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 vtn_type *sampled_type =
1079 vtn_value(b, w[2], vtn_value_type_type)->type;
1081 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1082 glsl_get_bit_size(sampled_type->type) != 32,
1083 "Sampled type of OpTypeImage must be a 32-bit scalar");
1085 enum glsl_sampler_dim dim;
1086 switch ((SpvDim)w[3]) {
1087 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1088 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1089 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1090 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1091 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1092 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1093 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1095 vtn_fail("Invalid SPIR-V image dimensionality");
1098 bool is_shadow = w[4];
1099 bool is_array = w[5];
1100 bool multisampled = w[6];
1101 unsigned sampled = w[7];
1102 SpvImageFormat format = w[8];
1105 val->type->access_qualifier = w[9];
1107 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1110 if (dim == GLSL_SAMPLER_DIM_2D)
1111 dim = GLSL_SAMPLER_DIM_MS;
1112 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1113 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1115 vtn_fail("Unsupported multisampled image type");
1118 val->type->image_format = translate_image_format(b, format);
1120 enum glsl_base_type sampled_base_type =
1121 glsl_get_base_type(sampled_type->type);
1123 val->type->sampled = true;
1124 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1126 } else if (sampled == 2) {
1127 vtn_assert(!is_shadow);
1128 val->type->sampled = false;
1129 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1131 vtn_fail("We need to know if the image will be sampled");
1136 case SpvOpTypeSampledImage:
1137 val->type->base_type = vtn_base_type_sampled_image;
1138 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1139 val->type->type = val->type->image->type;
1142 case SpvOpTypeSampler:
1143 /* The actual sampler type here doesn't really matter. It gets
1144 * thrown away the moment you combine it with an image. What really
1145 * matters is that it's a sampler type as opposed to an integer type
1146 * so the backend knows what to do.
1148 val->type->base_type = vtn_base_type_sampler;
1149 val->type->type = glsl_bare_sampler_type();
1152 case SpvOpTypeOpaque:
1153 case SpvOpTypeEvent:
1154 case SpvOpTypeDeviceEvent:
1155 case SpvOpTypeReserveId:
1156 case SpvOpTypeQueue:
1159 vtn_fail("Unhandled opcode");
1162 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1165 static nir_constant *
1166 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1168 nir_constant *c = rzalloc(b, nir_constant);
1170 /* For pointers and other typeless things, we have to return something but
1171 * it doesn't matter what.
1176 switch (glsl_get_base_type(type)) {
1178 case GLSL_TYPE_UINT:
1179 case GLSL_TYPE_INT16:
1180 case GLSL_TYPE_UINT16:
1181 case GLSL_TYPE_INT64:
1182 case GLSL_TYPE_UINT64:
1183 case GLSL_TYPE_BOOL:
1184 case GLSL_TYPE_FLOAT:
1185 case GLSL_TYPE_FLOAT16:
1186 case GLSL_TYPE_DOUBLE:
1187 /* Nothing to do here. It's already initialized to zero */
1190 case GLSL_TYPE_ARRAY:
1191 vtn_assert(glsl_get_length(type) > 0);
1192 c->num_elements = glsl_get_length(type);
1193 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1195 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1196 for (unsigned i = 1; i < c->num_elements; i++)
1197 c->elements[i] = c->elements[0];
1200 case GLSL_TYPE_STRUCT:
1201 c->num_elements = glsl_get_length(type);
1202 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1204 for (unsigned i = 0; i < c->num_elements; i++) {
1205 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1210 vtn_fail("Invalid type for null constant");
1217 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1218 int member, const struct vtn_decoration *dec,
1221 vtn_assert(member == -1);
1222 if (dec->decoration != SpvDecorationSpecId)
1225 struct spec_constant_value *const_value = data;
1227 for (unsigned i = 0; i < b->num_specializations; i++) {
1228 if (b->specializations[i].id == dec->literals[0]) {
1229 if (const_value->is_double)
1230 const_value->data64 = b->specializations[i].data64;
1232 const_value->data32 = b->specializations[i].data32;
1239 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1240 uint32_t const_value)
1242 struct spec_constant_value data;
1243 data.is_double = false;
1244 data.data32 = const_value;
1245 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1250 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1251 uint64_t const_value)
1253 struct spec_constant_value data;
1254 data.is_double = true;
1255 data.data64 = const_value;
1256 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1261 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1262 struct vtn_value *val,
1264 const struct vtn_decoration *dec,
1267 vtn_assert(member == -1);
1268 if (dec->decoration != SpvDecorationBuiltIn ||
1269 dec->literals[0] != SpvBuiltInWorkgroupSize)
1272 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1274 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1275 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1276 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1280 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1281 const uint32_t *w, unsigned count)
1283 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1284 val->constant = rzalloc(b, nir_constant);
1286 case SpvOpConstantTrue:
1287 case SpvOpConstantFalse:
1288 case SpvOpSpecConstantTrue:
1289 case SpvOpSpecConstantFalse: {
1290 vtn_fail_if(val->type->type != glsl_bool_type(),
1291 "Result type of %s must be OpTypeBool",
1292 spirv_op_to_string(opcode));
1294 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1295 opcode == SpvOpSpecConstantTrue);
1297 if (opcode == SpvOpSpecConstantTrue ||
1298 opcode == SpvOpSpecConstantFalse)
1299 int_val = get_specialization(b, val, int_val);
1301 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1305 case SpvOpConstant: {
1306 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1307 "Result type of %s must be a scalar",
1308 spirv_op_to_string(opcode));
1309 int bit_size = glsl_get_bit_size(val->type->type);
1312 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1315 val->constant->values->u32[0] = w[3];
1318 val->constant->values->u16[0] = w[3];
1321 vtn_fail("Unsupported SpvOpConstant bit size");
1326 case SpvOpSpecConstant: {
1327 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1328 "Result type of %s must be a scalar",
1329 spirv_op_to_string(opcode));
1330 int bit_size = glsl_get_bit_size(val->type->type);
1333 val->constant->values[0].u64[0] =
1334 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1337 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1340 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1343 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1348 case SpvOpSpecConstantComposite:
1349 case SpvOpConstantComposite: {
1350 unsigned elem_count = count - 3;
1351 vtn_fail_if(elem_count != val->type->length,
1352 "%s has %u constituents, expected %u",
1353 spirv_op_to_string(opcode), elem_count, val->type->length);
1355 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1356 for (unsigned i = 0; i < elem_count; i++)
1357 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1359 switch (val->type->base_type) {
1360 case vtn_base_type_vector: {
1361 assert(glsl_type_is_vector(val->type->type));
1362 int bit_size = glsl_get_bit_size(val->type->type);
1363 for (unsigned i = 0; i < elem_count; i++) {
1366 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1369 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1372 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1375 vtn_fail("Invalid SpvOpConstantComposite bit size");
1381 case vtn_base_type_matrix:
1382 assert(glsl_type_is_matrix(val->type->type));
1383 for (unsigned i = 0; i < elem_count; i++)
1384 val->constant->values[i] = elems[i]->values[0];
1387 case vtn_base_type_struct:
1388 case vtn_base_type_array:
1389 ralloc_steal(val->constant, elems);
1390 val->constant->num_elements = elem_count;
1391 val->constant->elements = elems;
1395 vtn_fail("Result type of %s must be a composite type",
1396 spirv_op_to_string(opcode));
1401 case SpvOpSpecConstantOp: {
1402 SpvOp opcode = get_specialization(b, val, w[3]);
1404 case SpvOpVectorShuffle: {
1405 struct vtn_value *v0 = &b->values[w[4]];
1406 struct vtn_value *v1 = &b->values[w[5]];
1408 vtn_assert(v0->value_type == vtn_value_type_constant ||
1409 v0->value_type == vtn_value_type_undef);
1410 vtn_assert(v1->value_type == vtn_value_type_constant ||
1411 v1->value_type == vtn_value_type_undef);
1413 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1414 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1416 vtn_assert(len0 + len1 < 16);
1418 unsigned bit_size = glsl_get_bit_size(val->type->type);
1419 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1420 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1422 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1423 (void)bit_size0; (void)bit_size1;
1425 if (bit_size == 64) {
1427 if (v0->value_type == vtn_value_type_constant) {
1428 for (unsigned i = 0; i < len0; i++)
1429 u64[i] = v0->constant->values[0].u64[i];
1431 if (v1->value_type == vtn_value_type_constant) {
1432 for (unsigned i = 0; i < len1; i++)
1433 u64[len0 + i] = v1->constant->values[0].u64[i];
1436 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1437 uint32_t comp = w[i + 6];
1438 /* If component is not used, set the value to a known constant
1439 * to detect if it is wrongly used.
1441 if (comp == (uint32_t)-1)
1442 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1444 val->constant->values[0].u64[j] = u64[comp];
1447 /* This is for both 32-bit and 16-bit values */
1449 if (v0->value_type == vtn_value_type_constant) {
1450 for (unsigned i = 0; i < len0; i++)
1451 u32[i] = v0->constant->values[0].u32[i];
1453 if (v1->value_type == vtn_value_type_constant) {
1454 for (unsigned i = 0; i < len1; i++)
1455 u32[len0 + i] = v1->constant->values[0].u32[i];
1458 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1459 uint32_t comp = w[i + 6];
1460 /* If component is not used, set the value to a known constant
1461 * to detect if it is wrongly used.
1463 if (comp == (uint32_t)-1)
1464 val->constant->values[0].u32[j] = 0xdeadbeef;
1466 val->constant->values[0].u32[j] = u32[comp];
1472 case SpvOpCompositeExtract:
1473 case SpvOpCompositeInsert: {
1474 struct vtn_value *comp;
1475 unsigned deref_start;
1476 struct nir_constant **c;
1477 if (opcode == SpvOpCompositeExtract) {
1478 comp = vtn_value(b, w[4], vtn_value_type_constant);
1480 c = &comp->constant;
1482 comp = vtn_value(b, w[5], vtn_value_type_constant);
1484 val->constant = nir_constant_clone(comp->constant,
1491 const struct vtn_type *type = comp->type;
1492 for (unsigned i = deref_start; i < count; i++) {
1493 vtn_fail_if(w[i] > type->length,
1494 "%uth index of %s is %u but the type has only "
1495 "%u elements", i - deref_start,
1496 spirv_op_to_string(opcode), w[i], type->length);
1498 switch (type->base_type) {
1499 case vtn_base_type_vector:
1501 type = type->array_element;
1504 case vtn_base_type_matrix:
1505 assert(col == 0 && elem == -1);
1508 type = type->array_element;
1511 case vtn_base_type_array:
1512 c = &(*c)->elements[w[i]];
1513 type = type->array_element;
1516 case vtn_base_type_struct:
1517 c = &(*c)->elements[w[i]];
1518 type = type->members[w[i]];
1522 vtn_fail("%s must only index into composite types",
1523 spirv_op_to_string(opcode));
1527 if (opcode == SpvOpCompositeExtract) {
1531 unsigned num_components = type->length;
1532 unsigned bit_size = glsl_get_bit_size(type->type);
1533 for (unsigned i = 0; i < num_components; i++)
1536 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1539 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1542 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1545 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1549 struct vtn_value *insert =
1550 vtn_value(b, w[4], vtn_value_type_constant);
1551 vtn_assert(insert->type == type);
1553 *c = insert->constant;
1555 unsigned num_components = type->length;
1556 unsigned bit_size = glsl_get_bit_size(type->type);
1557 for (unsigned i = 0; i < num_components; i++)
1560 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1563 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1566 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1569 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1578 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1579 nir_alu_type src_alu_type = dst_alu_type;
1580 unsigned num_components = glsl_get_vector_elements(val->type->type);
1583 vtn_assert(count <= 7);
1588 /* We have a source in a conversion */
1590 nir_get_nir_type_for_glsl_type(
1591 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1592 /* We use the bitsize of the conversion source to evaluate the opcode later */
1593 bit_size = glsl_get_bit_size(
1594 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1597 bit_size = glsl_get_bit_size(val->type->type);
1600 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1603 nir_const_value src[4];
1605 for (unsigned i = 0; i < count - 4; i++) {
1607 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1609 unsigned j = swap ? 1 - i : i;
1610 src[j] = c->values[0];
1613 val->constant->values[0] =
1614 nir_eval_const_opcode(op, num_components, bit_size, src);
1621 case SpvOpConstantNull:
1622 val->constant = vtn_null_constant(b, val->type->type);
1625 case SpvOpConstantSampler:
1626 vtn_fail("OpConstantSampler requires Kernel Capability");
1630 vtn_fail("Unhandled opcode");
1633 /* Now that we have the value, update the workgroup size if needed */
1634 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1638 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1639 const uint32_t *w, unsigned count)
1641 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1642 struct vtn_function *vtn_callee =
1643 vtn_value(b, w[3], vtn_value_type_function)->func;
1644 struct nir_function *callee = vtn_callee->impl->function;
1646 vtn_callee->referenced = true;
1648 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1649 for (unsigned i = 0; i < call->num_params; i++) {
1650 unsigned arg_id = w[4 + i];
1651 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1652 if (arg->value_type == vtn_value_type_pointer &&
1653 arg->pointer->ptr_type->type == NULL) {
1654 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1655 call->params[i] = nir_deref_var_clone(d, call);
1657 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1659 /* Make a temporary to store the argument in */
1661 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1662 call->params[i] = nir_deref_var_create(call, tmp);
1664 vtn_local_store(b, arg_ssa, call->params[i]);
1668 nir_variable *out_tmp = NULL;
1669 vtn_assert(res_type->type == callee->return_type);
1670 if (!glsl_type_is_void(callee->return_type)) {
1671 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1673 call->return_deref = nir_deref_var_create(call, out_tmp);
1676 nir_builder_instr_insert(&b->nb, &call->instr);
1678 if (glsl_type_is_void(callee->return_type)) {
1679 vtn_push_value(b, w[2], vtn_value_type_undef);
1681 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1685 struct vtn_ssa_value *
1686 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1688 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1691 if (!glsl_type_is_vector_or_scalar(type)) {
1692 unsigned elems = glsl_get_length(type);
1693 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1694 for (unsigned i = 0; i < elems; i++) {
1695 const struct glsl_type *child_type;
1697 switch (glsl_get_base_type(type)) {
1699 case GLSL_TYPE_UINT:
1700 case GLSL_TYPE_INT16:
1701 case GLSL_TYPE_UINT16:
1702 case GLSL_TYPE_INT64:
1703 case GLSL_TYPE_UINT64:
1704 case GLSL_TYPE_BOOL:
1705 case GLSL_TYPE_FLOAT:
1706 case GLSL_TYPE_FLOAT16:
1707 case GLSL_TYPE_DOUBLE:
1708 child_type = glsl_get_column_type(type);
1710 case GLSL_TYPE_ARRAY:
1711 child_type = glsl_get_array_element(type);
1713 case GLSL_TYPE_STRUCT:
1714 child_type = glsl_get_struct_field(type, i);
1717 vtn_fail("unkown base type");
1720 val->elems[i] = vtn_create_ssa_value(b, child_type);
1728 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1731 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1732 src.src_type = type;
1737 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1738 const uint32_t *w, unsigned count)
1740 if (opcode == SpvOpSampledImage) {
1741 struct vtn_value *val =
1742 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1743 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1744 val->sampled_image->type =
1745 vtn_value(b, w[1], vtn_value_type_type)->type;
1746 val->sampled_image->image =
1747 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1748 val->sampled_image->sampler =
1749 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1751 } else if (opcode == SpvOpImage) {
1752 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1753 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1754 if (src_val->value_type == vtn_value_type_sampled_image) {
1755 val->pointer = src_val->sampled_image->image;
1757 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1758 val->pointer = src_val->pointer;
1763 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1764 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1766 struct vtn_sampled_image sampled;
1767 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1768 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1769 sampled = *sampled_val->sampled_image;
1771 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1772 sampled.type = sampled_val->pointer->type;
1773 sampled.image = NULL;
1774 sampled.sampler = sampled_val->pointer;
1777 const struct glsl_type *image_type = sampled.type->type;
1778 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1779 const bool is_array = glsl_sampler_type_is_array(image_type);
1780 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1782 /* Figure out the base texture operation */
1785 case SpvOpImageSampleImplicitLod:
1786 case SpvOpImageSampleDrefImplicitLod:
1787 case SpvOpImageSampleProjImplicitLod:
1788 case SpvOpImageSampleProjDrefImplicitLod:
1789 texop = nir_texop_tex;
1792 case SpvOpImageSampleExplicitLod:
1793 case SpvOpImageSampleDrefExplicitLod:
1794 case SpvOpImageSampleProjExplicitLod:
1795 case SpvOpImageSampleProjDrefExplicitLod:
1796 texop = nir_texop_txl;
1799 case SpvOpImageFetch:
1800 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1801 texop = nir_texop_txf_ms;
1803 texop = nir_texop_txf;
1807 case SpvOpImageGather:
1808 case SpvOpImageDrefGather:
1809 texop = nir_texop_tg4;
1812 case SpvOpImageQuerySizeLod:
1813 case SpvOpImageQuerySize:
1814 texop = nir_texop_txs;
1817 case SpvOpImageQueryLod:
1818 texop = nir_texop_lod;
1821 case SpvOpImageQueryLevels:
1822 texop = nir_texop_query_levels;
1825 case SpvOpImageQuerySamples:
1826 texop = nir_texop_texture_samples;
1830 vtn_fail("Unhandled opcode");
1833 nir_tex_src srcs[8]; /* 8 should be enough */
1834 nir_tex_src *p = srcs;
1838 struct nir_ssa_def *coord;
1839 unsigned coord_components;
1841 case SpvOpImageSampleImplicitLod:
1842 case SpvOpImageSampleExplicitLod:
1843 case SpvOpImageSampleDrefImplicitLod:
1844 case SpvOpImageSampleDrefExplicitLod:
1845 case SpvOpImageSampleProjImplicitLod:
1846 case SpvOpImageSampleProjExplicitLod:
1847 case SpvOpImageSampleProjDrefImplicitLod:
1848 case SpvOpImageSampleProjDrefExplicitLod:
1849 case SpvOpImageFetch:
1850 case SpvOpImageGather:
1851 case SpvOpImageDrefGather:
1852 case SpvOpImageQueryLod: {
1853 /* All these types have the coordinate as their first real argument */
1854 switch (sampler_dim) {
1855 case GLSL_SAMPLER_DIM_1D:
1856 case GLSL_SAMPLER_DIM_BUF:
1857 coord_components = 1;
1859 case GLSL_SAMPLER_DIM_2D:
1860 case GLSL_SAMPLER_DIM_RECT:
1861 case GLSL_SAMPLER_DIM_MS:
1862 coord_components = 2;
1864 case GLSL_SAMPLER_DIM_3D:
1865 case GLSL_SAMPLER_DIM_CUBE:
1866 coord_components = 3;
1869 vtn_fail("Invalid sampler type");
1872 if (is_array && texop != nir_texop_lod)
1875 coord = vtn_ssa_value(b, w[idx++])->def;
1876 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1877 (1 << coord_components) - 1));
1878 p->src_type = nir_tex_src_coord;
1885 coord_components = 0;
1890 case SpvOpImageSampleProjImplicitLod:
1891 case SpvOpImageSampleProjExplicitLod:
1892 case SpvOpImageSampleProjDrefImplicitLod:
1893 case SpvOpImageSampleProjDrefExplicitLod:
1894 /* These have the projector as the last coordinate component */
1895 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1896 p->src_type = nir_tex_src_projector;
1904 unsigned gather_component = 0;
1906 case SpvOpImageSampleDrefImplicitLod:
1907 case SpvOpImageSampleDrefExplicitLod:
1908 case SpvOpImageSampleProjDrefImplicitLod:
1909 case SpvOpImageSampleProjDrefExplicitLod:
1910 case SpvOpImageDrefGather:
1911 /* These all have an explicit depth value as their next source */
1912 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
1915 case SpvOpImageGather:
1916 /* This has a component as its next source */
1918 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
1925 /* For OpImageQuerySizeLod, we always have an LOD */
1926 if (opcode == SpvOpImageQuerySizeLod)
1927 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1929 /* Now we need to handle some number of optional arguments */
1930 const struct vtn_ssa_value *gather_offsets = NULL;
1932 uint32_t operands = w[idx++];
1934 if (operands & SpvImageOperandsBiasMask) {
1935 vtn_assert(texop == nir_texop_tex);
1936 texop = nir_texop_txb;
1937 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1940 if (operands & SpvImageOperandsLodMask) {
1941 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
1942 texop == nir_texop_txs);
1943 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1946 if (operands & SpvImageOperandsGradMask) {
1947 vtn_assert(texop == nir_texop_txl);
1948 texop = nir_texop_txd;
1949 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
1950 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
1953 if (operands & SpvImageOperandsOffsetMask ||
1954 operands & SpvImageOperandsConstOffsetMask)
1955 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
1957 if (operands & SpvImageOperandsConstOffsetsMask) {
1958 gather_offsets = vtn_ssa_value(b, w[idx++]);
1959 (*p++) = (nir_tex_src){};
1962 if (operands & SpvImageOperandsSampleMask) {
1963 vtn_assert(texop == nir_texop_txf_ms);
1964 texop = nir_texop_txf_ms;
1965 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
1968 /* We should have now consumed exactly all of the arguments */
1969 vtn_assert(idx == count);
1971 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
1974 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
1976 instr->coord_components = coord_components;
1977 instr->sampler_dim = sampler_dim;
1978 instr->is_array = is_array;
1979 instr->is_shadow = is_shadow;
1980 instr->is_new_style_shadow =
1981 is_shadow && glsl_get_components(ret_type->type) == 1;
1982 instr->component = gather_component;
1984 switch (glsl_get_sampler_result_type(image_type)) {
1985 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
1986 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
1987 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
1988 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
1990 vtn_fail("Invalid base type for sampler result");
1993 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
1994 nir_deref_var *texture;
1995 if (sampled.image) {
1996 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
2002 instr->texture = nir_deref_var_clone(texture, instr);
2004 switch (instr->op) {
2010 /* These operations require a sampler */
2011 instr->sampler = nir_deref_var_clone(sampler, instr);
2014 case nir_texop_txf_ms:
2017 case nir_texop_query_levels:
2018 case nir_texop_texture_samples:
2019 case nir_texop_samples_identical:
2021 instr->sampler = NULL;
2023 case nir_texop_txf_ms_mcs:
2024 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2027 nir_ssa_dest_init(&instr->instr, &instr->dest,
2028 nir_tex_instr_dest_size(instr), 32, NULL);
2030 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2031 nir_tex_instr_dest_size(instr));
2034 nir_instr *instruction;
2035 if (gather_offsets) {
2036 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2037 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2038 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2040 /* Copy the current instruction 4x */
2041 for (uint32_t i = 1; i < 4; i++) {
2042 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2043 instrs[i]->op = instr->op;
2044 instrs[i]->coord_components = instr->coord_components;
2045 instrs[i]->sampler_dim = instr->sampler_dim;
2046 instrs[i]->is_array = instr->is_array;
2047 instrs[i]->is_shadow = instr->is_shadow;
2048 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2049 instrs[i]->component = instr->component;
2050 instrs[i]->dest_type = instr->dest_type;
2051 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
2052 instrs[i]->sampler = NULL;
2054 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2056 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2057 nir_tex_instr_dest_size(instr), 32, NULL);
2060 /* Fill in the last argument with the offset from the passed in offsets
2061 * and insert the instruction into the stream.
2063 for (uint32_t i = 0; i < 4; i++) {
2065 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2066 src.src_type = nir_tex_src_offset;
2067 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2068 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2071 /* Combine the results of the 4 instructions by taking their .w
2074 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2075 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2076 vec4->dest.write_mask = 0xf;
2077 for (uint32_t i = 0; i < 4; i++) {
2078 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2079 vec4->src[i].swizzle[0] = 3;
2081 def = &vec4->dest.dest.ssa;
2082 instruction = &vec4->instr;
2084 def = &instr->dest.ssa;
2085 instruction = &instr->instr;
2088 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2089 val->ssa->def = def;
2091 nir_builder_instr_insert(&b->nb, instruction);
2095 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2096 const uint32_t *w, nir_src *src)
2099 case SpvOpAtomicIIncrement:
2100 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2103 case SpvOpAtomicIDecrement:
2104 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2107 case SpvOpAtomicISub:
2109 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2112 case SpvOpAtomicCompareExchange:
2113 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2114 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2117 case SpvOpAtomicExchange:
2118 case SpvOpAtomicIAdd:
2119 case SpvOpAtomicSMin:
2120 case SpvOpAtomicUMin:
2121 case SpvOpAtomicSMax:
2122 case SpvOpAtomicUMax:
2123 case SpvOpAtomicAnd:
2125 case SpvOpAtomicXor:
2126 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2130 vtn_fail("Invalid SPIR-V atomic");
2134 static nir_ssa_def *
2135 get_image_coord(struct vtn_builder *b, uint32_t value)
2137 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2139 /* The image_load_store intrinsics assume a 4-dim coordinate */
2140 unsigned dim = glsl_get_vector_elements(coord->type);
2141 unsigned swizzle[4];
2142 for (unsigned i = 0; i < 4; i++)
2143 swizzle[i] = MIN2(i, dim - 1);
2145 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2149 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2150 const uint32_t *w, unsigned count)
2152 /* Just get this one out of the way */
2153 if (opcode == SpvOpImageTexelPointer) {
2154 struct vtn_value *val =
2155 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2156 val->image = ralloc(b, struct vtn_image_pointer);
2158 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2159 val->image->coord = get_image_coord(b, w[4]);
2160 val->image->sample = vtn_ssa_value(b, w[5])->def;
2164 struct vtn_image_pointer image;
2167 case SpvOpAtomicExchange:
2168 case SpvOpAtomicCompareExchange:
2169 case SpvOpAtomicCompareExchangeWeak:
2170 case SpvOpAtomicIIncrement:
2171 case SpvOpAtomicIDecrement:
2172 case SpvOpAtomicIAdd:
2173 case SpvOpAtomicISub:
2174 case SpvOpAtomicLoad:
2175 case SpvOpAtomicSMin:
2176 case SpvOpAtomicUMin:
2177 case SpvOpAtomicSMax:
2178 case SpvOpAtomicUMax:
2179 case SpvOpAtomicAnd:
2181 case SpvOpAtomicXor:
2182 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2185 case SpvOpAtomicStore:
2186 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2189 case SpvOpImageQuerySize:
2190 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2192 image.sample = NULL;
2195 case SpvOpImageRead:
2196 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2197 image.coord = get_image_coord(b, w[4]);
2199 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2200 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2201 image.sample = vtn_ssa_value(b, w[6])->def;
2203 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2207 case SpvOpImageWrite:
2208 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2209 image.coord = get_image_coord(b, w[2]);
2213 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2214 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2215 image.sample = vtn_ssa_value(b, w[5])->def;
2217 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2222 vtn_fail("Invalid image opcode");
2225 nir_intrinsic_op op;
2227 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2228 OP(ImageQuerySize, size)
2230 OP(ImageWrite, store)
2231 OP(AtomicLoad, load)
2232 OP(AtomicStore, store)
2233 OP(AtomicExchange, atomic_exchange)
2234 OP(AtomicCompareExchange, atomic_comp_swap)
2235 OP(AtomicIIncrement, atomic_add)
2236 OP(AtomicIDecrement, atomic_add)
2237 OP(AtomicIAdd, atomic_add)
2238 OP(AtomicISub, atomic_add)
2239 OP(AtomicSMin, atomic_min)
2240 OP(AtomicUMin, atomic_min)
2241 OP(AtomicSMax, atomic_max)
2242 OP(AtomicUMax, atomic_max)
2243 OP(AtomicAnd, atomic_and)
2244 OP(AtomicOr, atomic_or)
2245 OP(AtomicXor, atomic_xor)
2248 vtn_fail("Invalid image opcode");
2251 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2253 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2254 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2256 /* ImageQuerySize doesn't take any extra parameters */
2257 if (opcode != SpvOpImageQuerySize) {
2258 /* The image coordinate is always 4 components but we may not have that
2259 * many. Swizzle to compensate.
2262 for (unsigned i = 0; i < 4; i++)
2263 swiz[i] = i < image.coord->num_components ? i : 0;
2264 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2266 intrin->src[1] = nir_src_for_ssa(image.sample);
2270 case SpvOpAtomicLoad:
2271 case SpvOpImageQuerySize:
2272 case SpvOpImageRead:
2274 case SpvOpAtomicStore:
2275 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2277 case SpvOpImageWrite:
2278 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2281 case SpvOpAtomicCompareExchange:
2282 case SpvOpAtomicIIncrement:
2283 case SpvOpAtomicIDecrement:
2284 case SpvOpAtomicExchange:
2285 case SpvOpAtomicIAdd:
2286 case SpvOpAtomicISub:
2287 case SpvOpAtomicSMin:
2288 case SpvOpAtomicUMin:
2289 case SpvOpAtomicSMax:
2290 case SpvOpAtomicUMax:
2291 case SpvOpAtomicAnd:
2293 case SpvOpAtomicXor:
2294 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2298 vtn_fail("Invalid image opcode");
2301 if (opcode != SpvOpImageWrite) {
2302 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2303 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2305 unsigned dest_components =
2306 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2307 if (intrin->intrinsic == nir_intrinsic_image_size) {
2308 dest_components = intrin->num_components =
2309 glsl_get_vector_elements(type->type);
2312 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2313 dest_components, 32, NULL);
2315 nir_builder_instr_insert(&b->nb, &intrin->instr);
2317 val->ssa = vtn_create_ssa_value(b, type->type);
2318 val->ssa->def = &intrin->dest.ssa;
2320 nir_builder_instr_insert(&b->nb, &intrin->instr);
2324 static nir_intrinsic_op
2325 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2328 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2329 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2330 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2331 OP(AtomicExchange, atomic_exchange)
2332 OP(AtomicCompareExchange, atomic_comp_swap)
2333 OP(AtomicIIncrement, atomic_add)
2334 OP(AtomicIDecrement, atomic_add)
2335 OP(AtomicIAdd, atomic_add)
2336 OP(AtomicISub, atomic_add)
2337 OP(AtomicSMin, atomic_imin)
2338 OP(AtomicUMin, atomic_umin)
2339 OP(AtomicSMax, atomic_imax)
2340 OP(AtomicUMax, atomic_umax)
2341 OP(AtomicAnd, atomic_and)
2342 OP(AtomicOr, atomic_or)
2343 OP(AtomicXor, atomic_xor)
2346 vtn_fail("Invalid SSBO atomic");
2350 static nir_intrinsic_op
2351 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2354 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2355 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2356 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2357 OP(AtomicExchange, atomic_exchange)
2358 OP(AtomicCompareExchange, atomic_comp_swap)
2359 OP(AtomicIIncrement, atomic_add)
2360 OP(AtomicIDecrement, atomic_add)
2361 OP(AtomicIAdd, atomic_add)
2362 OP(AtomicISub, atomic_add)
2363 OP(AtomicSMin, atomic_imin)
2364 OP(AtomicUMin, atomic_umin)
2365 OP(AtomicSMax, atomic_imax)
2366 OP(AtomicUMax, atomic_umax)
2367 OP(AtomicAnd, atomic_and)
2368 OP(AtomicOr, atomic_or)
2369 OP(AtomicXor, atomic_xor)
2372 vtn_fail("Invalid shared atomic");
2376 static nir_intrinsic_op
2377 get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2380 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2381 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2382 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2383 OP(AtomicExchange, atomic_exchange)
2384 OP(AtomicCompareExchange, atomic_comp_swap)
2385 OP(AtomicIIncrement, atomic_add)
2386 OP(AtomicIDecrement, atomic_add)
2387 OP(AtomicIAdd, atomic_add)
2388 OP(AtomicISub, atomic_add)
2389 OP(AtomicSMin, atomic_imin)
2390 OP(AtomicUMin, atomic_umin)
2391 OP(AtomicSMax, atomic_imax)
2392 OP(AtomicUMax, atomic_umax)
2393 OP(AtomicAnd, atomic_and)
2394 OP(AtomicOr, atomic_or)
2395 OP(AtomicXor, atomic_xor)
2398 vtn_fail("Invalid shared atomic");
2403 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2404 const uint32_t *w, unsigned count)
2406 struct vtn_pointer *ptr;
2407 nir_intrinsic_instr *atomic;
2410 case SpvOpAtomicLoad:
2411 case SpvOpAtomicExchange:
2412 case SpvOpAtomicCompareExchange:
2413 case SpvOpAtomicCompareExchangeWeak:
2414 case SpvOpAtomicIIncrement:
2415 case SpvOpAtomicIDecrement:
2416 case SpvOpAtomicIAdd:
2417 case SpvOpAtomicISub:
2418 case SpvOpAtomicSMin:
2419 case SpvOpAtomicUMin:
2420 case SpvOpAtomicSMax:
2421 case SpvOpAtomicUMax:
2422 case SpvOpAtomicAnd:
2424 case SpvOpAtomicXor:
2425 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2428 case SpvOpAtomicStore:
2429 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2433 vtn_fail("Invalid SPIR-V atomic");
2437 SpvScope scope = w[4];
2438 SpvMemorySemanticsMask semantics = w[5];
2441 if (ptr->mode == vtn_variable_mode_workgroup &&
2442 !b->options->lower_workgroup_access_to_offsets) {
2443 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2444 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2445 nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
2446 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2447 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2450 case SpvOpAtomicLoad:
2451 atomic->num_components = glsl_get_vector_elements(deref_type);
2454 case SpvOpAtomicStore:
2455 atomic->num_components = glsl_get_vector_elements(deref_type);
2456 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2457 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2460 case SpvOpAtomicExchange:
2461 case SpvOpAtomicCompareExchange:
2462 case SpvOpAtomicCompareExchangeWeak:
2463 case SpvOpAtomicIIncrement:
2464 case SpvOpAtomicIDecrement:
2465 case SpvOpAtomicIAdd:
2466 case SpvOpAtomicISub:
2467 case SpvOpAtomicSMin:
2468 case SpvOpAtomicUMin:
2469 case SpvOpAtomicSMax:
2470 case SpvOpAtomicUMax:
2471 case SpvOpAtomicAnd:
2473 case SpvOpAtomicXor:
2474 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2478 vtn_fail("Invalid SPIR-V atomic");
2482 nir_ssa_def *offset, *index;
2483 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2485 nir_intrinsic_op op;
2486 if (ptr->mode == vtn_variable_mode_ssbo) {
2487 op = get_ssbo_nir_atomic_op(b, opcode);
2489 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2490 b->options->lower_workgroup_access_to_offsets);
2491 op = get_shared_nir_atomic_op(b, opcode);
2494 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2498 case SpvOpAtomicLoad:
2499 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2500 if (ptr->mode == vtn_variable_mode_ssbo)
2501 atomic->src[src++] = nir_src_for_ssa(index);
2502 atomic->src[src++] = nir_src_for_ssa(offset);
2505 case SpvOpAtomicStore:
2506 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2507 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2508 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2509 if (ptr->mode == vtn_variable_mode_ssbo)
2510 atomic->src[src++] = nir_src_for_ssa(index);
2511 atomic->src[src++] = nir_src_for_ssa(offset);
2514 case SpvOpAtomicExchange:
2515 case SpvOpAtomicCompareExchange:
2516 case SpvOpAtomicCompareExchangeWeak:
2517 case SpvOpAtomicIIncrement:
2518 case SpvOpAtomicIDecrement:
2519 case SpvOpAtomicIAdd:
2520 case SpvOpAtomicISub:
2521 case SpvOpAtomicSMin:
2522 case SpvOpAtomicUMin:
2523 case SpvOpAtomicSMax:
2524 case SpvOpAtomicUMax:
2525 case SpvOpAtomicAnd:
2527 case SpvOpAtomicXor:
2528 if (ptr->mode == vtn_variable_mode_ssbo)
2529 atomic->src[src++] = nir_src_for_ssa(index);
2530 atomic->src[src++] = nir_src_for_ssa(offset);
2531 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2535 vtn_fail("Invalid SPIR-V atomic");
2539 if (opcode != SpvOpAtomicStore) {
2540 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2542 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2543 glsl_get_vector_elements(type->type),
2544 glsl_get_bit_size(type->type), NULL);
2546 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2547 val->ssa = rzalloc(b, struct vtn_ssa_value);
2548 val->ssa->def = &atomic->dest.ssa;
2549 val->ssa->type = type->type;
2552 nir_builder_instr_insert(&b->nb, &atomic->instr);
2555 static nir_alu_instr *
2556 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2559 switch (num_components) {
2560 case 1: op = nir_op_fmov; break;
2561 case 2: op = nir_op_vec2; break;
2562 case 3: op = nir_op_vec3; break;
2563 case 4: op = nir_op_vec4; break;
2564 default: vtn_fail("bad vector size");
2567 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2568 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2570 vec->dest.write_mask = (1 << num_components) - 1;
2575 struct vtn_ssa_value *
2576 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2578 if (src->transposed)
2579 return src->transposed;
2581 struct vtn_ssa_value *dest =
2582 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2584 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2585 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2586 glsl_get_bit_size(src->type));
2587 if (glsl_type_is_vector_or_scalar(src->type)) {
2588 vec->src[0].src = nir_src_for_ssa(src->def);
2589 vec->src[0].swizzle[0] = i;
2591 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2592 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2593 vec->src[j].swizzle[0] = i;
2596 nir_builder_instr_insert(&b->nb, &vec->instr);
2597 dest->elems[i]->def = &vec->dest.dest.ssa;
2600 dest->transposed = src;
2606 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2608 unsigned swiz[4] = { index };
2609 return nir_swizzle(&b->nb, src, swiz, 1, true);
2613 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2616 nir_alu_instr *vec = create_vec(b, src->num_components,
2619 for (unsigned i = 0; i < src->num_components; i++) {
2621 vec->src[i].src = nir_src_for_ssa(insert);
2623 vec->src[i].src = nir_src_for_ssa(src);
2624 vec->src[i].swizzle[0] = i;
2628 nir_builder_instr_insert(&b->nb, &vec->instr);
2630 return &vec->dest.dest.ssa;
2634 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2637 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2638 for (unsigned i = 1; i < src->num_components; i++)
2639 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2640 vtn_vector_extract(b, src, i), dest);
2646 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2647 nir_ssa_def *insert, nir_ssa_def *index)
2649 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2650 for (unsigned i = 1; i < src->num_components; i++)
2651 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2652 vtn_vector_insert(b, src, insert, i), dest);
2657 static nir_ssa_def *
2658 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2659 nir_ssa_def *src0, nir_ssa_def *src1,
2660 const uint32_t *indices)
2662 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2664 for (unsigned i = 0; i < num_components; i++) {
2665 uint32_t index = indices[i];
2666 if (index == 0xffffffff) {
2668 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2669 } else if (index < src0->num_components) {
2670 vec->src[i].src = nir_src_for_ssa(src0);
2671 vec->src[i].swizzle[0] = index;
2673 vec->src[i].src = nir_src_for_ssa(src1);
2674 vec->src[i].swizzle[0] = index - src0->num_components;
2678 nir_builder_instr_insert(&b->nb, &vec->instr);
2680 return &vec->dest.dest.ssa;
2684 * Concatentates a number of vectors/scalars together to produce a vector
2686 static nir_ssa_def *
2687 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2688 unsigned num_srcs, nir_ssa_def **srcs)
2690 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2692 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2694 * "When constructing a vector, there must be at least two Constituent
2697 vtn_assert(num_srcs >= 2);
2699 unsigned dest_idx = 0;
2700 for (unsigned i = 0; i < num_srcs; i++) {
2701 nir_ssa_def *src = srcs[i];
2702 vtn_assert(dest_idx + src->num_components <= num_components);
2703 for (unsigned j = 0; j < src->num_components; j++) {
2704 vec->src[dest_idx].src = nir_src_for_ssa(src);
2705 vec->src[dest_idx].swizzle[0] = j;
2710 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2712 * "When constructing a vector, the total number of components in all
2713 * the operands must equal the number of components in Result Type."
2715 vtn_assert(dest_idx == num_components);
2717 nir_builder_instr_insert(&b->nb, &vec->instr);
2719 return &vec->dest.dest.ssa;
2722 static struct vtn_ssa_value *
2723 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2725 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2726 dest->type = src->type;
2728 if (glsl_type_is_vector_or_scalar(src->type)) {
2729 dest->def = src->def;
2731 unsigned elems = glsl_get_length(src->type);
2733 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2734 for (unsigned i = 0; i < elems; i++)
2735 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2741 static struct vtn_ssa_value *
2742 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2743 struct vtn_ssa_value *insert, const uint32_t *indices,
2744 unsigned num_indices)
2746 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2748 struct vtn_ssa_value *cur = dest;
2750 for (i = 0; i < num_indices - 1; i++) {
2751 cur = cur->elems[indices[i]];
2754 if (glsl_type_is_vector_or_scalar(cur->type)) {
2755 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2756 * the component granularity. In that case, the last index will be
2757 * the index to insert the scalar into the vector.
2760 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2762 cur->elems[indices[i]] = insert;
2768 static struct vtn_ssa_value *
2769 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2770 const uint32_t *indices, unsigned num_indices)
2772 struct vtn_ssa_value *cur = src;
2773 for (unsigned i = 0; i < num_indices; i++) {
2774 if (glsl_type_is_vector_or_scalar(cur->type)) {
2775 vtn_assert(i == num_indices - 1);
2776 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2777 * the component granularity. The last index will be the index of the
2778 * vector to extract.
2781 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2782 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2783 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2786 cur = cur->elems[indices[i]];
2794 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2795 const uint32_t *w, unsigned count)
2797 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2798 const struct glsl_type *type =
2799 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2800 val->ssa = vtn_create_ssa_value(b, type);
2803 case SpvOpVectorExtractDynamic:
2804 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2805 vtn_ssa_value(b, w[4])->def);
2808 case SpvOpVectorInsertDynamic:
2809 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2810 vtn_ssa_value(b, w[4])->def,
2811 vtn_ssa_value(b, w[5])->def);
2814 case SpvOpVectorShuffle:
2815 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2816 vtn_ssa_value(b, w[3])->def,
2817 vtn_ssa_value(b, w[4])->def,
2821 case SpvOpCompositeConstruct: {
2822 unsigned elems = count - 3;
2823 if (glsl_type_is_vector_or_scalar(type)) {
2824 nir_ssa_def *srcs[4];
2825 for (unsigned i = 0; i < elems; i++)
2826 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2828 vtn_vector_construct(b, glsl_get_vector_elements(type),
2831 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2832 for (unsigned i = 0; i < elems; i++)
2833 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2837 case SpvOpCompositeExtract:
2838 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2842 case SpvOpCompositeInsert:
2843 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2844 vtn_ssa_value(b, w[3]),
2848 case SpvOpCopyObject:
2849 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2853 vtn_fail("unknown composite operation");
2858 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
2859 const uint32_t *w, unsigned count)
2861 nir_intrinsic_op intrinsic_op;
2863 case SpvOpEmitVertex:
2864 case SpvOpEmitStreamVertex:
2865 intrinsic_op = nir_intrinsic_emit_vertex;
2867 case SpvOpEndPrimitive:
2868 case SpvOpEndStreamPrimitive:
2869 intrinsic_op = nir_intrinsic_end_primitive;
2871 case SpvOpMemoryBarrier:
2872 intrinsic_op = nir_intrinsic_memory_barrier;
2874 case SpvOpControlBarrier:
2875 intrinsic_op = nir_intrinsic_barrier;
2878 vtn_fail("unknown barrier instruction");
2881 nir_intrinsic_instr *intrin =
2882 nir_intrinsic_instr_create(b->shader, intrinsic_op);
2884 if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
2885 nir_intrinsic_set_stream_id(intrin, w[1]);
2887 nir_builder_instr_insert(&b->nb, &intrin->instr);
2891 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
2892 SpvExecutionMode mode)
2895 case SpvExecutionModeInputPoints:
2896 case SpvExecutionModeOutputPoints:
2897 return 0; /* GL_POINTS */
2898 case SpvExecutionModeInputLines:
2899 return 1; /* GL_LINES */
2900 case SpvExecutionModeInputLinesAdjacency:
2901 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2902 case SpvExecutionModeTriangles:
2903 return 4; /* GL_TRIANGLES */
2904 case SpvExecutionModeInputTrianglesAdjacency:
2905 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2906 case SpvExecutionModeQuads:
2907 return 7; /* GL_QUADS */
2908 case SpvExecutionModeIsolines:
2909 return 0x8E7A; /* GL_ISOLINES */
2910 case SpvExecutionModeOutputLineStrip:
2911 return 3; /* GL_LINE_STRIP */
2912 case SpvExecutionModeOutputTriangleStrip:
2913 return 5; /* GL_TRIANGLE_STRIP */
2915 vtn_fail("Invalid primitive type");
2920 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
2921 SpvExecutionMode mode)
2924 case SpvExecutionModeInputPoints:
2926 case SpvExecutionModeInputLines:
2928 case SpvExecutionModeInputLinesAdjacency:
2930 case SpvExecutionModeTriangles:
2932 case SpvExecutionModeInputTrianglesAdjacency:
2935 vtn_fail("Invalid GS input mode");
2939 static gl_shader_stage
2940 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
2943 case SpvExecutionModelVertex:
2944 return MESA_SHADER_VERTEX;
2945 case SpvExecutionModelTessellationControl:
2946 return MESA_SHADER_TESS_CTRL;
2947 case SpvExecutionModelTessellationEvaluation:
2948 return MESA_SHADER_TESS_EVAL;
2949 case SpvExecutionModelGeometry:
2950 return MESA_SHADER_GEOMETRY;
2951 case SpvExecutionModelFragment:
2952 return MESA_SHADER_FRAGMENT;
2953 case SpvExecutionModelGLCompute:
2954 return MESA_SHADER_COMPUTE;
2956 vtn_fail("Unsupported execution model");
2960 #define spv_check_supported(name, cap) do { \
2961 if (!(b->options && b->options->caps.name)) \
2962 vtn_warn("Unsupported SPIR-V capability: %s", \
2963 spirv_capability_to_string(cap)); \
2967 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
2968 const uint32_t *w, unsigned count)
2975 case SpvSourceLanguageUnknown: lang = "unknown"; break;
2976 case SpvSourceLanguageESSL: lang = "ESSL"; break;
2977 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
2978 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
2979 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
2980 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
2983 uint32_t version = w[2];
2986 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
2988 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
2992 case SpvOpSourceExtension:
2993 case SpvOpSourceContinued:
2994 case SpvOpExtension:
2995 /* Unhandled, but these are for debug so that's ok. */
2998 case SpvOpCapability: {
2999 SpvCapability cap = w[1];
3001 case SpvCapabilityMatrix:
3002 case SpvCapabilityShader:
3003 case SpvCapabilityGeometry:
3004 case SpvCapabilityGeometryPointSize:
3005 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3006 case SpvCapabilitySampledImageArrayDynamicIndexing:
3007 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3008 case SpvCapabilityStorageImageArrayDynamicIndexing:
3009 case SpvCapabilityImageRect:
3010 case SpvCapabilitySampledRect:
3011 case SpvCapabilitySampled1D:
3012 case SpvCapabilityImage1D:
3013 case SpvCapabilitySampledCubeArray:
3014 case SpvCapabilityImageCubeArray:
3015 case SpvCapabilitySampledBuffer:
3016 case SpvCapabilityImageBuffer:
3017 case SpvCapabilityImageQuery:
3018 case SpvCapabilityDerivativeControl:
3019 case SpvCapabilityInterpolationFunction:
3020 case SpvCapabilityMultiViewport:
3021 case SpvCapabilitySampleRateShading:
3022 case SpvCapabilityClipDistance:
3023 case SpvCapabilityCullDistance:
3024 case SpvCapabilityInputAttachment:
3025 case SpvCapabilityImageGatherExtended:
3026 case SpvCapabilityStorageImageExtendedFormats:
3029 case SpvCapabilityGeometryStreams:
3030 case SpvCapabilityLinkage:
3031 case SpvCapabilityVector16:
3032 case SpvCapabilityFloat16Buffer:
3033 case SpvCapabilityFloat16:
3034 case SpvCapabilityInt64Atomics:
3035 case SpvCapabilityAtomicStorage:
3036 case SpvCapabilityInt16:
3037 case SpvCapabilityStorageImageMultisample:
3038 case SpvCapabilityInt8:
3039 case SpvCapabilitySparseResidency:
3040 case SpvCapabilityMinLod:
3041 case SpvCapabilityTransformFeedback:
3042 vtn_warn("Unsupported SPIR-V capability: %s",
3043 spirv_capability_to_string(cap));
3046 case SpvCapabilityFloat64:
3047 spv_check_supported(float64, cap);
3049 case SpvCapabilityInt64:
3050 spv_check_supported(int64, cap);
3053 case SpvCapabilityAddresses:
3054 case SpvCapabilityKernel:
3055 case SpvCapabilityImageBasic:
3056 case SpvCapabilityImageReadWrite:
3057 case SpvCapabilityImageMipmap:
3058 case SpvCapabilityPipes:
3059 case SpvCapabilityGroups:
3060 case SpvCapabilityDeviceEnqueue:
3061 case SpvCapabilityLiteralSampler:
3062 case SpvCapabilityGenericPointer:
3063 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3064 spirv_capability_to_string(cap));
3067 case SpvCapabilityImageMSArray:
3068 spv_check_supported(image_ms_array, cap);
3071 case SpvCapabilityTessellation:
3072 case SpvCapabilityTessellationPointSize:
3073 spv_check_supported(tessellation, cap);
3076 case SpvCapabilityDrawParameters:
3077 spv_check_supported(draw_parameters, cap);
3080 case SpvCapabilityStorageImageReadWithoutFormat:
3081 spv_check_supported(image_read_without_format, cap);
3084 case SpvCapabilityStorageImageWriteWithoutFormat:
3085 spv_check_supported(image_write_without_format, cap);
3088 case SpvCapabilityMultiView:
3089 spv_check_supported(multiview, cap);
3092 case SpvCapabilityVariablePointersStorageBuffer:
3093 case SpvCapabilityVariablePointers:
3094 spv_check_supported(variable_pointers, cap);
3097 case SpvCapabilityStorageUniformBufferBlock16:
3098 case SpvCapabilityStorageUniform16:
3099 case SpvCapabilityStoragePushConstant16:
3100 case SpvCapabilityStorageInputOutput16:
3101 spv_check_supported(storage_16bit, cap);
3105 vtn_fail("Unhandled capability");
3110 case SpvOpExtInstImport:
3111 vtn_handle_extension(b, opcode, w, count);
3114 case SpvOpMemoryModel:
3115 vtn_assert(w[1] == SpvAddressingModelLogical);
3116 vtn_assert(w[2] == SpvMemoryModelSimple ||
3117 w[2] == SpvMemoryModelGLSL450);
3120 case SpvOpEntryPoint: {
3121 struct vtn_value *entry_point = &b->values[w[2]];
3122 /* Let this be a name label regardless */
3123 unsigned name_words;
3124 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3126 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3127 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3130 vtn_assert(b->entry_point == NULL);
3131 b->entry_point = entry_point;
3136 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3137 vtn_string_literal(b, &w[2], count - 2, NULL);
3141 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3144 case SpvOpMemberName:
3148 case SpvOpExecutionMode:
3149 case SpvOpDecorationGroup:
3151 case SpvOpMemberDecorate:
3152 case SpvOpGroupDecorate:
3153 case SpvOpGroupMemberDecorate:
3154 vtn_handle_decoration(b, opcode, w, count);
3158 return false; /* End of preamble */
3165 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3166 const struct vtn_decoration *mode, void *data)
3168 vtn_assert(b->entry_point == entry_point);
3170 switch(mode->exec_mode) {
3171 case SpvExecutionModeOriginUpperLeft:
3172 case SpvExecutionModeOriginLowerLeft:
3173 b->origin_upper_left =
3174 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3177 case SpvExecutionModeEarlyFragmentTests:
3178 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3179 b->shader->info.fs.early_fragment_tests = true;
3182 case SpvExecutionModeInvocations:
3183 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3184 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3187 case SpvExecutionModeDepthReplacing:
3188 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3189 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3191 case SpvExecutionModeDepthGreater:
3192 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3193 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3195 case SpvExecutionModeDepthLess:
3196 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3197 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3199 case SpvExecutionModeDepthUnchanged:
3200 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3201 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3204 case SpvExecutionModeLocalSize:
3205 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3206 b->shader->info.cs.local_size[0] = mode->literals[0];
3207 b->shader->info.cs.local_size[1] = mode->literals[1];
3208 b->shader->info.cs.local_size[2] = mode->literals[2];
3210 case SpvExecutionModeLocalSizeHint:
3211 break; /* Nothing to do with this */
3213 case SpvExecutionModeOutputVertices:
3214 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3215 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3216 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3218 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3219 b->shader->info.gs.vertices_out = mode->literals[0];
3223 case SpvExecutionModeInputPoints:
3224 case SpvExecutionModeInputLines:
3225 case SpvExecutionModeInputLinesAdjacency:
3226 case SpvExecutionModeTriangles:
3227 case SpvExecutionModeInputTrianglesAdjacency:
3228 case SpvExecutionModeQuads:
3229 case SpvExecutionModeIsolines:
3230 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3231 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3232 b->shader->info.tess.primitive_mode =
3233 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3235 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3236 b->shader->info.gs.vertices_in =
3237 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3241 case SpvExecutionModeOutputPoints:
3242 case SpvExecutionModeOutputLineStrip:
3243 case SpvExecutionModeOutputTriangleStrip:
3244 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3245 b->shader->info.gs.output_primitive =
3246 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3249 case SpvExecutionModeSpacingEqual:
3250 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3251 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3252 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3254 case SpvExecutionModeSpacingFractionalEven:
3255 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3256 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3257 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3259 case SpvExecutionModeSpacingFractionalOdd:
3260 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3261 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3262 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3264 case SpvExecutionModeVertexOrderCw:
3265 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3266 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3267 b->shader->info.tess.ccw = false;
3269 case SpvExecutionModeVertexOrderCcw:
3270 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3271 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3272 b->shader->info.tess.ccw = true;
3274 case SpvExecutionModePointMode:
3275 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3276 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3277 b->shader->info.tess.point_mode = true;
3280 case SpvExecutionModePixelCenterInteger:
3281 b->pixel_center_integer = true;
3284 case SpvExecutionModeXfb:
3285 vtn_fail("Unhandled execution mode");
3288 case SpvExecutionModeVecTypeHint:
3289 case SpvExecutionModeContractionOff:
3293 vtn_fail("Unhandled execution mode");
3298 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3299 const uint32_t *w, unsigned count)
3301 vtn_set_instruction_result_type(b, opcode, w, count);
3305 case SpvOpSourceContinued:
3306 case SpvOpSourceExtension:
3307 case SpvOpExtension:
3308 case SpvOpCapability:
3309 case SpvOpExtInstImport:
3310 case SpvOpMemoryModel:
3311 case SpvOpEntryPoint:
3312 case SpvOpExecutionMode:
3315 case SpvOpMemberName:
3316 case SpvOpDecorationGroup:
3318 case SpvOpMemberDecorate:
3319 case SpvOpGroupDecorate:
3320 case SpvOpGroupMemberDecorate:
3321 vtn_fail("Invalid opcode types and variables section");
3327 case SpvOpTypeFloat:
3328 case SpvOpTypeVector:
3329 case SpvOpTypeMatrix:
3330 case SpvOpTypeImage:
3331 case SpvOpTypeSampler:
3332 case SpvOpTypeSampledImage:
3333 case SpvOpTypeArray:
3334 case SpvOpTypeRuntimeArray:
3335 case SpvOpTypeStruct:
3336 case SpvOpTypeOpaque:
3337 case SpvOpTypePointer:
3338 case SpvOpTypeFunction:
3339 case SpvOpTypeEvent:
3340 case SpvOpTypeDeviceEvent:
3341 case SpvOpTypeReserveId:
3342 case SpvOpTypeQueue:
3344 vtn_handle_type(b, opcode, w, count);
3347 case SpvOpConstantTrue:
3348 case SpvOpConstantFalse:
3350 case SpvOpConstantComposite:
3351 case SpvOpConstantSampler:
3352 case SpvOpConstantNull:
3353 case SpvOpSpecConstantTrue:
3354 case SpvOpSpecConstantFalse:
3355 case SpvOpSpecConstant:
3356 case SpvOpSpecConstantComposite:
3357 case SpvOpSpecConstantOp:
3358 vtn_handle_constant(b, opcode, w, count);
3363 vtn_handle_variables(b, opcode, w, count);
3367 return false; /* End of preamble */
3374 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3375 const uint32_t *w, unsigned count)
3381 case SpvOpLoopMerge:
3382 case SpvOpSelectionMerge:
3383 /* This is handled by cfg pre-pass and walk_blocks */
3387 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3388 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3393 vtn_handle_extension(b, opcode, w, count);
3399 case SpvOpCopyMemory:
3400 case SpvOpCopyMemorySized:
3401 case SpvOpAccessChain:
3402 case SpvOpPtrAccessChain:
3403 case SpvOpInBoundsAccessChain:
3404 case SpvOpArrayLength:
3405 vtn_handle_variables(b, opcode, w, count);
3408 case SpvOpFunctionCall:
3409 vtn_handle_function_call(b, opcode, w, count);
3412 case SpvOpSampledImage:
3414 case SpvOpImageSampleImplicitLod:
3415 case SpvOpImageSampleExplicitLod:
3416 case SpvOpImageSampleDrefImplicitLod:
3417 case SpvOpImageSampleDrefExplicitLod:
3418 case SpvOpImageSampleProjImplicitLod:
3419 case SpvOpImageSampleProjExplicitLod:
3420 case SpvOpImageSampleProjDrefImplicitLod:
3421 case SpvOpImageSampleProjDrefExplicitLod:
3422 case SpvOpImageFetch:
3423 case SpvOpImageGather:
3424 case SpvOpImageDrefGather:
3425 case SpvOpImageQuerySizeLod:
3426 case SpvOpImageQueryLod:
3427 case SpvOpImageQueryLevels:
3428 case SpvOpImageQuerySamples:
3429 vtn_handle_texture(b, opcode, w, count);
3432 case SpvOpImageRead:
3433 case SpvOpImageWrite:
3434 case SpvOpImageTexelPointer:
3435 vtn_handle_image(b, opcode, w, count);
3438 case SpvOpImageQuerySize: {
3439 struct vtn_pointer *image =
3440 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3441 if (image->mode == vtn_variable_mode_image) {
3442 vtn_handle_image(b, opcode, w, count);
3444 vtn_assert(image->mode == vtn_variable_mode_sampler);
3445 vtn_handle_texture(b, opcode, w, count);
3450 case SpvOpAtomicLoad:
3451 case SpvOpAtomicExchange:
3452 case SpvOpAtomicCompareExchange:
3453 case SpvOpAtomicCompareExchangeWeak:
3454 case SpvOpAtomicIIncrement:
3455 case SpvOpAtomicIDecrement:
3456 case SpvOpAtomicIAdd:
3457 case SpvOpAtomicISub:
3458 case SpvOpAtomicSMin:
3459 case SpvOpAtomicUMin:
3460 case SpvOpAtomicSMax:
3461 case SpvOpAtomicUMax:
3462 case SpvOpAtomicAnd:
3464 case SpvOpAtomicXor: {
3465 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3466 if (pointer->value_type == vtn_value_type_image_pointer) {
3467 vtn_handle_image(b, opcode, w, count);
3469 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3470 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3475 case SpvOpAtomicStore: {
3476 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3477 if (pointer->value_type == vtn_value_type_image_pointer) {
3478 vtn_handle_image(b, opcode, w, count);
3480 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3481 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3487 /* Handle OpSelect up-front here because it needs to be able to handle
3488 * pointers and not just regular vectors and scalars.
3490 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3491 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3492 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3493 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3495 const struct glsl_type *sel_type;
3496 switch (res_val->type->base_type) {
3497 case vtn_base_type_scalar:
3498 sel_type = glsl_bool_type();
3500 case vtn_base_type_vector:
3501 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3503 case vtn_base_type_pointer:
3504 /* We need to have actual storage for pointer types */
3505 vtn_fail_if(res_val->type->type == NULL,
3506 "Invalid pointer result type for OpSelect");
3507 sel_type = glsl_bool_type();
3510 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3513 if (unlikely(sel_val->type->type != sel_type)) {
3514 if (sel_val->type->type == glsl_bool_type()) {
3515 /* This case is illegal but some older versions of GLSLang produce
3516 * it. The GLSLang issue was fixed on March 30, 2017:
3518 * https://github.com/KhronosGroup/glslang/issues/809
3520 * Unfortunately, there are applications in the wild which are
3521 * shipping with this bug so it isn't nice to fail on them so we
3522 * throw a warning instead. It's not actually a problem for us as
3523 * nir_builder will just splat the condition out which is most
3524 * likely what the client wanted anyway.
3526 vtn_warn("Condition type of OpSelect must have the same number "
3527 "of components as Result Type");
3529 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3530 "of Boolean type. It must have the same number of "
3531 "components as Result Type");
3535 vtn_fail_if(obj1_val->type != res_val->type ||
3536 obj2_val->type != res_val->type,
3537 "Object types must match the result type in OpSelect");
3539 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3540 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3541 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3542 vtn_ssa_value(b, w[4])->def,
3543 vtn_ssa_value(b, w[5])->def);
3544 vtn_push_ssa(b, w[2], res_type, ssa);
3553 case SpvOpConvertFToU:
3554 case SpvOpConvertFToS:
3555 case SpvOpConvertSToF:
3556 case SpvOpConvertUToF:
3560 case SpvOpQuantizeToF16:
3561 case SpvOpConvertPtrToU:
3562 case SpvOpConvertUToPtr:
3563 case SpvOpPtrCastToGeneric:
3564 case SpvOpGenericCastToPtr:
3570 case SpvOpSignBitSet:
3571 case SpvOpLessOrGreater:
3573 case SpvOpUnordered:
3588 case SpvOpVectorTimesScalar:
3590 case SpvOpIAddCarry:
3591 case SpvOpISubBorrow:
3592 case SpvOpUMulExtended:
3593 case SpvOpSMulExtended:
3594 case SpvOpShiftRightLogical:
3595 case SpvOpShiftRightArithmetic:
3596 case SpvOpShiftLeftLogical:
3597 case SpvOpLogicalEqual:
3598 case SpvOpLogicalNotEqual:
3599 case SpvOpLogicalOr:
3600 case SpvOpLogicalAnd:
3601 case SpvOpLogicalNot:
3602 case SpvOpBitwiseOr:
3603 case SpvOpBitwiseXor:
3604 case SpvOpBitwiseAnd:
3606 case SpvOpFOrdEqual:
3607 case SpvOpFUnordEqual:
3608 case SpvOpINotEqual:
3609 case SpvOpFOrdNotEqual:
3610 case SpvOpFUnordNotEqual:
3611 case SpvOpULessThan:
3612 case SpvOpSLessThan:
3613 case SpvOpFOrdLessThan:
3614 case SpvOpFUnordLessThan:
3615 case SpvOpUGreaterThan:
3616 case SpvOpSGreaterThan:
3617 case SpvOpFOrdGreaterThan:
3618 case SpvOpFUnordGreaterThan:
3619 case SpvOpULessThanEqual:
3620 case SpvOpSLessThanEqual:
3621 case SpvOpFOrdLessThanEqual:
3622 case SpvOpFUnordLessThanEqual:
3623 case SpvOpUGreaterThanEqual:
3624 case SpvOpSGreaterThanEqual:
3625 case SpvOpFOrdGreaterThanEqual:
3626 case SpvOpFUnordGreaterThanEqual:
3632 case SpvOpFwidthFine:
3633 case SpvOpDPdxCoarse:
3634 case SpvOpDPdyCoarse:
3635 case SpvOpFwidthCoarse:
3636 case SpvOpBitFieldInsert:
3637 case SpvOpBitFieldSExtract:
3638 case SpvOpBitFieldUExtract:
3639 case SpvOpBitReverse:
3641 case SpvOpTranspose:
3642 case SpvOpOuterProduct:
3643 case SpvOpMatrixTimesScalar:
3644 case SpvOpVectorTimesMatrix:
3645 case SpvOpMatrixTimesVector:
3646 case SpvOpMatrixTimesMatrix:
3647 vtn_handle_alu(b, opcode, w, count);
3650 case SpvOpVectorExtractDynamic:
3651 case SpvOpVectorInsertDynamic:
3652 case SpvOpVectorShuffle:
3653 case SpvOpCompositeConstruct:
3654 case SpvOpCompositeExtract:
3655 case SpvOpCompositeInsert:
3656 case SpvOpCopyObject:
3657 vtn_handle_composite(b, opcode, w, count);
3660 case SpvOpEmitVertex:
3661 case SpvOpEndPrimitive:
3662 case SpvOpEmitStreamVertex:
3663 case SpvOpEndStreamPrimitive:
3664 case SpvOpControlBarrier:
3665 case SpvOpMemoryBarrier:
3666 vtn_handle_barrier(b, opcode, w, count);
3670 vtn_fail("Unhandled opcode");
3677 spirv_to_nir(const uint32_t *words, size_t word_count,
3678 struct nir_spirv_specialization *spec, unsigned num_spec,
3679 gl_shader_stage stage, const char *entry_point_name,
3680 const struct spirv_to_nir_options *options,
3681 const nir_shader_compiler_options *nir_options)
3683 /* Initialize the stn_builder object */
3684 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3689 exec_list_make_empty(&b->functions);
3690 b->entry_point_stage = stage;
3691 b->entry_point_name = entry_point_name;
3692 b->options = options;
3694 /* See also _vtn_fail() */
3695 if (setjmp(b->fail_jump)) {
3700 const uint32_t *word_end = words + word_count;
3702 /* Handle the SPIR-V header (first 4 dwords) */
3703 vtn_assert(word_count > 5);
3705 vtn_assert(words[0] == SpvMagicNumber);
3706 vtn_assert(words[1] >= 0x10000);
3707 /* words[2] == generator magic */
3708 unsigned value_id_bound = words[3];
3709 vtn_assert(words[4] == 0);
3713 b->value_id_bound = value_id_bound;
3714 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3716 /* Handle all the preamble instructions */
3717 words = vtn_foreach_instruction(b, words, word_end,
3718 vtn_handle_preamble_instruction);
3720 if (b->entry_point == NULL) {
3721 vtn_fail("Entry point not found");
3726 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3728 /* Set shader info defaults */
3729 b->shader->info.gs.invocations = 1;
3731 /* Parse execution modes */
3732 vtn_foreach_execution_mode(b, b->entry_point,
3733 vtn_handle_execution_mode, NULL);
3735 b->specializations = spec;
3736 b->num_specializations = num_spec;
3738 /* Handle all variable, type, and constant instructions */
3739 words = vtn_foreach_instruction(b, words, word_end,
3740 vtn_handle_variable_or_type_instruction);
3742 /* Set types on all vtn_values */
3743 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
3745 vtn_build_cfg(b, words, word_end);
3747 assert(b->entry_point->value_type == vtn_value_type_function);
3748 b->entry_point->func->referenced = true;
3753 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3754 if (func->referenced && !func->emitted) {
3755 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3756 _mesa_key_pointer_equal);
3758 vtn_function_emit(b, func, vtn_handle_body_instruction);
3764 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
3765 nir_function *entry_point = b->entry_point->func->impl->function;
3766 vtn_assert(entry_point);
3768 /* Unparent the shader from the vtn_builder before we delete the builder */
3769 ralloc_steal(NULL, b->shader);