2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
37 vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
38 size_t spirv_offset, const char *message)
40 if (b->options->debug.func) {
41 b->options->debug.func(b->options->debug.private_data,
42 level, spirv_offset, message);
46 if (level >= NIR_SPIRV_DEBUG_LEVEL_WARNING)
47 fprintf(stderr, "%s\n", message);
52 vtn_logf(struct vtn_builder *b, enum nir_spirv_debug_level level,
53 size_t spirv_offset, const char *fmt, ...)
59 msg = ralloc_vasprintf(NULL, fmt, args);
62 vtn_log(b, level, spirv_offset, msg);
68 vtn_log_err(struct vtn_builder *b,
69 enum nir_spirv_debug_level level, const char *prefix,
70 const char *file, unsigned line,
71 const char *fmt, va_list args)
75 msg = ralloc_strdup(NULL, prefix);
78 ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
81 ralloc_asprintf_append(&msg, " ");
83 ralloc_vasprintf_append(&msg, fmt, args);
85 ralloc_asprintf_append(&msg, "\n %zu bytes into the SPIR-V binary",
89 ralloc_asprintf_append(&msg,
90 "\n in SPIR-V source file %s, line %d, col %d",
91 b->file, b->line, b->col);
94 vtn_log(b, level, b->spirv_offset, msg);
100 vtn_dump_shader(struct vtn_builder *b, const char *path, const char *prefix)
105 int len = snprintf(filename, sizeof(filename), "%s/%s-%d.spirv",
106 path, prefix, idx++);
107 if (len < 0 || len >= sizeof(filename))
110 FILE *f = fopen(filename, "w");
114 fwrite(b->spirv, sizeof(*b->spirv), b->spirv_word_count, f);
117 vtn_info("SPIR-V shader dumped to %s", filename);
121 _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
122 const char *fmt, ...)
127 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_WARNING, "SPIR-V WARNING:\n",
128 file, line, fmt, args);
133 _vtn_err(struct vtn_builder *b, const char *file, unsigned line,
134 const char *fmt, ...)
139 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V ERROR:\n",
140 file, line, fmt, args);
145 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
146 const char *fmt, ...)
151 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V parsing FAILED:\n",
152 file, line, fmt, args);
155 const char *dump_path = getenv("MESA_SPIRV_FAIL_DUMP_PATH");
157 vtn_dump_shader(b, dump_path, "fail");
159 longjmp(b->fail_jump, 1);
162 struct spec_constant_value {
170 static struct vtn_ssa_value *
171 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
173 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
176 if (glsl_type_is_vector_or_scalar(type)) {
177 unsigned num_components = glsl_get_vector_elements(val->type);
178 unsigned bit_size = glsl_get_bit_size(val->type);
179 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
181 unsigned elems = glsl_get_length(val->type);
182 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
183 if (glsl_type_is_matrix(type)) {
184 const struct glsl_type *elem_type =
185 glsl_vector_type(glsl_get_base_type(type),
186 glsl_get_vector_elements(type));
188 for (unsigned i = 0; i < elems; i++)
189 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
190 } else if (glsl_type_is_array(type)) {
191 const struct glsl_type *elem_type = glsl_get_array_element(type);
192 for (unsigned i = 0; i < elems; i++)
193 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
195 for (unsigned i = 0; i < elems; i++) {
196 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
197 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
205 static struct vtn_ssa_value *
206 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
207 const struct glsl_type *type)
209 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
214 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
217 switch (glsl_get_base_type(type)) {
220 case GLSL_TYPE_INT16:
221 case GLSL_TYPE_UINT16:
222 case GLSL_TYPE_UINT8:
224 case GLSL_TYPE_INT64:
225 case GLSL_TYPE_UINT64:
227 case GLSL_TYPE_FLOAT:
228 case GLSL_TYPE_FLOAT16:
229 case GLSL_TYPE_DOUBLE: {
230 int bit_size = glsl_get_bit_size(type);
231 if (glsl_type_is_vector_or_scalar(type)) {
232 unsigned num_components = glsl_get_vector_elements(val->type);
233 nir_load_const_instr *load =
234 nir_load_const_instr_create(b->shader, num_components, bit_size);
236 load->value = constant->values[0];
238 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
239 val->def = &load->def;
241 assert(glsl_type_is_matrix(type));
242 unsigned rows = glsl_get_vector_elements(val->type);
243 unsigned columns = glsl_get_matrix_columns(val->type);
244 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
246 for (unsigned i = 0; i < columns; i++) {
247 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
248 col_val->type = glsl_get_column_type(val->type);
249 nir_load_const_instr *load =
250 nir_load_const_instr_create(b->shader, rows, bit_size);
252 load->value = constant->values[i];
254 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
255 col_val->def = &load->def;
257 val->elems[i] = col_val;
263 case GLSL_TYPE_ARRAY: {
264 unsigned elems = glsl_get_length(val->type);
265 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
266 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
267 for (unsigned i = 0; i < elems; i++)
268 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
273 case GLSL_TYPE_STRUCT: {
274 unsigned elems = glsl_get_length(val->type);
275 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
276 for (unsigned i = 0; i < elems; i++) {
277 const struct glsl_type *elem_type =
278 glsl_get_struct_field(val->type, i);
279 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
286 vtn_fail("bad constant type");
292 struct vtn_ssa_value *
293 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
295 struct vtn_value *val = vtn_untyped_value(b, value_id);
296 switch (val->value_type) {
297 case vtn_value_type_undef:
298 return vtn_undef_ssa_value(b, val->type->type);
300 case vtn_value_type_constant:
301 return vtn_const_ssa_value(b, val->constant, val->type->type);
303 case vtn_value_type_ssa:
306 case vtn_value_type_pointer:
307 vtn_assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
308 struct vtn_ssa_value *ssa =
309 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
310 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
314 vtn_fail("Invalid type for an SSA value");
319 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
320 unsigned word_count, unsigned *words_used)
322 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
324 /* Ammount of space taken by the string (including the null) */
325 unsigned len = strlen(dup) + 1;
326 *words_used = DIV_ROUND_UP(len, sizeof(*words));
332 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
333 const uint32_t *end, vtn_instruction_handler handler)
339 const uint32_t *w = start;
341 SpvOp opcode = w[0] & SpvOpCodeMask;
342 unsigned count = w[0] >> SpvWordCountShift;
343 vtn_assert(count >= 1 && w + count <= end);
345 b->spirv_offset = (uint8_t *)w - (uint8_t *)b->spirv;
349 break; /* Do nothing */
352 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
364 if (!handler(b, opcode, w, count))
382 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
383 const uint32_t *w, unsigned count)
386 case SpvOpExtInstImport: {
387 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
388 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
389 val->ext_handler = vtn_handle_glsl450_instruction;
390 } else if ((strcmp((const char *)&w[2], "SPV_AMD_gcn_shader") == 0)
391 && (b->options && b->options->caps.gcn_shader)) {
392 val->ext_handler = vtn_handle_amd_gcn_shader_instruction;
393 } else if ((strcmp((const char *)&w[2], "SPV_AMD_shader_trinary_minmax") == 0)
394 && (b->options && b->options->caps.trinary_minmax)) {
395 val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
397 vtn_fail("Unsupported extension");
403 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
404 bool handled = val->ext_handler(b, w[4], w, count);
410 vtn_fail("Unhandled opcode");
415 _foreach_decoration_helper(struct vtn_builder *b,
416 struct vtn_value *base_value,
418 struct vtn_value *value,
419 vtn_decoration_foreach_cb cb, void *data)
421 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
423 if (dec->scope == VTN_DEC_DECORATION) {
424 member = parent_member;
425 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
426 vtn_fail_if(value->value_type != vtn_value_type_type ||
427 value->type->base_type != vtn_base_type_struct,
428 "OpMemberDecorate and OpGroupMemberDecorate are only "
429 "allowed on OpTypeStruct");
430 /* This means we haven't recursed yet */
431 assert(value == base_value);
433 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
435 vtn_fail_if(member >= base_value->type->length,
436 "OpMemberDecorate specifies member %d but the "
437 "OpTypeStruct has only %u members",
438 member, base_value->type->length);
440 /* Not a decoration */
441 assert(dec->scope == VTN_DEC_EXECUTION_MODE);
446 assert(dec->group->value_type == vtn_value_type_decoration_group);
447 _foreach_decoration_helper(b, base_value, member, dec->group,
450 cb(b, base_value, member, dec, data);
455 /** Iterates (recursively if needed) over all of the decorations on a value
457 * This function iterates over all of the decorations applied to a given
458 * value. If it encounters a decoration group, it recurses into the group
459 * and iterates over all of those decorations as well.
462 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
463 vtn_decoration_foreach_cb cb, void *data)
465 _foreach_decoration_helper(b, value, -1, value, cb, data);
469 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
470 vtn_execution_mode_foreach_cb cb, void *data)
472 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
473 if (dec->scope != VTN_DEC_EXECUTION_MODE)
476 assert(dec->group == NULL);
477 cb(b, value, dec, data);
482 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
483 const uint32_t *w, unsigned count)
485 const uint32_t *w_end = w + count;
486 const uint32_t target = w[1];
490 case SpvOpDecorationGroup:
491 vtn_push_value(b, target, vtn_value_type_decoration_group);
495 case SpvOpMemberDecorate:
496 case SpvOpExecutionMode: {
497 struct vtn_value *val = vtn_untyped_value(b, target);
499 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
502 dec->scope = VTN_DEC_DECORATION;
504 case SpvOpMemberDecorate:
505 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
506 vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
507 "Member argument of OpMemberDecorate too large");
509 case SpvOpExecutionMode:
510 dec->scope = VTN_DEC_EXECUTION_MODE;
513 unreachable("Invalid decoration opcode");
515 dec->decoration = *(w++);
518 /* Link into the list */
519 dec->next = val->decoration;
520 val->decoration = dec;
524 case SpvOpGroupMemberDecorate:
525 case SpvOpGroupDecorate: {
526 struct vtn_value *group =
527 vtn_value(b, target, vtn_value_type_decoration_group);
529 for (; w < w_end; w++) {
530 struct vtn_value *val = vtn_untyped_value(b, *w);
531 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
534 if (opcode == SpvOpGroupDecorate) {
535 dec->scope = VTN_DEC_DECORATION;
537 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
538 vtn_fail_if(dec->scope < 0, /* Check for overflow */
539 "Member argument of OpGroupMemberDecorate too large");
542 /* Link into the list */
543 dec->next = val->decoration;
544 val->decoration = dec;
550 unreachable("Unhandled opcode");
554 struct member_decoration_ctx {
556 struct glsl_struct_field *fields;
557 struct vtn_type *type;
560 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
561 * OpStore, or OpCopyMemory between them without breaking anything.
562 * Technically, the SPIR-V rules require the exact same type ID but this lets
563 * us internally be a bit looser.
566 vtn_types_compatible(struct vtn_builder *b,
567 struct vtn_type *t1, struct vtn_type *t2)
569 if (t1->id == t2->id)
572 if (t1->base_type != t2->base_type)
575 switch (t1->base_type) {
576 case vtn_base_type_void:
577 case vtn_base_type_scalar:
578 case vtn_base_type_vector:
579 case vtn_base_type_matrix:
580 case vtn_base_type_image:
581 case vtn_base_type_sampler:
582 case vtn_base_type_sampled_image:
583 return t1->type == t2->type;
585 case vtn_base_type_array:
586 return t1->length == t2->length &&
587 vtn_types_compatible(b, t1->array_element, t2->array_element);
589 case vtn_base_type_pointer:
590 return vtn_types_compatible(b, t1->deref, t2->deref);
592 case vtn_base_type_struct:
593 if (t1->length != t2->length)
596 for (unsigned i = 0; i < t1->length; i++) {
597 if (!vtn_types_compatible(b, t1->members[i], t2->members[i]))
602 case vtn_base_type_function:
603 /* This case shouldn't get hit since you can't copy around function
604 * types. Just require them to be identical.
609 vtn_fail("Invalid base type");
612 /* does a shallow copy of a vtn_type */
614 static struct vtn_type *
615 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
617 struct vtn_type *dest = ralloc(b, struct vtn_type);
620 switch (src->base_type) {
621 case vtn_base_type_void:
622 case vtn_base_type_scalar:
623 case vtn_base_type_vector:
624 case vtn_base_type_matrix:
625 case vtn_base_type_array:
626 case vtn_base_type_pointer:
627 case vtn_base_type_image:
628 case vtn_base_type_sampler:
629 case vtn_base_type_sampled_image:
630 /* Nothing more to do */
633 case vtn_base_type_struct:
634 dest->members = ralloc_array(b, struct vtn_type *, src->length);
635 memcpy(dest->members, src->members,
636 src->length * sizeof(src->members[0]));
638 dest->offsets = ralloc_array(b, unsigned, src->length);
639 memcpy(dest->offsets, src->offsets,
640 src->length * sizeof(src->offsets[0]));
643 case vtn_base_type_function:
644 dest->params = ralloc_array(b, struct vtn_type *, src->length);
645 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
652 static struct vtn_type *
653 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
655 type->members[member] = vtn_type_copy(b, type->members[member]);
656 type = type->members[member];
658 /* We may have an array of matrices.... Oh, joy! */
659 while (glsl_type_is_array(type->type)) {
660 type->array_element = vtn_type_copy(b, type->array_element);
661 type = type->array_element;
664 vtn_assert(glsl_type_is_matrix(type->type));
670 struct_member_decoration_cb(struct vtn_builder *b,
671 struct vtn_value *val, int member,
672 const struct vtn_decoration *dec, void *void_ctx)
674 struct member_decoration_ctx *ctx = void_ctx;
679 assert(member < ctx->num_fields);
681 switch (dec->decoration) {
682 case SpvDecorationNonWritable:
683 case SpvDecorationNonReadable:
684 case SpvDecorationRelaxedPrecision:
685 case SpvDecorationVolatile:
686 case SpvDecorationCoherent:
687 case SpvDecorationUniform:
688 break; /* FIXME: Do nothing with this for now. */
689 case SpvDecorationNoPerspective:
690 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
692 case SpvDecorationFlat:
693 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
695 case SpvDecorationCentroid:
696 ctx->fields[member].centroid = true;
698 case SpvDecorationSample:
699 ctx->fields[member].sample = true;
701 case SpvDecorationStream:
702 /* Vulkan only allows one GS stream */
703 vtn_assert(dec->literals[0] == 0);
705 case SpvDecorationLocation:
706 ctx->fields[member].location = dec->literals[0];
708 case SpvDecorationComponent:
709 break; /* FIXME: What should we do with these? */
710 case SpvDecorationBuiltIn:
711 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
712 ctx->type->members[member]->is_builtin = true;
713 ctx->type->members[member]->builtin = dec->literals[0];
714 ctx->type->builtin_block = true;
716 case SpvDecorationOffset:
717 ctx->type->offsets[member] = dec->literals[0];
719 case SpvDecorationMatrixStride:
720 /* Handled as a second pass */
722 case SpvDecorationColMajor:
723 break; /* Nothing to do here. Column-major is the default. */
724 case SpvDecorationRowMajor:
725 mutable_matrix_member(b, ctx->type, member)->row_major = true;
728 case SpvDecorationPatch:
731 case SpvDecorationSpecId:
732 case SpvDecorationBlock:
733 case SpvDecorationBufferBlock:
734 case SpvDecorationArrayStride:
735 case SpvDecorationGLSLShared:
736 case SpvDecorationGLSLPacked:
737 case SpvDecorationInvariant:
738 case SpvDecorationRestrict:
739 case SpvDecorationAliased:
740 case SpvDecorationConstant:
741 case SpvDecorationIndex:
742 case SpvDecorationBinding:
743 case SpvDecorationDescriptorSet:
744 case SpvDecorationLinkageAttributes:
745 case SpvDecorationNoContraction:
746 case SpvDecorationInputAttachmentIndex:
747 vtn_warn("Decoration not allowed on struct members: %s",
748 spirv_decoration_to_string(dec->decoration));
751 case SpvDecorationXfbBuffer:
752 case SpvDecorationXfbStride:
753 vtn_warn("Vulkan does not have transform feedback");
756 case SpvDecorationCPacked:
757 case SpvDecorationSaturatedConversion:
758 case SpvDecorationFuncParamAttr:
759 case SpvDecorationFPRoundingMode:
760 case SpvDecorationFPFastMathMode:
761 case SpvDecorationAlignment:
762 vtn_warn("Decoration only allowed for CL-style kernels: %s",
763 spirv_decoration_to_string(dec->decoration));
767 vtn_fail("Unhandled decoration");
771 /* Matrix strides are handled as a separate pass because we need to know
772 * whether the matrix is row-major or not first.
775 struct_member_matrix_stride_cb(struct vtn_builder *b,
776 struct vtn_value *val, int member,
777 const struct vtn_decoration *dec,
780 if (dec->decoration != SpvDecorationMatrixStride)
783 vtn_fail_if(member < 0,
784 "The MatrixStride decoration is only allowed on members "
787 struct member_decoration_ctx *ctx = void_ctx;
789 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
790 if (mat_type->row_major) {
791 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
792 mat_type->stride = mat_type->array_element->stride;
793 mat_type->array_element->stride = dec->literals[0];
795 vtn_assert(mat_type->array_element->stride > 0);
796 mat_type->stride = dec->literals[0];
801 type_decoration_cb(struct vtn_builder *b,
802 struct vtn_value *val, int member,
803 const struct vtn_decoration *dec, void *ctx)
805 struct vtn_type *type = val->type;
808 /* This should have been handled by OpTypeStruct */
809 assert(val->type->base_type == vtn_base_type_struct);
810 assert(member >= 0 && member < val->type->length);
814 switch (dec->decoration) {
815 case SpvDecorationArrayStride:
816 vtn_assert(type->base_type == vtn_base_type_matrix ||
817 type->base_type == vtn_base_type_array ||
818 type->base_type == vtn_base_type_pointer);
819 type->stride = dec->literals[0];
821 case SpvDecorationBlock:
822 vtn_assert(type->base_type == vtn_base_type_struct);
825 case SpvDecorationBufferBlock:
826 vtn_assert(type->base_type == vtn_base_type_struct);
827 type->buffer_block = true;
829 case SpvDecorationGLSLShared:
830 case SpvDecorationGLSLPacked:
831 /* Ignore these, since we get explicit offsets anyways */
834 case SpvDecorationRowMajor:
835 case SpvDecorationColMajor:
836 case SpvDecorationMatrixStride:
837 case SpvDecorationBuiltIn:
838 case SpvDecorationNoPerspective:
839 case SpvDecorationFlat:
840 case SpvDecorationPatch:
841 case SpvDecorationCentroid:
842 case SpvDecorationSample:
843 case SpvDecorationVolatile:
844 case SpvDecorationCoherent:
845 case SpvDecorationNonWritable:
846 case SpvDecorationNonReadable:
847 case SpvDecorationUniform:
848 case SpvDecorationStream:
849 case SpvDecorationLocation:
850 case SpvDecorationComponent:
851 case SpvDecorationOffset:
852 case SpvDecorationXfbBuffer:
853 case SpvDecorationXfbStride:
854 vtn_warn("Decoration only allowed for struct members: %s",
855 spirv_decoration_to_string(dec->decoration));
858 case SpvDecorationRelaxedPrecision:
859 case SpvDecorationSpecId:
860 case SpvDecorationInvariant:
861 case SpvDecorationRestrict:
862 case SpvDecorationAliased:
863 case SpvDecorationConstant:
864 case SpvDecorationIndex:
865 case SpvDecorationBinding:
866 case SpvDecorationDescriptorSet:
867 case SpvDecorationLinkageAttributes:
868 case SpvDecorationNoContraction:
869 case SpvDecorationInputAttachmentIndex:
870 vtn_warn("Decoration not allowed on types: %s",
871 spirv_decoration_to_string(dec->decoration));
874 case SpvDecorationCPacked:
875 case SpvDecorationSaturatedConversion:
876 case SpvDecorationFuncParamAttr:
877 case SpvDecorationFPRoundingMode:
878 case SpvDecorationFPFastMathMode:
879 case SpvDecorationAlignment:
880 vtn_warn("Decoration only allowed for CL-style kernels: %s",
881 spirv_decoration_to_string(dec->decoration));
885 vtn_fail("Unhandled decoration");
890 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
893 case SpvImageFormatUnknown: return 0; /* GL_NONE */
894 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
895 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
896 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
897 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
898 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
899 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
900 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
901 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
902 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
903 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
904 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
905 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
906 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
907 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
908 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
909 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
910 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
911 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
912 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
913 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
914 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
915 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
916 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
917 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
918 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
919 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
920 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
921 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
922 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
923 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
924 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
925 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
926 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
927 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
928 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
929 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
930 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
931 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
932 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
934 vtn_fail("Invalid image format");
938 static struct vtn_type *
939 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
940 uint32_t *size_out, uint32_t *align_out)
942 switch (type->base_type) {
943 case vtn_base_type_scalar: {
944 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
945 *size_out = comp_size;
946 *align_out = comp_size;
950 case vtn_base_type_vector: {
951 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
952 unsigned align_comps = type->length == 3 ? 4 : type->length;
953 *size_out = comp_size * type->length,
954 *align_out = comp_size * align_comps;
958 case vtn_base_type_matrix:
959 case vtn_base_type_array: {
960 /* We're going to add an array stride */
961 type = vtn_type_copy(b, type);
962 uint32_t elem_size, elem_align;
963 type->array_element = vtn_type_layout_std430(b, type->array_element,
964 &elem_size, &elem_align);
965 type->stride = vtn_align_u32(elem_size, elem_align);
966 *size_out = type->stride * type->length;
967 *align_out = elem_align;
971 case vtn_base_type_struct: {
972 /* We're going to add member offsets */
973 type = vtn_type_copy(b, type);
976 for (unsigned i = 0; i < type->length; i++) {
977 uint32_t mem_size, mem_align;
978 type->members[i] = vtn_type_layout_std430(b, type->members[i],
979 &mem_size, &mem_align);
980 offset = vtn_align_u32(offset, mem_align);
981 type->offsets[i] = offset;
983 align = MAX2(align, mem_align);
991 unreachable("Invalid SPIR-V type for std430");
996 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
997 const uint32_t *w, unsigned count)
999 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
1001 val->type = rzalloc(b, struct vtn_type);
1002 val->type->id = w[1];
1006 val->type->base_type = vtn_base_type_void;
1007 val->type->type = glsl_void_type();
1010 val->type->base_type = vtn_base_type_scalar;
1011 val->type->type = glsl_bool_type();
1012 val->type->length = 1;
1014 case SpvOpTypeInt: {
1015 int bit_size = w[2];
1016 const bool signedness = w[3];
1017 val->type->base_type = vtn_base_type_scalar;
1020 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
1023 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
1026 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
1029 val->type->type = (signedness ? glsl_int8_t_type() : glsl_uint8_t_type());
1032 vtn_fail("Invalid int bit size");
1034 val->type->length = 1;
1038 case SpvOpTypeFloat: {
1039 int bit_size = w[2];
1040 val->type->base_type = vtn_base_type_scalar;
1043 val->type->type = glsl_float16_t_type();
1046 val->type->type = glsl_float_type();
1049 val->type->type = glsl_double_type();
1052 vtn_fail("Invalid float bit size");
1054 val->type->length = 1;
1058 case SpvOpTypeVector: {
1059 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1060 unsigned elems = w[3];
1062 vtn_fail_if(base->base_type != vtn_base_type_scalar,
1063 "Base type for OpTypeVector must be a scalar");
1064 vtn_fail_if((elems < 2 || elems > 4) && (elems != 8) && (elems != 16),
1065 "Invalid component count for OpTypeVector");
1067 val->type->base_type = vtn_base_type_vector;
1068 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
1069 val->type->length = elems;
1070 val->type->stride = glsl_get_bit_size(base->type) / 8;
1071 val->type->array_element = base;
1075 case SpvOpTypeMatrix: {
1076 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1077 unsigned columns = w[3];
1079 vtn_fail_if(base->base_type != vtn_base_type_vector,
1080 "Base type for OpTypeMatrix must be a vector");
1081 vtn_fail_if(columns < 2 || columns > 4,
1082 "Invalid column count for OpTypeMatrix");
1084 val->type->base_type = vtn_base_type_matrix;
1085 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
1086 glsl_get_vector_elements(base->type),
1088 vtn_fail_if(glsl_type_is_error(val->type->type),
1089 "Unsupported base type for OpTypeMatrix");
1090 assert(!glsl_type_is_error(val->type->type));
1091 val->type->length = columns;
1092 val->type->array_element = base;
1093 val->type->row_major = false;
1094 val->type->stride = 0;
1098 case SpvOpTypeRuntimeArray:
1099 case SpvOpTypeArray: {
1100 struct vtn_type *array_element =
1101 vtn_value(b, w[2], vtn_value_type_type)->type;
1103 if (opcode == SpvOpTypeRuntimeArray) {
1104 /* A length of 0 is used to denote unsized arrays */
1105 val->type->length = 0;
1108 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
1111 val->type->base_type = vtn_base_type_array;
1112 val->type->type = glsl_array_type(array_element->type, val->type->length);
1113 val->type->array_element = array_element;
1114 val->type->stride = 0;
1118 case SpvOpTypeStruct: {
1119 unsigned num_fields = count - 2;
1120 val->type->base_type = vtn_base_type_struct;
1121 val->type->length = num_fields;
1122 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
1123 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1125 NIR_VLA(struct glsl_struct_field, fields, count);
1126 for (unsigned i = 0; i < num_fields; i++) {
1127 val->type->members[i] =
1128 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1129 fields[i] = (struct glsl_struct_field) {
1130 .type = val->type->members[i]->type,
1131 .name = ralloc_asprintf(b, "field%d", i),
1136 struct member_decoration_ctx ctx = {
1137 .num_fields = num_fields,
1142 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1143 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1145 const char *name = val->name ? val->name : "struct";
1147 val->type->type = glsl_struct_type(fields, num_fields, name);
1151 case SpvOpTypeFunction: {
1152 val->type->base_type = vtn_base_type_function;
1153 val->type->type = NULL;
1155 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1157 const unsigned num_params = count - 3;
1158 val->type->length = num_params;
1159 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1160 for (unsigned i = 0; i < count - 3; i++) {
1161 val->type->params[i] =
1162 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1167 case SpvOpTypePointer: {
1168 SpvStorageClass storage_class = w[2];
1169 struct vtn_type *deref_type =
1170 vtn_value(b, w[3], vtn_value_type_type)->type;
1172 val->type->base_type = vtn_base_type_pointer;
1173 val->type->storage_class = storage_class;
1174 val->type->deref = deref_type;
1176 if (storage_class == SpvStorageClassUniform ||
1177 storage_class == SpvStorageClassStorageBuffer) {
1178 /* These can actually be stored to nir_variables and used as SSA
1179 * values so they need a real glsl_type.
1181 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1184 if (storage_class == SpvStorageClassWorkgroup &&
1185 b->options->lower_workgroup_access_to_offsets) {
1186 uint32_t size, align;
1187 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1189 val->type->length = size;
1190 val->type->align = align;
1191 /* These can actually be stored to nir_variables and used as SSA
1192 * values so they need a real glsl_type.
1194 val->type->type = glsl_uint_type();
1199 case SpvOpTypeImage: {
1200 val->type->base_type = vtn_base_type_image;
1202 const struct vtn_type *sampled_type =
1203 vtn_value(b, w[2], vtn_value_type_type)->type;
1205 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1206 glsl_get_bit_size(sampled_type->type) != 32,
1207 "Sampled type of OpTypeImage must be a 32-bit scalar");
1209 enum glsl_sampler_dim dim;
1210 switch ((SpvDim)w[3]) {
1211 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1212 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1213 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1214 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1215 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1216 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1217 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1219 vtn_fail("Invalid SPIR-V image dimensionality");
1222 bool is_shadow = w[4];
1223 bool is_array = w[5];
1224 bool multisampled = w[6];
1225 unsigned sampled = w[7];
1226 SpvImageFormat format = w[8];
1229 val->type->access_qualifier = w[9];
1231 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1234 if (dim == GLSL_SAMPLER_DIM_2D)
1235 dim = GLSL_SAMPLER_DIM_MS;
1236 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1237 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1239 vtn_fail("Unsupported multisampled image type");
1242 val->type->image_format = translate_image_format(b, format);
1244 enum glsl_base_type sampled_base_type =
1245 glsl_get_base_type(sampled_type->type);
1247 val->type->sampled = true;
1248 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1250 } else if (sampled == 2) {
1251 vtn_assert(!is_shadow);
1252 val->type->sampled = false;
1253 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1255 vtn_fail("We need to know if the image will be sampled");
1260 case SpvOpTypeSampledImage:
1261 val->type->base_type = vtn_base_type_sampled_image;
1262 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1263 val->type->type = val->type->image->type;
1266 case SpvOpTypeSampler:
1267 /* The actual sampler type here doesn't really matter. It gets
1268 * thrown away the moment you combine it with an image. What really
1269 * matters is that it's a sampler type as opposed to an integer type
1270 * so the backend knows what to do.
1272 val->type->base_type = vtn_base_type_sampler;
1273 val->type->type = glsl_bare_sampler_type();
1276 case SpvOpTypeOpaque:
1277 case SpvOpTypeEvent:
1278 case SpvOpTypeDeviceEvent:
1279 case SpvOpTypeReserveId:
1280 case SpvOpTypeQueue:
1283 vtn_fail("Unhandled opcode");
1286 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1289 static nir_constant *
1290 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1292 nir_constant *c = rzalloc(b, nir_constant);
1294 /* For pointers and other typeless things, we have to return something but
1295 * it doesn't matter what.
1300 switch (glsl_get_base_type(type)) {
1302 case GLSL_TYPE_UINT:
1303 case GLSL_TYPE_INT16:
1304 case GLSL_TYPE_UINT16:
1305 case GLSL_TYPE_UINT8:
1306 case GLSL_TYPE_INT8:
1307 case GLSL_TYPE_INT64:
1308 case GLSL_TYPE_UINT64:
1309 case GLSL_TYPE_BOOL:
1310 case GLSL_TYPE_FLOAT:
1311 case GLSL_TYPE_FLOAT16:
1312 case GLSL_TYPE_DOUBLE:
1313 /* Nothing to do here. It's already initialized to zero */
1316 case GLSL_TYPE_ARRAY:
1317 vtn_assert(glsl_get_length(type) > 0);
1318 c->num_elements = glsl_get_length(type);
1319 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1321 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1322 for (unsigned i = 1; i < c->num_elements; i++)
1323 c->elements[i] = c->elements[0];
1326 case GLSL_TYPE_STRUCT:
1327 c->num_elements = glsl_get_length(type);
1328 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1330 for (unsigned i = 0; i < c->num_elements; i++) {
1331 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1336 vtn_fail("Invalid type for null constant");
1343 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1344 int member, const struct vtn_decoration *dec,
1347 vtn_assert(member == -1);
1348 if (dec->decoration != SpvDecorationSpecId)
1351 struct spec_constant_value *const_value = data;
1353 for (unsigned i = 0; i < b->num_specializations; i++) {
1354 if (b->specializations[i].id == dec->literals[0]) {
1355 if (const_value->is_double)
1356 const_value->data64 = b->specializations[i].data64;
1358 const_value->data32 = b->specializations[i].data32;
1365 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1366 uint32_t const_value)
1368 struct spec_constant_value data;
1369 data.is_double = false;
1370 data.data32 = const_value;
1371 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1376 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1377 uint64_t const_value)
1379 struct spec_constant_value data;
1380 data.is_double = true;
1381 data.data64 = const_value;
1382 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1387 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1388 struct vtn_value *val,
1390 const struct vtn_decoration *dec,
1393 vtn_assert(member == -1);
1394 if (dec->decoration != SpvDecorationBuiltIn ||
1395 dec->literals[0] != SpvBuiltInWorkgroupSize)
1398 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1400 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1401 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1402 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1406 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1407 const uint32_t *w, unsigned count)
1409 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1410 val->constant = rzalloc(b, nir_constant);
1412 case SpvOpConstantTrue:
1413 case SpvOpConstantFalse:
1414 case SpvOpSpecConstantTrue:
1415 case SpvOpSpecConstantFalse: {
1416 vtn_fail_if(val->type->type != glsl_bool_type(),
1417 "Result type of %s must be OpTypeBool",
1418 spirv_op_to_string(opcode));
1420 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1421 opcode == SpvOpSpecConstantTrue);
1423 if (opcode == SpvOpSpecConstantTrue ||
1424 opcode == SpvOpSpecConstantFalse)
1425 int_val = get_specialization(b, val, int_val);
1427 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1431 case SpvOpConstant: {
1432 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1433 "Result type of %s must be a scalar",
1434 spirv_op_to_string(opcode));
1435 int bit_size = glsl_get_bit_size(val->type->type);
1438 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1441 val->constant->values->u32[0] = w[3];
1444 val->constant->values->u16[0] = w[3];
1447 val->constant->values->u8[0] = w[3];
1450 vtn_fail("Unsupported SpvOpConstant bit size");
1455 case SpvOpSpecConstant: {
1456 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1457 "Result type of %s must be a scalar",
1458 spirv_op_to_string(opcode));
1459 int bit_size = glsl_get_bit_size(val->type->type);
1462 val->constant->values[0].u64[0] =
1463 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1466 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1469 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1472 val->constant->values[0].u8[0] = get_specialization(b, val, w[3]);
1475 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1480 case SpvOpSpecConstantComposite:
1481 case SpvOpConstantComposite: {
1482 unsigned elem_count = count - 3;
1483 vtn_fail_if(elem_count != val->type->length,
1484 "%s has %u constituents, expected %u",
1485 spirv_op_to_string(opcode), elem_count, val->type->length);
1487 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1488 for (unsigned i = 0; i < elem_count; i++)
1489 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1491 switch (val->type->base_type) {
1492 case vtn_base_type_vector: {
1493 assert(glsl_type_is_vector(val->type->type));
1494 int bit_size = glsl_get_bit_size(val->type->type);
1495 for (unsigned i = 0; i < elem_count; i++) {
1498 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1501 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1504 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1507 val->constant->values[0].u8[i] = elems[i]->values[0].u8[0];
1510 vtn_fail("Invalid SpvOpConstantComposite bit size");
1516 case vtn_base_type_matrix:
1517 assert(glsl_type_is_matrix(val->type->type));
1518 for (unsigned i = 0; i < elem_count; i++)
1519 val->constant->values[i] = elems[i]->values[0];
1522 case vtn_base_type_struct:
1523 case vtn_base_type_array:
1524 ralloc_steal(val->constant, elems);
1525 val->constant->num_elements = elem_count;
1526 val->constant->elements = elems;
1530 vtn_fail("Result type of %s must be a composite type",
1531 spirv_op_to_string(opcode));
1536 case SpvOpSpecConstantOp: {
1537 SpvOp opcode = get_specialization(b, val, w[3]);
1539 case SpvOpVectorShuffle: {
1540 struct vtn_value *v0 = &b->values[w[4]];
1541 struct vtn_value *v1 = &b->values[w[5]];
1543 vtn_assert(v0->value_type == vtn_value_type_constant ||
1544 v0->value_type == vtn_value_type_undef);
1545 vtn_assert(v1->value_type == vtn_value_type_constant ||
1546 v1->value_type == vtn_value_type_undef);
1548 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1549 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1551 vtn_assert(len0 + len1 < 16);
1553 unsigned bit_size = glsl_get_bit_size(val->type->type);
1554 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1555 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1557 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1558 (void)bit_size0; (void)bit_size1;
1560 if (bit_size == 64) {
1562 if (v0->value_type == vtn_value_type_constant) {
1563 for (unsigned i = 0; i < len0; i++)
1564 u64[i] = v0->constant->values[0].u64[i];
1566 if (v1->value_type == vtn_value_type_constant) {
1567 for (unsigned i = 0; i < len1; i++)
1568 u64[len0 + i] = v1->constant->values[0].u64[i];
1571 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1572 uint32_t comp = w[i + 6];
1573 /* If component is not used, set the value to a known constant
1574 * to detect if it is wrongly used.
1576 if (comp == (uint32_t)-1)
1577 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1579 val->constant->values[0].u64[j] = u64[comp];
1582 /* This is for both 32-bit and 16-bit values */
1584 if (v0->value_type == vtn_value_type_constant) {
1585 for (unsigned i = 0; i < len0; i++)
1586 u32[i] = v0->constant->values[0].u32[i];
1588 if (v1->value_type == vtn_value_type_constant) {
1589 for (unsigned i = 0; i < len1; i++)
1590 u32[len0 + i] = v1->constant->values[0].u32[i];
1593 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1594 uint32_t comp = w[i + 6];
1595 /* If component is not used, set the value to a known constant
1596 * to detect if it is wrongly used.
1598 if (comp == (uint32_t)-1)
1599 val->constant->values[0].u32[j] = 0xdeadbeef;
1601 val->constant->values[0].u32[j] = u32[comp];
1607 case SpvOpCompositeExtract:
1608 case SpvOpCompositeInsert: {
1609 struct vtn_value *comp;
1610 unsigned deref_start;
1611 struct nir_constant **c;
1612 if (opcode == SpvOpCompositeExtract) {
1613 comp = vtn_value(b, w[4], vtn_value_type_constant);
1615 c = &comp->constant;
1617 comp = vtn_value(b, w[5], vtn_value_type_constant);
1619 val->constant = nir_constant_clone(comp->constant,
1626 const struct vtn_type *type = comp->type;
1627 for (unsigned i = deref_start; i < count; i++) {
1628 vtn_fail_if(w[i] > type->length,
1629 "%uth index of %s is %u but the type has only "
1630 "%u elements", i - deref_start,
1631 spirv_op_to_string(opcode), w[i], type->length);
1633 switch (type->base_type) {
1634 case vtn_base_type_vector:
1636 type = type->array_element;
1639 case vtn_base_type_matrix:
1640 assert(col == 0 && elem == -1);
1643 type = type->array_element;
1646 case vtn_base_type_array:
1647 c = &(*c)->elements[w[i]];
1648 type = type->array_element;
1651 case vtn_base_type_struct:
1652 c = &(*c)->elements[w[i]];
1653 type = type->members[w[i]];
1657 vtn_fail("%s must only index into composite types",
1658 spirv_op_to_string(opcode));
1662 if (opcode == SpvOpCompositeExtract) {
1666 unsigned num_components = type->length;
1667 unsigned bit_size = glsl_get_bit_size(type->type);
1668 for (unsigned i = 0; i < num_components; i++)
1671 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1674 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1677 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1680 val->constant->values[0].u8[i] = (*c)->values[col].u8[elem + i];
1683 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1687 struct vtn_value *insert =
1688 vtn_value(b, w[4], vtn_value_type_constant);
1689 vtn_assert(insert->type == type);
1691 *c = insert->constant;
1693 unsigned num_components = type->length;
1694 unsigned bit_size = glsl_get_bit_size(type->type);
1695 for (unsigned i = 0; i < num_components; i++)
1698 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1701 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1704 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1707 (*c)->values[col].u8[elem + i] = insert->constant->values[0].u8[i];
1710 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1719 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1720 nir_alu_type src_alu_type = dst_alu_type;
1721 unsigned num_components = glsl_get_vector_elements(val->type->type);
1724 vtn_assert(count <= 7);
1729 /* We have a source in a conversion */
1731 nir_get_nir_type_for_glsl_type(
1732 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1733 /* We use the bitsize of the conversion source to evaluate the opcode later */
1734 bit_size = glsl_get_bit_size(
1735 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1738 bit_size = glsl_get_bit_size(val->type->type);
1741 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1742 nir_alu_type_get_type_size(src_alu_type),
1743 nir_alu_type_get_type_size(dst_alu_type));
1744 nir_const_value src[4];
1746 for (unsigned i = 0; i < count - 4; i++) {
1748 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1750 unsigned j = swap ? 1 - i : i;
1751 src[j] = c->values[0];
1754 val->constant->values[0] =
1755 nir_eval_const_opcode(op, num_components, bit_size, src);
1762 case SpvOpConstantNull:
1763 val->constant = vtn_null_constant(b, val->type->type);
1766 case SpvOpConstantSampler:
1767 vtn_fail("OpConstantSampler requires Kernel Capability");
1771 vtn_fail("Unhandled opcode");
1774 /* Now that we have the value, update the workgroup size if needed */
1775 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1779 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1780 const uint32_t *w, unsigned count)
1782 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1783 struct vtn_function *vtn_callee =
1784 vtn_value(b, w[3], vtn_value_type_function)->func;
1785 struct nir_function *callee = vtn_callee->impl->function;
1787 vtn_callee->referenced = true;
1789 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1790 for (unsigned i = 0; i < call->num_params; i++) {
1791 unsigned arg_id = w[4 + i];
1792 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1793 if (arg->value_type == vtn_value_type_pointer &&
1794 arg->pointer->ptr_type->type == NULL) {
1795 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1796 call->params[i] = nir_deref_var_clone(d, call);
1798 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1800 /* Make a temporary to store the argument in */
1802 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1803 call->params[i] = nir_deref_var_create(call, tmp);
1805 vtn_local_store(b, arg_ssa, call->params[i]);
1809 nir_variable *out_tmp = NULL;
1810 vtn_assert(res_type->type == callee->return_type);
1811 if (!glsl_type_is_void(callee->return_type)) {
1812 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1814 call->return_deref = nir_deref_var_create(call, out_tmp);
1817 nir_builder_instr_insert(&b->nb, &call->instr);
1819 if (glsl_type_is_void(callee->return_type)) {
1820 vtn_push_value(b, w[2], vtn_value_type_undef);
1822 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1826 struct vtn_ssa_value *
1827 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1829 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1832 if (!glsl_type_is_vector_or_scalar(type)) {
1833 unsigned elems = glsl_get_length(type);
1834 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1835 for (unsigned i = 0; i < elems; i++) {
1836 const struct glsl_type *child_type;
1838 switch (glsl_get_base_type(type)) {
1840 case GLSL_TYPE_UINT:
1841 case GLSL_TYPE_INT16:
1842 case GLSL_TYPE_UINT16:
1843 case GLSL_TYPE_UINT8:
1844 case GLSL_TYPE_INT8:
1845 case GLSL_TYPE_INT64:
1846 case GLSL_TYPE_UINT64:
1847 case GLSL_TYPE_BOOL:
1848 case GLSL_TYPE_FLOAT:
1849 case GLSL_TYPE_FLOAT16:
1850 case GLSL_TYPE_DOUBLE:
1851 child_type = glsl_get_column_type(type);
1853 case GLSL_TYPE_ARRAY:
1854 child_type = glsl_get_array_element(type);
1856 case GLSL_TYPE_STRUCT:
1857 child_type = glsl_get_struct_field(type, i);
1860 vtn_fail("unkown base type");
1863 val->elems[i] = vtn_create_ssa_value(b, child_type);
1871 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1874 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1875 src.src_type = type;
1880 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1881 const uint32_t *w, unsigned count)
1883 if (opcode == SpvOpSampledImage) {
1884 struct vtn_value *val =
1885 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1886 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1887 val->sampled_image->type =
1888 vtn_value(b, w[1], vtn_value_type_type)->type;
1889 val->sampled_image->image =
1890 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1891 val->sampled_image->sampler =
1892 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1894 } else if (opcode == SpvOpImage) {
1895 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1896 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1897 if (src_val->value_type == vtn_value_type_sampled_image) {
1898 val->pointer = src_val->sampled_image->image;
1900 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1901 val->pointer = src_val->pointer;
1906 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1907 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1909 struct vtn_sampled_image sampled;
1910 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1911 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1912 sampled = *sampled_val->sampled_image;
1914 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1915 sampled.type = sampled_val->pointer->type;
1916 sampled.image = NULL;
1917 sampled.sampler = sampled_val->pointer;
1920 const struct glsl_type *image_type = sampled.type->type;
1921 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1922 const bool is_array = glsl_sampler_type_is_array(image_type);
1924 /* Figure out the base texture operation */
1927 case SpvOpImageSampleImplicitLod:
1928 case SpvOpImageSampleDrefImplicitLod:
1929 case SpvOpImageSampleProjImplicitLod:
1930 case SpvOpImageSampleProjDrefImplicitLod:
1931 texop = nir_texop_tex;
1934 case SpvOpImageSampleExplicitLod:
1935 case SpvOpImageSampleDrefExplicitLod:
1936 case SpvOpImageSampleProjExplicitLod:
1937 case SpvOpImageSampleProjDrefExplicitLod:
1938 texop = nir_texop_txl;
1941 case SpvOpImageFetch:
1942 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1943 texop = nir_texop_txf_ms;
1945 texop = nir_texop_txf;
1949 case SpvOpImageGather:
1950 case SpvOpImageDrefGather:
1951 texop = nir_texop_tg4;
1954 case SpvOpImageQuerySizeLod:
1955 case SpvOpImageQuerySize:
1956 texop = nir_texop_txs;
1959 case SpvOpImageQueryLod:
1960 texop = nir_texop_lod;
1963 case SpvOpImageQueryLevels:
1964 texop = nir_texop_query_levels;
1967 case SpvOpImageQuerySamples:
1968 texop = nir_texop_texture_samples;
1972 vtn_fail("Unhandled opcode");
1975 nir_tex_src srcs[8]; /* 8 should be enough */
1976 nir_tex_src *p = srcs;
1980 struct nir_ssa_def *coord;
1981 unsigned coord_components;
1983 case SpvOpImageSampleImplicitLod:
1984 case SpvOpImageSampleExplicitLod:
1985 case SpvOpImageSampleDrefImplicitLod:
1986 case SpvOpImageSampleDrefExplicitLod:
1987 case SpvOpImageSampleProjImplicitLod:
1988 case SpvOpImageSampleProjExplicitLod:
1989 case SpvOpImageSampleProjDrefImplicitLod:
1990 case SpvOpImageSampleProjDrefExplicitLod:
1991 case SpvOpImageFetch:
1992 case SpvOpImageGather:
1993 case SpvOpImageDrefGather:
1994 case SpvOpImageQueryLod: {
1995 /* All these types have the coordinate as their first real argument */
1996 switch (sampler_dim) {
1997 case GLSL_SAMPLER_DIM_1D:
1998 case GLSL_SAMPLER_DIM_BUF:
1999 coord_components = 1;
2001 case GLSL_SAMPLER_DIM_2D:
2002 case GLSL_SAMPLER_DIM_RECT:
2003 case GLSL_SAMPLER_DIM_MS:
2004 coord_components = 2;
2006 case GLSL_SAMPLER_DIM_3D:
2007 case GLSL_SAMPLER_DIM_CUBE:
2008 coord_components = 3;
2011 vtn_fail("Invalid sampler type");
2014 if (is_array && texop != nir_texop_lod)
2017 coord = vtn_ssa_value(b, w[idx++])->def;
2018 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
2019 (1 << coord_components) - 1));
2020 p->src_type = nir_tex_src_coord;
2027 coord_components = 0;
2032 case SpvOpImageSampleProjImplicitLod:
2033 case SpvOpImageSampleProjExplicitLod:
2034 case SpvOpImageSampleProjDrefImplicitLod:
2035 case SpvOpImageSampleProjDrefExplicitLod:
2036 /* These have the projector as the last coordinate component */
2037 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
2038 p->src_type = nir_tex_src_projector;
2046 bool is_shadow = false;
2047 unsigned gather_component = 0;
2049 case SpvOpImageSampleDrefImplicitLod:
2050 case SpvOpImageSampleDrefExplicitLod:
2051 case SpvOpImageSampleProjDrefImplicitLod:
2052 case SpvOpImageSampleProjDrefExplicitLod:
2053 case SpvOpImageDrefGather:
2054 /* These all have an explicit depth value as their next source */
2056 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
2059 case SpvOpImageGather:
2060 /* This has a component as its next source */
2062 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
2069 /* For OpImageQuerySizeLod, we always have an LOD */
2070 if (opcode == SpvOpImageQuerySizeLod)
2071 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2073 /* Now we need to handle some number of optional arguments */
2074 const struct vtn_ssa_value *gather_offsets = NULL;
2076 uint32_t operands = w[idx++];
2078 if (operands & SpvImageOperandsBiasMask) {
2079 vtn_assert(texop == nir_texop_tex);
2080 texop = nir_texop_txb;
2081 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
2084 if (operands & SpvImageOperandsLodMask) {
2085 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
2086 texop == nir_texop_txs);
2087 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2090 if (operands & SpvImageOperandsGradMask) {
2091 vtn_assert(texop == nir_texop_txl);
2092 texop = nir_texop_txd;
2093 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
2094 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
2097 if (operands & SpvImageOperandsOffsetMask ||
2098 operands & SpvImageOperandsConstOffsetMask)
2099 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
2101 if (operands & SpvImageOperandsConstOffsetsMask) {
2102 nir_tex_src none = {0};
2103 gather_offsets = vtn_ssa_value(b, w[idx++]);
2107 if (operands & SpvImageOperandsSampleMask) {
2108 vtn_assert(texop == nir_texop_txf_ms);
2109 texop = nir_texop_txf_ms;
2110 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
2113 /* We should have now consumed exactly all of the arguments */
2114 vtn_assert(idx == count);
2116 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
2119 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
2121 instr->coord_components = coord_components;
2122 instr->sampler_dim = sampler_dim;
2123 instr->is_array = is_array;
2124 instr->is_shadow = is_shadow;
2125 instr->is_new_style_shadow =
2126 is_shadow && glsl_get_components(ret_type->type) == 1;
2127 instr->component = gather_component;
2129 switch (glsl_get_sampler_result_type(image_type)) {
2130 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
2131 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
2132 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
2133 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
2135 vtn_fail("Invalid base type for sampler result");
2138 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
2139 nir_deref_var *texture;
2140 if (sampled.image) {
2141 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
2147 instr->texture = nir_deref_var_clone(texture, instr);
2149 switch (instr->op) {
2155 /* These operations require a sampler */
2156 instr->sampler = nir_deref_var_clone(sampler, instr);
2159 case nir_texop_txf_ms:
2162 case nir_texop_query_levels:
2163 case nir_texop_texture_samples:
2164 case nir_texop_samples_identical:
2166 instr->sampler = NULL;
2168 case nir_texop_txf_ms_mcs:
2169 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2172 nir_ssa_dest_init(&instr->instr, &instr->dest,
2173 nir_tex_instr_dest_size(instr), 32, NULL);
2175 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2176 nir_tex_instr_dest_size(instr));
2179 nir_instr *instruction;
2180 if (gather_offsets) {
2181 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2182 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2183 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2185 /* Copy the current instruction 4x */
2186 for (uint32_t i = 1; i < 4; i++) {
2187 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2188 instrs[i]->op = instr->op;
2189 instrs[i]->coord_components = instr->coord_components;
2190 instrs[i]->sampler_dim = instr->sampler_dim;
2191 instrs[i]->is_array = instr->is_array;
2192 instrs[i]->is_shadow = instr->is_shadow;
2193 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2194 instrs[i]->component = instr->component;
2195 instrs[i]->dest_type = instr->dest_type;
2196 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
2197 instrs[i]->sampler = NULL;
2199 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2201 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2202 nir_tex_instr_dest_size(instr), 32, NULL);
2205 /* Fill in the last argument with the offset from the passed in offsets
2206 * and insert the instruction into the stream.
2208 for (uint32_t i = 0; i < 4; i++) {
2210 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2211 src.src_type = nir_tex_src_offset;
2212 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2213 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2216 /* Combine the results of the 4 instructions by taking their .w
2219 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2220 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2221 vec4->dest.write_mask = 0xf;
2222 for (uint32_t i = 0; i < 4; i++) {
2223 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2224 vec4->src[i].swizzle[0] = 3;
2226 def = &vec4->dest.dest.ssa;
2227 instruction = &vec4->instr;
2229 def = &instr->dest.ssa;
2230 instruction = &instr->instr;
2233 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2234 val->ssa->def = def;
2236 nir_builder_instr_insert(&b->nb, instruction);
2240 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2241 const uint32_t *w, nir_src *src)
2244 case SpvOpAtomicIIncrement:
2245 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2248 case SpvOpAtomicIDecrement:
2249 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2252 case SpvOpAtomicISub:
2254 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2257 case SpvOpAtomicCompareExchange:
2258 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2259 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2262 case SpvOpAtomicExchange:
2263 case SpvOpAtomicIAdd:
2264 case SpvOpAtomicSMin:
2265 case SpvOpAtomicUMin:
2266 case SpvOpAtomicSMax:
2267 case SpvOpAtomicUMax:
2268 case SpvOpAtomicAnd:
2270 case SpvOpAtomicXor:
2271 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2275 vtn_fail("Invalid SPIR-V atomic");
2279 static nir_ssa_def *
2280 get_image_coord(struct vtn_builder *b, uint32_t value)
2282 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2284 /* The image_load_store intrinsics assume a 4-dim coordinate */
2285 unsigned dim = glsl_get_vector_elements(coord->type);
2286 unsigned swizzle[4];
2287 for (unsigned i = 0; i < 4; i++)
2288 swizzle[i] = MIN2(i, dim - 1);
2290 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2294 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2295 const uint32_t *w, unsigned count)
2297 /* Just get this one out of the way */
2298 if (opcode == SpvOpImageTexelPointer) {
2299 struct vtn_value *val =
2300 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2301 val->image = ralloc(b, struct vtn_image_pointer);
2303 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2304 val->image->coord = get_image_coord(b, w[4]);
2305 val->image->sample = vtn_ssa_value(b, w[5])->def;
2309 struct vtn_image_pointer image;
2312 case SpvOpAtomicExchange:
2313 case SpvOpAtomicCompareExchange:
2314 case SpvOpAtomicCompareExchangeWeak:
2315 case SpvOpAtomicIIncrement:
2316 case SpvOpAtomicIDecrement:
2317 case SpvOpAtomicIAdd:
2318 case SpvOpAtomicISub:
2319 case SpvOpAtomicLoad:
2320 case SpvOpAtomicSMin:
2321 case SpvOpAtomicUMin:
2322 case SpvOpAtomicSMax:
2323 case SpvOpAtomicUMax:
2324 case SpvOpAtomicAnd:
2326 case SpvOpAtomicXor:
2327 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2330 case SpvOpAtomicStore:
2331 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2334 case SpvOpImageQuerySize:
2335 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2337 image.sample = NULL;
2340 case SpvOpImageRead:
2341 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2342 image.coord = get_image_coord(b, w[4]);
2344 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2345 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2346 image.sample = vtn_ssa_value(b, w[6])->def;
2348 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2352 case SpvOpImageWrite:
2353 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2354 image.coord = get_image_coord(b, w[2]);
2358 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2359 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2360 image.sample = vtn_ssa_value(b, w[5])->def;
2362 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2367 vtn_fail("Invalid image opcode");
2370 nir_intrinsic_op op;
2372 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_var_##N; break;
2373 OP(ImageQuerySize, size)
2375 OP(ImageWrite, store)
2376 OP(AtomicLoad, load)
2377 OP(AtomicStore, store)
2378 OP(AtomicExchange, atomic_exchange)
2379 OP(AtomicCompareExchange, atomic_comp_swap)
2380 OP(AtomicIIncrement, atomic_add)
2381 OP(AtomicIDecrement, atomic_add)
2382 OP(AtomicIAdd, atomic_add)
2383 OP(AtomicISub, atomic_add)
2384 OP(AtomicSMin, atomic_min)
2385 OP(AtomicUMin, atomic_min)
2386 OP(AtomicSMax, atomic_max)
2387 OP(AtomicUMax, atomic_max)
2388 OP(AtomicAnd, atomic_and)
2389 OP(AtomicOr, atomic_or)
2390 OP(AtomicXor, atomic_xor)
2393 vtn_fail("Invalid image opcode");
2396 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2398 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2399 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2401 /* ImageQuerySize doesn't take any extra parameters */
2402 if (opcode != SpvOpImageQuerySize) {
2403 /* The image coordinate is always 4 components but we may not have that
2404 * many. Swizzle to compensate.
2407 for (unsigned i = 0; i < 4; i++)
2408 swiz[i] = i < image.coord->num_components ? i : 0;
2409 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2411 intrin->src[1] = nir_src_for_ssa(image.sample);
2415 case SpvOpAtomicLoad:
2416 case SpvOpImageQuerySize:
2417 case SpvOpImageRead:
2419 case SpvOpAtomicStore:
2420 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2422 case SpvOpImageWrite:
2423 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2426 case SpvOpAtomicCompareExchange:
2427 case SpvOpAtomicIIncrement:
2428 case SpvOpAtomicIDecrement:
2429 case SpvOpAtomicExchange:
2430 case SpvOpAtomicIAdd:
2431 case SpvOpAtomicISub:
2432 case SpvOpAtomicSMin:
2433 case SpvOpAtomicUMin:
2434 case SpvOpAtomicSMax:
2435 case SpvOpAtomicUMax:
2436 case SpvOpAtomicAnd:
2438 case SpvOpAtomicXor:
2439 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2443 vtn_fail("Invalid image opcode");
2446 if (opcode != SpvOpImageWrite) {
2447 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2448 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2450 unsigned dest_components = nir_intrinsic_dest_components(intrin);
2451 if (intrin->intrinsic == nir_intrinsic_image_var_size) {
2452 dest_components = intrin->num_components =
2453 glsl_get_vector_elements(type->type);
2456 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2457 dest_components, 32, NULL);
2459 nir_builder_instr_insert(&b->nb, &intrin->instr);
2461 val->ssa = vtn_create_ssa_value(b, type->type);
2462 val->ssa->def = &intrin->dest.ssa;
2464 nir_builder_instr_insert(&b->nb, &intrin->instr);
2468 static nir_intrinsic_op
2469 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2472 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2473 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2474 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2475 OP(AtomicExchange, atomic_exchange)
2476 OP(AtomicCompareExchange, atomic_comp_swap)
2477 OP(AtomicIIncrement, atomic_add)
2478 OP(AtomicIDecrement, atomic_add)
2479 OP(AtomicIAdd, atomic_add)
2480 OP(AtomicISub, atomic_add)
2481 OP(AtomicSMin, atomic_imin)
2482 OP(AtomicUMin, atomic_umin)
2483 OP(AtomicSMax, atomic_imax)
2484 OP(AtomicUMax, atomic_umax)
2485 OP(AtomicAnd, atomic_and)
2486 OP(AtomicOr, atomic_or)
2487 OP(AtomicXor, atomic_xor)
2490 vtn_fail("Invalid SSBO atomic");
2494 static nir_intrinsic_op
2495 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2498 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2499 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2500 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2501 OP(AtomicExchange, atomic_exchange)
2502 OP(AtomicCompareExchange, atomic_comp_swap)
2503 OP(AtomicIIncrement, atomic_add)
2504 OP(AtomicIDecrement, atomic_add)
2505 OP(AtomicIAdd, atomic_add)
2506 OP(AtomicISub, atomic_add)
2507 OP(AtomicSMin, atomic_imin)
2508 OP(AtomicUMin, atomic_umin)
2509 OP(AtomicSMax, atomic_imax)
2510 OP(AtomicUMax, atomic_umax)
2511 OP(AtomicAnd, atomic_and)
2512 OP(AtomicOr, atomic_or)
2513 OP(AtomicXor, atomic_xor)
2516 vtn_fail("Invalid shared atomic");
2520 static nir_intrinsic_op
2521 get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2524 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2525 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2526 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2527 OP(AtomicExchange, atomic_exchange)
2528 OP(AtomicCompareExchange, atomic_comp_swap)
2529 OP(AtomicIIncrement, atomic_add)
2530 OP(AtomicIDecrement, atomic_add)
2531 OP(AtomicIAdd, atomic_add)
2532 OP(AtomicISub, atomic_add)
2533 OP(AtomicSMin, atomic_imin)
2534 OP(AtomicUMin, atomic_umin)
2535 OP(AtomicSMax, atomic_imax)
2536 OP(AtomicUMax, atomic_umax)
2537 OP(AtomicAnd, atomic_and)
2538 OP(AtomicOr, atomic_or)
2539 OP(AtomicXor, atomic_xor)
2542 vtn_fail("Invalid shared atomic");
2547 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2548 const uint32_t *w, unsigned count)
2550 struct vtn_pointer *ptr;
2551 nir_intrinsic_instr *atomic;
2554 case SpvOpAtomicLoad:
2555 case SpvOpAtomicExchange:
2556 case SpvOpAtomicCompareExchange:
2557 case SpvOpAtomicCompareExchangeWeak:
2558 case SpvOpAtomicIIncrement:
2559 case SpvOpAtomicIDecrement:
2560 case SpvOpAtomicIAdd:
2561 case SpvOpAtomicISub:
2562 case SpvOpAtomicSMin:
2563 case SpvOpAtomicUMin:
2564 case SpvOpAtomicSMax:
2565 case SpvOpAtomicUMax:
2566 case SpvOpAtomicAnd:
2568 case SpvOpAtomicXor:
2569 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2572 case SpvOpAtomicStore:
2573 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2577 vtn_fail("Invalid SPIR-V atomic");
2581 SpvScope scope = w[4];
2582 SpvMemorySemanticsMask semantics = w[5];
2585 if (ptr->mode == vtn_variable_mode_workgroup &&
2586 !b->options->lower_workgroup_access_to_offsets) {
2587 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2588 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2589 nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
2590 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2591 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2594 case SpvOpAtomicLoad:
2595 atomic->num_components = glsl_get_vector_elements(deref_type);
2598 case SpvOpAtomicStore:
2599 atomic->num_components = glsl_get_vector_elements(deref_type);
2600 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2601 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2604 case SpvOpAtomicExchange:
2605 case SpvOpAtomicCompareExchange:
2606 case SpvOpAtomicCompareExchangeWeak:
2607 case SpvOpAtomicIIncrement:
2608 case SpvOpAtomicIDecrement:
2609 case SpvOpAtomicIAdd:
2610 case SpvOpAtomicISub:
2611 case SpvOpAtomicSMin:
2612 case SpvOpAtomicUMin:
2613 case SpvOpAtomicSMax:
2614 case SpvOpAtomicUMax:
2615 case SpvOpAtomicAnd:
2617 case SpvOpAtomicXor:
2618 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2622 vtn_fail("Invalid SPIR-V atomic");
2626 nir_ssa_def *offset, *index;
2627 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2629 nir_intrinsic_op op;
2630 if (ptr->mode == vtn_variable_mode_ssbo) {
2631 op = get_ssbo_nir_atomic_op(b, opcode);
2633 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2634 b->options->lower_workgroup_access_to_offsets);
2635 op = get_shared_nir_atomic_op(b, opcode);
2638 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2642 case SpvOpAtomicLoad:
2643 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2644 if (ptr->mode == vtn_variable_mode_ssbo)
2645 atomic->src[src++] = nir_src_for_ssa(index);
2646 atomic->src[src++] = nir_src_for_ssa(offset);
2649 case SpvOpAtomicStore:
2650 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2651 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2652 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2653 if (ptr->mode == vtn_variable_mode_ssbo)
2654 atomic->src[src++] = nir_src_for_ssa(index);
2655 atomic->src[src++] = nir_src_for_ssa(offset);
2658 case SpvOpAtomicExchange:
2659 case SpvOpAtomicCompareExchange:
2660 case SpvOpAtomicCompareExchangeWeak:
2661 case SpvOpAtomicIIncrement:
2662 case SpvOpAtomicIDecrement:
2663 case SpvOpAtomicIAdd:
2664 case SpvOpAtomicISub:
2665 case SpvOpAtomicSMin:
2666 case SpvOpAtomicUMin:
2667 case SpvOpAtomicSMax:
2668 case SpvOpAtomicUMax:
2669 case SpvOpAtomicAnd:
2671 case SpvOpAtomicXor:
2672 if (ptr->mode == vtn_variable_mode_ssbo)
2673 atomic->src[src++] = nir_src_for_ssa(index);
2674 atomic->src[src++] = nir_src_for_ssa(offset);
2675 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2679 vtn_fail("Invalid SPIR-V atomic");
2683 if (opcode != SpvOpAtomicStore) {
2684 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2686 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2687 glsl_get_vector_elements(type->type),
2688 glsl_get_bit_size(type->type), NULL);
2690 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2691 val->ssa = rzalloc(b, struct vtn_ssa_value);
2692 val->ssa->def = &atomic->dest.ssa;
2693 val->ssa->type = type->type;
2696 nir_builder_instr_insert(&b->nb, &atomic->instr);
2699 static nir_alu_instr *
2700 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2703 switch (num_components) {
2704 case 1: op = nir_op_fmov; break;
2705 case 2: op = nir_op_vec2; break;
2706 case 3: op = nir_op_vec3; break;
2707 case 4: op = nir_op_vec4; break;
2708 default: vtn_fail("bad vector size");
2711 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2712 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2714 vec->dest.write_mask = (1 << num_components) - 1;
2719 struct vtn_ssa_value *
2720 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2722 if (src->transposed)
2723 return src->transposed;
2725 struct vtn_ssa_value *dest =
2726 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2728 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2729 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2730 glsl_get_bit_size(src->type));
2731 if (glsl_type_is_vector_or_scalar(src->type)) {
2732 vec->src[0].src = nir_src_for_ssa(src->def);
2733 vec->src[0].swizzle[0] = i;
2735 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2736 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2737 vec->src[j].swizzle[0] = i;
2740 nir_builder_instr_insert(&b->nb, &vec->instr);
2741 dest->elems[i]->def = &vec->dest.dest.ssa;
2744 dest->transposed = src;
2750 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2752 unsigned swiz[4] = { index };
2753 return nir_swizzle(&b->nb, src, swiz, 1, true);
2757 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2760 nir_alu_instr *vec = create_vec(b, src->num_components,
2763 for (unsigned i = 0; i < src->num_components; i++) {
2765 vec->src[i].src = nir_src_for_ssa(insert);
2767 vec->src[i].src = nir_src_for_ssa(src);
2768 vec->src[i].swizzle[0] = i;
2772 nir_builder_instr_insert(&b->nb, &vec->instr);
2774 return &vec->dest.dest.ssa;
2778 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2781 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2782 for (unsigned i = 1; i < src->num_components; i++)
2783 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2784 vtn_vector_extract(b, src, i), dest);
2790 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2791 nir_ssa_def *insert, nir_ssa_def *index)
2793 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2794 for (unsigned i = 1; i < src->num_components; i++)
2795 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2796 vtn_vector_insert(b, src, insert, i), dest);
2801 static nir_ssa_def *
2802 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2803 nir_ssa_def *src0, nir_ssa_def *src1,
2804 const uint32_t *indices)
2806 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2808 for (unsigned i = 0; i < num_components; i++) {
2809 uint32_t index = indices[i];
2810 if (index == 0xffffffff) {
2812 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2813 } else if (index < src0->num_components) {
2814 vec->src[i].src = nir_src_for_ssa(src0);
2815 vec->src[i].swizzle[0] = index;
2817 vec->src[i].src = nir_src_for_ssa(src1);
2818 vec->src[i].swizzle[0] = index - src0->num_components;
2822 nir_builder_instr_insert(&b->nb, &vec->instr);
2824 return &vec->dest.dest.ssa;
2828 * Concatentates a number of vectors/scalars together to produce a vector
2830 static nir_ssa_def *
2831 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2832 unsigned num_srcs, nir_ssa_def **srcs)
2834 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2836 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2838 * "When constructing a vector, there must be at least two Constituent
2841 vtn_assert(num_srcs >= 2);
2843 unsigned dest_idx = 0;
2844 for (unsigned i = 0; i < num_srcs; i++) {
2845 nir_ssa_def *src = srcs[i];
2846 vtn_assert(dest_idx + src->num_components <= num_components);
2847 for (unsigned j = 0; j < src->num_components; j++) {
2848 vec->src[dest_idx].src = nir_src_for_ssa(src);
2849 vec->src[dest_idx].swizzle[0] = j;
2854 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2856 * "When constructing a vector, the total number of components in all
2857 * the operands must equal the number of components in Result Type."
2859 vtn_assert(dest_idx == num_components);
2861 nir_builder_instr_insert(&b->nb, &vec->instr);
2863 return &vec->dest.dest.ssa;
2866 static struct vtn_ssa_value *
2867 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2869 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2870 dest->type = src->type;
2872 if (glsl_type_is_vector_or_scalar(src->type)) {
2873 dest->def = src->def;
2875 unsigned elems = glsl_get_length(src->type);
2877 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2878 for (unsigned i = 0; i < elems; i++)
2879 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2885 static struct vtn_ssa_value *
2886 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2887 struct vtn_ssa_value *insert, const uint32_t *indices,
2888 unsigned num_indices)
2890 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2892 struct vtn_ssa_value *cur = dest;
2894 for (i = 0; i < num_indices - 1; i++) {
2895 cur = cur->elems[indices[i]];
2898 if (glsl_type_is_vector_or_scalar(cur->type)) {
2899 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2900 * the component granularity. In that case, the last index will be
2901 * the index to insert the scalar into the vector.
2904 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2906 cur->elems[indices[i]] = insert;
2912 static struct vtn_ssa_value *
2913 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2914 const uint32_t *indices, unsigned num_indices)
2916 struct vtn_ssa_value *cur = src;
2917 for (unsigned i = 0; i < num_indices; i++) {
2918 if (glsl_type_is_vector_or_scalar(cur->type)) {
2919 vtn_assert(i == num_indices - 1);
2920 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2921 * the component granularity. The last index will be the index of the
2922 * vector to extract.
2925 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2926 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2927 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2930 cur = cur->elems[indices[i]];
2938 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2939 const uint32_t *w, unsigned count)
2941 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2942 const struct glsl_type *type =
2943 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2944 val->ssa = vtn_create_ssa_value(b, type);
2947 case SpvOpVectorExtractDynamic:
2948 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2949 vtn_ssa_value(b, w[4])->def);
2952 case SpvOpVectorInsertDynamic:
2953 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2954 vtn_ssa_value(b, w[4])->def,
2955 vtn_ssa_value(b, w[5])->def);
2958 case SpvOpVectorShuffle:
2959 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2960 vtn_ssa_value(b, w[3])->def,
2961 vtn_ssa_value(b, w[4])->def,
2965 case SpvOpCompositeConstruct: {
2966 unsigned elems = count - 3;
2968 if (glsl_type_is_vector_or_scalar(type)) {
2969 nir_ssa_def *srcs[4];
2970 for (unsigned i = 0; i < elems; i++)
2971 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2973 vtn_vector_construct(b, glsl_get_vector_elements(type),
2976 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2977 for (unsigned i = 0; i < elems; i++)
2978 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2982 case SpvOpCompositeExtract:
2983 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2987 case SpvOpCompositeInsert:
2988 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2989 vtn_ssa_value(b, w[3]),
2993 case SpvOpCopyObject:
2994 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2998 vtn_fail("unknown composite operation");
3003 vtn_emit_barrier(struct vtn_builder *b, nir_intrinsic_op op)
3005 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
3006 nir_builder_instr_insert(&b->nb, &intrin->instr);
3010 vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
3011 SpvMemorySemanticsMask semantics)
3013 static const SpvMemorySemanticsMask all_memory_semantics =
3014 SpvMemorySemanticsUniformMemoryMask |
3015 SpvMemorySemanticsWorkgroupMemoryMask |
3016 SpvMemorySemanticsAtomicCounterMemoryMask |
3017 SpvMemorySemanticsImageMemoryMask;
3019 /* If we're not actually doing a memory barrier, bail */
3020 if (!(semantics & all_memory_semantics))
3023 /* GL and Vulkan don't have these */
3024 vtn_assert(scope != SpvScopeCrossDevice);
3026 if (scope == SpvScopeSubgroup)
3027 return; /* Nothing to do here */
3029 if (scope == SpvScopeWorkgroup) {
3030 vtn_emit_barrier(b, nir_intrinsic_group_memory_barrier);
3034 /* There's only two scopes thing left */
3035 vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice);
3037 if ((semantics & all_memory_semantics) == all_memory_semantics) {
3038 vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
3042 /* Issue a bunch of more specific barriers */
3043 uint32_t bits = semantics;
3045 SpvMemorySemanticsMask semantic = 1 << u_bit_scan(&bits);
3047 case SpvMemorySemanticsUniformMemoryMask:
3048 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer);
3050 case SpvMemorySemanticsWorkgroupMemoryMask:
3051 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared);
3053 case SpvMemorySemanticsAtomicCounterMemoryMask:
3054 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter);
3056 case SpvMemorySemanticsImageMemoryMask:
3057 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image);
3066 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
3067 const uint32_t *w, unsigned count)
3070 case SpvOpEmitVertex:
3071 case SpvOpEmitStreamVertex:
3072 case SpvOpEndPrimitive:
3073 case SpvOpEndStreamPrimitive: {
3074 nir_intrinsic_op intrinsic_op;
3076 case SpvOpEmitVertex:
3077 case SpvOpEmitStreamVertex:
3078 intrinsic_op = nir_intrinsic_emit_vertex;
3080 case SpvOpEndPrimitive:
3081 case SpvOpEndStreamPrimitive:
3082 intrinsic_op = nir_intrinsic_end_primitive;
3085 unreachable("Invalid opcode");
3088 nir_intrinsic_instr *intrin =
3089 nir_intrinsic_instr_create(b->shader, intrinsic_op);
3092 case SpvOpEmitStreamVertex:
3093 case SpvOpEndStreamPrimitive:
3094 nir_intrinsic_set_stream_id(intrin, w[1]);
3100 nir_builder_instr_insert(&b->nb, &intrin->instr);
3104 case SpvOpMemoryBarrier: {
3105 SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
3106 SpvMemorySemanticsMask semantics =
3107 vtn_constant_value(b, w[2])->values[0].u32[0];
3108 vtn_emit_memory_barrier(b, scope, semantics);
3112 case SpvOpControlBarrier: {
3113 SpvScope execution_scope =
3114 vtn_constant_value(b, w[1])->values[0].u32[0];
3115 if (execution_scope == SpvScopeWorkgroup)
3116 vtn_emit_barrier(b, nir_intrinsic_barrier);
3118 SpvScope memory_scope =
3119 vtn_constant_value(b, w[2])->values[0].u32[0];
3120 SpvMemorySemanticsMask memory_semantics =
3121 vtn_constant_value(b, w[3])->values[0].u32[0];
3122 vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
3127 unreachable("unknown barrier instruction");
3132 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
3133 SpvExecutionMode mode)
3136 case SpvExecutionModeInputPoints:
3137 case SpvExecutionModeOutputPoints:
3138 return 0; /* GL_POINTS */
3139 case SpvExecutionModeInputLines:
3140 return 1; /* GL_LINES */
3141 case SpvExecutionModeInputLinesAdjacency:
3142 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3143 case SpvExecutionModeTriangles:
3144 return 4; /* GL_TRIANGLES */
3145 case SpvExecutionModeInputTrianglesAdjacency:
3146 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3147 case SpvExecutionModeQuads:
3148 return 7; /* GL_QUADS */
3149 case SpvExecutionModeIsolines:
3150 return 0x8E7A; /* GL_ISOLINES */
3151 case SpvExecutionModeOutputLineStrip:
3152 return 3; /* GL_LINE_STRIP */
3153 case SpvExecutionModeOutputTriangleStrip:
3154 return 5; /* GL_TRIANGLE_STRIP */
3156 vtn_fail("Invalid primitive type");
3161 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
3162 SpvExecutionMode mode)
3165 case SpvExecutionModeInputPoints:
3167 case SpvExecutionModeInputLines:
3169 case SpvExecutionModeInputLinesAdjacency:
3171 case SpvExecutionModeTriangles:
3173 case SpvExecutionModeInputTrianglesAdjacency:
3176 vtn_fail("Invalid GS input mode");
3180 static gl_shader_stage
3181 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
3184 case SpvExecutionModelVertex:
3185 return MESA_SHADER_VERTEX;
3186 case SpvExecutionModelTessellationControl:
3187 return MESA_SHADER_TESS_CTRL;
3188 case SpvExecutionModelTessellationEvaluation:
3189 return MESA_SHADER_TESS_EVAL;
3190 case SpvExecutionModelGeometry:
3191 return MESA_SHADER_GEOMETRY;
3192 case SpvExecutionModelFragment:
3193 return MESA_SHADER_FRAGMENT;
3194 case SpvExecutionModelGLCompute:
3195 return MESA_SHADER_COMPUTE;
3197 vtn_fail("Unsupported execution model");
3201 #define spv_check_supported(name, cap) do { \
3202 if (!(b->options && b->options->caps.name)) \
3203 vtn_warn("Unsupported SPIR-V capability: %s", \
3204 spirv_capability_to_string(cap)); \
3209 vtn_handle_entry_point(struct vtn_builder *b, const uint32_t *w,
3212 struct vtn_value *entry_point = &b->values[w[2]];
3213 /* Let this be a name label regardless */
3214 unsigned name_words;
3215 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3217 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3218 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3221 vtn_assert(b->entry_point == NULL);
3222 b->entry_point = entry_point;
3226 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
3227 const uint32_t *w, unsigned count)
3234 case SpvSourceLanguageUnknown: lang = "unknown"; break;
3235 case SpvSourceLanguageESSL: lang = "ESSL"; break;
3236 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
3237 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
3238 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
3239 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
3242 uint32_t version = w[2];
3245 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
3247 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
3251 case SpvOpSourceExtension:
3252 case SpvOpSourceContinued:
3253 case SpvOpExtension:
3254 case SpvOpModuleProcessed:
3255 /* Unhandled, but these are for debug so that's ok. */
3258 case SpvOpCapability: {
3259 SpvCapability cap = w[1];
3261 case SpvCapabilityMatrix:
3262 case SpvCapabilityShader:
3263 case SpvCapabilityGeometry:
3264 case SpvCapabilityGeometryPointSize:
3265 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3266 case SpvCapabilitySampledImageArrayDynamicIndexing:
3267 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3268 case SpvCapabilityStorageImageArrayDynamicIndexing:
3269 case SpvCapabilityImageRect:
3270 case SpvCapabilitySampledRect:
3271 case SpvCapabilitySampled1D:
3272 case SpvCapabilityImage1D:
3273 case SpvCapabilitySampledCubeArray:
3274 case SpvCapabilityImageCubeArray:
3275 case SpvCapabilitySampledBuffer:
3276 case SpvCapabilityImageBuffer:
3277 case SpvCapabilityImageQuery:
3278 case SpvCapabilityDerivativeControl:
3279 case SpvCapabilityInterpolationFunction:
3280 case SpvCapabilityMultiViewport:
3281 case SpvCapabilitySampleRateShading:
3282 case SpvCapabilityClipDistance:
3283 case SpvCapabilityCullDistance:
3284 case SpvCapabilityInputAttachment:
3285 case SpvCapabilityImageGatherExtended:
3286 case SpvCapabilityStorageImageExtendedFormats:
3289 case SpvCapabilityGeometryStreams:
3290 case SpvCapabilityLinkage:
3291 case SpvCapabilityVector16:
3292 case SpvCapabilityFloat16Buffer:
3293 case SpvCapabilityFloat16:
3294 case SpvCapabilityInt64Atomics:
3295 case SpvCapabilityAtomicStorage:
3296 case SpvCapabilityStorageImageMultisample:
3297 case SpvCapabilityInt8:
3298 case SpvCapabilitySparseResidency:
3299 case SpvCapabilityMinLod:
3300 case SpvCapabilityTransformFeedback:
3301 vtn_warn("Unsupported SPIR-V capability: %s",
3302 spirv_capability_to_string(cap));
3305 case SpvCapabilityFloat64:
3306 spv_check_supported(float64, cap);
3308 case SpvCapabilityInt64:
3309 spv_check_supported(int64, cap);
3311 case SpvCapabilityInt16:
3312 spv_check_supported(int16, cap);
3315 case SpvCapabilityAddresses:
3316 case SpvCapabilityKernel:
3317 case SpvCapabilityImageBasic:
3318 case SpvCapabilityImageReadWrite:
3319 case SpvCapabilityImageMipmap:
3320 case SpvCapabilityPipes:
3321 case SpvCapabilityGroups:
3322 case SpvCapabilityDeviceEnqueue:
3323 case SpvCapabilityLiteralSampler:
3324 case SpvCapabilityGenericPointer:
3325 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3326 spirv_capability_to_string(cap));
3329 case SpvCapabilityImageMSArray:
3330 spv_check_supported(image_ms_array, cap);
3333 case SpvCapabilityTessellation:
3334 case SpvCapabilityTessellationPointSize:
3335 spv_check_supported(tessellation, cap);
3338 case SpvCapabilityDrawParameters:
3339 spv_check_supported(draw_parameters, cap);
3342 case SpvCapabilityStorageImageReadWithoutFormat:
3343 spv_check_supported(image_read_without_format, cap);
3346 case SpvCapabilityStorageImageWriteWithoutFormat:
3347 spv_check_supported(image_write_without_format, cap);
3350 case SpvCapabilityDeviceGroup:
3351 spv_check_supported(device_group, cap);
3354 case SpvCapabilityMultiView:
3355 spv_check_supported(multiview, cap);
3358 case SpvCapabilityGroupNonUniform:
3359 spv_check_supported(subgroup_basic, cap);
3362 case SpvCapabilityGroupNonUniformVote:
3363 spv_check_supported(subgroup_vote, cap);
3366 case SpvCapabilitySubgroupBallotKHR:
3367 case SpvCapabilityGroupNonUniformBallot:
3368 spv_check_supported(subgroup_ballot, cap);
3371 case SpvCapabilityGroupNonUniformShuffle:
3372 case SpvCapabilityGroupNonUniformShuffleRelative:
3373 spv_check_supported(subgroup_shuffle, cap);
3376 case SpvCapabilityGroupNonUniformQuad:
3377 spv_check_supported(subgroup_quad, cap);
3380 case SpvCapabilityGroupNonUniformArithmetic:
3381 case SpvCapabilityGroupNonUniformClustered:
3382 spv_check_supported(subgroup_arithmetic, cap);
3385 case SpvCapabilityVariablePointersStorageBuffer:
3386 case SpvCapabilityVariablePointers:
3387 spv_check_supported(variable_pointers, cap);
3390 case SpvCapabilityStorageUniformBufferBlock16:
3391 case SpvCapabilityStorageUniform16:
3392 case SpvCapabilityStoragePushConstant16:
3393 case SpvCapabilityStorageInputOutput16:
3394 spv_check_supported(storage_16bit, cap);
3397 case SpvCapabilityShaderViewportIndexLayerEXT:
3398 spv_check_supported(shader_viewport_index_layer, cap);
3401 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT:
3402 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT:
3403 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT:
3404 spv_check_supported(descriptor_array_dynamic_indexing, cap);
3407 case SpvCapabilityRuntimeDescriptorArrayEXT:
3408 spv_check_supported(runtime_descriptor_array, cap);
3411 case SpvCapabilityStencilExportEXT:
3412 spv_check_supported(stencil_export, cap);
3416 vtn_fail("Unhandled capability");
3421 case SpvOpExtInstImport:
3422 vtn_handle_extension(b, opcode, w, count);
3425 case SpvOpMemoryModel:
3426 vtn_assert(w[1] == SpvAddressingModelLogical);
3427 vtn_assert(w[2] == SpvMemoryModelSimple ||
3428 w[2] == SpvMemoryModelGLSL450);
3431 case SpvOpEntryPoint:
3432 vtn_handle_entry_point(b, w, count);
3436 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3437 vtn_string_literal(b, &w[2], count - 2, NULL);
3441 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3444 case SpvOpMemberName:
3448 case SpvOpExecutionMode:
3449 case SpvOpDecorationGroup:
3451 case SpvOpMemberDecorate:
3452 case SpvOpGroupDecorate:
3453 case SpvOpGroupMemberDecorate:
3454 vtn_handle_decoration(b, opcode, w, count);
3458 return false; /* End of preamble */
3465 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3466 const struct vtn_decoration *mode, void *data)
3468 vtn_assert(b->entry_point == entry_point);
3470 switch(mode->exec_mode) {
3471 case SpvExecutionModeOriginUpperLeft:
3472 case SpvExecutionModeOriginLowerLeft:
3473 b->origin_upper_left =
3474 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3477 case SpvExecutionModeEarlyFragmentTests:
3478 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3479 b->shader->info.fs.early_fragment_tests = true;
3482 case SpvExecutionModeInvocations:
3483 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3484 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3487 case SpvExecutionModeDepthReplacing:
3488 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3489 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3491 case SpvExecutionModeDepthGreater:
3492 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3493 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3495 case SpvExecutionModeDepthLess:
3496 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3497 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3499 case SpvExecutionModeDepthUnchanged:
3500 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3501 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3504 case SpvExecutionModeLocalSize:
3505 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3506 b->shader->info.cs.local_size[0] = mode->literals[0];
3507 b->shader->info.cs.local_size[1] = mode->literals[1];
3508 b->shader->info.cs.local_size[2] = mode->literals[2];
3510 case SpvExecutionModeLocalSizeHint:
3511 break; /* Nothing to do with this */
3513 case SpvExecutionModeOutputVertices:
3514 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3515 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3516 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3518 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3519 b->shader->info.gs.vertices_out = mode->literals[0];
3523 case SpvExecutionModeInputPoints:
3524 case SpvExecutionModeInputLines:
3525 case SpvExecutionModeInputLinesAdjacency:
3526 case SpvExecutionModeTriangles:
3527 case SpvExecutionModeInputTrianglesAdjacency:
3528 case SpvExecutionModeQuads:
3529 case SpvExecutionModeIsolines:
3530 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3531 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3532 b->shader->info.tess.primitive_mode =
3533 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3535 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3536 b->shader->info.gs.vertices_in =
3537 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3541 case SpvExecutionModeOutputPoints:
3542 case SpvExecutionModeOutputLineStrip:
3543 case SpvExecutionModeOutputTriangleStrip:
3544 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3545 b->shader->info.gs.output_primitive =
3546 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3549 case SpvExecutionModeSpacingEqual:
3550 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3551 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3552 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3554 case SpvExecutionModeSpacingFractionalEven:
3555 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3556 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3557 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3559 case SpvExecutionModeSpacingFractionalOdd:
3560 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3561 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3562 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3564 case SpvExecutionModeVertexOrderCw:
3565 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3566 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3567 b->shader->info.tess.ccw = false;
3569 case SpvExecutionModeVertexOrderCcw:
3570 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3571 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3572 b->shader->info.tess.ccw = true;
3574 case SpvExecutionModePointMode:
3575 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3576 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3577 b->shader->info.tess.point_mode = true;
3580 case SpvExecutionModePixelCenterInteger:
3581 b->pixel_center_integer = true;
3584 case SpvExecutionModeXfb:
3585 vtn_fail("Unhandled execution mode");
3588 case SpvExecutionModeVecTypeHint:
3589 case SpvExecutionModeContractionOff:
3592 case SpvExecutionModeStencilRefReplacingEXT:
3593 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3597 vtn_fail("Unhandled execution mode");
3602 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3603 const uint32_t *w, unsigned count)
3605 vtn_set_instruction_result_type(b, opcode, w, count);
3609 case SpvOpSourceContinued:
3610 case SpvOpSourceExtension:
3611 case SpvOpExtension:
3612 case SpvOpCapability:
3613 case SpvOpExtInstImport:
3614 case SpvOpMemoryModel:
3615 case SpvOpEntryPoint:
3616 case SpvOpExecutionMode:
3619 case SpvOpMemberName:
3620 case SpvOpDecorationGroup:
3622 case SpvOpMemberDecorate:
3623 case SpvOpGroupDecorate:
3624 case SpvOpGroupMemberDecorate:
3625 vtn_fail("Invalid opcode types and variables section");
3631 case SpvOpTypeFloat:
3632 case SpvOpTypeVector:
3633 case SpvOpTypeMatrix:
3634 case SpvOpTypeImage:
3635 case SpvOpTypeSampler:
3636 case SpvOpTypeSampledImage:
3637 case SpvOpTypeArray:
3638 case SpvOpTypeRuntimeArray:
3639 case SpvOpTypeStruct:
3640 case SpvOpTypeOpaque:
3641 case SpvOpTypePointer:
3642 case SpvOpTypeFunction:
3643 case SpvOpTypeEvent:
3644 case SpvOpTypeDeviceEvent:
3645 case SpvOpTypeReserveId:
3646 case SpvOpTypeQueue:
3648 vtn_handle_type(b, opcode, w, count);
3651 case SpvOpConstantTrue:
3652 case SpvOpConstantFalse:
3654 case SpvOpConstantComposite:
3655 case SpvOpConstantSampler:
3656 case SpvOpConstantNull:
3657 case SpvOpSpecConstantTrue:
3658 case SpvOpSpecConstantFalse:
3659 case SpvOpSpecConstant:
3660 case SpvOpSpecConstantComposite:
3661 case SpvOpSpecConstantOp:
3662 vtn_handle_constant(b, opcode, w, count);
3667 vtn_handle_variables(b, opcode, w, count);
3671 return false; /* End of preamble */
3678 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3679 const uint32_t *w, unsigned count)
3685 case SpvOpLoopMerge:
3686 case SpvOpSelectionMerge:
3687 /* This is handled by cfg pre-pass and walk_blocks */
3691 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3692 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3697 vtn_handle_extension(b, opcode, w, count);
3703 case SpvOpCopyMemory:
3704 case SpvOpCopyMemorySized:
3705 case SpvOpAccessChain:
3706 case SpvOpPtrAccessChain:
3707 case SpvOpInBoundsAccessChain:
3708 case SpvOpArrayLength:
3709 vtn_handle_variables(b, opcode, w, count);
3712 case SpvOpFunctionCall:
3713 vtn_handle_function_call(b, opcode, w, count);
3716 case SpvOpSampledImage:
3718 case SpvOpImageSampleImplicitLod:
3719 case SpvOpImageSampleExplicitLod:
3720 case SpvOpImageSampleDrefImplicitLod:
3721 case SpvOpImageSampleDrefExplicitLod:
3722 case SpvOpImageSampleProjImplicitLod:
3723 case SpvOpImageSampleProjExplicitLod:
3724 case SpvOpImageSampleProjDrefImplicitLod:
3725 case SpvOpImageSampleProjDrefExplicitLod:
3726 case SpvOpImageFetch:
3727 case SpvOpImageGather:
3728 case SpvOpImageDrefGather:
3729 case SpvOpImageQuerySizeLod:
3730 case SpvOpImageQueryLod:
3731 case SpvOpImageQueryLevels:
3732 case SpvOpImageQuerySamples:
3733 vtn_handle_texture(b, opcode, w, count);
3736 case SpvOpImageRead:
3737 case SpvOpImageWrite:
3738 case SpvOpImageTexelPointer:
3739 vtn_handle_image(b, opcode, w, count);
3742 case SpvOpImageQuerySize: {
3743 struct vtn_pointer *image =
3744 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3745 if (image->mode == vtn_variable_mode_image) {
3746 vtn_handle_image(b, opcode, w, count);
3748 vtn_assert(image->mode == vtn_variable_mode_sampler);
3749 vtn_handle_texture(b, opcode, w, count);
3754 case SpvOpAtomicLoad:
3755 case SpvOpAtomicExchange:
3756 case SpvOpAtomicCompareExchange:
3757 case SpvOpAtomicCompareExchangeWeak:
3758 case SpvOpAtomicIIncrement:
3759 case SpvOpAtomicIDecrement:
3760 case SpvOpAtomicIAdd:
3761 case SpvOpAtomicISub:
3762 case SpvOpAtomicSMin:
3763 case SpvOpAtomicUMin:
3764 case SpvOpAtomicSMax:
3765 case SpvOpAtomicUMax:
3766 case SpvOpAtomicAnd:
3768 case SpvOpAtomicXor: {
3769 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3770 if (pointer->value_type == vtn_value_type_image_pointer) {
3771 vtn_handle_image(b, opcode, w, count);
3773 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3774 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3779 case SpvOpAtomicStore: {
3780 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3781 if (pointer->value_type == vtn_value_type_image_pointer) {
3782 vtn_handle_image(b, opcode, w, count);
3784 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3785 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3791 /* Handle OpSelect up-front here because it needs to be able to handle
3792 * pointers and not just regular vectors and scalars.
3794 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3795 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3796 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3797 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3799 const struct glsl_type *sel_type;
3800 switch (res_val->type->base_type) {
3801 case vtn_base_type_scalar:
3802 sel_type = glsl_bool_type();
3804 case vtn_base_type_vector:
3805 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3807 case vtn_base_type_pointer:
3808 /* We need to have actual storage for pointer types */
3809 vtn_fail_if(res_val->type->type == NULL,
3810 "Invalid pointer result type for OpSelect");
3811 sel_type = glsl_bool_type();
3814 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3817 if (unlikely(sel_val->type->type != sel_type)) {
3818 if (sel_val->type->type == glsl_bool_type()) {
3819 /* This case is illegal but some older versions of GLSLang produce
3820 * it. The GLSLang issue was fixed on March 30, 2017:
3822 * https://github.com/KhronosGroup/glslang/issues/809
3824 * Unfortunately, there are applications in the wild which are
3825 * shipping with this bug so it isn't nice to fail on them so we
3826 * throw a warning instead. It's not actually a problem for us as
3827 * nir_builder will just splat the condition out which is most
3828 * likely what the client wanted anyway.
3830 vtn_warn("Condition type of OpSelect must have the same number "
3831 "of components as Result Type");
3833 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3834 "of Boolean type. It must have the same number of "
3835 "components as Result Type");
3839 vtn_fail_if(obj1_val->type != res_val->type ||
3840 obj2_val->type != res_val->type,
3841 "Object types must match the result type in OpSelect");
3843 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3844 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3845 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3846 vtn_ssa_value(b, w[4])->def,
3847 vtn_ssa_value(b, w[5])->def);
3848 vtn_push_ssa(b, w[2], res_type, ssa);
3857 case SpvOpConvertFToU:
3858 case SpvOpConvertFToS:
3859 case SpvOpConvertSToF:
3860 case SpvOpConvertUToF:
3864 case SpvOpQuantizeToF16:
3865 case SpvOpConvertPtrToU:
3866 case SpvOpConvertUToPtr:
3867 case SpvOpPtrCastToGeneric:
3868 case SpvOpGenericCastToPtr:
3874 case SpvOpSignBitSet:
3875 case SpvOpLessOrGreater:
3877 case SpvOpUnordered:
3892 case SpvOpVectorTimesScalar:
3894 case SpvOpIAddCarry:
3895 case SpvOpISubBorrow:
3896 case SpvOpUMulExtended:
3897 case SpvOpSMulExtended:
3898 case SpvOpShiftRightLogical:
3899 case SpvOpShiftRightArithmetic:
3900 case SpvOpShiftLeftLogical:
3901 case SpvOpLogicalEqual:
3902 case SpvOpLogicalNotEqual:
3903 case SpvOpLogicalOr:
3904 case SpvOpLogicalAnd:
3905 case SpvOpLogicalNot:
3906 case SpvOpBitwiseOr:
3907 case SpvOpBitwiseXor:
3908 case SpvOpBitwiseAnd:
3910 case SpvOpFOrdEqual:
3911 case SpvOpFUnordEqual:
3912 case SpvOpINotEqual:
3913 case SpvOpFOrdNotEqual:
3914 case SpvOpFUnordNotEqual:
3915 case SpvOpULessThan:
3916 case SpvOpSLessThan:
3917 case SpvOpFOrdLessThan:
3918 case SpvOpFUnordLessThan:
3919 case SpvOpUGreaterThan:
3920 case SpvOpSGreaterThan:
3921 case SpvOpFOrdGreaterThan:
3922 case SpvOpFUnordGreaterThan:
3923 case SpvOpULessThanEqual:
3924 case SpvOpSLessThanEqual:
3925 case SpvOpFOrdLessThanEqual:
3926 case SpvOpFUnordLessThanEqual:
3927 case SpvOpUGreaterThanEqual:
3928 case SpvOpSGreaterThanEqual:
3929 case SpvOpFOrdGreaterThanEqual:
3930 case SpvOpFUnordGreaterThanEqual:
3936 case SpvOpFwidthFine:
3937 case SpvOpDPdxCoarse:
3938 case SpvOpDPdyCoarse:
3939 case SpvOpFwidthCoarse:
3940 case SpvOpBitFieldInsert:
3941 case SpvOpBitFieldSExtract:
3942 case SpvOpBitFieldUExtract:
3943 case SpvOpBitReverse:
3945 case SpvOpTranspose:
3946 case SpvOpOuterProduct:
3947 case SpvOpMatrixTimesScalar:
3948 case SpvOpVectorTimesMatrix:
3949 case SpvOpMatrixTimesVector:
3950 case SpvOpMatrixTimesMatrix:
3951 vtn_handle_alu(b, opcode, w, count);
3954 case SpvOpVectorExtractDynamic:
3955 case SpvOpVectorInsertDynamic:
3956 case SpvOpVectorShuffle:
3957 case SpvOpCompositeConstruct:
3958 case SpvOpCompositeExtract:
3959 case SpvOpCompositeInsert:
3960 case SpvOpCopyObject:
3961 vtn_handle_composite(b, opcode, w, count);
3964 case SpvOpEmitVertex:
3965 case SpvOpEndPrimitive:
3966 case SpvOpEmitStreamVertex:
3967 case SpvOpEndStreamPrimitive:
3968 case SpvOpControlBarrier:
3969 case SpvOpMemoryBarrier:
3970 vtn_handle_barrier(b, opcode, w, count);
3973 case SpvOpGroupNonUniformElect:
3974 case SpvOpGroupNonUniformAll:
3975 case SpvOpGroupNonUniformAny:
3976 case SpvOpGroupNonUniformAllEqual:
3977 case SpvOpGroupNonUniformBroadcast:
3978 case SpvOpGroupNonUniformBroadcastFirst:
3979 case SpvOpGroupNonUniformBallot:
3980 case SpvOpGroupNonUniformInverseBallot:
3981 case SpvOpGroupNonUniformBallotBitExtract:
3982 case SpvOpGroupNonUniformBallotBitCount:
3983 case SpvOpGroupNonUniformBallotFindLSB:
3984 case SpvOpGroupNonUniformBallotFindMSB:
3985 case SpvOpGroupNonUniformShuffle:
3986 case SpvOpGroupNonUniformShuffleXor:
3987 case SpvOpGroupNonUniformShuffleUp:
3988 case SpvOpGroupNonUniformShuffleDown:
3989 case SpvOpGroupNonUniformIAdd:
3990 case SpvOpGroupNonUniformFAdd:
3991 case SpvOpGroupNonUniformIMul:
3992 case SpvOpGroupNonUniformFMul:
3993 case SpvOpGroupNonUniformSMin:
3994 case SpvOpGroupNonUniformUMin:
3995 case SpvOpGroupNonUniformFMin:
3996 case SpvOpGroupNonUniformSMax:
3997 case SpvOpGroupNonUniformUMax:
3998 case SpvOpGroupNonUniformFMax:
3999 case SpvOpGroupNonUniformBitwiseAnd:
4000 case SpvOpGroupNonUniformBitwiseOr:
4001 case SpvOpGroupNonUniformBitwiseXor:
4002 case SpvOpGroupNonUniformLogicalAnd:
4003 case SpvOpGroupNonUniformLogicalOr:
4004 case SpvOpGroupNonUniformLogicalXor:
4005 case SpvOpGroupNonUniformQuadBroadcast:
4006 case SpvOpGroupNonUniformQuadSwap:
4007 vtn_handle_subgroup(b, opcode, w, count);
4011 vtn_fail("Unhandled opcode");
4018 vtn_create_builder(const uint32_t *words, size_t word_count,
4019 gl_shader_stage stage, const char *entry_point_name,
4020 const struct spirv_to_nir_options *options)
4022 /* Initialize the vtn_builder object */
4023 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
4025 b->spirv_word_count = word_count;
4029 exec_list_make_empty(&b->functions);
4030 b->entry_point_stage = stage;
4031 b->entry_point_name = entry_point_name;
4032 b->options = options;
4035 * Handle the SPIR-V header (first 5 dwords).
4036 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4038 if (word_count <= 5)
4041 if (words[0] != SpvMagicNumber) {
4042 vtn_err("words[0] was 0x%x, want 0x%x", words[0], SpvMagicNumber);
4045 if (words[1] < 0x10000) {
4046 vtn_err("words[1] was 0x%x, want >= 0x10000", words[1]);
4050 /* words[2] == generator magic */
4051 unsigned value_id_bound = words[3];
4052 if (words[4] != 0) {
4053 vtn_err("words[4] was %u, want 0", words[4]);
4057 b->value_id_bound = value_id_bound;
4058 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
4067 spirv_to_nir(const uint32_t *words, size_t word_count,
4068 struct nir_spirv_specialization *spec, unsigned num_spec,
4069 gl_shader_stage stage, const char *entry_point_name,
4070 const struct spirv_to_nir_options *options,
4071 const nir_shader_compiler_options *nir_options)
4074 const uint32_t *word_end = words + word_count;
4076 struct vtn_builder *b = vtn_create_builder(words, word_count,
4077 stage, entry_point_name,
4083 /* See also _vtn_fail() */
4084 if (setjmp(b->fail_jump)) {
4089 /* Skip the SPIR-V header, handled at vtn_create_builder */
4092 /* Handle all the preamble instructions */
4093 words = vtn_foreach_instruction(b, words, word_end,
4094 vtn_handle_preamble_instruction);
4096 if (b->entry_point == NULL) {
4097 vtn_fail("Entry point not found");
4102 b->shader = nir_shader_create(b, stage, nir_options, NULL);
4104 /* Set shader info defaults */
4105 b->shader->info.gs.invocations = 1;
4107 /* Parse execution modes */
4108 vtn_foreach_execution_mode(b, b->entry_point,
4109 vtn_handle_execution_mode, NULL);
4111 b->specializations = spec;
4112 b->num_specializations = num_spec;
4114 /* Handle all variable, type, and constant instructions */
4115 words = vtn_foreach_instruction(b, words, word_end,
4116 vtn_handle_variable_or_type_instruction);
4118 /* Set types on all vtn_values */
4119 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
4121 vtn_build_cfg(b, words, word_end);
4123 assert(b->entry_point->value_type == vtn_value_type_function);
4124 b->entry_point->func->referenced = true;
4129 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
4130 if (func->referenced && !func->emitted) {
4131 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
4132 _mesa_key_pointer_equal);
4134 vtn_function_emit(b, func, vtn_handle_body_instruction);
4140 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
4141 nir_function *entry_point = b->entry_point->func->impl->function;
4142 vtn_assert(entry_point);
4144 /* Unparent the shader from the vtn_builder before we delete the builder */
4145 ralloc_steal(NULL, b->shader);