val->type;
692 switch (dec->decoration) {
693 case SpvDecorationArrayStride:
694 vtn_assert(type->base_type == vtn_base_type_matrix ||
695 type->base_type == vtn_base_type_array ||
696 type->base_type == vtn_base_type_pointer);
697 type->stride = dec->literals[0];
699 case SpvDecorationBlock:
700 vtn_assert(type->base_type == vtn_base_type_struct);
703 case SpvDecorationBufferBlock:
704 vtn_assert(type->base_type == vtn_base_type_struct);
705 type->buffer_block = true;
707 case SpvDecorationGLSLShared:
708 case SpvDecorationGLSLPacked:
709 /* Ignore these, since we get explicit offsets anyways */
712 case SpvDecorationRowMajor:
713 case SpvDecorationColMajor:
714 case SpvDecorationMatrixStride:
715 case SpvDecorationBuiltIn:
716 case SpvDecorationNoPerspective:
717 case SpvDecorationFlat:
718 case SpvDecorationPatch:
719 case SpvDecorationCentroid:
720 case SpvDecorationSample:
721 case SpvDecorationVolatile:
722 case SpvDecorationCoherent:
723 case SpvDecorationNonWritable:
724 case SpvDecorationNonReadable:
725 case SpvDecorationUniform:
726 case SpvDecorationStream:
727 case SpvDecorationLocation:
728 case SpvDecorationComponent:
729 case SpvDecorationOffset:
730 case SpvDecorationXfbBuffer:
731 case SpvDecorationXfbStride:
732 vtn_warn("Decoration only allowed for struct members: %s",
733 spirv_decoration_to_string(dec->decoration));
736 case SpvDecorationRelaxedPrecision:
737 case SpvDecorationSpecId:
738 case SpvDecorationInvariant:
739 case SpvDecorationRestrict:
740 case SpvDecorationAliased:
741 case SpvDecorationConstant:
742 case SpvDecorationIndex:
743 case SpvDecorationBinding:
744 case SpvDecorationDescriptorSet:
745 case SpvDecorationLinkageAttributes:
746 case SpvDecorationNoContraction:
747 case SpvDecorationInputAttachmentIndex:
748 vtn_warn("Decoration not allowed on types: %s",
749 spirv_decoration_to_string(dec->decoration));
752 case SpvDecorationCPacked:
753 case SpvDecorationSaturatedConversion:
754 case SpvDecorationFuncParamAttr:
755 case SpvDecorationFPRoundingMode:
756 case SpvDecorationFPFastMathMode:
757 case SpvDecorationAlignment:
758 vtn_warn("Decoration only allowed for CL-style kernels: %s",
759 spirv_decoration_to_string(dec->decoration));
763 vtn_fail("Unhandled decoration");
768 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
771 case SpvImageFormatUnknown: return 0; /* GL_NONE */
772 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
773 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
774 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
775 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
776 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
777 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
778 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
779 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
780 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
781 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
782 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
783 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
784 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
785 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
786 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
787 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
788 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
789 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
790 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
791 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
792 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
793 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
794 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
795 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
796 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
797 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
798 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
799 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
800 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
801 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
802 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
803 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
804 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
805 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
806 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
807 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
808 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
809 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
810 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
812 vtn_fail("Invalid image format");
816 static struct vtn_type *
817 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
818 uint32_t *size_out, uint32_t *align_out)
820 switch (type->base_type) {
821 case vtn_base_type_scalar: {
822 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
823 *size_out = comp_size;
824 *align_out = comp_size;
828 case vtn_base_type_vector: {
829 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
830 assert(type->length > 0 && type->length <= 4);
831 unsigned align_comps = type->length == 3 ? 4 : type->length;
832 *size_out = comp_size * type->length,
833 *align_out = comp_size * align_comps;
837 case vtn_base_type_matrix:
838 case vtn_base_type_array: {
839 /* We're going to add an array stride */
840 type = vtn_type_copy(b, type);
841 uint32_t elem_size, elem_align;
842 type->array_element = vtn_type_layout_std430(b, type->array_element,
843 &elem_size, &elem_align);
844 type->stride = vtn_align_u32(elem_size, elem_align);
845 *size_out = type->stride * type->length;
846 *align_out = elem_align;
850 case vtn_base_type_struct: {
851 /* We're going to add member offsets */
852 type = vtn_type_copy(b, type);
855 for (unsigned i = 0; i < type->length; i++) {
856 uint32_t mem_size, mem_align;
857 type->members[i] = vtn_type_layout_std430(b, type->members[i],
858 &mem_size, &mem_align);
859 offset = vtn_align_u32(offset, mem_align);
860 type->offsets[i] = offset;
862 align = MAX2(align, mem_align);
870 unreachable("Invalid SPIR-V type for std430");
875 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
876 const uint32_t *w, unsigned count)
878 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
880 val->type = rzalloc(b, struct vtn_type);
881 val->type->val = val;
885 val->type->base_type = vtn_base_type_void;
886 val->type->type = glsl_void_type();
889 val->type->base_type = vtn_base_type_scalar;
890 val->type->type = glsl_bool_type();
891 val->type->length = 1;
895 const bool signedness = w[3];
896 val->type->base_type = vtn_base_type_scalar;
899 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
902 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
905 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
908 vtn_fail("Invalid int bit size");
910 val->type->length = 1;
914 case SpvOpTypeFloat: {
916 val->type->base_type = vtn_base_type_scalar;
919 val->type->type = glsl_float16_t_type();
922 val->type->type = glsl_float_type();
925 val->type->type = glsl_double_type();
928 vtn_fail("Invalid float bit size");
930 val->type->length = 1;
934 case SpvOpTypeVector: {
935 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
936 unsigned elems = w[3];
938 vtn_fail_if(base->base_type != vtn_base_type_scalar,
939 "Base type for OpTypeVector must be a scalar");
940 vtn_fail_if(elems < 2 || elems > 4,
941 "Invalid component count for OpTypeVector");
943 val->type->base_type = vtn_base_type_vector;
944 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
945 val->type->length = elems;
946 val->type->stride = glsl_get_bit_size(base->type) / 8;
947 val->type->array_element = base;
951 case SpvOpTypeMatrix: {
952 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
953 unsigned columns = w[3];
955 vtn_fail_if(base->base_type != vtn_base_type_vector,
956 "Base type for OpTypeMatrix must be a vector");
957 vtn_fail_if(columns < 2 || columns > 4,
958 "Invalid column count for OpTypeMatrix");
960 val->type->base_type = vtn_base_type_matrix;
961 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
962 glsl_get_vector_elements(base->type),
964 vtn_fail_if(glsl_type_is_error(val->type->type),
965 "Unsupported base type for OpTypeMatrix");
966 assert(!glsl_type_is_error(val->type->type));
967 val->type->length = columns;
968 val->type->array_element = base;
969 val->type->row_major = false;
970 val->type->stride = 0;
974 case SpvOpTypeRuntimeArray:
975 case SpvOpTypeArray: {
976 struct vtn_type *array_element =
977 vtn_value(b, w[2], vtn_value_type_type)->type;
979 if (opcode == SpvOpTypeRuntimeArray) {
980 /* A length of 0 is used to denote unsized arrays */
981 val->type->length = 0;
984 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
987 val->type->base_type = vtn_base_type_array;
988 val->type->type = glsl_array_type(array_element->type, val->type->length);
989 val->type->array_element = array_element;
990 val->type->stride = 0;
994 case SpvOpTypeStruct: {
995 unsigned num_fields = count - 2;
996 val->type->base_type = vtn_base_type_struct;
997 val->type->length = num_fields;
998 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
999 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1001 NIR_VLA(struct glsl_struct_field, fields, count);
1002 for (unsigned i = 0; i < num_fields; i++) {
1003 val->type->members[i] =
1004 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1005 fields[i] = (struct glsl_struct_field) {
1006 .type = val->type->members[i]->type,
1007 .name = ralloc_asprintf(b, "field%d", i),
1012 struct member_decoration_ctx ctx = {
1013 .num_fields = num_fields,
1018 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1019 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1021 const char *name = val->name ? val->name : "struct";
1023 val->type->type = glsl_struct_type(fields, num_fields, name);
1027 case SpvOpTypeFunction: {
1028 val->type->base_type = vtn_base_type_function;
1029 val->type->type = NULL;
1031 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1033 const unsigned num_params = count - 3;
1034 val->type->length = num_params;
1035 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1036 for (unsigned i = 0; i < count - 3; i++) {
1037 val->type->params[i] =
1038 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1043 case SpvOpTypePointer: {
1044 SpvStorageClass storage_class = w[2];
1045 struct vtn_type *deref_type =
1046 vtn_value(b, w[3], vtn_value_type_type)->type;
1048 val->type->base_type = vtn_base_type_pointer;
1049 val->type->storage_class = storage_class;
1050 val->type->deref = deref_type;
1052 if (storage_class == SpvStorageClassUniform ||
1053 storage_class == SpvStorageClassStorageBuffer) {
1054 /* These can actually be stored to nir_variables and used as SSA
1055 * values so they need a real glsl_type.
1057 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1060 if (storage_class == SpvStorageClassWorkgroup &&
1061 b->options->lower_workgroup_access_to_offsets) {
1062 uint32_t size, align;
1063 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1065 val->type->length = size;
1066 val->type->align = align;
1067 /* These can actually be stored to nir_variables and used as SSA
1068 * values so they need a real glsl_type.
1070 val->type->type = glsl_uint_type();
1075 case SpvOpTypeImage: {
1076 val->type->base_type = vtn_base_type_image;
1078 const struct glsl_type *sampled_type =
1079 vtn_value(b, w[2], vtn_value_type_type)->type->type;
1081 vtn_assert(glsl_type_is_vector_or_scalar(sampled_type));
1083 enum glsl_sampler_dim dim;
1084 switch ((SpvDim)w[3]) {
1085 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1086 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1087 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1088 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1089 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1090 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1091 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1093 vtn_fail("Invalid SPIR-V Sampler dimension");
1096 bool is_shadow = w[4];
1097 bool is_array = w[5];
1098 bool multisampled = w[6];
1099 unsigned sampled = w[7];
1100 SpvImageFormat format = w[8];
1103 val->type->access_qualifier = w[9];
1105 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1108 if (dim == GLSL_SAMPLER_DIM_2D)
1109 dim = GLSL_SAMPLER_DIM_MS;
1110 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1111 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1113 vtn_fail("Unsupported multisampled image type");
1116 val->type->image_format = translate_image_format(b, format);
1119 val->type->sampled = true;
1120 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1121 glsl_get_base_type(sampled_type));
1122 } else if (sampled == 2) {
1123 vtn_assert(!is_shadow);
1124 val->type->sampled = false;
1125 val->type->type = glsl_image_type(dim, is_array,
1126 glsl_get_base_type(sampled_type));
1128 vtn_fail("We need to know if the image will be sampled");
1133 case SpvOpTypeSampledImage:
1134 val->type->base_type = vtn_base_type_sampled_image;
1135 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1136 val->type->type = val->type->image->type;
1139 case SpvOpTypeSampler:
1140 /* The actual sampler type here doesn't really matter. It gets
1141 * thrown away the moment you combine it with an image. What really
1142 * matters is that it's a sampler type as opposed to an integer type
1143 * so the backend knows what to do.
1145 val->type->base_type = vtn_base_type_sampler;
1146 val->type->type = glsl_bare_sampler_type();
1149 case SpvOpTypeOpaque:
1150 case SpvOpTypeEvent:
1151 case SpvOpTypeDeviceEvent:
1152 case SpvOpTypeReserveId:
1153 case SpvOpTypeQueue:
1156 vtn_fail("Unhandled opcode");
1159 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1162 static nir_constant *
1163 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1165 nir_constant *c = rzalloc(b, nir_constant);
1167 /* For pointers and other typeless things, we have to return something but
1168 * it doesn't matter what.
1173 switch (glsl_get_base_type(type)) {
1175 case GLSL_TYPE_UINT:
1176 case GLSL_TYPE_INT16:
1177 case GLSL_TYPE_UINT16:
1178 case GLSL_TYPE_INT64:
1179 case GLSL_TYPE_UINT64:
1180 case GLSL_TYPE_BOOL:
1181 case GLSL_TYPE_FLOAT:
1182 case GLSL_TYPE_FLOAT16:
1183 case GLSL_TYPE_DOUBLE:
1184 /* Nothing to do here. It's already initialized to zero */
1187 case GLSL_TYPE_ARRAY:
1188 vtn_assert(glsl_get_length(type) > 0);
1189 c->num_elements = glsl_get_length(type);
1190 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1192 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1193 for (unsigned i = 1; i < c->num_elements; i++)
1194 c->elements[i] = c->elements[0];
1197 case GLSL_TYPE_STRUCT:
1198 c->num_elements = glsl_get_length(type);
1199 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1201 for (unsigned i = 0; i < c->num_elements; i++) {
1202 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1207 vtn_fail("Invalid type for null constant");
1214 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1215 int member, const struct vtn_decoration *dec,
1218 vtn_assert(member == -1);
1219 if (dec->decoration != SpvDecorationSpecId)
1222 struct spec_constant_value *const_value = data;
1224 for (unsigned i = 0; i < b->num_specializations; i++) {
1225 if (b->specializations[i].id == dec->literals[0]) {
1226 if (const_value->is_double)
1227 const_value->data64 = b->specializations[i].data64;
1229 const_value->data32 = b->specializations[i].data32;
1236 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1237 uint32_t const_value)
1239 struct spec_constant_value data;
1240 data.is_double = false;
1241 data.data32 = const_value;
1242 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1247 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1248 uint64_t const_value)
1250 struct spec_constant_value data;
1251 data.is_double = true;
1252 data.data64 = const_value;
1253 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1258 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1259 struct vtn_value *val,
1261 const struct vtn_decoration *dec,
1264 vtn_assert(member == -1);
1265 if (dec->decoration != SpvDecorationBuiltIn ||
1266 dec->literals[0] != SpvBuiltInWorkgroupSize)
1269 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1271 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1272 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1273 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1277 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1278 const uint32_t *w, unsigned count)
1280 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1281 val->constant = rzalloc(b, nir_constant);
1283 case SpvOpConstantTrue:
1284 case SpvOpConstantFalse:
1285 case SpvOpSpecConstantTrue:
1286 case SpvOpSpecConstantFalse: {
1287 vtn_fail_if(val->type->type != glsl_bool_type(),
1288 "Result type of %s must be OpTypeBool",
1289 spirv_op_to_string(opcode));
1291 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1292 opcode == SpvOpSpecConstantTrue);
1294 if (opcode == SpvOpSpecConstantTrue ||
1295 opcode == SpvOpSpecConstantFalse)
1296 int_val = get_specialization(b, val, int_val);
1298 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1302 case SpvOpConstant: {
1303 vtn_assert(glsl_type_is_scalar(val->type->type));
1304 int bit_size = glsl_get_bit_size(val->type->type);
1307 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1310 val->constant->values->u32[0] = w[3];
1313 val->constant->values->u16[0] = w[3];
1316 vtn_fail("Unsupported SpvOpConstant bit size");
1320 case SpvOpSpecConstant: {
1321 vtn_assert(glsl_type_is_scalar(val->type->type));
1322 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1323 int bit_size = glsl_get_bit_size(val->type->type);
1326 val->constant->values[0].u64[0] =
1327 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1330 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1333 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1336 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1340 case SpvOpSpecConstantComposite:
1341 case SpvOpConstantComposite: {
1342 unsigned elem_count = count - 3;
1343 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1344 for (unsigned i = 0; i < elem_count; i++)
1345 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1347 switch (glsl_get_base_type(val->type->type)) {
1348 case GLSL_TYPE_UINT:
1350 case GLSL_TYPE_UINT16:
1351 case GLSL_TYPE_INT16:
1352 case GLSL_TYPE_UINT64:
1353 case GLSL_TYPE_INT64:
1354 case GLSL_TYPE_FLOAT:
1355 case GLSL_TYPE_FLOAT16:
1356 case GLSL_TYPE_BOOL:
1357 case GLSL_TYPE_DOUBLE: {
1358 int bit_size = glsl_get_bit_size(val->type->type);
1359 if (glsl_type_is_matrix(val->type->type)) {
1360 vtn_assert(glsl_get_matrix_columns(val->type->type) == elem_count);
1361 for (unsigned i = 0; i < elem_count; i++)
1362 val->constant->values[i] = elems[i]->values[0];
1364 vtn_assert(glsl_type_is_vector(val->type->type));
1365 vtn_assert(glsl_get_vector_elements(val->type->type) == elem_count);
1366 for (unsigned i = 0; i < elem_count; i++) {
1369 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1372 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1375 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1378 vtn_fail("Invalid SpvOpConstantComposite bit size");
1385 case GLSL_TYPE_STRUCT:
1386 case GLSL_TYPE_ARRAY:
1387 ralloc_steal(val->constant, elems);
1388 val->constant->num_elements = elem_count;
1389 val->constant->elements = elems;
1393 vtn_fail("Unsupported type for constants");
1398 case SpvOpSpecConstantOp: {
1399 SpvOp opcode = get_specialization(b, val, w[3]);
1401 case SpvOpVectorShuffle: {
1402 struct vtn_value *v0 = &b->values[w[4]];
1403 struct vtn_value *v1 = &b->values[w[5]];
1405 vtn_assert(v0->value_type == vtn_value_type_constant ||
1406 v0->value_type == vtn_value_type_undef);
1407 vtn_assert(v1->value_type == vtn_value_type_constant ||
1408 v1->value_type == vtn_value_type_undef);
1410 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1411 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1413 vtn_assert(len0 + len1 < 16);
1415 unsigned bit_size = glsl_get_bit_size(val->type->type);
1416 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1417 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1419 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1420 (void)bit_size0; (void)bit_size1;
1422 if (bit_size == 64) {
1424 if (v0->value_type == vtn_value_type_constant) {
1425 for (unsigned i = 0; i < len0; i++)
1426 u64[i] = v0->constant->values[0].u64[i];
1428 if (v1->value_type == vtn_value_type_constant) {
1429 for (unsigned i = 0; i < len1; i++)
1430 u64[len0 + i] = v1->constant->values[0].u64[i];
1433 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1434 uint32_t comp = w[i + 6];
1435 /* If component is not used, set the value to a known constant
1436 * to detect if it is wrongly used.
1438 if (comp == (uint32_t)-1)
1439 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1441 val->constant->values[0].u64[j] = u64[comp];
1444 /* This is for both 32-bit and 16-bit values */
1446 if (v0->value_type == vtn_value_type_constant) {
1447 for (unsigned i = 0; i < len0; i++)
1448 u32[i] = v0->constant->values[0].u32[i];
1450 if (v1->value_type == vtn_value_type_constant) {
1451 for (unsigned i = 0; i < len1; i++)
1452 u32[len0 + i] = v1->constant->values[0].u32[i];
1455 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1456 uint32_t comp = w[i + 6];
1457 /* If component is not used, set the value to a known constant
1458 * to detect if it is wrongly used.
1460 if (comp == (uint32_t)-1)
1461 val->constant->values[0].u32[j] = 0xdeadbeef;
1463 val->constant->values[0].u32[j] = u32[comp];
1469 case SpvOpCompositeExtract:
1470 case SpvOpCompositeInsert: {
1471 struct vtn_value *comp;
1472 unsigned deref_start;
1473 struct nir_constant **c;
1474 if (opcode == SpvOpCompositeExtract) {
1475 comp = vtn_value(b, w[4], vtn_value_type_constant);
1477 c = &comp->constant;
1479 comp = vtn_value(b, w[5], vtn_value_type_constant);
1481 val->constant = nir_constant_clone(comp->constant,
1488 const struct glsl_type *type = comp->type->type;
1489 for (unsigned i = deref_start; i < count; i++) {
1490 switch (glsl_get_base_type(type)) {
1491 case GLSL_TYPE_UINT:
1493 case GLSL_TYPE_UINT16:
1494 case GLSL_TYPE_INT16:
1495 case GLSL_TYPE_UINT64:
1496 case GLSL_TYPE_INT64:
1497 case GLSL_TYPE_FLOAT:
1498 case GLSL_TYPE_FLOAT16:
1499 case GLSL_TYPE_DOUBLE:
1500 case GLSL_TYPE_BOOL:
1501 /* If we hit this granularity, we're picking off an element */
1502 if (glsl_type_is_matrix(type)) {
1503 vtn_assert(col == 0 && elem == -1);
1506 type = glsl_get_column_type(type);
1508 vtn_assert(elem <= 0 && glsl_type_is_vector(type));
1510 type = glsl_scalar_type(glsl_get_base_type(type));
1514 case GLSL_TYPE_ARRAY:
1515 c = &(*c)->elements[w[i]];
1516 type = glsl_get_array_element(type);
1519 case GLSL_TYPE_STRUCT:
1520 c = &(*c)->elements[w[i]];
1521 type = glsl_get_struct_field(type, w[i]);
1525 vtn_fail("Invalid constant type");
1529 if (opcode == SpvOpCompositeExtract) {
1533 unsigned num_components = glsl_get_vector_elements(type);
1534 unsigned bit_size = glsl_get_bit_size(type);
1535 for (unsigned i = 0; i < num_components; i++)
1538 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1541 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1544 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1547 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1551 struct vtn_value *insert =
1552 vtn_value(b, w[4], vtn_value_type_constant);
1553 vtn_assert(insert->type->type == type);
1555 *c = insert->constant;
1557 unsigned num_components = glsl_get_vector_elements(type);
1558 unsigned bit_size = glsl_get_bit_size(type);
1559 for (unsigned i = 0; i < num_components; i++)
1562 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1565 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1568 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1571 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1580 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1581 nir_alu_type src_alu_type = dst_alu_type;
1582 unsigned num_components = glsl_get_vector_elements(val->type->type);
1585 vtn_assert(count <= 7);
1590 /* We have a source in a conversion */
1592 nir_get_nir_type_for_glsl_type(
1593 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1594 /* We use the bitsize of the conversion source to evaluate the opcode later */
1595 bit_size = glsl_get_bit_size(
1596 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1599 bit_size = glsl_get_bit_size(val->type->type);
1602 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1605 nir_const_value src[4];
1607 for (unsigned i = 0; i < count - 4; i++) {
1609 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1611 unsigned j = swap ? 1 - i : i;
1612 src[j] = c->values[0];
1615 val->constant->values[0] =
1616 nir_eval_const_opcode(op, num_components, bit_size, src);
1623 case SpvOpConstantNull:
1624 val->constant = vtn_null_constant(b, val->type->type);
1627 case SpvOpConstantSampler:
1628 vtn_fail("OpConstantSampler requires Kernel Capability");
1632 vtn_fail("Unhandled opcode");
1635 /* Now that we have the value, update the workgroup size if needed */
1636 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1640 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1641 const uint32_t *w, unsigned count)
1643 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1644 struct vtn_function *vtn_callee =
1645 vtn_value(b, w[3], vtn_value_type_function)->func;
1646 struct nir_function *callee = vtn_callee->impl->function;
1648 vtn_callee->referenced = true;
1650 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1651 for (unsigned i = 0; i < call->num_params; i++) {
1652 unsigned arg_id = w[4 + i];
1653 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1654 if (arg->value_type == vtn_value_type_pointer &&
1655 arg->pointer->ptr_type->type == NULL) {
1656 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1657 call->params[i] = nir_deref_var_clone(d, call);
1659 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1661 /* Make a temporary to store the argument in */
1663 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1664 call->params[i] = nir_deref_var_create(call, tmp);
1666 vtn_local_store(b, arg_ssa, call->params[i]);
1670 nir_variable *out_tmp = NULL;
1671 vtn_assert(res_type->type == callee->return_type);
1672 if (!glsl_type_is_void(callee->return_type)) {
1673 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1675 call->return_deref = nir_deref_var_create(call, out_tmp);
1678 nir_builder_instr_insert(&b->nb, &call->instr);
1680 if (glsl_type_is_void(callee->return_type)) {
1681 vtn_push_value(b, w[2], vtn_value_type_undef);
1683 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1687 struct vtn_ssa_value *
1688 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1690 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1693 if (!glsl_type_is_vector_or_scalar(type)) {
1694 unsigned elems = glsl_get_length(type);
1695 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1696 for (unsigned i = 0; i < elems; i++) {
1697 const struct glsl_type *child_type;
1699 switch (glsl_get_base_type(type)) {
1701 case GLSL_TYPE_UINT:
1702 case GLSL_TYPE_INT16:
1703 case GLSL_TYPE_UINT16:
1704 case GLSL_TYPE_INT64:
1705 case GLSL_TYPE_UINT64:
1706 case GLSL_TYPE_BOOL:
1707 case GLSL_TYPE_FLOAT:
1708 case GLSL_TYPE_FLOAT16:
1709 case GLSL_TYPE_DOUBLE:
1710 child_type = glsl_get_column_type(type);
1712 case GLSL_TYPE_ARRAY:
1713 child_type = glsl_get_array_element(type);
1715 case GLSL_TYPE_STRUCT:
1716 child_type = glsl_get_struct_field(type, i);
1719 vtn_fail("unkown base type");
1722 val->elems[i] = vtn_create_ssa_value(b, child_type);
1730 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1733 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1734 src.src_type = type;
1739 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1740 const uint32_t *w, unsigned count)
1742 if (opcode == SpvOpSampledImage) {
1743 struct vtn_value *val =
1744 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1745 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1746 val->sampled_image->type =
1747 vtn_value(b, w[1], vtn_value_type_type)->type;
1748 val->sampled_image->image =
1749 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1750 val->sampled_image->sampler =
1751 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1753 } else if (opcode == SpvOpImage) {
1754 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1755 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1756 if (src_val->value_type == vtn_value_type_sampled_image) {
1757 val->pointer = src_val->sampled_image->image;
1759 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1760 val->pointer = src_val->pointer;
1765 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1766 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1768 struct vtn_sampled_image sampled;
1769 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1770 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1771 sampled = *sampled_val->sampled_image;
1773 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1774 sampled.type = sampled_val->pointer->type;
1775 sampled.image = NULL;
1776 sampled.sampler = sampled_val->pointer;
1779 const struct glsl_type *image_type = sampled.type->type;
1780 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1781 const bool is_array = glsl_sampler_type_is_array(image_type);
1782 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1784 /* Figure out the base texture operation */
1787 case SpvOpImageSampleImplicitLod:
1788 case SpvOpImageSampleDrefImplicitLod:
1789 case SpvOpImageSampleProjImplicitLod:
1790 case SpvOpImageSampleProjDrefImplicitLod:
1791 texop = nir_texop_tex;
1794 case SpvOpImageSampleExplicitLod:
1795 case SpvOpImageSampleDrefExplicitLod:
1796 case SpvOpImageSampleProjExplicitLod:
1797 case SpvOpImageSampleProjDrefExplicitLod:
1798 texop = nir_texop_txl;
1801 case SpvOpImageFetch:
1802 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1803 texop = nir_texop_txf_ms;
1805 texop = nir_texop_txf;
1809 case SpvOpImageGather:
1810 case SpvOpImageDrefGather:
1811 texop = nir_texop_tg4;
1814 case SpvOpImageQuerySizeLod:
1815 case SpvOpImageQuerySize:
1816 texop = nir_texop_txs;
1819 case SpvOpImageQueryLod:
1820 texop = nir_texop_lod;
1823 case SpvOpImageQueryLevels:
1824 texop = nir_texop_query_levels;
1827 case SpvOpImageQuerySamples:
1828 texop = nir_texop_texture_samples;
1832 vtn_fail("Unhandled opcode");
1835 nir_tex_src srcs[8]; /* 8 should be enough */
1836 nir_tex_src *p = srcs;
1840 struct nir_ssa_def *coord;
1841 unsigned coord_components;
1843 case SpvOpImageSampleImplicitLod:
1844 case SpvOpImageSampleExplicitLod:
1845 case SpvOpImageSampleDrefImplicitLod:
1846 case SpvOpImageSampleDrefExplicitLod:
1847 case SpvOpImageSampleProjImplicitLod:
1848 case SpvOpImageSampleProjExplicitLod:
1849 case SpvOpImageSampleProjDrefImplicitLod:
1850 case SpvOpImageSampleProjDrefExplicitLod:
1851 case SpvOpImageFetch:
1852 case SpvOpImageGather:
1853 case SpvOpImageDrefGather:
1854 case SpvOpImageQueryLod: {
1855 /* All these types have the coordinate as their first real argument */
1856 switch (sampler_dim) {
1857 case GLSL_SAMPLER_DIM_1D:
1858 case GLSL_SAMPLER_DIM_BUF:
1859 coord_components = 1;
1861 case GLSL_SAMPLER_DIM_2D:
1862 case GLSL_SAMPLER_DIM_RECT:
1863 case GLSL_SAMPLER_DIM_MS:
1864 coord_components = 2;
1866 case GLSL_SAMPLER_DIM_3D:
1867 case GLSL_SAMPLER_DIM_CUBE:
1868 coord_components = 3;
1871 vtn_fail("Invalid sampler type");
1874 if (is_array && texop != nir_texop_lod)
1877 coord = vtn_ssa_value(b, w[idx++])->def;
1878 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1879 (1 << coord_components) - 1));
1880 p->src_type = nir_tex_src_coord;
1887 coord_components = 0;
1892 case SpvOpImageSampleProjImplicitLod:
1893 case SpvOpImageSampleProjExplicitLod:
1894 case SpvOpImageSampleProjDrefImplicitLod:
1895 case SpvOpImageSampleProjDrefExplicitLod:
1896 /* These have the projector as the last coordinate component */
1897 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1898 p->src_type = nir_tex_src_projector;
1906 unsigned gather_component = 0;
1908 case SpvOpImageSampleDrefImplicitLod:
1909 case SpvOpImageSampleDrefExplicitLod:
1910 case SpvOpImageSampleProjDrefImplicitLod:
1911 case SpvOpImageSampleProjDrefExplicitLod:
1912 case SpvOpImageDrefGather:
1913 /* These all have an explicit depth value as their next source */
1914 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
1917 case SpvOpImageGather:
1918 /* This has a component as its next source */
1920 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
1927 /* For OpImageQuerySizeLod, we always have an LOD */
1928 if (opcode == SpvOpImageQuerySizeLod)
1929 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1931 /* Now we need to handle some number of optional arguments */
1932 const struct vtn_ssa_value *gather_offsets = NULL;
1934 uint32_t operands = w[idx++];
1936 if (operands & SpvImageOperandsBiasMask) {
1937 vtn_assert(texop == nir_texop_tex);
1938 texop = nir_texop_txb;
1939 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1942 if (operands & SpvImageOperandsLodMask) {
1943 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
1944 texop == nir_texop_txs);
1945 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1948 if (operands & SpvImageOperandsGradMask) {
1949 vtn_assert(texop == nir_texop_txl);
1950 texop = nir_texop_txd;
1951 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
1952 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
1955 if (operands & SpvImageOperandsOffsetMask ||
1956 operands & SpvImageOperandsConstOffsetMask)
1957 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
1959 if (operands & SpvImageOperandsConstOffsetsMask) {
1960 gather_offsets = vtn_ssa_value(b, w[idx++]);
1961 (*p++) = (nir_tex_src){};
1964 if (operands & SpvImageOperandsSampleMask) {
1965 vtn_assert(texop == nir_texop_txf_ms);
1966 texop = nir_texop_txf_ms;
1967 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
1970 /* We should have now consumed exactly all of the arguments */
1971 vtn_assert(idx == count);
1973 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
1976 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
1978 instr->coord_components = coord_components;
1979 instr->sampler_dim = sampler_dim;
1980 instr->is_array = is_array;
1981 instr->is_shadow = is_shadow;
1982 instr->is_new_style_shadow =
1983 is_shadow && glsl_get_components(ret_type->type) == 1;
1984 instr->component = gather_component;
1986 switch (glsl_get_sampler_result_type(image_type)) {
1987 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
1988 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
1989 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
1990 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
1992 vtn_fail("Invalid base type for sampler result");
1995 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
1996 nir_deref_var *texture;
1997 if (sampled.image) {
1998 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
2004 instr->texture = nir_deref_var_clone(texture, instr);
2006 switch (instr->op) {
2012 /* These operations require a sampler */
2013 instr->sampler = nir_deref_var_clone(sampler, instr);
2016 case nir_texop_txf_ms:
2019 case nir_texop_query_levels:
2020 case nir_texop_texture_samples:
2021 case nir_texop_samples_identical:
2023 instr->sampler = NULL;
2025 case nir_texop_txf_ms_mcs:
2026 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2029 nir_ssa_dest_init(&instr->instr, &instr->dest,
2030 nir_tex_instr_dest_size(instr), 32, NULL);
2032 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2033 nir_tex_instr_dest_size(instr));
2036 nir_instr *instruction;
2037 if (gather_offsets) {
2038 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2039 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2040 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2042 /* Copy the current instruction 4x */
2043 for (uint32_t i = 1; i < 4; i++) {
2044 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2045 instrs[i]->op = instr->op;
2046 instrs[i]->coord_components = instr->coord_components;
2047 instrs[i]->sampler_dim = instr->sampler_dim;
2048 instrs[i]->is_array = instr->is_array;
2049 instrs[i]->is_shadow = instr->is_shadow;
2050 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2051 instrs[i]->component = instr->component;
2052 instrs[i]->dest_type = instr->dest_type;
2053 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
2054 instrs[i]->sampler = NULL;
2056 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2058 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2059 nir_tex_instr_dest_size(instr), 32, NULL);
2062 /* Fill in the last argument with the offset from the passed in offsets
2063 * and insert the instruction into the stream.
2065 for (uint32_t i = 0; i < 4; i++) {
2067 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2068 src.src_type = nir_tex_src_offset;
2069 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2070 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2073 /* Combine the results of the 4 instructions by taking their .w
2076 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2077 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2078 vec4->dest.write_mask = 0xf;
2079 for (uint32_t i = 0; i < 4; i++) {
2080 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2081 vec4->src[i].swizzle[0] = 3;
2083 def = &vec4->dest.dest.ssa;
2084 instruction = &vec4->instr;
2086 def = &instr->dest.ssa;
2087 instruction = &instr->instr;
2090 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2091 val->ssa->def = def;
2093 nir_builder_instr_insert(&b->nb, instruction);
2097 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2098 const uint32_t *w, nir_src *src)
2101 case SpvOpAtomicIIncrement:
2102 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2105 case SpvOpAtomicIDecrement:
2106 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2109 case SpvOpAtomicISub:
2111 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2114 case SpvOpAtomicCompareExchange:
2115 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2116 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2119 case SpvOpAtomicExchange:
2120 case SpvOpAtomicIAdd:
2121 case SpvOpAtomicSMin:
2122 case SpvOpAtomicUMin:
2123 case SpvOpAtomicSMax:
2124 case SpvOpAtomicUMax:
2125 case SpvOpAtomicAnd:
2127 case SpvOpAtomicXor:
2128 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2132 vtn_fail("Invalid SPIR-V atomic");
2136 static nir_ssa_def *
2137 get_image_coord(struct vtn_builder *b, uint32_t value)
2139 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2141 /* The image_load_store intrinsics assume a 4-dim coordinate */
2142 unsigned dim = glsl_get_vector_elements(coord->type);
2143 unsigned swizzle[4];
2144 for (unsigned i = 0; i < 4; i++)
2145 swizzle[i] = MIN2(i, dim - 1);
2147 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2151 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2152 const uint32_t *w, unsigned count)
2154 /* Just get this one out of the way */
2155 if (opcode == SpvOpImageTexelPointer) {
2156 struct vtn_value *val =
2157 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2158 val->image = ralloc(b, struct vtn_image_pointer);
2160 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2161 val->image->coord = get_image_coord(b, w[4]);
2162 val->image->sample = vtn_ssa_value(b, w[5])->def;
2166 struct vtn_image_pointer image;
2169 case SpvOpAtomicExchange:
2170 case SpvOpAtomicCompareExchange:
2171 case SpvOpAtomicCompareExchangeWeak:
2172 case SpvOpAtomicIIncrement:
2173 case SpvOpAtomicIDecrement:
2174 case SpvOpAtomicIAdd:
2175 case SpvOpAtomicISub:
2176 case SpvOpAtomicLoad:
2177 case SpvOpAtomicSMin:
2178 case SpvOpAtomicUMin:
2179 case SpvOpAtomicSMax:
2180 case SpvOpAtomicUMax:
2181 case SpvOpAtomicAnd:
2183 case SpvOpAtomicXor:
2184 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2187 case SpvOpAtomicStore:
2188 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2191 case SpvOpImageQuerySize:
2192 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2194 image.sample = NULL;
2197 case SpvOpImageRead:
2198 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2199 image.coord = get_image_coord(b, w[4]);
2201 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2202 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2203 image.sample = vtn_ssa_value(b, w[6])->def;
2205 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2209 case SpvOpImageWrite:
2210 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2211 image.coord = get_image_coord(b, w[2]);
2215 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2216 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2217 image.sample = vtn_ssa_value(b, w[5])->def;
2219 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2224 vtn_fail("Invalid image opcode");
2227 nir_intrinsic_op op;
2229 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2230 OP(ImageQuerySize, size)
2232 OP(ImageWrite, store)
2233 OP(AtomicLoad, load)
2234 OP(AtomicStore, store)
2235 OP(AtomicExchange, atomic_exchange)
2236 OP(AtomicCompareExchange, atomic_comp_swap)
2237 OP(AtomicIIncrement, atomic_add)
2238 OP(AtomicIDecrement, atomic_add)
2239 OP(AtomicIAdd, atomic_add)
2240 OP(AtomicISub, atomic_add)
2241 OP(AtomicSMin, atomic_min)
2242 OP(AtomicUMin, atomic_min)
2243 OP(AtomicSMax, atomic_max)
2244 OP(AtomicUMax, atomic_max)
2245 OP(AtomicAnd, atomic_and)
2246 OP(AtomicOr, atomic_or)
2247 OP(AtomicXor, atomic_xor)
2250 vtn_fail("Invalid image opcode");
2253 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2255 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2256 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2258 /* ImageQuerySize doesn't take any extra parameters */
2259 if (opcode != SpvOpImageQuerySize) {
2260 /* The image coordinate is always 4 components but we may not have that
2261 * many. Swizzle to compensate.
2264 for (unsigned i = 0; i < 4; i++)
2265 swiz[i] = i < image.coord->num_components ? i : 0;
2266 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2268 intrin->src[1] = nir_src_for_ssa(image.sample);
2272 case SpvOpAtomicLoad:
2273 case SpvOpImageQuerySize:
2274 case SpvOpImageRead:
2276 case SpvOpAtomicStore:
2277 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2279 case SpvOpImageWrite:
2280 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2283 case SpvOpAtomicCompareExchange:
2284 case SpvOpAtomicIIncrement:
2285 case SpvOpAtomicIDecrement:
2286 case SpvOpAtomicExchange:
2287 case SpvOpAtomicIAdd:
2288 case SpvOpAtomicISub:
2289 case SpvOpAtomicSMin:
2290 case SpvOpAtomicUMin:
2291 case SpvOpAtomicSMax:
2292 case SpvOpAtomicUMax:
2293 case SpvOpAtomicAnd:
2295 case SpvOpAtomicXor:
2296 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2300 vtn_fail("Invalid image opcode");
2303 if (opcode != SpvOpImageWrite) {
2304 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2305 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2307 unsigned dest_components =
2308 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2309 if (intrin->intrinsic == nir_intrinsic_image_size) {
2310 dest_components = intrin->num_components =
2311 glsl_get_vector_elements(type->type);
2314 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2315 dest_components, 32, NULL);
2317 nir_builder_instr_insert(&b->nb, &intrin->instr);
2319 val->ssa = vtn_create_ssa_value(b, type->type);
2320 val->ssa->def = &intrin->dest.ssa;
2322 nir_builder_instr_insert(&b->nb, &intrin->instr);
2326 static nir_intrinsic_op
2327 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2330 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2331 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2332 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2333 OP(AtomicExchange, atomic_exchange)
2334 OP(AtomicCompareExchange, atomic_comp_swap)
2335 OP(AtomicIIncrement, atomic_add)
2336 OP(AtomicIDecrement, atomic_add)
2337 OP(AtomicIAdd, atomic_add)
2338 OP(AtomicISub, atomic_add)
2339 OP(AtomicSMin, atomic_imin)
2340 OP(AtomicUMin, atomic_umin)
2341 OP(AtomicSMax, atomic_imax)
2342 OP(AtomicUMax, atomic_umax)
2343 OP(AtomicAnd, atomic_and)
2344 OP(AtomicOr, atomic_or)
2345 OP(AtomicXor, atomic_xor)
2348 vtn_fail("Invalid SSBO atomic");
2352 static nir_intrinsic_op
2353 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2356 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2357 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2358 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2359 OP(AtomicExchange, atomic_exchange)
2360 OP(AtomicCompareExchange, atomic_comp_swap)
2361 OP(AtomicIIncrement, atomic_add)
2362 OP(AtomicIDecrement, atomic_add)
2363 OP(AtomicIAdd, atomic_add)
2364 OP(AtomicISub, atomic_add)
2365 OP(AtomicSMin, atomic_imin)
2366 OP(AtomicUMin, atomic_umin)
2367 OP(AtomicSMax, atomic_imax)
2368 OP(AtomicUMax, atomic_umax)
2369 OP(AtomicAnd, atomic_and)
2370 OP(AtomicOr, atomic_or)
2371 OP(AtomicXor, atomic_xor)
2374 vtn_fail("Invalid shared atomic");
2378 static nir_intrinsic_op
2379 get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2382 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2383 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2384 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2385 OP(AtomicExchange, atomic_exchange)
2386 OP(AtomicCompareExchange, atomic_comp_swap)
2387 OP(AtomicIIncrement, atomic_add)
2388 OP(AtomicIDecrement, atomic_add)
2389 OP(AtomicIAdd, atomic_add)
2390 OP(AtomicISub, atomic_add)
2391 OP(AtomicSMin, atomic_imin)
2392 OP(AtomicUMin, atomic_umin)
2393 OP(AtomicSMax, atomic_imax)
2394 OP(AtomicUMax, atomic_umax)
2395 OP(AtomicAnd, atomic_and)
2396 OP(AtomicOr, atomic_or)
2397 OP(AtomicXor, atomic_xor)
2400 vtn_fail("Invalid shared atomic");
2405 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2406 const uint32_t *w, unsigned count)
2408 struct vtn_pointer *ptr;
2409 nir_intrinsic_instr *atomic;
2412 case SpvOpAtomicLoad:
2413 case SpvOpAtomicExchange:
2414 case SpvOpAtomicCompareExchange:
2415 case SpvOpAtomicCompareExchangeWeak:
2416 case SpvOpAtomicIIncrement:
2417 case SpvOpAtomicIDecrement:
2418 case SpvOpAtomicIAdd:
2419 case SpvOpAtomicISub:
2420 case SpvOpAtomicSMin:
2421 case SpvOpAtomicUMin:
2422 case SpvOpAtomicSMax:
2423 case SpvOpAtomicUMax:
2424 case SpvOpAtomicAnd:
2426 case SpvOpAtomicXor:
2427 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2430 case SpvOpAtomicStore:
2431 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2435 vtn_fail("Invalid SPIR-V atomic");
2439 SpvScope scope = w[4];
2440 SpvMemorySemanticsMask semantics = w[5];
2443 if (ptr->mode == vtn_variable_mode_workgroup &&
2444 !b->options->lower_workgroup_access_to_offsets) {
2445 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2446 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2447 nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
2448 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2449 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2452 case SpvOpAtomicLoad:
2453 atomic->num_components = glsl_get_vector_elements(deref_type);
2456 case SpvOpAtomicStore:
2457 atomic->num_components = glsl_get_vector_elements(deref_type);
2458 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2459 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2462 case SpvOpAtomicExchange:
2463 case SpvOpAtomicCompareExchange:
2464 case SpvOpAtomicCompareExchangeWeak:
2465 case SpvOpAtomicIIncrement:
2466 case SpvOpAtomicIDecrement:
2467 case SpvOpAtomicIAdd:
2468 case SpvOpAtomicISub:
2469 case SpvOpAtomicSMin:
2470 case SpvOpAtomicUMin:
2471 case SpvOpAtomicSMax:
2472 case SpvOpAtomicUMax:
2473 case SpvOpAtomicAnd:
2475 case SpvOpAtomicXor:
2476 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2480 vtn_fail("Invalid SPIR-V atomic");
2484 nir_ssa_def *offset, *index;
2485 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2487 nir_intrinsic_op op;
2488 if (ptr->mode == vtn_variable_mode_ssbo) {
2489 op = get_ssbo_nir_atomic_op(b, opcode);
2491 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2492 b->options->lower_workgroup_access_to_offsets);
2493 op = get_shared_nir_atomic_op(b, opcode);
2496 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2500 case SpvOpAtomicLoad:
2501 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2502 if (ptr->mode == vtn_variable_mode_ssbo)
2503 atomic->src[src++] = nir_src_for_ssa(index);
2504 atomic->src[src++] = nir_src_for_ssa(offset);
2507 case SpvOpAtomicStore:
2508 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2509 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2510 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2511 if (ptr->mode == vtn_variable_mode_ssbo)
2512 atomic->src[src++] = nir_src_for_ssa(index);
2513 atomic->src[src++] = nir_src_for_ssa(offset);
2516 case SpvOpAtomicExchange:
2517 case SpvOpAtomicCompareExchange:
2518 case SpvOpAtomicCompareExchangeWeak:
2519 case SpvOpAtomicIIncrement:
2520 case SpvOpAtomicIDecrement:
2521 case SpvOpAtomicIAdd:
2522 case SpvOpAtomicISub:
2523 case SpvOpAtomicSMin:
2524 case SpvOpAtomicUMin:
2525 case SpvOpAtomicSMax:
2526 case SpvOpAtomicUMax:
2527 case SpvOpAtomicAnd:
2529 case SpvOpAtomicXor:
2530 if (ptr->mode == vtn_variable_mode_ssbo)
2531 atomic->src[src++] = nir_src_for_ssa(index);
2532 atomic->src[src++] = nir_src_for_ssa(offset);
2533 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2537 vtn_fail("Invalid SPIR-V atomic");
2541 if (opcode != SpvOpAtomicStore) {
2542 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2544 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2545 glsl_get_vector_elements(type->type),
2546 glsl_get_bit_size(type->type), NULL);
2548 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2549 val->ssa = rzalloc(b, struct vtn_ssa_value);
2550 val->ssa->def = &atomic->dest.ssa;
2551 val->ssa->type = type->type;
2554 nir_builder_instr_insert(&b->nb, &atomic->instr);
2557 static nir_alu_instr *
2558 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2561 switch (num_components) {
2562 case 1: op = nir_op_fmov; break;
2563 case 2: op = nir_op_vec2; break;
2564 case 3: op = nir_op_vec3; break;
2565 case 4: op = nir_op_vec4; break;
2566 default: vtn_fail("bad vector size");
2569 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2570 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2572 vec->dest.write_mask = (1 << num_components) - 1;
2577 struct vtn_ssa_value *
2578 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2580 if (src->transposed)
2581 return src->transposed;
2583 struct vtn_ssa_value *dest =
2584 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2586 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2587 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2588 glsl_get_bit_size(src->type));
2589 if (glsl_type_is_vector_or_scalar(src->type)) {
2590 vec->src[0].src = nir_src_for_ssa(src->def);
2591 vec->src[0].swizzle[0] = i;
2593 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2594 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2595 vec->src[j].swizzle[0] = i;
2598 nir_builder_instr_insert(&b->nb, &vec->instr);
2599 dest->elems[i]->def = &vec->dest.dest.ssa;
2602 dest->transposed = src;
2608 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2610 unsigned swiz[4] = { index };
2611 return nir_swizzle(&b->nb, src, swiz, 1, true);
2615 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2618 nir_alu_instr *vec = create_vec(b, src->num_components,
2621 for (unsigned i = 0; i < src->num_components; i++) {
2623 vec->src[i].src = nir_src_for_ssa(insert);
2625 vec->src[i].src = nir_src_for_ssa(src);
2626 vec->src[i].swizzle[0] = i;
2630 nir_builder_instr_insert(&b->nb, &vec->instr);
2632 return &vec->dest.dest.ssa;
2636 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2639 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2640 for (unsigned i = 1; i < src->num_components; i++)
2641 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2642 vtn_vector_extract(b, src, i), dest);
2648 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2649 nir_ssa_def *insert, nir_ssa_def *index)
2651 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2652 for (unsigned i = 1; i < src->num_components; i++)
2653 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2654 vtn_vector_insert(b, src, insert, i), dest);
2659 static nir_ssa_def *
2660 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2661 nir_ssa_def *src0, nir_ssa_def *src1,
2662 const uint32_t *indices)
2664 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2666 for (unsigned i = 0; i < num_components; i++) {
2667 uint32_t index = indices[i];
2668 if (index == 0xffffffff) {
2670 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2671 } else if (index < src0->num_components) {
2672 vec->src[i].src = nir_src_for_ssa(src0);
2673 vec->src[i].swizzle[0] = index;
2675 vec->src[i].src = nir_src_for_ssa(src1);
2676 vec->src[i].swizzle[0] = index - src0->num_components;
2680 nir_builder_instr_insert(&b->nb, &vec->instr);
2682 return &vec->dest.dest.ssa;
2686 * Concatentates a number of vectors/scalars together to produce a vector
2688 static nir_ssa_def *
2689 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2690 unsigned num_srcs, nir_ssa_def **srcs)
2692 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2694 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2696 * "When constructing a vector, there must be at least two Constituent
2699 vtn_assert(num_srcs >= 2);
2701 unsigned dest_idx = 0;
2702 for (unsigned i = 0; i < num_srcs; i++) {
2703 nir_ssa_def *src = srcs[i];
2704 vtn_assert(dest_idx + src->num_components <= num_components);
2705 for (unsigned j = 0; j < src->num_components; j++) {
2706 vec->src[dest_idx].src = nir_src_for_ssa(src);
2707 vec->src[dest_idx].swizzle[0] = j;
2712 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2714 * "When constructing a vector, the total number of components in all
2715 * the operands must equal the number of components in Result Type."
2717 vtn_assert(dest_idx == num_components);
2719 nir_builder_instr_insert(&b->nb, &vec->instr);
2721 return &vec->dest.dest.ssa;
2724 static struct vtn_ssa_value *
2725 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2727 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2728 dest->type = src->type;
2730 if (glsl_type_is_vector_or_scalar(src->type)) {
2731 dest->def = src->def;
2733 unsigned elems = glsl_get_length(src->type);
2735 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2736 for (unsigned i = 0; i < elems; i++)
2737 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2743 static struct vtn_ssa_value *
2744 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2745 struct vtn_ssa_value *insert, const uint32_t *indices,
2746 unsigned num_indices)
2748 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2750 struct vtn_ssa_value *cur = dest;
2752 for (i = 0; i < num_indices - 1; i++) {
2753 cur = cur->elems[indices[i]];
2756 if (glsl_type_is_vector_or_scalar(cur->type)) {
2757 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2758 * the component granularity. In that case, the last index will be
2759 * the index to insert the scalar into the vector.
2762 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2764 cur->elems[indices[i]] = insert;
2770 static struct vtn_ssa_value *
2771 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2772 const uint32_t *indices, unsigned num_indices)
2774 struct vtn_ssa_value *cur = src;
2775 for (unsigned i = 0; i < num_indices; i++) {
2776 if (glsl_type_is_vector_or_scalar(cur->type)) {
2777 vtn_assert(i == num_indices - 1);
2778 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2779 * the component granularity. The last index will be the index of the
2780 * vector to extract.
2783 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2784 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2785 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2788 cur = cur->elems[indices[i]];
2796 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2797 const uint32_t *w, unsigned count)
2799 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2800 const struct glsl_type *type =
2801 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2802 val->ssa = vtn_create_ssa_value(b, type);
2805 case SpvOpVectorExtractDynamic:
2806 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2807 vtn_ssa_value(b, w[4])->def);
2810 case SpvOpVectorInsertDynamic:
2811 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2812 vtn_ssa_value(b, w[4])->def,
2813 vtn_ssa_value(b, w[5])->def);
2816 case SpvOpVectorShuffle:
2817 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2818 vtn_ssa_value(b, w[3])->def,
2819 vtn_ssa_value(b, w[4])->def,
2823 case SpvOpCompositeConstruct: {
2824 unsigned elems = count - 3;
2825 if (glsl_type_is_vector_or_scalar(type)) {
2826 nir_ssa_def *srcs[4];
2827 for (unsigned i = 0; i < elems; i++)
2828 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2830 vtn_vector_construct(b, glsl_get_vector_elements(type),
2833 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2834 for (unsigned i = 0; i < elems; i++)
2835 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2839 case SpvOpCompositeExtract:
2840 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2844 case SpvOpCompositeInsert:
2845 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2846 vtn_ssa_value(b, w[3]),
2850 case SpvOpCopyObject:
2851 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2855 vtn_fail("unknown composite operation");
2860 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
2861 const uint32_t *w, unsigned count)
2863 nir_intrinsic_op intrinsic_op;
2865 case SpvOpEmitVertex:
2866 case SpvOpEmitStreamVertex:
2867 intrinsic_op = nir_intrinsic_emit_vertex;
2869 case SpvOpEndPrimitive:
2870 case SpvOpEndStreamPrimitive:
2871 intrinsic_op = nir_intrinsic_end_primitive;
2873 case SpvOpMemoryBarrier:
2874 intrinsic_op = nir_intrinsic_memory_barrier;
2876 case SpvOpControlBarrier:
2877 intrinsic_op = nir_intrinsic_barrier;
2880 vtn_fail("unknown barrier instruction");
2883 nir_intrinsic_instr *intrin =
2884 nir_intrinsic_instr_create(b->shader, intrinsic_op);
2886 if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
2887 nir_intrinsic_set_stream_id(intrin, w[1]);
2889 nir_builder_instr_insert(&b->nb, &intrin->instr);
2893 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
2894 SpvExecutionMode mode)
2897 case SpvExecutionModeInputPoints:
2898 case SpvExecutionModeOutputPoints:
2899 return 0; /* GL_POINTS */
2900 case SpvExecutionModeInputLines:
2901 return 1; /* GL_LINES */
2902 case SpvExecutionModeInputLinesAdjacency:
2903 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2904 case SpvExecutionModeTriangles:
2905 return 4; /* GL_TRIANGLES */
2906 case SpvExecutionModeInputTrianglesAdjacency:
2907 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2908 case SpvExecutionModeQuads:
2909 return 7; /* GL_QUADS */
2910 case SpvExecutionModeIsolines:
2911 return 0x8E7A; /* GL_ISOLINES */
2912 case SpvExecutionModeOutputLineStrip:
2913 return 3; /* GL_LINE_STRIP */
2914 case SpvExecutionModeOutputTriangleStrip:
2915 return 5; /* GL_TRIANGLE_STRIP */
2917 vtn_fail("Invalid primitive type");
2922 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
2923 SpvExecutionMode mode)
2926 case SpvExecutionModeInputPoints:
2928 case SpvExecutionModeInputLines:
2930 case SpvExecutionModeInputLinesAdjacency:
2932 case SpvExecutionModeTriangles:
2934 case SpvExecutionModeInputTrianglesAdjacency:
2937 vtn_fail("Invalid GS input mode");
2941 static gl_shader_stage
2942 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
2945 case SpvExecutionModelVertex:
2946 return MESA_SHADER_VERTEX;
2947 case SpvExecutionModelTessellationControl:
2948 return MESA_SHADER_TESS_CTRL;
2949 case SpvExecutionModelTessellationEvaluation:
2950 return MESA_SHADER_TESS_EVAL;
2951 case SpvExecutionModelGeometry:
2952 return MESA_SHADER_GEOMETRY;
2953 case SpvExecutionModelFragment:
2954 return MESA_SHADER_FRAGMENT;
2955 case SpvExecutionModelGLCompute:
2956 return MESA_SHADER_COMPUTE;
2958 vtn_fail("Unsupported execution model");
2962 #define spv_check_supported(name, cap) do { \
2963 if (!(b->options && b->options->caps.name)) \
2964 vtn_warn("Unsupported SPIR-V capability: %s", \
2965 spirv_capability_to_string(cap)); \
2969 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
2970 const uint32_t *w, unsigned count)
2977 case SpvSourceLanguageUnknown: lang = "unknown"; break;
2978 case SpvSourceLanguageESSL: lang = "ESSL"; break;
2979 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
2980 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
2981 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
2982 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
2985 uint32_t version = w[2];
2988 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
2990 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
2994 case SpvOpSourceExtension:
2995 case SpvOpSourceContinued:
2996 case SpvOpExtension:
2997 /* Unhandled, but these are for debug so that's ok. */
3000 case SpvOpCapability: {
3001 SpvCapability cap = w[1];
3003 case SpvCapabilityMatrix:
3004 case SpvCapabilityShader:
3005 case SpvCapabilityGeometry:
3006 case SpvCapabilityGeometryPointSize:
3007 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3008 case SpvCapabilitySampledImageArrayDynamicIndexing:
3009 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3010 case SpvCapabilityStorageImageArrayDynamicIndexing:
3011 case SpvCapabilityImageRect:
3012 case SpvCapabilitySampledRect:
3013 case SpvCapabilitySampled1D:
3014 case SpvCapabilityImage1D:
3015 case SpvCapabilitySampledCubeArray:
3016 case SpvCapabilityImageCubeArray:
3017 case SpvCapabilitySampledBuffer:
3018 case SpvCapabilityImageBuffer:
3019 case SpvCapabilityImageQuery:
3020 case SpvCapabilityDerivativeControl:
3021 case SpvCapabilityInterpolationFunction:
3022 case SpvCapabilityMultiViewport:
3023 case SpvCapabilitySampleRateShading:
3024 case SpvCapabilityClipDistance:
3025 case SpvCapabilityCullDistance:
3026 case SpvCapabilityInputAttachment:
3027 case SpvCapabilityImageGatherExtended:
3028 case SpvCapabilityStorageImageExtendedFormats:
3031 case SpvCapabilityGeometryStreams:
3032 case SpvCapabilityLinkage:
3033 case SpvCapabilityVector16:
3034 case SpvCapabilityFloat16Buffer:
3035 case SpvCapabilityFloat16:
3036 case SpvCapabilityInt64Atomics:
3037 case SpvCapabilityAtomicStorage:
3038 case SpvCapabilityInt16:
3039 case SpvCapabilityStorageImageMultisample:
3040 case SpvCapabilityInt8:
3041 case SpvCapabilitySparseResidency:
3042 case SpvCapabilityMinLod:
3043 case SpvCapabilityTransformFeedback:
3044 vtn_warn("Unsupported SPIR-V capability: %s",
3045 spirv_capability_to_string(cap));
3048 case SpvCapabilityFloat64:
3049 spv_check_supported(float64, cap);
3051 case SpvCapabilityInt64:
3052 spv_check_supported(int64, cap);
3055 case SpvCapabilityAddresses:
3056 case SpvCapabilityKernel:
3057 case SpvCapabilityImageBasic:
3058 case SpvCapabilityImageReadWrite:
3059 case SpvCapabilityImageMipmap:
3060 case SpvCapabilityPipes:
3061 case SpvCapabilityGroups:
3062 case SpvCapabilityDeviceEnqueue:
3063 case SpvCapabilityLiteralSampler:
3064 case SpvCapabilityGenericPointer:
3065 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3066 spirv_capability_to_string(cap));
3069 case SpvCapabilityImageMSArray:
3070 spv_check_supported(image_ms_array, cap);
3073 case SpvCapabilityTessellation:
3074 case SpvCapabilityTessellationPointSize:
3075 spv_check_supported(tessellation, cap);
3078 case SpvCapabilityDrawParameters:
3079 spv_check_supported(draw_parameters, cap);
3082 case SpvCapabilityStorageImageReadWithoutFormat:
3083 spv_check_supported(image_read_without_format, cap);
3086 case SpvCapabilityStorageImageWriteWithoutFormat:
3087 spv_check_supported(image_write_without_format, cap);
3090 case SpvCapabilityMultiView:
3091 spv_check_supported(multiview, cap);
3094 case SpvCapabilityVariablePointersStorageBuffer:
3095 case SpvCapabilityVariablePointers:
3096 spv_check_supported(variable_pointers, cap);
3099 case SpvCapabilityStorageUniformBufferBlock16:
3100 case SpvCapabilityStorageUniform16:
3101 case SpvCapabilityStoragePushConstant16:
3102 case SpvCapabilityStorageInputOutput16:
3103 spv_check_supported(storage_16bit, cap);
3107 vtn_fail("Unhandled capability");
3112 case SpvOpExtInstImport:
3113 vtn_handle_extension(b, opcode, w, count);
3116 case SpvOpMemoryModel:
3117 vtn_assert(w[1] == SpvAddressingModelLogical);
3118 vtn_assert(w[2] == SpvMemoryModelSimple ||
3119 w[2] == SpvMemoryModelGLSL450);
3122 case SpvOpEntryPoint: {
3123 struct vtn_value *entry_point = &b->values[w[2]];
3124 /* Let this be a name label regardless */
3125 unsigned name_words;
3126 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3128 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3129 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3132 vtn_assert(b->entry_point == NULL);
3133 b->entry_point = entry_point;
3138 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3139 vtn_string_literal(b, &w[2], count - 2, NULL);
3143 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3146 case SpvOpMemberName:
3150 case SpvOpExecutionMode:
3151 case SpvOpDecorationGroup:
3153 case SpvOpMemberDecorate:
3154 case SpvOpGroupDecorate:
3155 case SpvOpGroupMemberDecorate:
3156 vtn_handle_decoration(b, opcode, w, count);
3160 return false; /* End of preamble */
3167 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3168 const struct vtn_decoration *mode, void *data)
3170 vtn_assert(b->entry_point == entry_point);
3172 switch(mode->exec_mode) {
3173 case SpvExecutionModeOriginUpperLeft:
3174 case SpvExecutionModeOriginLowerLeft:
3175 b->origin_upper_left =
3176 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3179 case SpvExecutionModeEarlyFragmentTests:
3180 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3181 b->shader->info.fs.early_fragment_tests = true;
3184 case SpvExecutionModeInvocations:
3185 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3186 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3189 case SpvExecutionModeDepthReplacing:
3190 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3191 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3193 case SpvExecutionModeDepthGreater:
3194 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3195 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3197 case SpvExecutionModeDepthLess:
3198 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3199 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3201 case SpvExecutionModeDepthUnchanged:
3202 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3203 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3206 case SpvExecutionModeLocalSize:
3207 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3208 b->shader->info.cs.local_size[0] = mode->literals[0];
3209 b->shader->info.cs.local_size[1] = mode->literals[1];
3210 b->shader->info.cs.local_size[2] = mode->literals[2];
3212 case SpvExecutionModeLocalSizeHint:
3213 break; /* Nothing to do with this */
3215 case SpvExecutionModeOutputVertices:
3216 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3217 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3218 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3220 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3221 b->shader->info.gs.vertices_out = mode->literals[0];
3225 case SpvExecutionModeInputPoints:
3226 case SpvExecutionModeInputLines:
3227 case SpvExecutionModeInputLinesAdjacency:
3228 case SpvExecutionModeTriangles:
3229 case SpvExecutionModeInputTrianglesAdjacency:
3230 case SpvExecutionModeQuads:
3231 case SpvExecutionModeIsolines:
3232 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3233 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3234 b->shader->info.tess.primitive_mode =
3235 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3237 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3238 b->shader->info.gs.vertices_in =
3239 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3243 case SpvExecutionModeOutputPoints:
3244 case SpvExecutionModeOutputLineStrip:
3245 case SpvExecutionModeOutputTriangleStrip:
3246 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3247 b->shader->info.gs.output_primitive =
3248 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3251 case SpvExecutionModeSpacingEqual:
3252 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3253 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3254 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3256 case SpvExecutionModeSpacingFractionalEven:
3257 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3258 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3259 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3261 case SpvExecutionModeSpacingFractionalOdd:
3262 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3263 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3264 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3266 case SpvExecutionModeVertexOrderCw:
3267 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3268 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3269 b->shader->info.tess.ccw = false;
3271 case SpvExecutionModeVertexOrderCcw:
3272 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3273 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3274 b->shader->info.tess.ccw = true;
3276 case SpvExecutionModePointMode:
3277 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3278 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3279 b->shader->info.tess.point_mode = true;
3282 case SpvExecutionModePixelCenterInteger:
3283 b->pixel_center_integer = true;
3286 case SpvExecutionModeXfb:
3287 vtn_fail("Unhandled execution mode");
3290 case SpvExecutionModeVecTypeHint:
3291 case SpvExecutionModeContractionOff:
3295 vtn_fail("Unhandled execution mode");
3300 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3301 const uint32_t *w, unsigned count)
3303 vtn_set_instruction_result_type(b, opcode, w, count);
3307 case SpvOpSourceContinued:
3308 case SpvOpSourceExtension:
3309 case SpvOpExtension:
3310 case SpvOpCapability:
3311 case SpvOpExtInstImport:
3312 case SpvOpMemoryModel:
3313 case SpvOpEntryPoint:
3314 case SpvOpExecutionMode:
3317 case SpvOpMemberName:
3318 case SpvOpDecorationGroup:
3320 case SpvOpMemberDecorate:
3321 case SpvOpGroupDecorate:
3322 case SpvOpGroupMemberDecorate:
3323 vtn_fail("Invalid opcode types and variables section");
3329 case SpvOpTypeFloat:
3330 case SpvOpTypeVector:
3331 case SpvOpTypeMatrix:
3332 case SpvOpTypeImage:
3333 case SpvOpTypeSampler:
3334 case SpvOpTypeSampledImage:
3335 case SpvOpTypeArray:
3336 case SpvOpTypeRuntimeArray:
3337 case SpvOpTypeStruct:
3338 case SpvOpTypeOpaque:
3339 case SpvOpTypePointer:
3340 case SpvOpTypeFunction:
3341 case SpvOpTypeEvent:
3342 case SpvOpTypeDeviceEvent:
3343 case SpvOpTypeReserveId:
3344 case SpvOpTypeQueue:
3346 vtn_handle_type(b, opcode, w, count);
3349 case SpvOpConstantTrue:
3350 case SpvOpConstantFalse:
3352 case SpvOpConstantComposite:
3353 case SpvOpConstantSampler:
3354 case SpvOpConstantNull:
3355 case SpvOpSpecConstantTrue:
3356 case SpvOpSpecConstantFalse:
3357 case SpvOpSpecConstant:
3358 case SpvOpSpecConstantComposite:
3359 case SpvOpSpecConstantOp:
3360 vtn_handle_constant(b, opcode, w, count);
3365 vtn_handle_variables(b, opcode, w, count);
3369 return false; /* End of preamble */
3376 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3377 const uint32_t *w, unsigned count)
3383 case SpvOpLoopMerge:
3384 case SpvOpSelectionMerge:
3385 /* This is handled by cfg pre-pass and walk_blocks */
3389 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3390 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3395 vtn_handle_extension(b, opcode, w, count);
3401 case SpvOpCopyMemory:
3402 case SpvOpCopyMemorySized:
3403 case SpvOpAccessChain:
3404 case SpvOpPtrAccessChain:
3405 case SpvOpInBoundsAccessChain:
3406 case SpvOpArrayLength:
3407 vtn_handle_variables(b, opcode, w, count);
3410 case SpvOpFunctionCall:
3411 vtn_handle_function_call(b, opcode, w, count);
3414 case SpvOpSampledImage:
3416 case SpvOpImageSampleImplicitLod:
3417 case SpvOpImageSampleExplicitLod:
3418 case SpvOpImageSampleDrefImplicitLod:
3419 case SpvOpImageSampleDrefExplicitLod:
3420 case SpvOpImageSampleProjImplicitLod:
3421 case SpvOpImageSampleProjExplicitLod:
3422 case SpvOpImageSampleProjDrefImplicitLod:
3423 case SpvOpImageSampleProjDrefExplicitLod:
3424 case SpvOpImageFetch:
3425 case SpvOpImageGather:
3426 case SpvOpImageDrefGather:
3427 case SpvOpImageQuerySizeLod:
3428 case SpvOpImageQueryLod:
3429 case SpvOpImageQueryLevels:
3430 case SpvOpImageQuerySamples:
3431 vtn_handle_texture(b, opcode, w, count);
3434 case SpvOpImageRead:
3435 case SpvOpImageWrite:
3436 case SpvOpImageTexelPointer:
3437 vtn_handle_image(b, opcode, w, count);
3440 case SpvOpImageQuerySize: {
3441 struct vtn_pointer *image =
3442 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3443 if (image->mode == vtn_variable_mode_image) {
3444 vtn_handle_image(b, opcode, w, count);
3446 vtn_assert(image->mode == vtn_variable_mode_sampler);
3447 vtn_handle_texture(b, opcode, w, count);
3452 case SpvOpAtomicLoad:
3453 case SpvOpAtomicExchange:
3454 case SpvOpAtomicCompareExchange:
3455 case SpvOpAtomicCompareExchangeWeak:
3456 case SpvOpAtomicIIncrement:
3457 case SpvOpAtomicIDecrement:
3458 case SpvOpAtomicIAdd:
3459 case SpvOpAtomicISub:
3460 case SpvOpAtomicSMin:
3461 case SpvOpAtomicUMin:
3462 case SpvOpAtomicSMax:
3463 case SpvOpAtomicUMax:
3464 case SpvOpAtomicAnd:
3466 case SpvOpAtomicXor: {
3467 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3468 if (pointer->value_type == vtn_value_type_image_pointer) {
3469 vtn_handle_image(b, opcode, w, count);
3471 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3472 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3477 case SpvOpAtomicStore: {
3478 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3479 if (pointer->value_type == vtn_value_type_image_pointer) {
3480 vtn_handle_image(b, opcode, w, count);
3482 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3483 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3489 /* Handle OpSelect up-front here because it needs to be able to handle
3490 * pointers and not just regular vectors and scalars.
3492 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3493 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3494 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3495 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3497 const struct glsl_type *sel_type;
3498 switch (res_val->type->base_type) {
3499 case vtn_base_type_scalar:
3500 sel_type = glsl_bool_type();
3502 case vtn_base_type_vector:
3503 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3505 case vtn_base_type_pointer:
3506 /* We need to have actual storage for pointer types */
3507 vtn_fail_if(res_val->type->type == NULL,
3508 "Invalid pointer result type for OpSelect");
3509 sel_type = glsl_bool_type();
3512 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3515 if (unlikely(sel_val->type->type != sel_type)) {
3516 if (sel_val->type->type == glsl_bool_type()) {
3517 /* This case is illegal but some older versions of GLSLang produce
3518 * it. The GLSLang issue was fixed on March 30, 2017:
3520 * https://github.com/KhronosGroup/glslang/issues/809
3522 * Unfortunately, there are applications in the wild which are
3523 * shipping with this bug so it isn't nice to fail on them so we
3524 * throw a warning instead. It's not actually a problem for us as
3525 * nir_builder will just splat the condition out which is most
3526 * likely what the client wanted anyway.
3528 vtn_warn("Condition type of OpSelect must have the same number "
3529 "of components as Result Type");
3531 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3532 "of Boolean type. It must have the same number of "
3533 "components as Result Type");
3537 vtn_fail_if(obj1_val->type != res_val->type ||
3538 obj2_val->type != res_val->type,
3539 "Object types must match the result type in OpSelect");
3541 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3542 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3543 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3544 vtn_ssa_value(b, w[4])->def,
3545 vtn_ssa_value(b, w[5])->def);
3546 vtn_push_ssa(b, w[2], res_type, ssa);
3555 case SpvOpConvertFToU:
3556 case SpvOpConvertFToS:
3557 case SpvOpConvertSToF:
3558 case SpvOpConvertUToF:
3562 case SpvOpQuantizeToF16:
3563 case SpvOpConvertPtrToU:
3564 case SpvOpConvertUToPtr:
3565 case SpvOpPtrCastToGeneric:
3566 case SpvOpGenericCastToPtr:
3572 case SpvOpSignBitSet:
3573 case SpvOpLessOrGreater:
3575 case SpvOpUnordered:
3590 case SpvOpVectorTimesScalar:
3592 case SpvOpIAddCarry:
3593 case SpvOpISubBorrow:
3594 case SpvOpUMulExtended:
3595 case SpvOpSMulExtended:
3596 case SpvOpShiftRightLogical:
3597 case SpvOpShiftRightArithmetic:
3598 case SpvOpShiftLeftLogical:
3599 case SpvOpLogicalEqual:
3600 case SpvOpLogicalNotEqual:
3601 case SpvOpLogicalOr:
3602 case SpvOpLogicalAnd:
3603 case SpvOpLogicalNot:
3604 case SpvOpBitwiseOr:
3605 case SpvOpBitwiseXor:
3606 case SpvOpBitwiseAnd:
3608 case SpvOpFOrdEqual:
3609 case SpvOpFUnordEqual:
3610 case SpvOpINotEqual:
3611 case SpvOpFOrdNotEqual:
3612 case SpvOpFUnordNotEqual:
3613 case SpvOpULessThan:
3614 case SpvOpSLessThan:
3615 case SpvOpFOrdLessThan:
3616 case SpvOpFUnordLessThan:
3617 case SpvOpUGreaterThan:
3618 case SpvOpSGreaterThan:
3619 case SpvOpFOrdGreaterThan:
3620 case SpvOpFUnordGreaterThan:
3621 case SpvOpULessThanEqual:
3622 case SpvOpSLessThanEqual:
3623 case SpvOpFOrdLessThanEqual:
3624 case SpvOpFUnordLessThanEqual:
3625 case SpvOpUGreaterThanEqual:
3626 case SpvOpSGreaterThanEqual:
3627 case SpvOpFOrdGreaterThanEqual:
3628 case SpvOpFUnordGreaterThanEqual:
3634 case SpvOpFwidthFine:
3635 case SpvOpDPdxCoarse:
3636 case SpvOpDPdyCoarse:
3637 case SpvOpFwidthCoarse:
3638 case SpvOpBitFieldInsert:
3639 case SpvOpBitFieldSExtract:
3640 case SpvOpBitFieldUExtract:
3641 case SpvOpBitReverse:
3643 case SpvOpTranspose:
3644 case SpvOpOuterProduct:
3645 case SpvOpMatrixTimesScalar:
3646 case SpvOpVectorTimesMatrix:
3647 case SpvOpMatrixTimesVector:
3648 case SpvOpMatrixTimesMatrix:
3649 vtn_handle_alu(b, opcode, w, count);
3652 case SpvOpVectorExtractDynamic:
3653 case SpvOpVectorInsertDynamic:
3654 case SpvOpVectorShuffle:
3655 case SpvOpCompositeConstruct:
3656 case SpvOpCompositeExtract:
3657 case SpvOpCompositeInsert:
3658 case SpvOpCopyObject:
3659 vtn_handle_composite(b, opcode, w, count);
3662 case SpvOpEmitVertex:
3663 case SpvOpEndPrimitive:
3664 case SpvOpEmitStreamVertex:
3665 case SpvOpEndStreamPrimitive:
3666 case SpvOpControlBarrier:
3667 case SpvOpMemoryBarrier:
3668 vtn_handle_barrier(b, opcode, w, count);
3672 vtn_fail("Unhandled opcode");
3679 spirv_to_nir(const uint32_t *words, size_t word_count,
3680 struct nir_spirv_specialization *spec, unsigned num_spec,
3681 gl_shader_stage stage, const char *entry_point_name,
3682 const struct spirv_to_nir_options *options,
3683 const nir_shader_compiler_options *nir_options)
3685 /* Initialize the stn_builder object */
3686 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3691 exec_list_make_empty(&b->functions);
3692 b->entry_point_stage = stage;
3693 b->entry_point_name = entry_point_name;
3694 b->options = options;
3696 /* See also _vtn_fail() */
3697 if (setjmp(b->fail_jump)) {
3702 const uint32_t *word_end = words + word_count;
3704 /* Handle the SPIR-V header (first 4 dwords) */
3705 vtn_assert(word_count > 5);
3707 vtn_assert(words[0] == SpvMagicNumber);
3708 vtn_assert(words[1] >= 0x10000);
3709 /* words[2] == generator magic */
3710 unsigned value_id_bound = words[3];
3711 vtn_assert(words[4] == 0);
3715 b->value_id_bound = value_id_bound;
3716 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3718 /* Handle all the preamble instructions */
3719 words = vtn_foreach_instruction(b, words, word_end,
3720 vtn_handle_preamble_instruction);
3722 if (b->entry_point == NULL) {
3723 vtn_fail("Entry point not found");
3728 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3730 /* Set shader info defaults */
3731 b->shader->info.gs.invocations = 1;
3733 /* Parse execution modes */
3734 vtn_foreach_execution_mode(b, b->entry_point,
3735 vtn_handle_execution_mode, NULL);
3737 b->specializations = spec;
3738 b->num_specializations = num_spec;
3740 /* Handle all variable, type, and constant instructions */
3741 words = vtn_foreach_instruction(b, words, word_end,
3742 vtn_handle_variable_or_type_instruction);
3744 /* Set types on all vtn_values */
3745 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
3747 vtn_build_cfg(b, words, word_end);
3749 assert(b->entry_point->value_type == vtn_value_type_function);
3750 b->entry_point->func->referenced = true;
3755 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3756 if (func->referenced && !func->emitted) {
3757 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3758 _mesa_key_pointer_equal);
3760 vtn_function_emit(b, func, vtn_handle_body_instruction);
3766 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
3767 nir_function *entry_point = b->entry_point->func->impl->function;
3768 vtn_assert(entry_point);
3770 /* Unparent the shader from the vtn_builder before we delete the builder */
3771 ralloc_steal(NULL, b->shader);