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 "nir/nir_deref.h"
33 #include "spirv_info.h"
38 vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
39 size_t spirv_offset, const char *message)
41 if (b->options->debug.func) {
42 b->options->debug.func(b->options->debug.private_data,
43 level, spirv_offset, message);
47 if (level >= NIR_SPIRV_DEBUG_LEVEL_WARNING)
48 fprintf(stderr, "%s\n", message);
53 vtn_logf(struct vtn_builder *b, enum nir_spirv_debug_level level,
54 size_t spirv_offset, const char *fmt, ...)
60 msg = ralloc_vasprintf(NULL, fmt, args);
63 vtn_log(b, level, spirv_offset, msg);
69 vtn_log_err(struct vtn_builder *b,
70 enum nir_spirv_debug_level level, const char *prefix,
71 const char *file, unsigned line,
72 const char *fmt, va_list args)
76 msg = ralloc_strdup(NULL, prefix);
79 ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
82 ralloc_asprintf_append(&msg, " ");
84 ralloc_vasprintf_append(&msg, fmt, args);
86 ralloc_asprintf_append(&msg, "\n %zu bytes into the SPIR-V binary",
90 ralloc_asprintf_append(&msg,
91 "\n in SPIR-V source file %s, line %d, col %d",
92 b->file, b->line, b->col);
95 vtn_log(b, level, b->spirv_offset, msg);
101 vtn_dump_shader(struct vtn_builder *b, const char *path, const char *prefix)
106 int len = snprintf(filename, sizeof(filename), "%s/%s-%d.spirv",
107 path, prefix, idx++);
108 if (len < 0 || len >= sizeof(filename))
111 FILE *f = fopen(filename, "w");
115 fwrite(b->spirv, sizeof(*b->spirv), b->spirv_word_count, f);
118 vtn_info("SPIR-V shader dumped to %s", filename);
122 _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
123 const char *fmt, ...)
128 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_WARNING, "SPIR-V WARNING:\n",
129 file, line, fmt, args);
134 _vtn_err(struct vtn_builder *b, const char *file, unsigned line,
135 const char *fmt, ...)
140 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V ERROR:\n",
141 file, line, fmt, args);
146 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
147 const char *fmt, ...)
152 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V parsing FAILED:\n",
153 file, line, fmt, args);
156 const char *dump_path = getenv("MESA_SPIRV_FAIL_DUMP_PATH");
158 vtn_dump_shader(b, dump_path, "fail");
160 longjmp(b->fail_jump, 1);
163 struct spec_constant_value {
171 static struct vtn_ssa_value *
172 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
174 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
177 if (glsl_type_is_vector_or_scalar(type)) {
178 unsigned num_components = glsl_get_vector_elements(val->type);
179 unsigned bit_size = glsl_get_bit_size(val->type);
180 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
182 unsigned elems = glsl_get_length(val->type);
183 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
184 if (glsl_type_is_matrix(type)) {
185 const struct glsl_type *elem_type =
186 glsl_vector_type(glsl_get_base_type(type),
187 glsl_get_vector_elements(type));
189 for (unsigned i = 0; i < elems; i++)
190 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
191 } else if (glsl_type_is_array(type)) {
192 const struct glsl_type *elem_type = glsl_get_array_element(type);
193 for (unsigned i = 0; i < elems; i++)
194 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
196 for (unsigned i = 0; i < elems; i++) {
197 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
198 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
206 static struct vtn_ssa_value *
207 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
208 const struct glsl_type *type)
210 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
215 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
218 switch (glsl_get_base_type(type)) {
221 case GLSL_TYPE_INT16:
222 case GLSL_TYPE_UINT16:
223 case GLSL_TYPE_UINT8:
225 case GLSL_TYPE_INT64:
226 case GLSL_TYPE_UINT64:
228 case GLSL_TYPE_FLOAT:
229 case GLSL_TYPE_FLOAT16:
230 case GLSL_TYPE_DOUBLE: {
231 int bit_size = glsl_get_bit_size(type);
232 if (glsl_type_is_vector_or_scalar(type)) {
233 unsigned num_components = glsl_get_vector_elements(val->type);
234 nir_load_const_instr *load =
235 nir_load_const_instr_create(b->shader, num_components, bit_size);
237 load->value = constant->values[0];
239 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
240 val->def = &load->def;
242 assert(glsl_type_is_matrix(type));
243 unsigned rows = glsl_get_vector_elements(val->type);
244 unsigned columns = glsl_get_matrix_columns(val->type);
245 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
247 for (unsigned i = 0; i < columns; i++) {
248 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
249 col_val->type = glsl_get_column_type(val->type);
250 nir_load_const_instr *load =
251 nir_load_const_instr_create(b->shader, rows, bit_size);
253 load->value = constant->values[i];
255 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
256 col_val->def = &load->def;
258 val->elems[i] = col_val;
264 case GLSL_TYPE_ARRAY: {
265 unsigned elems = glsl_get_length(val->type);
266 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
267 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
268 for (unsigned i = 0; i < elems; i++)
269 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
274 case GLSL_TYPE_STRUCT: {
275 unsigned elems = glsl_get_length(val->type);
276 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
277 for (unsigned i = 0; i < elems; i++) {
278 const struct glsl_type *elem_type =
279 glsl_get_struct_field(val->type, i);
280 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
287 vtn_fail("bad constant type");
293 struct vtn_ssa_value *
294 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
296 struct vtn_value *val = vtn_untyped_value(b, value_id);
297 switch (val->value_type) {
298 case vtn_value_type_undef:
299 return vtn_undef_ssa_value(b, val->type->type);
301 case vtn_value_type_constant:
302 return vtn_const_ssa_value(b, val->constant, val->type->type);
304 case vtn_value_type_ssa:
307 case vtn_value_type_pointer:
308 vtn_assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
309 struct vtn_ssa_value *ssa =
310 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
311 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
315 vtn_fail("Invalid type for an SSA value");
320 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
321 unsigned word_count, unsigned *words_used)
323 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
325 /* Ammount of space taken by the string (including the null) */
326 unsigned len = strlen(dup) + 1;
327 *words_used = DIV_ROUND_UP(len, sizeof(*words));
333 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
334 const uint32_t *end, vtn_instruction_handler handler)
340 const uint32_t *w = start;
342 SpvOp opcode = w[0] & SpvOpCodeMask;
343 unsigned count = w[0] >> SpvWordCountShift;
344 vtn_assert(count >= 1 && w + count <= end);
346 b->spirv_offset = (uint8_t *)w - (uint8_t *)b->spirv;
350 break; /* Do nothing */
353 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
365 if (!handler(b, opcode, w, count))
383 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
384 const uint32_t *w, unsigned count)
386 const char *ext = (const char *)&w[2];
388 case SpvOpExtInstImport: {
389 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
390 if (strcmp(ext, "GLSL.std.450") == 0) {
391 val->ext_handler = vtn_handle_glsl450_instruction;
392 } else if ((strcmp(ext, "SPV_AMD_gcn_shader") == 0)
393 && (b->options && b->options->caps.gcn_shader)) {
394 val->ext_handler = vtn_handle_amd_gcn_shader_instruction;
395 } else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0)
396 && (b->options && b->options->caps.trinary_minmax)) {
397 val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
399 vtn_fail("Unsupported extension: %s", ext);
405 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
406 bool handled = val->ext_handler(b, w[4], w, count);
412 vtn_fail("Unhandled opcode");
417 _foreach_decoration_helper(struct vtn_builder *b,
418 struct vtn_value *base_value,
420 struct vtn_value *value,
421 vtn_decoration_foreach_cb cb, void *data)
423 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
425 if (dec->scope == VTN_DEC_DECORATION) {
426 member = parent_member;
427 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
428 vtn_fail_if(value->value_type != vtn_value_type_type ||
429 value->type->base_type != vtn_base_type_struct,
430 "OpMemberDecorate and OpGroupMemberDecorate are only "
431 "allowed on OpTypeStruct");
432 /* This means we haven't recursed yet */
433 assert(value == base_value);
435 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
437 vtn_fail_if(member >= base_value->type->length,
438 "OpMemberDecorate specifies member %d but the "
439 "OpTypeStruct has only %u members",
440 member, base_value->type->length);
442 /* Not a decoration */
443 assert(dec->scope == VTN_DEC_EXECUTION_MODE);
448 assert(dec->group->value_type == vtn_value_type_decoration_group);
449 _foreach_decoration_helper(b, base_value, member, dec->group,
452 cb(b, base_value, member, dec, data);
457 /** Iterates (recursively if needed) over all of the decorations on a value
459 * This function iterates over all of the decorations applied to a given
460 * value. If it encounters a decoration group, it recurses into the group
461 * and iterates over all of those decorations as well.
464 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
465 vtn_decoration_foreach_cb cb, void *data)
467 _foreach_decoration_helper(b, value, -1, value, cb, data);
471 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
472 vtn_execution_mode_foreach_cb cb, void *data)
474 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
475 if (dec->scope != VTN_DEC_EXECUTION_MODE)
478 assert(dec->group == NULL);
479 cb(b, value, dec, data);
484 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
485 const uint32_t *w, unsigned count)
487 const uint32_t *w_end = w + count;
488 const uint32_t target = w[1];
492 case SpvOpDecorationGroup:
493 vtn_push_value(b, target, vtn_value_type_decoration_group);
497 case SpvOpMemberDecorate:
498 case SpvOpExecutionMode: {
499 struct vtn_value *val = vtn_untyped_value(b, target);
501 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
504 dec->scope = VTN_DEC_DECORATION;
506 case SpvOpMemberDecorate:
507 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
508 vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
509 "Member argument of OpMemberDecorate too large");
511 case SpvOpExecutionMode:
512 dec->scope = VTN_DEC_EXECUTION_MODE;
515 unreachable("Invalid decoration opcode");
517 dec->decoration = *(w++);
520 /* Link into the list */
521 dec->next = val->decoration;
522 val->decoration = dec;
526 case SpvOpGroupMemberDecorate:
527 case SpvOpGroupDecorate: {
528 struct vtn_value *group =
529 vtn_value(b, target, vtn_value_type_decoration_group);
531 for (; w < w_end; w++) {
532 struct vtn_value *val = vtn_untyped_value(b, *w);
533 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
536 if (opcode == SpvOpGroupDecorate) {
537 dec->scope = VTN_DEC_DECORATION;
539 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
540 vtn_fail_if(dec->scope < 0, /* Check for overflow */
541 "Member argument of OpGroupMemberDecorate too large");
544 /* Link into the list */
545 dec->next = val->decoration;
546 val->decoration = dec;
552 unreachable("Unhandled opcode");
556 struct member_decoration_ctx {
558 struct glsl_struct_field *fields;
559 struct vtn_type *type;
562 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
563 * OpStore, or OpCopyMemory between them without breaking anything.
564 * Technically, the SPIR-V rules require the exact same type ID but this lets
565 * us internally be a bit looser.
568 vtn_types_compatible(struct vtn_builder *b,
569 struct vtn_type *t1, struct vtn_type *t2)
571 if (t1->id == t2->id)
574 if (t1->base_type != t2->base_type)
577 switch (t1->base_type) {
578 case vtn_base_type_void:
579 case vtn_base_type_scalar:
580 case vtn_base_type_vector:
581 case vtn_base_type_matrix:
582 case vtn_base_type_image:
583 case vtn_base_type_sampler:
584 case vtn_base_type_sampled_image:
585 return t1->type == t2->type;
587 case vtn_base_type_array:
588 return t1->length == t2->length &&
589 vtn_types_compatible(b, t1->array_element, t2->array_element);
591 case vtn_base_type_pointer:
592 return vtn_types_compatible(b, t1->deref, t2->deref);
594 case vtn_base_type_struct:
595 if (t1->length != t2->length)
598 for (unsigned i = 0; i < t1->length; i++) {
599 if (!vtn_types_compatible(b, t1->members[i], t2->members[i]))
604 case vtn_base_type_function:
605 /* This case shouldn't get hit since you can't copy around function
606 * types. Just require them to be identical.
611 vtn_fail("Invalid base type");
614 /* does a shallow copy of a vtn_type */
616 static struct vtn_type *
617 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
619 struct vtn_type *dest = ralloc(b, struct vtn_type);
622 switch (src->base_type) {
623 case vtn_base_type_void:
624 case vtn_base_type_scalar:
625 case vtn_base_type_vector:
626 case vtn_base_type_matrix:
627 case vtn_base_type_array:
628 case vtn_base_type_pointer:
629 case vtn_base_type_image:
630 case vtn_base_type_sampler:
631 case vtn_base_type_sampled_image:
632 /* Nothing more to do */
635 case vtn_base_type_struct:
636 dest->members = ralloc_array(b, struct vtn_type *, src->length);
637 memcpy(dest->members, src->members,
638 src->length * sizeof(src->members[0]));
640 dest->offsets = ralloc_array(b, unsigned, src->length);
641 memcpy(dest->offsets, src->offsets,
642 src->length * sizeof(src->offsets[0]));
645 case vtn_base_type_function:
646 dest->params = ralloc_array(b, struct vtn_type *, src->length);
647 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
654 static struct vtn_type *
655 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
657 type->members[member] = vtn_type_copy(b, type->members[member]);
658 type = type->members[member];
660 /* We may have an array of matrices.... Oh, joy! */
661 while (glsl_type_is_array(type->type)) {
662 type->array_element = vtn_type_copy(b, type->array_element);
663 type = type->array_element;
666 vtn_assert(glsl_type_is_matrix(type->type));
672 struct_member_decoration_cb(struct vtn_builder *b,
673 struct vtn_value *val, int member,
674 const struct vtn_decoration *dec, void *void_ctx)
676 struct member_decoration_ctx *ctx = void_ctx;
681 assert(member < ctx->num_fields);
683 switch (dec->decoration) {
684 case SpvDecorationNonWritable:
685 case SpvDecorationNonReadable:
686 case SpvDecorationRelaxedPrecision:
687 case SpvDecorationVolatile:
688 case SpvDecorationCoherent:
689 case SpvDecorationUniform:
690 break; /* FIXME: Do nothing with this for now. */
691 case SpvDecorationNoPerspective:
692 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
694 case SpvDecorationFlat:
695 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
697 case SpvDecorationCentroid:
698 ctx->fields[member].centroid = true;
700 case SpvDecorationSample:
701 ctx->fields[member].sample = true;
703 case SpvDecorationStream:
704 /* Vulkan only allows one GS stream */
705 vtn_assert(dec->literals[0] == 0);
707 case SpvDecorationLocation:
708 ctx->fields[member].location = dec->literals[0];
710 case SpvDecorationComponent:
711 break; /* FIXME: What should we do with these? */
712 case SpvDecorationBuiltIn:
713 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
714 ctx->type->members[member]->is_builtin = true;
715 ctx->type->members[member]->builtin = dec->literals[0];
716 ctx->type->builtin_block = true;
718 case SpvDecorationOffset:
719 ctx->type->offsets[member] = dec->literals[0];
721 case SpvDecorationMatrixStride:
722 /* Handled as a second pass */
724 case SpvDecorationColMajor:
725 break; /* Nothing to do here. Column-major is the default. */
726 case SpvDecorationRowMajor:
727 mutable_matrix_member(b, ctx->type, member)->row_major = true;
730 case SpvDecorationPatch:
733 case SpvDecorationSpecId:
734 case SpvDecorationBlock:
735 case SpvDecorationBufferBlock:
736 case SpvDecorationArrayStride:
737 case SpvDecorationGLSLShared:
738 case SpvDecorationGLSLPacked:
739 case SpvDecorationInvariant:
740 case SpvDecorationRestrict:
741 case SpvDecorationAliased:
742 case SpvDecorationConstant:
743 case SpvDecorationIndex:
744 case SpvDecorationBinding:
745 case SpvDecorationDescriptorSet:
746 case SpvDecorationLinkageAttributes:
747 case SpvDecorationNoContraction:
748 case SpvDecorationInputAttachmentIndex:
749 vtn_warn("Decoration not allowed on struct members: %s",
750 spirv_decoration_to_string(dec->decoration));
753 case SpvDecorationXfbBuffer:
754 case SpvDecorationXfbStride:
755 vtn_warn("Vulkan does not have transform feedback");
758 case SpvDecorationCPacked:
759 case SpvDecorationSaturatedConversion:
760 case SpvDecorationFuncParamAttr:
761 case SpvDecorationFPRoundingMode:
762 case SpvDecorationFPFastMathMode:
763 case SpvDecorationAlignment:
764 vtn_warn("Decoration only allowed for CL-style kernels: %s",
765 spirv_decoration_to_string(dec->decoration));
769 vtn_fail("Unhandled decoration");
773 /* Matrix strides are handled as a separate pass because we need to know
774 * whether the matrix is row-major or not first.
777 struct_member_matrix_stride_cb(struct vtn_builder *b,
778 struct vtn_value *val, int member,
779 const struct vtn_decoration *dec,
782 if (dec->decoration != SpvDecorationMatrixStride)
785 vtn_fail_if(member < 0,
786 "The MatrixStride decoration is only allowed on members "
789 struct member_decoration_ctx *ctx = void_ctx;
791 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
792 if (mat_type->row_major) {
793 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
794 mat_type->stride = mat_type->array_element->stride;
795 mat_type->array_element->stride = dec->literals[0];
797 vtn_assert(mat_type->array_element->stride > 0);
798 mat_type->stride = dec->literals[0];
803 type_decoration_cb(struct vtn_builder *b,
804 struct vtn_value *val, int member,
805 const struct vtn_decoration *dec, void *ctx)
807 struct vtn_type *type = val->type;
810 /* This should have been handled by OpTypeStruct */
811 assert(val->type->base_type == vtn_base_type_struct);
812 assert(member >= 0 && member < val->type->length);
816 switch (dec->decoration) {
817 case SpvDecorationArrayStride:
818 vtn_assert(type->base_type == vtn_base_type_matrix ||
819 type->base_type == vtn_base_type_array ||
820 type->base_type == vtn_base_type_pointer);
821 type->stride = dec->literals[0];
823 case SpvDecorationBlock:
824 vtn_assert(type->base_type == vtn_base_type_struct);
827 case SpvDecorationBufferBlock:
828 vtn_assert(type->base_type == vtn_base_type_struct);
829 type->buffer_block = true;
831 case SpvDecorationGLSLShared:
832 case SpvDecorationGLSLPacked:
833 /* Ignore these, since we get explicit offsets anyways */
836 case SpvDecorationRowMajor:
837 case SpvDecorationColMajor:
838 case SpvDecorationMatrixStride:
839 case SpvDecorationBuiltIn:
840 case SpvDecorationNoPerspective:
841 case SpvDecorationFlat:
842 case SpvDecorationPatch:
843 case SpvDecorationCentroid:
844 case SpvDecorationSample:
845 case SpvDecorationVolatile:
846 case SpvDecorationCoherent:
847 case SpvDecorationNonWritable:
848 case SpvDecorationNonReadable:
849 case SpvDecorationUniform:
850 case SpvDecorationLocation:
851 case SpvDecorationComponent:
852 case SpvDecorationOffset:
853 case SpvDecorationXfbBuffer:
854 case SpvDecorationXfbStride:
855 vtn_warn("Decoration only allowed for struct members: %s",
856 spirv_decoration_to_string(dec->decoration));
859 case SpvDecorationStream:
860 /* We don't need to do anything here, as stream is filled up when
861 * aplying the decoration to a variable, just check that if it is not a
862 * struct member, it should be a struct.
864 vtn_assert(type->base_type == vtn_base_type_struct);
867 case SpvDecorationRelaxedPrecision:
868 case SpvDecorationSpecId:
869 case SpvDecorationInvariant:
870 case SpvDecorationRestrict:
871 case SpvDecorationAliased:
872 case SpvDecorationConstant:
873 case SpvDecorationIndex:
874 case SpvDecorationBinding:
875 case SpvDecorationDescriptorSet:
876 case SpvDecorationLinkageAttributes:
877 case SpvDecorationNoContraction:
878 case SpvDecorationInputAttachmentIndex:
879 vtn_warn("Decoration not allowed on types: %s",
880 spirv_decoration_to_string(dec->decoration));
883 case SpvDecorationCPacked:
884 case SpvDecorationSaturatedConversion:
885 case SpvDecorationFuncParamAttr:
886 case SpvDecorationFPRoundingMode:
887 case SpvDecorationFPFastMathMode:
888 case SpvDecorationAlignment:
889 vtn_warn("Decoration only allowed for CL-style kernels: %s",
890 spirv_decoration_to_string(dec->decoration));
894 vtn_fail("Unhandled decoration");
899 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
902 case SpvImageFormatUnknown: return 0; /* GL_NONE */
903 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
904 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
905 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
906 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
907 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
908 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
909 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
910 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
911 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
912 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
913 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
914 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
915 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
916 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
917 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
918 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
919 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
920 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
921 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
922 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
923 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
924 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
925 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
926 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
927 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
928 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
929 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
930 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
931 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
932 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
933 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
934 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
935 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
936 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
937 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
938 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
939 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
940 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
941 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
943 vtn_fail("Invalid image format");
947 static struct vtn_type *
948 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
949 uint32_t *size_out, uint32_t *align_out)
951 switch (type->base_type) {
952 case vtn_base_type_scalar: {
953 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
954 *size_out = comp_size;
955 *align_out = comp_size;
959 case vtn_base_type_vector: {
960 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
961 unsigned align_comps = type->length == 3 ? 4 : type->length;
962 *size_out = comp_size * type->length,
963 *align_out = comp_size * align_comps;
967 case vtn_base_type_matrix:
968 case vtn_base_type_array: {
969 /* We're going to add an array stride */
970 type = vtn_type_copy(b, type);
971 uint32_t elem_size, elem_align;
972 type->array_element = vtn_type_layout_std430(b, type->array_element,
973 &elem_size, &elem_align);
974 type->stride = vtn_align_u32(elem_size, elem_align);
975 *size_out = type->stride * type->length;
976 *align_out = elem_align;
980 case vtn_base_type_struct: {
981 /* We're going to add member offsets */
982 type = vtn_type_copy(b, type);
985 for (unsigned i = 0; i < type->length; i++) {
986 uint32_t mem_size, mem_align;
987 type->members[i] = vtn_type_layout_std430(b, type->members[i],
988 &mem_size, &mem_align);
989 offset = vtn_align_u32(offset, mem_align);
990 type->offsets[i] = offset;
992 align = MAX2(align, mem_align);
1000 unreachable("Invalid SPIR-V type for std430");
1005 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
1006 const uint32_t *w, unsigned count)
1008 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
1010 val->type = rzalloc(b, struct vtn_type);
1011 val->type->id = w[1];
1015 val->type->base_type = vtn_base_type_void;
1016 val->type->type = glsl_void_type();
1019 val->type->base_type = vtn_base_type_scalar;
1020 val->type->type = glsl_bool_type();
1021 val->type->length = 1;
1023 case SpvOpTypeInt: {
1024 int bit_size = w[2];
1025 const bool signedness = w[3];
1026 val->type->base_type = vtn_base_type_scalar;
1029 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
1032 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
1035 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
1038 val->type->type = (signedness ? glsl_int8_t_type() : glsl_uint8_t_type());
1041 vtn_fail("Invalid int bit size");
1043 val->type->length = 1;
1047 case SpvOpTypeFloat: {
1048 int bit_size = w[2];
1049 val->type->base_type = vtn_base_type_scalar;
1052 val->type->type = glsl_float16_t_type();
1055 val->type->type = glsl_float_type();
1058 val->type->type = glsl_double_type();
1061 vtn_fail("Invalid float bit size");
1063 val->type->length = 1;
1067 case SpvOpTypeVector: {
1068 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1069 unsigned elems = w[3];
1071 vtn_fail_if(base->base_type != vtn_base_type_scalar,
1072 "Base type for OpTypeVector must be a scalar");
1073 vtn_fail_if((elems < 2 || elems > 4) && (elems != 8) && (elems != 16),
1074 "Invalid component count for OpTypeVector");
1076 val->type->base_type = vtn_base_type_vector;
1077 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
1078 val->type->length = elems;
1079 val->type->stride = glsl_get_bit_size(base->type) / 8;
1080 val->type->array_element = base;
1084 case SpvOpTypeMatrix: {
1085 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1086 unsigned columns = w[3];
1088 vtn_fail_if(base->base_type != vtn_base_type_vector,
1089 "Base type for OpTypeMatrix must be a vector");
1090 vtn_fail_if(columns < 2 || columns > 4,
1091 "Invalid column count for OpTypeMatrix");
1093 val->type->base_type = vtn_base_type_matrix;
1094 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
1095 glsl_get_vector_elements(base->type),
1097 vtn_fail_if(glsl_type_is_error(val->type->type),
1098 "Unsupported base type for OpTypeMatrix");
1099 assert(!glsl_type_is_error(val->type->type));
1100 val->type->length = columns;
1101 val->type->array_element = base;
1102 val->type->row_major = false;
1103 val->type->stride = 0;
1107 case SpvOpTypeRuntimeArray:
1108 case SpvOpTypeArray: {
1109 struct vtn_type *array_element =
1110 vtn_value(b, w[2], vtn_value_type_type)->type;
1112 if (opcode == SpvOpTypeRuntimeArray) {
1113 /* A length of 0 is used to denote unsized arrays */
1114 val->type->length = 0;
1117 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
1120 val->type->base_type = vtn_base_type_array;
1121 val->type->type = glsl_array_type(array_element->type, val->type->length);
1122 val->type->array_element = array_element;
1123 val->type->stride = 0;
1127 case SpvOpTypeStruct: {
1128 unsigned num_fields = count - 2;
1129 val->type->base_type = vtn_base_type_struct;
1130 val->type->length = num_fields;
1131 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
1132 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1134 NIR_VLA(struct glsl_struct_field, fields, count);
1135 for (unsigned i = 0; i < num_fields; i++) {
1136 val->type->members[i] =
1137 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1138 fields[i] = (struct glsl_struct_field) {
1139 .type = val->type->members[i]->type,
1140 .name = ralloc_asprintf(b, "field%d", i),
1145 struct member_decoration_ctx ctx = {
1146 .num_fields = num_fields,
1151 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1152 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1154 const char *name = val->name ? val->name : "struct";
1156 val->type->type = glsl_struct_type(fields, num_fields, name);
1160 case SpvOpTypeFunction: {
1161 val->type->base_type = vtn_base_type_function;
1162 val->type->type = NULL;
1164 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1166 const unsigned num_params = count - 3;
1167 val->type->length = num_params;
1168 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1169 for (unsigned i = 0; i < count - 3; i++) {
1170 val->type->params[i] =
1171 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1176 case SpvOpTypePointer: {
1177 SpvStorageClass storage_class = w[2];
1178 struct vtn_type *deref_type =
1179 vtn_value(b, w[3], vtn_value_type_type)->type;
1181 val->type->base_type = vtn_base_type_pointer;
1182 val->type->storage_class = storage_class;
1183 val->type->deref = deref_type;
1185 if (storage_class == SpvStorageClassUniform ||
1186 storage_class == SpvStorageClassStorageBuffer) {
1187 /* These can actually be stored to nir_variables and used as SSA
1188 * values so they need a real glsl_type.
1190 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1193 if (storage_class == SpvStorageClassPushConstant) {
1194 /* These can actually be stored to nir_variables and used as SSA
1195 * values so they need a real glsl_type.
1197 val->type->type = glsl_uint_type();
1200 if (storage_class == SpvStorageClassWorkgroup &&
1201 b->options->lower_workgroup_access_to_offsets) {
1202 uint32_t size, align;
1203 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1205 val->type->length = size;
1206 val->type->align = align;
1207 /* These can actually be stored to nir_variables and used as SSA
1208 * values so they need a real glsl_type.
1210 val->type->type = glsl_uint_type();
1215 case SpvOpTypeImage: {
1216 val->type->base_type = vtn_base_type_image;
1218 const struct vtn_type *sampled_type =
1219 vtn_value(b, w[2], vtn_value_type_type)->type;
1221 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1222 glsl_get_bit_size(sampled_type->type) != 32,
1223 "Sampled type of OpTypeImage must be a 32-bit scalar");
1225 enum glsl_sampler_dim dim;
1226 switch ((SpvDim)w[3]) {
1227 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1228 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1229 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1230 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1231 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1232 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1233 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1235 vtn_fail("Invalid SPIR-V image dimensionality");
1238 bool is_shadow = w[4];
1239 bool is_array = w[5];
1240 bool multisampled = w[6];
1241 unsigned sampled = w[7];
1242 SpvImageFormat format = w[8];
1245 val->type->access_qualifier = w[9];
1247 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1250 if (dim == GLSL_SAMPLER_DIM_2D)
1251 dim = GLSL_SAMPLER_DIM_MS;
1252 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1253 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1255 vtn_fail("Unsupported multisampled image type");
1258 val->type->image_format = translate_image_format(b, format);
1260 enum glsl_base_type sampled_base_type =
1261 glsl_get_base_type(sampled_type->type);
1263 val->type->sampled = true;
1264 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1266 } else if (sampled == 2) {
1267 vtn_assert(!is_shadow);
1268 val->type->sampled = false;
1269 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1271 vtn_fail("We need to know if the image will be sampled");
1276 case SpvOpTypeSampledImage:
1277 val->type->base_type = vtn_base_type_sampled_image;
1278 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1279 val->type->type = val->type->image->type;
1282 case SpvOpTypeSampler:
1283 /* The actual sampler type here doesn't really matter. It gets
1284 * thrown away the moment you combine it with an image. What really
1285 * matters is that it's a sampler type as opposed to an integer type
1286 * so the backend knows what to do.
1288 val->type->base_type = vtn_base_type_sampler;
1289 val->type->type = glsl_bare_sampler_type();
1292 case SpvOpTypeOpaque:
1293 case SpvOpTypeEvent:
1294 case SpvOpTypeDeviceEvent:
1295 case SpvOpTypeReserveId:
1296 case SpvOpTypeQueue:
1299 vtn_fail("Unhandled opcode");
1302 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1305 static nir_constant *
1306 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1308 nir_constant *c = rzalloc(b, nir_constant);
1310 /* For pointers and other typeless things, we have to return something but
1311 * it doesn't matter what.
1316 switch (glsl_get_base_type(type)) {
1318 case GLSL_TYPE_UINT:
1319 case GLSL_TYPE_INT16:
1320 case GLSL_TYPE_UINT16:
1321 case GLSL_TYPE_UINT8:
1322 case GLSL_TYPE_INT8:
1323 case GLSL_TYPE_INT64:
1324 case GLSL_TYPE_UINT64:
1325 case GLSL_TYPE_BOOL:
1326 case GLSL_TYPE_FLOAT:
1327 case GLSL_TYPE_FLOAT16:
1328 case GLSL_TYPE_DOUBLE:
1329 /* Nothing to do here. It's already initialized to zero */
1332 case GLSL_TYPE_ARRAY:
1333 vtn_assert(glsl_get_length(type) > 0);
1334 c->num_elements = glsl_get_length(type);
1335 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1337 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1338 for (unsigned i = 1; i < c->num_elements; i++)
1339 c->elements[i] = c->elements[0];
1342 case GLSL_TYPE_STRUCT:
1343 c->num_elements = glsl_get_length(type);
1344 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1346 for (unsigned i = 0; i < c->num_elements; i++) {
1347 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1352 vtn_fail("Invalid type for null constant");
1359 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1360 int member, const struct vtn_decoration *dec,
1363 vtn_assert(member == -1);
1364 if (dec->decoration != SpvDecorationSpecId)
1367 struct spec_constant_value *const_value = data;
1369 for (unsigned i = 0; i < b->num_specializations; i++) {
1370 if (b->specializations[i].id == dec->literals[0]) {
1371 if (const_value->is_double)
1372 const_value->data64 = b->specializations[i].data64;
1374 const_value->data32 = b->specializations[i].data32;
1381 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1382 uint32_t const_value)
1384 struct spec_constant_value data;
1385 data.is_double = false;
1386 data.data32 = const_value;
1387 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1392 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1393 uint64_t const_value)
1395 struct spec_constant_value data;
1396 data.is_double = true;
1397 data.data64 = const_value;
1398 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1403 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1404 struct vtn_value *val,
1406 const struct vtn_decoration *dec,
1409 vtn_assert(member == -1);
1410 if (dec->decoration != SpvDecorationBuiltIn ||
1411 dec->literals[0] != SpvBuiltInWorkgroupSize)
1414 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1416 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1417 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1418 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1422 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1423 const uint32_t *w, unsigned count)
1425 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1426 val->constant = rzalloc(b, nir_constant);
1428 case SpvOpConstantTrue:
1429 case SpvOpConstantFalse:
1430 case SpvOpSpecConstantTrue:
1431 case SpvOpSpecConstantFalse: {
1432 vtn_fail_if(val->type->type != glsl_bool_type(),
1433 "Result type of %s must be OpTypeBool",
1434 spirv_op_to_string(opcode));
1436 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1437 opcode == SpvOpSpecConstantTrue);
1439 if (opcode == SpvOpSpecConstantTrue ||
1440 opcode == SpvOpSpecConstantFalse)
1441 int_val = get_specialization(b, val, int_val);
1443 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1447 case SpvOpConstant: {
1448 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1449 "Result type of %s must be a scalar",
1450 spirv_op_to_string(opcode));
1451 int bit_size = glsl_get_bit_size(val->type->type);
1454 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1457 val->constant->values->u32[0] = w[3];
1460 val->constant->values->u16[0] = w[3];
1463 val->constant->values->u8[0] = w[3];
1466 vtn_fail("Unsupported SpvOpConstant bit size");
1471 case SpvOpSpecConstant: {
1472 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1473 "Result type of %s must be a scalar",
1474 spirv_op_to_string(opcode));
1475 int bit_size = glsl_get_bit_size(val->type->type);
1478 val->constant->values[0].u64[0] =
1479 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1482 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1485 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1488 val->constant->values[0].u8[0] = get_specialization(b, val, w[3]);
1491 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1496 case SpvOpSpecConstantComposite:
1497 case SpvOpConstantComposite: {
1498 unsigned elem_count = count - 3;
1499 vtn_fail_if(elem_count != val->type->length,
1500 "%s has %u constituents, expected %u",
1501 spirv_op_to_string(opcode), elem_count, val->type->length);
1503 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1504 for (unsigned i = 0; i < elem_count; i++) {
1505 struct vtn_value *val = vtn_untyped_value(b, w[i + 3]);
1507 if (val->value_type == vtn_value_type_constant) {
1508 elems[i] = val->constant;
1510 vtn_fail_if(val->value_type != vtn_value_type_undef,
1511 "only constants or undefs allowed for "
1512 "SpvOpConstantComposite");
1513 /* to make it easier, just insert a NULL constant for now */
1514 elems[i] = vtn_null_constant(b, val->type->type);
1518 switch (val->type->base_type) {
1519 case vtn_base_type_vector: {
1520 assert(glsl_type_is_vector(val->type->type));
1521 int bit_size = glsl_get_bit_size(val->type->type);
1522 for (unsigned i = 0; i < elem_count; i++) {
1525 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1528 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1531 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1534 val->constant->values[0].u8[i] = elems[i]->values[0].u8[0];
1537 vtn_fail("Invalid SpvOpConstantComposite bit size");
1543 case vtn_base_type_matrix:
1544 assert(glsl_type_is_matrix(val->type->type));
1545 for (unsigned i = 0; i < elem_count; i++)
1546 val->constant->values[i] = elems[i]->values[0];
1549 case vtn_base_type_struct:
1550 case vtn_base_type_array:
1551 ralloc_steal(val->constant, elems);
1552 val->constant->num_elements = elem_count;
1553 val->constant->elements = elems;
1557 vtn_fail("Result type of %s must be a composite type",
1558 spirv_op_to_string(opcode));
1563 case SpvOpSpecConstantOp: {
1564 SpvOp opcode = get_specialization(b, val, w[3]);
1566 case SpvOpVectorShuffle: {
1567 struct vtn_value *v0 = &b->values[w[4]];
1568 struct vtn_value *v1 = &b->values[w[5]];
1570 vtn_assert(v0->value_type == vtn_value_type_constant ||
1571 v0->value_type == vtn_value_type_undef);
1572 vtn_assert(v1->value_type == vtn_value_type_constant ||
1573 v1->value_type == vtn_value_type_undef);
1575 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1576 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1578 vtn_assert(len0 + len1 < 16);
1580 unsigned bit_size = glsl_get_bit_size(val->type->type);
1581 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1582 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1584 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1585 (void)bit_size0; (void)bit_size1;
1587 if (bit_size == 64) {
1589 if (v0->value_type == vtn_value_type_constant) {
1590 for (unsigned i = 0; i < len0; i++)
1591 u64[i] = v0->constant->values[0].u64[i];
1593 if (v1->value_type == vtn_value_type_constant) {
1594 for (unsigned i = 0; i < len1; i++)
1595 u64[len0 + i] = v1->constant->values[0].u64[i];
1598 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1599 uint32_t comp = w[i + 6];
1600 /* If component is not used, set the value to a known constant
1601 * to detect if it is wrongly used.
1603 if (comp == (uint32_t)-1)
1604 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1606 val->constant->values[0].u64[j] = u64[comp];
1609 /* This is for both 32-bit and 16-bit values */
1611 if (v0->value_type == vtn_value_type_constant) {
1612 for (unsigned i = 0; i < len0; i++)
1613 u32[i] = v0->constant->values[0].u32[i];
1615 if (v1->value_type == vtn_value_type_constant) {
1616 for (unsigned i = 0; i < len1; i++)
1617 u32[len0 + i] = v1->constant->values[0].u32[i];
1620 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1621 uint32_t comp = w[i + 6];
1622 /* If component is not used, set the value to a known constant
1623 * to detect if it is wrongly used.
1625 if (comp == (uint32_t)-1)
1626 val->constant->values[0].u32[j] = 0xdeadbeef;
1628 val->constant->values[0].u32[j] = u32[comp];
1634 case SpvOpCompositeExtract:
1635 case SpvOpCompositeInsert: {
1636 struct vtn_value *comp;
1637 unsigned deref_start;
1638 struct nir_constant **c;
1639 if (opcode == SpvOpCompositeExtract) {
1640 comp = vtn_value(b, w[4], vtn_value_type_constant);
1642 c = &comp->constant;
1644 comp = vtn_value(b, w[5], vtn_value_type_constant);
1646 val->constant = nir_constant_clone(comp->constant,
1653 const struct vtn_type *type = comp->type;
1654 for (unsigned i = deref_start; i < count; i++) {
1655 vtn_fail_if(w[i] > type->length,
1656 "%uth index of %s is %u but the type has only "
1657 "%u elements", i - deref_start,
1658 spirv_op_to_string(opcode), w[i], type->length);
1660 switch (type->base_type) {
1661 case vtn_base_type_vector:
1663 type = type->array_element;
1666 case vtn_base_type_matrix:
1667 assert(col == 0 && elem == -1);
1670 type = type->array_element;
1673 case vtn_base_type_array:
1674 c = &(*c)->elements[w[i]];
1675 type = type->array_element;
1678 case vtn_base_type_struct:
1679 c = &(*c)->elements[w[i]];
1680 type = type->members[w[i]];
1684 vtn_fail("%s must only index into composite types",
1685 spirv_op_to_string(opcode));
1689 if (opcode == SpvOpCompositeExtract) {
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 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1701 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1704 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1707 val->constant->values[0].u8[i] = (*c)->values[col].u8[elem + i];
1710 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1714 struct vtn_value *insert =
1715 vtn_value(b, w[4], vtn_value_type_constant);
1716 vtn_assert(insert->type == type);
1718 *c = insert->constant;
1720 unsigned num_components = type->length;
1721 unsigned bit_size = glsl_get_bit_size(type->type);
1722 for (unsigned i = 0; i < num_components; i++)
1725 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1728 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1731 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1734 (*c)->values[col].u8[elem + i] = insert->constant->values[0].u8[i];
1737 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1746 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1747 nir_alu_type src_alu_type = dst_alu_type;
1748 unsigned num_components = glsl_get_vector_elements(val->type->type);
1751 vtn_assert(count <= 7);
1756 /* We have a source in a conversion */
1758 nir_get_nir_type_for_glsl_type(
1759 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1760 /* We use the bitsize of the conversion source to evaluate the opcode later */
1761 bit_size = glsl_get_bit_size(
1762 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1765 bit_size = glsl_get_bit_size(val->type->type);
1768 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1769 nir_alu_type_get_type_size(src_alu_type),
1770 nir_alu_type_get_type_size(dst_alu_type));
1771 nir_const_value src[4];
1773 for (unsigned i = 0; i < count - 4; i++) {
1775 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1777 unsigned j = swap ? 1 - i : i;
1778 src[j] = c->values[0];
1781 val->constant->values[0] =
1782 nir_eval_const_opcode(op, num_components, bit_size, src);
1789 case SpvOpConstantNull:
1790 val->constant = vtn_null_constant(b, val->type->type);
1793 case SpvOpConstantSampler:
1794 vtn_fail("OpConstantSampler requires Kernel Capability");
1798 vtn_fail("Unhandled opcode");
1801 /* Now that we have the value, update the workgroup size if needed */
1802 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1806 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1807 const uint32_t *w, unsigned count)
1809 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1810 struct vtn_function *vtn_callee =
1811 vtn_value(b, w[3], vtn_value_type_function)->func;
1812 struct nir_function *callee = vtn_callee->impl->function;
1814 vtn_callee->referenced = true;
1816 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1818 unsigned param_idx = 0;
1820 nir_deref_instr *ret_deref = NULL;
1821 struct vtn_type *ret_type = vtn_callee->type->return_type;
1822 if (ret_type->base_type != vtn_base_type_void) {
1823 nir_variable *ret_tmp =
1824 nir_local_variable_create(b->nb.impl, ret_type->type, "return_tmp");
1825 ret_deref = nir_build_deref_var(&b->nb, ret_tmp);
1826 call->params[param_idx++] = nir_src_for_ssa(&ret_deref->dest.ssa);
1829 for (unsigned i = 0; i < vtn_callee->type->length; i++) {
1830 struct vtn_type *arg_type = vtn_callee->type->params[i];
1831 unsigned arg_id = w[4 + i];
1833 if (arg_type->base_type == vtn_base_type_sampled_image) {
1834 struct vtn_sampled_image *sampled_image =
1835 vtn_value(b, arg_id, vtn_value_type_sampled_image)->sampled_image;
1837 call->params[param_idx++] =
1838 nir_src_for_ssa(&sampled_image->image->deref->dest.ssa);
1839 call->params[param_idx++] =
1840 nir_src_for_ssa(&sampled_image->sampler->deref->dest.ssa);
1841 } else if (arg_type->base_type == vtn_base_type_pointer ||
1842 arg_type->base_type == vtn_base_type_image ||
1843 arg_type->base_type == vtn_base_type_sampler) {
1844 struct vtn_pointer *pointer =
1845 vtn_value(b, arg_id, vtn_value_type_pointer)->pointer;
1846 call->params[param_idx++] =
1847 nir_src_for_ssa(vtn_pointer_to_ssa(b, pointer));
1849 /* This is a regular SSA value and we need a temporary */
1851 nir_local_variable_create(b->nb.impl, arg_type->type, "arg_tmp");
1852 nir_deref_instr *tmp_deref = nir_build_deref_var(&b->nb, tmp);
1853 vtn_local_store(b, vtn_ssa_value(b, arg_id), tmp_deref);
1854 call->params[param_idx++] = nir_src_for_ssa(&tmp_deref->dest.ssa);
1857 assert(param_idx == call->num_params);
1859 nir_builder_instr_insert(&b->nb, &call->instr);
1861 if (ret_type->base_type == vtn_base_type_void) {
1862 vtn_push_value(b, w[2], vtn_value_type_undef);
1864 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, ret_deref));
1868 struct vtn_ssa_value *
1869 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1871 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1874 if (!glsl_type_is_vector_or_scalar(type)) {
1875 unsigned elems = glsl_get_length(type);
1876 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1877 for (unsigned i = 0; i < elems; i++) {
1878 const struct glsl_type *child_type;
1880 switch (glsl_get_base_type(type)) {
1882 case GLSL_TYPE_UINT:
1883 case GLSL_TYPE_INT16:
1884 case GLSL_TYPE_UINT16:
1885 case GLSL_TYPE_UINT8:
1886 case GLSL_TYPE_INT8:
1887 case GLSL_TYPE_INT64:
1888 case GLSL_TYPE_UINT64:
1889 case GLSL_TYPE_BOOL:
1890 case GLSL_TYPE_FLOAT:
1891 case GLSL_TYPE_FLOAT16:
1892 case GLSL_TYPE_DOUBLE:
1893 child_type = glsl_get_column_type(type);
1895 case GLSL_TYPE_ARRAY:
1896 child_type = glsl_get_array_element(type);
1898 case GLSL_TYPE_STRUCT:
1899 child_type = glsl_get_struct_field(type, i);
1902 vtn_fail("unkown base type");
1905 val->elems[i] = vtn_create_ssa_value(b, child_type);
1913 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1916 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1917 src.src_type = type;
1922 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1923 const uint32_t *w, unsigned count)
1925 if (opcode == SpvOpSampledImage) {
1926 struct vtn_value *val =
1927 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1928 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1929 val->sampled_image->type =
1930 vtn_value(b, w[1], vtn_value_type_type)->type;
1931 val->sampled_image->image =
1932 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1933 val->sampled_image->sampler =
1934 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1936 } else if (opcode == SpvOpImage) {
1937 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1938 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1939 if (src_val->value_type == vtn_value_type_sampled_image) {
1940 val->pointer = src_val->sampled_image->image;
1942 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1943 val->pointer = src_val->pointer;
1948 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1949 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1951 struct vtn_sampled_image sampled;
1952 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1953 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1954 sampled = *sampled_val->sampled_image;
1956 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1957 sampled.type = sampled_val->pointer->type;
1958 sampled.image = NULL;
1959 sampled.sampler = sampled_val->pointer;
1962 const struct glsl_type *image_type = sampled.type->type;
1963 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1964 const bool is_array = glsl_sampler_type_is_array(image_type);
1966 /* Figure out the base texture operation */
1969 case SpvOpImageSampleImplicitLod:
1970 case SpvOpImageSampleDrefImplicitLod:
1971 case SpvOpImageSampleProjImplicitLod:
1972 case SpvOpImageSampleProjDrefImplicitLod:
1973 texop = nir_texop_tex;
1976 case SpvOpImageSampleExplicitLod:
1977 case SpvOpImageSampleDrefExplicitLod:
1978 case SpvOpImageSampleProjExplicitLod:
1979 case SpvOpImageSampleProjDrefExplicitLod:
1980 texop = nir_texop_txl;
1983 case SpvOpImageFetch:
1984 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1985 texop = nir_texop_txf_ms;
1987 texop = nir_texop_txf;
1991 case SpvOpImageGather:
1992 case SpvOpImageDrefGather:
1993 texop = nir_texop_tg4;
1996 case SpvOpImageQuerySizeLod:
1997 case SpvOpImageQuerySize:
1998 texop = nir_texop_txs;
2001 case SpvOpImageQueryLod:
2002 texop = nir_texop_lod;
2005 case SpvOpImageQueryLevels:
2006 texop = nir_texop_query_levels;
2009 case SpvOpImageQuerySamples:
2010 texop = nir_texop_texture_samples;
2014 vtn_fail("Unhandled opcode");
2017 nir_tex_src srcs[10]; /* 10 should be enough */
2018 nir_tex_src *p = srcs;
2020 nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
2021 nir_deref_instr *texture =
2022 sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
2024 p->src = nir_src_for_ssa(&texture->dest.ssa);
2025 p->src_type = nir_tex_src_texture_deref;
2034 /* These operations require a sampler */
2035 p->src = nir_src_for_ssa(&sampler->dest.ssa);
2036 p->src_type = nir_tex_src_sampler_deref;
2040 case nir_texop_txf_ms:
2043 case nir_texop_query_levels:
2044 case nir_texop_texture_samples:
2045 case nir_texop_samples_identical:
2048 case nir_texop_txf_ms_mcs:
2049 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2054 struct nir_ssa_def *coord;
2055 unsigned coord_components;
2057 case SpvOpImageSampleImplicitLod:
2058 case SpvOpImageSampleExplicitLod:
2059 case SpvOpImageSampleDrefImplicitLod:
2060 case SpvOpImageSampleDrefExplicitLod:
2061 case SpvOpImageSampleProjImplicitLod:
2062 case SpvOpImageSampleProjExplicitLod:
2063 case SpvOpImageSampleProjDrefImplicitLod:
2064 case SpvOpImageSampleProjDrefExplicitLod:
2065 case SpvOpImageFetch:
2066 case SpvOpImageGather:
2067 case SpvOpImageDrefGather:
2068 case SpvOpImageQueryLod: {
2069 /* All these types have the coordinate as their first real argument */
2070 switch (sampler_dim) {
2071 case GLSL_SAMPLER_DIM_1D:
2072 case GLSL_SAMPLER_DIM_BUF:
2073 coord_components = 1;
2075 case GLSL_SAMPLER_DIM_2D:
2076 case GLSL_SAMPLER_DIM_RECT:
2077 case GLSL_SAMPLER_DIM_MS:
2078 coord_components = 2;
2080 case GLSL_SAMPLER_DIM_3D:
2081 case GLSL_SAMPLER_DIM_CUBE:
2082 coord_components = 3;
2085 vtn_fail("Invalid sampler type");
2088 if (is_array && texop != nir_texop_lod)
2091 coord = vtn_ssa_value(b, w[idx++])->def;
2092 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
2093 (1 << coord_components) - 1));
2094 p->src_type = nir_tex_src_coord;
2101 coord_components = 0;
2106 case SpvOpImageSampleProjImplicitLod:
2107 case SpvOpImageSampleProjExplicitLod:
2108 case SpvOpImageSampleProjDrefImplicitLod:
2109 case SpvOpImageSampleProjDrefExplicitLod:
2110 /* These have the projector as the last coordinate component */
2111 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
2112 p->src_type = nir_tex_src_projector;
2120 bool is_shadow = false;
2121 unsigned gather_component = 0;
2123 case SpvOpImageSampleDrefImplicitLod:
2124 case SpvOpImageSampleDrefExplicitLod:
2125 case SpvOpImageSampleProjDrefImplicitLod:
2126 case SpvOpImageSampleProjDrefExplicitLod:
2127 case SpvOpImageDrefGather:
2128 /* These all have an explicit depth value as their next source */
2130 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
2133 case SpvOpImageGather:
2134 /* This has a component as its next source */
2136 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
2143 /* For OpImageQuerySizeLod, we always have an LOD */
2144 if (opcode == SpvOpImageQuerySizeLod)
2145 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2147 /* Now we need to handle some number of optional arguments */
2148 const struct vtn_ssa_value *gather_offsets = NULL;
2150 uint32_t operands = w[idx++];
2152 if (operands & SpvImageOperandsBiasMask) {
2153 vtn_assert(texop == nir_texop_tex);
2154 texop = nir_texop_txb;
2155 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
2158 if (operands & SpvImageOperandsLodMask) {
2159 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
2160 texop == nir_texop_txs);
2161 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2164 if (operands & SpvImageOperandsGradMask) {
2165 vtn_assert(texop == nir_texop_txl);
2166 texop = nir_texop_txd;
2167 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
2168 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
2171 if (operands & SpvImageOperandsOffsetMask ||
2172 operands & SpvImageOperandsConstOffsetMask)
2173 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
2175 if (operands & SpvImageOperandsConstOffsetsMask) {
2176 nir_tex_src none = {0};
2177 gather_offsets = vtn_ssa_value(b, w[idx++]);
2181 if (operands & SpvImageOperandsSampleMask) {
2182 vtn_assert(texop == nir_texop_txf_ms);
2183 texop = nir_texop_txf_ms;
2184 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
2187 /* We should have now consumed exactly all of the arguments */
2188 vtn_assert(idx == count);
2190 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
2193 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
2195 instr->coord_components = coord_components;
2196 instr->sampler_dim = sampler_dim;
2197 instr->is_array = is_array;
2198 instr->is_shadow = is_shadow;
2199 instr->is_new_style_shadow =
2200 is_shadow && glsl_get_components(ret_type->type) == 1;
2201 instr->component = gather_component;
2203 switch (glsl_get_sampler_result_type(image_type)) {
2204 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
2205 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
2206 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
2207 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
2209 vtn_fail("Invalid base type for sampler result");
2212 nir_ssa_dest_init(&instr->instr, &instr->dest,
2213 nir_tex_instr_dest_size(instr), 32, NULL);
2215 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2216 nir_tex_instr_dest_size(instr));
2219 nir_instr *instruction;
2220 if (gather_offsets) {
2221 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2222 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2223 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2225 /* Copy the current instruction 4x */
2226 for (uint32_t i = 1; i < 4; i++) {
2227 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2228 instrs[i]->op = instr->op;
2229 instrs[i]->coord_components = instr->coord_components;
2230 instrs[i]->sampler_dim = instr->sampler_dim;
2231 instrs[i]->is_array = instr->is_array;
2232 instrs[i]->is_shadow = instr->is_shadow;
2233 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2234 instrs[i]->component = instr->component;
2235 instrs[i]->dest_type = instr->dest_type;
2237 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2239 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2240 nir_tex_instr_dest_size(instr), 32, NULL);
2243 /* Fill in the last argument with the offset from the passed in offsets
2244 * and insert the instruction into the stream.
2246 for (uint32_t i = 0; i < 4; i++) {
2248 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2249 src.src_type = nir_tex_src_offset;
2250 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2251 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2254 /* Combine the results of the 4 instructions by taking their .w
2257 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2258 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2259 vec4->dest.write_mask = 0xf;
2260 for (uint32_t i = 0; i < 4; i++) {
2261 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2262 vec4->src[i].swizzle[0] = 3;
2264 def = &vec4->dest.dest.ssa;
2265 instruction = &vec4->instr;
2267 def = &instr->dest.ssa;
2268 instruction = &instr->instr;
2271 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2272 val->ssa->def = def;
2274 nir_builder_instr_insert(&b->nb, instruction);
2278 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2279 const uint32_t *w, nir_src *src)
2282 case SpvOpAtomicIIncrement:
2283 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2286 case SpvOpAtomicIDecrement:
2287 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2290 case SpvOpAtomicISub:
2292 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2295 case SpvOpAtomicCompareExchange:
2296 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2297 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2300 case SpvOpAtomicExchange:
2301 case SpvOpAtomicIAdd:
2302 case SpvOpAtomicSMin:
2303 case SpvOpAtomicUMin:
2304 case SpvOpAtomicSMax:
2305 case SpvOpAtomicUMax:
2306 case SpvOpAtomicAnd:
2308 case SpvOpAtomicXor:
2309 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2313 vtn_fail("Invalid SPIR-V atomic");
2317 static nir_ssa_def *
2318 get_image_coord(struct vtn_builder *b, uint32_t value)
2320 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2322 /* The image_load_store intrinsics assume a 4-dim coordinate */
2323 unsigned dim = glsl_get_vector_elements(coord->type);
2324 unsigned swizzle[4];
2325 for (unsigned i = 0; i < 4; i++)
2326 swizzle[i] = MIN2(i, dim - 1);
2328 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2331 static nir_ssa_def *
2332 expand_to_vec4(nir_builder *b, nir_ssa_def *value)
2334 if (value->num_components == 4)
2338 for (unsigned i = 0; i < 4; i++)
2339 swiz[i] = i < value->num_components ? i : 0;
2340 return nir_swizzle(b, value, swiz, 4, false);
2344 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2345 const uint32_t *w, unsigned count)
2347 /* Just get this one out of the way */
2348 if (opcode == SpvOpImageTexelPointer) {
2349 struct vtn_value *val =
2350 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2351 val->image = ralloc(b, struct vtn_image_pointer);
2353 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2354 val->image->coord = get_image_coord(b, w[4]);
2355 val->image->sample = vtn_ssa_value(b, w[5])->def;
2359 struct vtn_image_pointer image;
2362 case SpvOpAtomicExchange:
2363 case SpvOpAtomicCompareExchange:
2364 case SpvOpAtomicCompareExchangeWeak:
2365 case SpvOpAtomicIIncrement:
2366 case SpvOpAtomicIDecrement:
2367 case SpvOpAtomicIAdd:
2368 case SpvOpAtomicISub:
2369 case SpvOpAtomicLoad:
2370 case SpvOpAtomicSMin:
2371 case SpvOpAtomicUMin:
2372 case SpvOpAtomicSMax:
2373 case SpvOpAtomicUMax:
2374 case SpvOpAtomicAnd:
2376 case SpvOpAtomicXor:
2377 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2380 case SpvOpAtomicStore:
2381 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2384 case SpvOpImageQuerySize:
2385 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2387 image.sample = NULL;
2390 case SpvOpImageRead:
2391 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2392 image.coord = get_image_coord(b, w[4]);
2394 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2395 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2396 image.sample = vtn_ssa_value(b, w[6])->def;
2398 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2402 case SpvOpImageWrite:
2403 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2404 image.coord = get_image_coord(b, w[2]);
2408 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2409 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2410 image.sample = vtn_ssa_value(b, w[5])->def;
2412 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2417 vtn_fail("Invalid image opcode");
2420 nir_intrinsic_op op;
2422 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2423 OP(ImageQuerySize, size)
2425 OP(ImageWrite, store)
2426 OP(AtomicLoad, load)
2427 OP(AtomicStore, store)
2428 OP(AtomicExchange, atomic_exchange)
2429 OP(AtomicCompareExchange, atomic_comp_swap)
2430 OP(AtomicIIncrement, atomic_add)
2431 OP(AtomicIDecrement, atomic_add)
2432 OP(AtomicIAdd, atomic_add)
2433 OP(AtomicISub, atomic_add)
2434 OP(AtomicSMin, atomic_min)
2435 OP(AtomicUMin, atomic_min)
2436 OP(AtomicSMax, atomic_max)
2437 OP(AtomicUMax, atomic_max)
2438 OP(AtomicAnd, atomic_and)
2439 OP(AtomicOr, atomic_or)
2440 OP(AtomicXor, atomic_xor)
2443 vtn_fail("Invalid image opcode");
2446 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2448 nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image.image);
2449 intrin->src[0] = nir_src_for_ssa(&image_deref->dest.ssa);
2451 /* ImageQuerySize doesn't take any extra parameters */
2452 if (opcode != SpvOpImageQuerySize) {
2453 /* The image coordinate is always 4 components but we may not have that
2454 * many. Swizzle to compensate.
2456 intrin->src[1] = nir_src_for_ssa(expand_to_vec4(&b->nb, image.coord));
2457 intrin->src[2] = nir_src_for_ssa(image.sample);
2461 case SpvOpAtomicLoad:
2462 case SpvOpImageQuerySize:
2463 case SpvOpImageRead:
2465 case SpvOpAtomicStore:
2466 case SpvOpImageWrite: {
2467 const uint32_t value_id = opcode == SpvOpAtomicStore ? w[4] : w[3];
2468 nir_ssa_def *value = vtn_ssa_value(b, value_id)->def;
2469 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2470 intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value));
2474 case SpvOpAtomicCompareExchange:
2475 case SpvOpAtomicIIncrement:
2476 case SpvOpAtomicIDecrement:
2477 case SpvOpAtomicExchange:
2478 case SpvOpAtomicIAdd:
2479 case SpvOpAtomicISub:
2480 case SpvOpAtomicSMin:
2481 case SpvOpAtomicUMin:
2482 case SpvOpAtomicSMax:
2483 case SpvOpAtomicUMax:
2484 case SpvOpAtomicAnd:
2486 case SpvOpAtomicXor:
2487 fill_common_atomic_sources(b, opcode, w, &intrin->src[3]);
2491 vtn_fail("Invalid image opcode");
2494 if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) {
2495 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2496 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2498 unsigned dest_components = glsl_get_vector_elements(type->type);
2499 intrin->num_components = nir_intrinsic_infos[op].dest_components;
2500 if (intrin->num_components == 0)
2501 intrin->num_components = dest_components;
2503 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2504 intrin->num_components, 32, NULL);
2506 nir_builder_instr_insert(&b->nb, &intrin->instr);
2508 nir_ssa_def *result = &intrin->dest.ssa;
2509 if (intrin->num_components != dest_components)
2510 result = nir_channels(&b->nb, result, (1 << dest_components) - 1);
2512 val->ssa = vtn_create_ssa_value(b, type->type);
2513 val->ssa->def = result;
2515 nir_builder_instr_insert(&b->nb, &intrin->instr);
2519 static nir_intrinsic_op
2520 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2523 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2524 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2525 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2526 OP(AtomicExchange, atomic_exchange)
2527 OP(AtomicCompareExchange, atomic_comp_swap)
2528 OP(AtomicIIncrement, atomic_add)
2529 OP(AtomicIDecrement, atomic_add)
2530 OP(AtomicIAdd, atomic_add)
2531 OP(AtomicISub, atomic_add)
2532 OP(AtomicSMin, atomic_imin)
2533 OP(AtomicUMin, atomic_umin)
2534 OP(AtomicSMax, atomic_imax)
2535 OP(AtomicUMax, atomic_umax)
2536 OP(AtomicAnd, atomic_and)
2537 OP(AtomicOr, atomic_or)
2538 OP(AtomicXor, atomic_xor)
2541 vtn_fail("Invalid SSBO atomic");
2545 static nir_intrinsic_op
2546 get_uniform_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2549 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2550 OP(AtomicLoad, read_deref)
2551 OP(AtomicExchange, exchange)
2552 OP(AtomicCompareExchange, comp_swap)
2553 OP(AtomicIIncrement, inc_deref)
2554 OP(AtomicIDecrement, post_dec_deref)
2555 OP(AtomicIAdd, add_deref)
2556 OP(AtomicISub, add_deref)
2557 OP(AtomicUMin, min_deref)
2558 OP(AtomicUMax, max_deref)
2559 OP(AtomicAnd, and_deref)
2560 OP(AtomicOr, or_deref)
2561 OP(AtomicXor, xor_deref)
2564 /* We left the following out: AtomicStore, AtomicSMin and
2565 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2566 * moment Atomic Counter support is needed for ARB_spirv support, so is
2567 * only need to support GLSL Atomic Counters that are uints and don't
2568 * allow direct storage.
2570 unreachable("Invalid uniform atomic");
2574 static nir_intrinsic_op
2575 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2578 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2579 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2580 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2581 OP(AtomicExchange, atomic_exchange)
2582 OP(AtomicCompareExchange, atomic_comp_swap)
2583 OP(AtomicIIncrement, atomic_add)
2584 OP(AtomicIDecrement, atomic_add)
2585 OP(AtomicIAdd, atomic_add)
2586 OP(AtomicISub, atomic_add)
2587 OP(AtomicSMin, atomic_imin)
2588 OP(AtomicUMin, atomic_umin)
2589 OP(AtomicSMax, atomic_imax)
2590 OP(AtomicUMax, atomic_umax)
2591 OP(AtomicAnd, atomic_and)
2592 OP(AtomicOr, atomic_or)
2593 OP(AtomicXor, atomic_xor)
2596 vtn_fail("Invalid shared atomic");
2600 static nir_intrinsic_op
2601 get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2604 case SpvOpAtomicLoad: return nir_intrinsic_load_deref;
2605 case SpvOpAtomicStore: return nir_intrinsic_store_deref;
2606 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2607 OP(AtomicExchange, atomic_exchange)
2608 OP(AtomicCompareExchange, atomic_comp_swap)
2609 OP(AtomicIIncrement, atomic_add)
2610 OP(AtomicIDecrement, atomic_add)
2611 OP(AtomicIAdd, atomic_add)
2612 OP(AtomicISub, atomic_add)
2613 OP(AtomicSMin, atomic_imin)
2614 OP(AtomicUMin, atomic_umin)
2615 OP(AtomicSMax, atomic_imax)
2616 OP(AtomicUMax, atomic_umax)
2617 OP(AtomicAnd, atomic_and)
2618 OP(AtomicOr, atomic_or)
2619 OP(AtomicXor, atomic_xor)
2622 vtn_fail("Invalid shared atomic");
2627 * Handles shared atomics, ssbo atomics and atomic counters.
2630 vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
2631 const uint32_t *w, unsigned count)
2633 struct vtn_pointer *ptr;
2634 nir_intrinsic_instr *atomic;
2637 case SpvOpAtomicLoad:
2638 case SpvOpAtomicExchange:
2639 case SpvOpAtomicCompareExchange:
2640 case SpvOpAtomicCompareExchangeWeak:
2641 case SpvOpAtomicIIncrement:
2642 case SpvOpAtomicIDecrement:
2643 case SpvOpAtomicIAdd:
2644 case SpvOpAtomicISub:
2645 case SpvOpAtomicSMin:
2646 case SpvOpAtomicUMin:
2647 case SpvOpAtomicSMax:
2648 case SpvOpAtomicUMax:
2649 case SpvOpAtomicAnd:
2651 case SpvOpAtomicXor:
2652 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2655 case SpvOpAtomicStore:
2656 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2660 vtn_fail("Invalid SPIR-V atomic");
2664 SpvScope scope = w[4];
2665 SpvMemorySemanticsMask semantics = w[5];
2668 /* uniform as "atomic counter uniform" */
2669 if (ptr->mode == vtn_variable_mode_uniform) {
2670 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
2671 const struct glsl_type *deref_type = deref->type;
2672 nir_intrinsic_op op = get_uniform_nir_atomic_op(b, opcode);
2673 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2674 atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
2676 /* SSBO needs to initialize index/offset. In this case we don't need to,
2677 * as that info is already stored on the ptr->var->var nir_variable (see
2678 * vtn_create_variable)
2682 case SpvOpAtomicLoad:
2683 atomic->num_components = glsl_get_vector_elements(deref_type);
2686 case SpvOpAtomicStore:
2687 atomic->num_components = glsl_get_vector_elements(deref_type);
2688 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2691 case SpvOpAtomicExchange:
2692 case SpvOpAtomicCompareExchange:
2693 case SpvOpAtomicCompareExchangeWeak:
2694 case SpvOpAtomicIIncrement:
2695 case SpvOpAtomicIDecrement:
2696 case SpvOpAtomicIAdd:
2697 case SpvOpAtomicISub:
2698 case SpvOpAtomicSMin:
2699 case SpvOpAtomicUMin:
2700 case SpvOpAtomicSMax:
2701 case SpvOpAtomicUMax:
2702 case SpvOpAtomicAnd:
2704 case SpvOpAtomicXor:
2705 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2706 * atomic counter uniforms doesn't have sources
2711 unreachable("Invalid SPIR-V atomic");
2714 } else if (ptr->mode == vtn_variable_mode_workgroup &&
2715 !b->options->lower_workgroup_access_to_offsets) {
2716 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
2717 const struct glsl_type *deref_type = deref->type;
2718 nir_intrinsic_op op = get_deref_nir_atomic_op(b, opcode);
2719 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2720 atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
2723 case SpvOpAtomicLoad:
2724 atomic->num_components = glsl_get_vector_elements(deref_type);
2727 case SpvOpAtomicStore:
2728 atomic->num_components = glsl_get_vector_elements(deref_type);
2729 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2730 atomic->src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2733 case SpvOpAtomicExchange:
2734 case SpvOpAtomicCompareExchange:
2735 case SpvOpAtomicCompareExchangeWeak:
2736 case SpvOpAtomicIIncrement:
2737 case SpvOpAtomicIDecrement:
2738 case SpvOpAtomicIAdd:
2739 case SpvOpAtomicISub:
2740 case SpvOpAtomicSMin:
2741 case SpvOpAtomicUMin:
2742 case SpvOpAtomicSMax:
2743 case SpvOpAtomicUMax:
2744 case SpvOpAtomicAnd:
2746 case SpvOpAtomicXor:
2747 fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
2751 vtn_fail("Invalid SPIR-V atomic");
2755 nir_ssa_def *offset, *index;
2756 offset = vtn_pointer_to_offset(b, ptr, &index);
2758 nir_intrinsic_op op;
2759 if (ptr->mode == vtn_variable_mode_ssbo) {
2760 op = get_ssbo_nir_atomic_op(b, opcode);
2762 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2763 b->options->lower_workgroup_access_to_offsets);
2764 op = get_shared_nir_atomic_op(b, opcode);
2767 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2771 case SpvOpAtomicLoad:
2772 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2773 if (ptr->mode == vtn_variable_mode_ssbo)
2774 atomic->src[src++] = nir_src_for_ssa(index);
2775 atomic->src[src++] = nir_src_for_ssa(offset);
2778 case SpvOpAtomicStore:
2779 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2780 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2781 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2782 if (ptr->mode == vtn_variable_mode_ssbo)
2783 atomic->src[src++] = nir_src_for_ssa(index);
2784 atomic->src[src++] = nir_src_for_ssa(offset);
2787 case SpvOpAtomicExchange:
2788 case SpvOpAtomicCompareExchange:
2789 case SpvOpAtomicCompareExchangeWeak:
2790 case SpvOpAtomicIIncrement:
2791 case SpvOpAtomicIDecrement:
2792 case SpvOpAtomicIAdd:
2793 case SpvOpAtomicISub:
2794 case SpvOpAtomicSMin:
2795 case SpvOpAtomicUMin:
2796 case SpvOpAtomicSMax:
2797 case SpvOpAtomicUMax:
2798 case SpvOpAtomicAnd:
2800 case SpvOpAtomicXor:
2801 if (ptr->mode == vtn_variable_mode_ssbo)
2802 atomic->src[src++] = nir_src_for_ssa(index);
2803 atomic->src[src++] = nir_src_for_ssa(offset);
2804 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2808 vtn_fail("Invalid SPIR-V atomic");
2812 if (opcode != SpvOpAtomicStore) {
2813 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2815 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2816 glsl_get_vector_elements(type->type),
2817 glsl_get_bit_size(type->type), NULL);
2819 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2820 val->ssa = rzalloc(b, struct vtn_ssa_value);
2821 val->ssa->def = &atomic->dest.ssa;
2822 val->ssa->type = type->type;
2825 nir_builder_instr_insert(&b->nb, &atomic->instr);
2828 static nir_alu_instr *
2829 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2832 switch (num_components) {
2833 case 1: op = nir_op_imov; break;
2834 case 2: op = nir_op_vec2; break;
2835 case 3: op = nir_op_vec3; break;
2836 case 4: op = nir_op_vec4; break;
2837 default: vtn_fail("bad vector size");
2840 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2841 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2843 vec->dest.write_mask = (1 << num_components) - 1;
2848 struct vtn_ssa_value *
2849 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2851 if (src->transposed)
2852 return src->transposed;
2854 struct vtn_ssa_value *dest =
2855 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2857 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2858 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2859 glsl_get_bit_size(src->type));
2860 if (glsl_type_is_vector_or_scalar(src->type)) {
2861 vec->src[0].src = nir_src_for_ssa(src->def);
2862 vec->src[0].swizzle[0] = i;
2864 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2865 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2866 vec->src[j].swizzle[0] = i;
2869 nir_builder_instr_insert(&b->nb, &vec->instr);
2870 dest->elems[i]->def = &vec->dest.dest.ssa;
2873 dest->transposed = src;
2879 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2881 return nir_channel(&b->nb, src, index);
2885 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2888 nir_alu_instr *vec = create_vec(b, src->num_components,
2891 for (unsigned i = 0; i < src->num_components; i++) {
2893 vec->src[i].src = nir_src_for_ssa(insert);
2895 vec->src[i].src = nir_src_for_ssa(src);
2896 vec->src[i].swizzle[0] = i;
2900 nir_builder_instr_insert(&b->nb, &vec->instr);
2902 return &vec->dest.dest.ssa;
2906 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2909 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2910 for (unsigned i = 1; i < src->num_components; i++)
2911 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2912 vtn_vector_extract(b, src, i), dest);
2918 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2919 nir_ssa_def *insert, nir_ssa_def *index)
2921 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2922 for (unsigned i = 1; i < src->num_components; i++)
2923 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2924 vtn_vector_insert(b, src, insert, i), dest);
2929 static nir_ssa_def *
2930 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2931 nir_ssa_def *src0, nir_ssa_def *src1,
2932 const uint32_t *indices)
2934 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2936 for (unsigned i = 0; i < num_components; i++) {
2937 uint32_t index = indices[i];
2938 if (index == 0xffffffff) {
2940 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2941 } else if (index < src0->num_components) {
2942 vec->src[i].src = nir_src_for_ssa(src0);
2943 vec->src[i].swizzle[0] = index;
2945 vec->src[i].src = nir_src_for_ssa(src1);
2946 vec->src[i].swizzle[0] = index - src0->num_components;
2950 nir_builder_instr_insert(&b->nb, &vec->instr);
2952 return &vec->dest.dest.ssa;
2956 * Concatentates a number of vectors/scalars together to produce a vector
2958 static nir_ssa_def *
2959 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2960 unsigned num_srcs, nir_ssa_def **srcs)
2962 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2964 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2966 * "When constructing a vector, there must be at least two Constituent
2969 vtn_assert(num_srcs >= 2);
2971 unsigned dest_idx = 0;
2972 for (unsigned i = 0; i < num_srcs; i++) {
2973 nir_ssa_def *src = srcs[i];
2974 vtn_assert(dest_idx + src->num_components <= num_components);
2975 for (unsigned j = 0; j < src->num_components; j++) {
2976 vec->src[dest_idx].src = nir_src_for_ssa(src);
2977 vec->src[dest_idx].swizzle[0] = j;
2982 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2984 * "When constructing a vector, the total number of components in all
2985 * the operands must equal the number of components in Result Type."
2987 vtn_assert(dest_idx == num_components);
2989 nir_builder_instr_insert(&b->nb, &vec->instr);
2991 return &vec->dest.dest.ssa;
2994 static struct vtn_ssa_value *
2995 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2997 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2998 dest->type = src->type;
3000 if (glsl_type_is_vector_or_scalar(src->type)) {
3001 dest->def = src->def;
3003 unsigned elems = glsl_get_length(src->type);
3005 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
3006 for (unsigned i = 0; i < elems; i++)
3007 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
3013 static struct vtn_ssa_value *
3014 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
3015 struct vtn_ssa_value *insert, const uint32_t *indices,
3016 unsigned num_indices)
3018 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
3020 struct vtn_ssa_value *cur = dest;
3022 for (i = 0; i < num_indices - 1; i++) {
3023 cur = cur->elems[indices[i]];
3026 if (glsl_type_is_vector_or_scalar(cur->type)) {
3027 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3028 * the component granularity. In that case, the last index will be
3029 * the index to insert the scalar into the vector.
3032 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
3034 cur->elems[indices[i]] = insert;
3040 static struct vtn_ssa_value *
3041 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
3042 const uint32_t *indices, unsigned num_indices)
3044 struct vtn_ssa_value *cur = src;
3045 for (unsigned i = 0; i < num_indices; i++) {
3046 if (glsl_type_is_vector_or_scalar(cur->type)) {
3047 vtn_assert(i == num_indices - 1);
3048 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3049 * the component granularity. The last index will be the index of the
3050 * vector to extract.
3053 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
3054 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
3055 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
3058 cur = cur->elems[indices[i]];
3066 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
3067 const uint32_t *w, unsigned count)
3069 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
3070 const struct glsl_type *type =
3071 vtn_value(b, w[1], vtn_value_type_type)->type->type;
3072 val->ssa = vtn_create_ssa_value(b, type);
3075 case SpvOpVectorExtractDynamic:
3076 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
3077 vtn_ssa_value(b, w[4])->def);
3080 case SpvOpVectorInsertDynamic:
3081 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
3082 vtn_ssa_value(b, w[4])->def,
3083 vtn_ssa_value(b, w[5])->def);
3086 case SpvOpVectorShuffle:
3087 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
3088 vtn_ssa_value(b, w[3])->def,
3089 vtn_ssa_value(b, w[4])->def,
3093 case SpvOpCompositeConstruct: {
3094 unsigned elems = count - 3;
3096 if (glsl_type_is_vector_or_scalar(type)) {
3097 nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS];
3098 for (unsigned i = 0; i < elems; i++)
3099 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
3101 vtn_vector_construct(b, glsl_get_vector_elements(type),
3104 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
3105 for (unsigned i = 0; i < elems; i++)
3106 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
3110 case SpvOpCompositeExtract:
3111 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
3115 case SpvOpCompositeInsert:
3116 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
3117 vtn_ssa_value(b, w[3]),
3121 case SpvOpCopyObject:
3122 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
3126 vtn_fail("unknown composite operation");
3131 vtn_emit_barrier(struct vtn_builder *b, nir_intrinsic_op op)
3133 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
3134 nir_builder_instr_insert(&b->nb, &intrin->instr);
3138 vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
3139 SpvMemorySemanticsMask semantics)
3141 static const SpvMemorySemanticsMask all_memory_semantics =
3142 SpvMemorySemanticsUniformMemoryMask |
3143 SpvMemorySemanticsWorkgroupMemoryMask |
3144 SpvMemorySemanticsAtomicCounterMemoryMask |
3145 SpvMemorySemanticsImageMemoryMask;
3147 /* If we're not actually doing a memory barrier, bail */
3148 if (!(semantics & all_memory_semantics))
3151 /* GL and Vulkan don't have these */
3152 vtn_assert(scope != SpvScopeCrossDevice);
3154 if (scope == SpvScopeSubgroup)
3155 return; /* Nothing to do here */
3157 if (scope == SpvScopeWorkgroup) {
3158 vtn_emit_barrier(b, nir_intrinsic_group_memory_barrier);
3162 /* There's only two scopes thing left */
3163 vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice);
3165 if ((semantics & all_memory_semantics) == all_memory_semantics) {
3166 vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
3170 /* Issue a bunch of more specific barriers */
3171 uint32_t bits = semantics;
3173 SpvMemorySemanticsMask semantic = 1 << u_bit_scan(&bits);
3175 case SpvMemorySemanticsUniformMemoryMask:
3176 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer);
3178 case SpvMemorySemanticsWorkgroupMemoryMask:
3179 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared);
3181 case SpvMemorySemanticsAtomicCounterMemoryMask:
3182 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter);
3184 case SpvMemorySemanticsImageMemoryMask:
3185 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image);
3194 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
3195 const uint32_t *w, unsigned count)
3198 case SpvOpEmitVertex:
3199 case SpvOpEmitStreamVertex:
3200 case SpvOpEndPrimitive:
3201 case SpvOpEndStreamPrimitive: {
3202 nir_intrinsic_op intrinsic_op;
3204 case SpvOpEmitVertex:
3205 case SpvOpEmitStreamVertex:
3206 intrinsic_op = nir_intrinsic_emit_vertex;
3208 case SpvOpEndPrimitive:
3209 case SpvOpEndStreamPrimitive:
3210 intrinsic_op = nir_intrinsic_end_primitive;
3213 unreachable("Invalid opcode");
3216 nir_intrinsic_instr *intrin =
3217 nir_intrinsic_instr_create(b->shader, intrinsic_op);
3220 case SpvOpEmitStreamVertex:
3221 case SpvOpEndStreamPrimitive: {
3222 unsigned stream = vtn_constant_value(b, w[1])->values[0].u32[0];
3223 nir_intrinsic_set_stream_id(intrin, stream);
3231 nir_builder_instr_insert(&b->nb, &intrin->instr);
3235 case SpvOpMemoryBarrier: {
3236 SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
3237 SpvMemorySemanticsMask semantics =
3238 vtn_constant_value(b, w[2])->values[0].u32[0];
3239 vtn_emit_memory_barrier(b, scope, semantics);
3243 case SpvOpControlBarrier: {
3244 SpvScope execution_scope =
3245 vtn_constant_value(b, w[1])->values[0].u32[0];
3246 if (execution_scope == SpvScopeWorkgroup)
3247 vtn_emit_barrier(b, nir_intrinsic_barrier);
3249 SpvScope memory_scope =
3250 vtn_constant_value(b, w[2])->values[0].u32[0];
3251 SpvMemorySemanticsMask memory_semantics =
3252 vtn_constant_value(b, w[3])->values[0].u32[0];
3253 vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
3258 unreachable("unknown barrier instruction");
3263 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
3264 SpvExecutionMode mode)
3267 case SpvExecutionModeInputPoints:
3268 case SpvExecutionModeOutputPoints:
3269 return 0; /* GL_POINTS */
3270 case SpvExecutionModeInputLines:
3271 return 1; /* GL_LINES */
3272 case SpvExecutionModeInputLinesAdjacency:
3273 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3274 case SpvExecutionModeTriangles:
3275 return 4; /* GL_TRIANGLES */
3276 case SpvExecutionModeInputTrianglesAdjacency:
3277 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3278 case SpvExecutionModeQuads:
3279 return 7; /* GL_QUADS */
3280 case SpvExecutionModeIsolines:
3281 return 0x8E7A; /* GL_ISOLINES */
3282 case SpvExecutionModeOutputLineStrip:
3283 return 3; /* GL_LINE_STRIP */
3284 case SpvExecutionModeOutputTriangleStrip:
3285 return 5; /* GL_TRIANGLE_STRIP */
3287 vtn_fail("Invalid primitive type");
3292 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
3293 SpvExecutionMode mode)
3296 case SpvExecutionModeInputPoints:
3298 case SpvExecutionModeInputLines:
3300 case SpvExecutionModeInputLinesAdjacency:
3302 case SpvExecutionModeTriangles:
3304 case SpvExecutionModeInputTrianglesAdjacency:
3307 vtn_fail("Invalid GS input mode");
3311 static gl_shader_stage
3312 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
3315 case SpvExecutionModelVertex:
3316 return MESA_SHADER_VERTEX;
3317 case SpvExecutionModelTessellationControl:
3318 return MESA_SHADER_TESS_CTRL;
3319 case SpvExecutionModelTessellationEvaluation:
3320 return MESA_SHADER_TESS_EVAL;
3321 case SpvExecutionModelGeometry:
3322 return MESA_SHADER_GEOMETRY;
3323 case SpvExecutionModelFragment:
3324 return MESA_SHADER_FRAGMENT;
3325 case SpvExecutionModelGLCompute:
3326 return MESA_SHADER_COMPUTE;
3328 vtn_fail("Unsupported execution model");
3332 #define spv_check_supported(name, cap) do { \
3333 if (!(b->options && b->options->caps.name)) \
3334 vtn_warn("Unsupported SPIR-V capability: %s", \
3335 spirv_capability_to_string(cap)); \
3340 vtn_handle_entry_point(struct vtn_builder *b, const uint32_t *w,
3343 struct vtn_value *entry_point = &b->values[w[2]];
3344 /* Let this be a name label regardless */
3345 unsigned name_words;
3346 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3348 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3349 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3352 vtn_assert(b->entry_point == NULL);
3353 b->entry_point = entry_point;
3357 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
3358 const uint32_t *w, unsigned count)
3365 case SpvSourceLanguageUnknown: lang = "unknown"; break;
3366 case SpvSourceLanguageESSL: lang = "ESSL"; break;
3367 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
3368 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
3369 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
3370 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
3373 uint32_t version = w[2];
3376 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
3378 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
3382 case SpvOpSourceExtension:
3383 case SpvOpSourceContinued:
3384 case SpvOpExtension:
3385 case SpvOpModuleProcessed:
3386 /* Unhandled, but these are for debug so that's ok. */
3389 case SpvOpCapability: {
3390 SpvCapability cap = w[1];
3392 case SpvCapabilityMatrix:
3393 case SpvCapabilityShader:
3394 case SpvCapabilityGeometry:
3395 case SpvCapabilityGeometryPointSize:
3396 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3397 case SpvCapabilitySampledImageArrayDynamicIndexing:
3398 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3399 case SpvCapabilityStorageImageArrayDynamicIndexing:
3400 case SpvCapabilityImageRect:
3401 case SpvCapabilitySampledRect:
3402 case SpvCapabilitySampled1D:
3403 case SpvCapabilityImage1D:
3404 case SpvCapabilitySampledCubeArray:
3405 case SpvCapabilityImageCubeArray:
3406 case SpvCapabilitySampledBuffer:
3407 case SpvCapabilityImageBuffer:
3408 case SpvCapabilityImageQuery:
3409 case SpvCapabilityDerivativeControl:
3410 case SpvCapabilityInterpolationFunction:
3411 case SpvCapabilityMultiViewport:
3412 case SpvCapabilitySampleRateShading:
3413 case SpvCapabilityClipDistance:
3414 case SpvCapabilityCullDistance:
3415 case SpvCapabilityInputAttachment:
3416 case SpvCapabilityImageGatherExtended:
3417 case SpvCapabilityStorageImageExtendedFormats:
3420 case SpvCapabilityLinkage:
3421 case SpvCapabilityVector16:
3422 case SpvCapabilityFloat16Buffer:
3423 case SpvCapabilityFloat16:
3424 case SpvCapabilityInt64Atomics:
3425 case SpvCapabilityStorageImageMultisample:
3426 case SpvCapabilityInt8:
3427 case SpvCapabilitySparseResidency:
3428 case SpvCapabilityMinLod:
3429 vtn_warn("Unsupported SPIR-V capability: %s",
3430 spirv_capability_to_string(cap));
3433 case SpvCapabilityAtomicStorage:
3434 spv_check_supported(atomic_storage, cap);
3437 case SpvCapabilityFloat64:
3438 spv_check_supported(float64, cap);
3440 case SpvCapabilityInt64:
3441 spv_check_supported(int64, cap);
3443 case SpvCapabilityInt16:
3444 spv_check_supported(int16, cap);
3447 case SpvCapabilityTransformFeedback:
3448 spv_check_supported(transform_feedback, cap);
3451 case SpvCapabilityGeometryStreams:
3452 spv_check_supported(geometry_streams, cap);
3455 case SpvCapabilityAddresses:
3456 case SpvCapabilityKernel:
3457 case SpvCapabilityImageBasic:
3458 case SpvCapabilityImageReadWrite:
3459 case SpvCapabilityImageMipmap:
3460 case SpvCapabilityPipes:
3461 case SpvCapabilityGroups:
3462 case SpvCapabilityDeviceEnqueue:
3463 case SpvCapabilityLiteralSampler:
3464 case SpvCapabilityGenericPointer:
3465 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3466 spirv_capability_to_string(cap));
3469 case SpvCapabilityImageMSArray:
3470 spv_check_supported(image_ms_array, cap);
3473 case SpvCapabilityTessellation:
3474 case SpvCapabilityTessellationPointSize:
3475 spv_check_supported(tessellation, cap);
3478 case SpvCapabilityDrawParameters:
3479 spv_check_supported(draw_parameters, cap);
3482 case SpvCapabilityStorageImageReadWithoutFormat:
3483 spv_check_supported(image_read_without_format, cap);
3486 case SpvCapabilityStorageImageWriteWithoutFormat:
3487 spv_check_supported(image_write_without_format, cap);
3490 case SpvCapabilityDeviceGroup:
3491 spv_check_supported(device_group, cap);
3494 case SpvCapabilityMultiView:
3495 spv_check_supported(multiview, cap);
3498 case SpvCapabilityGroupNonUniform:
3499 spv_check_supported(subgroup_basic, cap);
3502 case SpvCapabilityGroupNonUniformVote:
3503 spv_check_supported(subgroup_vote, cap);
3506 case SpvCapabilitySubgroupBallotKHR:
3507 case SpvCapabilityGroupNonUniformBallot:
3508 spv_check_supported(subgroup_ballot, cap);
3511 case SpvCapabilityGroupNonUniformShuffle:
3512 case SpvCapabilityGroupNonUniformShuffleRelative:
3513 spv_check_supported(subgroup_shuffle, cap);
3516 case SpvCapabilityGroupNonUniformQuad:
3517 spv_check_supported(subgroup_quad, cap);
3520 case SpvCapabilityGroupNonUniformArithmetic:
3521 case SpvCapabilityGroupNonUniformClustered:
3522 spv_check_supported(subgroup_arithmetic, cap);
3525 case SpvCapabilityVariablePointersStorageBuffer:
3526 case SpvCapabilityVariablePointers:
3527 spv_check_supported(variable_pointers, cap);
3530 case SpvCapabilityStorageUniformBufferBlock16:
3531 case SpvCapabilityStorageUniform16:
3532 case SpvCapabilityStoragePushConstant16:
3533 case SpvCapabilityStorageInputOutput16:
3534 spv_check_supported(storage_16bit, cap);
3537 case SpvCapabilityShaderViewportIndexLayerEXT:
3538 spv_check_supported(shader_viewport_index_layer, cap);
3541 case SpvCapabilityStorageBuffer8BitAccess:
3542 case SpvCapabilityUniformAndStorageBuffer8BitAccess:
3543 case SpvCapabilityStoragePushConstant8:
3544 spv_check_supported(storage_8bit, cap);
3547 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT:
3548 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT:
3549 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT:
3550 spv_check_supported(descriptor_array_dynamic_indexing, cap);
3553 case SpvCapabilityRuntimeDescriptorArrayEXT:
3554 spv_check_supported(runtime_descriptor_array, cap);
3557 case SpvCapabilityStencilExportEXT:
3558 spv_check_supported(stencil_export, cap);
3561 case SpvCapabilitySampleMaskPostDepthCoverage:
3562 spv_check_supported(post_depth_coverage, cap);
3566 vtn_fail("Unhandled capability");
3571 case SpvOpExtInstImport:
3572 vtn_handle_extension(b, opcode, w, count);
3575 case SpvOpMemoryModel:
3576 vtn_assert(w[1] == SpvAddressingModelLogical);
3577 vtn_assert(w[2] == SpvMemoryModelSimple ||
3578 w[2] == SpvMemoryModelGLSL450);
3581 case SpvOpEntryPoint:
3582 vtn_handle_entry_point(b, w, count);
3586 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3587 vtn_string_literal(b, &w[2], count - 2, NULL);
3591 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3594 case SpvOpMemberName:
3598 case SpvOpExecutionMode:
3599 case SpvOpDecorationGroup:
3601 case SpvOpMemberDecorate:
3602 case SpvOpGroupDecorate:
3603 case SpvOpGroupMemberDecorate:
3604 vtn_handle_decoration(b, opcode, w, count);
3608 return false; /* End of preamble */
3615 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3616 const struct vtn_decoration *mode, void *data)
3618 vtn_assert(b->entry_point == entry_point);
3620 switch(mode->exec_mode) {
3621 case SpvExecutionModeOriginUpperLeft:
3622 case SpvExecutionModeOriginLowerLeft:
3623 b->origin_upper_left =
3624 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3627 case SpvExecutionModeEarlyFragmentTests:
3628 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3629 b->shader->info.fs.early_fragment_tests = true;
3632 case SpvExecutionModePostDepthCoverage:
3633 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3634 b->shader->info.fs.post_depth_coverage = true;
3637 case SpvExecutionModeInvocations:
3638 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3639 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3642 case SpvExecutionModeDepthReplacing:
3643 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3644 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3646 case SpvExecutionModeDepthGreater:
3647 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3648 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3650 case SpvExecutionModeDepthLess:
3651 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3652 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3654 case SpvExecutionModeDepthUnchanged:
3655 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3656 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3659 case SpvExecutionModeLocalSize:
3660 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3661 b->shader->info.cs.local_size[0] = mode->literals[0];
3662 b->shader->info.cs.local_size[1] = mode->literals[1];
3663 b->shader->info.cs.local_size[2] = mode->literals[2];
3665 case SpvExecutionModeLocalSizeHint:
3666 break; /* Nothing to do with this */
3668 case SpvExecutionModeOutputVertices:
3669 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3670 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3671 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3673 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3674 b->shader->info.gs.vertices_out = mode->literals[0];
3678 case SpvExecutionModeInputPoints:
3679 case SpvExecutionModeInputLines:
3680 case SpvExecutionModeInputLinesAdjacency:
3681 case SpvExecutionModeTriangles:
3682 case SpvExecutionModeInputTrianglesAdjacency:
3683 case SpvExecutionModeQuads:
3684 case SpvExecutionModeIsolines:
3685 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3686 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3687 b->shader->info.tess.primitive_mode =
3688 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3690 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3691 b->shader->info.gs.vertices_in =
3692 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3696 case SpvExecutionModeOutputPoints:
3697 case SpvExecutionModeOutputLineStrip:
3698 case SpvExecutionModeOutputTriangleStrip:
3699 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3700 b->shader->info.gs.output_primitive =
3701 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3704 case SpvExecutionModeSpacingEqual:
3705 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3706 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3707 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3709 case SpvExecutionModeSpacingFractionalEven:
3710 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3711 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3712 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3714 case SpvExecutionModeSpacingFractionalOdd:
3715 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3716 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3717 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3719 case SpvExecutionModeVertexOrderCw:
3720 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3721 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3722 b->shader->info.tess.ccw = false;
3724 case SpvExecutionModeVertexOrderCcw:
3725 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3726 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3727 b->shader->info.tess.ccw = true;
3729 case SpvExecutionModePointMode:
3730 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3731 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3732 b->shader->info.tess.point_mode = true;
3735 case SpvExecutionModePixelCenterInteger:
3736 b->pixel_center_integer = true;
3739 case SpvExecutionModeXfb:
3740 b->shader->info.has_transform_feedback_varyings = true;
3743 case SpvExecutionModeVecTypeHint:
3744 case SpvExecutionModeContractionOff:
3747 case SpvExecutionModeStencilRefReplacingEXT:
3748 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3752 vtn_fail("Unhandled execution mode");
3757 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3758 const uint32_t *w, unsigned count)
3760 vtn_set_instruction_result_type(b, opcode, w, count);
3764 case SpvOpSourceContinued:
3765 case SpvOpSourceExtension:
3766 case SpvOpExtension:
3767 case SpvOpCapability:
3768 case SpvOpExtInstImport:
3769 case SpvOpMemoryModel:
3770 case SpvOpEntryPoint:
3771 case SpvOpExecutionMode:
3774 case SpvOpMemberName:
3775 case SpvOpDecorationGroup:
3777 case SpvOpMemberDecorate:
3778 case SpvOpGroupDecorate:
3779 case SpvOpGroupMemberDecorate:
3780 vtn_fail("Invalid opcode types and variables section");
3786 case SpvOpTypeFloat:
3787 case SpvOpTypeVector:
3788 case SpvOpTypeMatrix:
3789 case SpvOpTypeImage:
3790 case SpvOpTypeSampler:
3791 case SpvOpTypeSampledImage:
3792 case SpvOpTypeArray:
3793 case SpvOpTypeRuntimeArray:
3794 case SpvOpTypeStruct:
3795 case SpvOpTypeOpaque:
3796 case SpvOpTypePointer:
3797 case SpvOpTypeFunction:
3798 case SpvOpTypeEvent:
3799 case SpvOpTypeDeviceEvent:
3800 case SpvOpTypeReserveId:
3801 case SpvOpTypeQueue:
3803 vtn_handle_type(b, opcode, w, count);
3806 case SpvOpConstantTrue:
3807 case SpvOpConstantFalse:
3809 case SpvOpConstantComposite:
3810 case SpvOpConstantSampler:
3811 case SpvOpConstantNull:
3812 case SpvOpSpecConstantTrue:
3813 case SpvOpSpecConstantFalse:
3814 case SpvOpSpecConstant:
3815 case SpvOpSpecConstantComposite:
3816 case SpvOpSpecConstantOp:
3817 vtn_handle_constant(b, opcode, w, count);
3822 vtn_handle_variables(b, opcode, w, count);
3826 return false; /* End of preamble */
3833 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3834 const uint32_t *w, unsigned count)
3840 case SpvOpLoopMerge:
3841 case SpvOpSelectionMerge:
3842 /* This is handled by cfg pre-pass and walk_blocks */
3846 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3847 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3852 vtn_handle_extension(b, opcode, w, count);
3858 case SpvOpCopyMemory:
3859 case SpvOpCopyMemorySized:
3860 case SpvOpAccessChain:
3861 case SpvOpPtrAccessChain:
3862 case SpvOpInBoundsAccessChain:
3863 case SpvOpArrayLength:
3864 vtn_handle_variables(b, opcode, w, count);
3867 case SpvOpFunctionCall:
3868 vtn_handle_function_call(b, opcode, w, count);
3871 case SpvOpSampledImage:
3873 case SpvOpImageSampleImplicitLod:
3874 case SpvOpImageSampleExplicitLod:
3875 case SpvOpImageSampleDrefImplicitLod:
3876 case SpvOpImageSampleDrefExplicitLod:
3877 case SpvOpImageSampleProjImplicitLod:
3878 case SpvOpImageSampleProjExplicitLod:
3879 case SpvOpImageSampleProjDrefImplicitLod:
3880 case SpvOpImageSampleProjDrefExplicitLod:
3881 case SpvOpImageFetch:
3882 case SpvOpImageGather:
3883 case SpvOpImageDrefGather:
3884 case SpvOpImageQuerySizeLod:
3885 case SpvOpImageQueryLod:
3886 case SpvOpImageQueryLevels:
3887 case SpvOpImageQuerySamples:
3888 vtn_handle_texture(b, opcode, w, count);
3891 case SpvOpImageRead:
3892 case SpvOpImageWrite:
3893 case SpvOpImageTexelPointer:
3894 vtn_handle_image(b, opcode, w, count);
3897 case SpvOpImageQuerySize: {
3898 struct vtn_pointer *image =
3899 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3900 if (glsl_type_is_image(image->type->type)) {
3901 vtn_handle_image(b, opcode, w, count);
3903 vtn_assert(glsl_type_is_sampler(image->type->type));
3904 vtn_handle_texture(b, opcode, w, count);
3909 case SpvOpAtomicLoad:
3910 case SpvOpAtomicExchange:
3911 case SpvOpAtomicCompareExchange:
3912 case SpvOpAtomicCompareExchangeWeak:
3913 case SpvOpAtomicIIncrement:
3914 case SpvOpAtomicIDecrement:
3915 case SpvOpAtomicIAdd:
3916 case SpvOpAtomicISub:
3917 case SpvOpAtomicSMin:
3918 case SpvOpAtomicUMin:
3919 case SpvOpAtomicSMax:
3920 case SpvOpAtomicUMax:
3921 case SpvOpAtomicAnd:
3923 case SpvOpAtomicXor: {
3924 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3925 if (pointer->value_type == vtn_value_type_image_pointer) {
3926 vtn_handle_image(b, opcode, w, count);
3928 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3929 vtn_handle_atomics(b, opcode, w, count);
3934 case SpvOpAtomicStore: {
3935 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3936 if (pointer->value_type == vtn_value_type_image_pointer) {
3937 vtn_handle_image(b, opcode, w, count);
3939 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3940 vtn_handle_atomics(b, opcode, w, count);
3946 /* Handle OpSelect up-front here because it needs to be able to handle
3947 * pointers and not just regular vectors and scalars.
3949 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3950 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3951 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3952 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3954 const struct glsl_type *sel_type;
3955 switch (res_val->type->base_type) {
3956 case vtn_base_type_scalar:
3957 sel_type = glsl_bool_type();
3959 case vtn_base_type_vector:
3960 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3962 case vtn_base_type_pointer:
3963 /* We need to have actual storage for pointer types */
3964 vtn_fail_if(res_val->type->type == NULL,
3965 "Invalid pointer result type for OpSelect");
3966 sel_type = glsl_bool_type();
3969 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3972 if (unlikely(sel_val->type->type != sel_type)) {
3973 if (sel_val->type->type == glsl_bool_type()) {
3974 /* This case is illegal but some older versions of GLSLang produce
3975 * it. The GLSLang issue was fixed on March 30, 2017:
3977 * https://github.com/KhronosGroup/glslang/issues/809
3979 * Unfortunately, there are applications in the wild which are
3980 * shipping with this bug so it isn't nice to fail on them so we
3981 * throw a warning instead. It's not actually a problem for us as
3982 * nir_builder will just splat the condition out which is most
3983 * likely what the client wanted anyway.
3985 vtn_warn("Condition type of OpSelect must have the same number "
3986 "of components as Result Type");
3988 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3989 "of Boolean type. It must have the same number of "
3990 "components as Result Type");
3994 vtn_fail_if(obj1_val->type != res_val->type ||
3995 obj2_val->type != res_val->type,
3996 "Object types must match the result type in OpSelect");
3998 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3999 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
4000 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
4001 vtn_ssa_value(b, w[4])->def,
4002 vtn_ssa_value(b, w[5])->def);
4003 vtn_push_ssa(b, w[2], res_type, ssa);
4012 case SpvOpConvertFToU:
4013 case SpvOpConvertFToS:
4014 case SpvOpConvertSToF:
4015 case SpvOpConvertUToF:
4019 case SpvOpQuantizeToF16:
4020 case SpvOpConvertPtrToU:
4021 case SpvOpConvertUToPtr:
4022 case SpvOpPtrCastToGeneric:
4023 case SpvOpGenericCastToPtr:
4029 case SpvOpSignBitSet:
4030 case SpvOpLessOrGreater:
4032 case SpvOpUnordered:
4047 case SpvOpVectorTimesScalar:
4049 case SpvOpIAddCarry:
4050 case SpvOpISubBorrow:
4051 case SpvOpUMulExtended:
4052 case SpvOpSMulExtended:
4053 case SpvOpShiftRightLogical:
4054 case SpvOpShiftRightArithmetic:
4055 case SpvOpShiftLeftLogical:
4056 case SpvOpLogicalEqual:
4057 case SpvOpLogicalNotEqual:
4058 case SpvOpLogicalOr:
4059 case SpvOpLogicalAnd:
4060 case SpvOpLogicalNot:
4061 case SpvOpBitwiseOr:
4062 case SpvOpBitwiseXor:
4063 case SpvOpBitwiseAnd:
4065 case SpvOpFOrdEqual:
4066 case SpvOpFUnordEqual:
4067 case SpvOpINotEqual:
4068 case SpvOpFOrdNotEqual:
4069 case SpvOpFUnordNotEqual:
4070 case SpvOpULessThan:
4071 case SpvOpSLessThan:
4072 case SpvOpFOrdLessThan:
4073 case SpvOpFUnordLessThan:
4074 case SpvOpUGreaterThan:
4075 case SpvOpSGreaterThan:
4076 case SpvOpFOrdGreaterThan:
4077 case SpvOpFUnordGreaterThan:
4078 case SpvOpULessThanEqual:
4079 case SpvOpSLessThanEqual:
4080 case SpvOpFOrdLessThanEqual:
4081 case SpvOpFUnordLessThanEqual:
4082 case SpvOpUGreaterThanEqual:
4083 case SpvOpSGreaterThanEqual:
4084 case SpvOpFOrdGreaterThanEqual:
4085 case SpvOpFUnordGreaterThanEqual:
4091 case SpvOpFwidthFine:
4092 case SpvOpDPdxCoarse:
4093 case SpvOpDPdyCoarse:
4094 case SpvOpFwidthCoarse:
4095 case SpvOpBitFieldInsert:
4096 case SpvOpBitFieldSExtract:
4097 case SpvOpBitFieldUExtract:
4098 case SpvOpBitReverse:
4100 case SpvOpTranspose:
4101 case SpvOpOuterProduct:
4102 case SpvOpMatrixTimesScalar:
4103 case SpvOpVectorTimesMatrix:
4104 case SpvOpMatrixTimesVector:
4105 case SpvOpMatrixTimesMatrix:
4106 vtn_handle_alu(b, opcode, w, count);
4109 case SpvOpVectorExtractDynamic:
4110 case SpvOpVectorInsertDynamic:
4111 case SpvOpVectorShuffle:
4112 case SpvOpCompositeConstruct:
4113 case SpvOpCompositeExtract:
4114 case SpvOpCompositeInsert:
4115 case SpvOpCopyObject:
4116 vtn_handle_composite(b, opcode, w, count);
4119 case SpvOpEmitVertex:
4120 case SpvOpEndPrimitive:
4121 case SpvOpEmitStreamVertex:
4122 case SpvOpEndStreamPrimitive:
4123 case SpvOpControlBarrier:
4124 case SpvOpMemoryBarrier:
4125 vtn_handle_barrier(b, opcode, w, count);
4128 case SpvOpGroupNonUniformElect:
4129 case SpvOpGroupNonUniformAll:
4130 case SpvOpGroupNonUniformAny:
4131 case SpvOpGroupNonUniformAllEqual:
4132 case SpvOpGroupNonUniformBroadcast:
4133 case SpvOpGroupNonUniformBroadcastFirst:
4134 case SpvOpGroupNonUniformBallot:
4135 case SpvOpGroupNonUniformInverseBallot:
4136 case SpvOpGroupNonUniformBallotBitExtract:
4137 case SpvOpGroupNonUniformBallotBitCount:
4138 case SpvOpGroupNonUniformBallotFindLSB:
4139 case SpvOpGroupNonUniformBallotFindMSB:
4140 case SpvOpGroupNonUniformShuffle:
4141 case SpvOpGroupNonUniformShuffleXor:
4142 case SpvOpGroupNonUniformShuffleUp:
4143 case SpvOpGroupNonUniformShuffleDown:
4144 case SpvOpGroupNonUniformIAdd:
4145 case SpvOpGroupNonUniformFAdd:
4146 case SpvOpGroupNonUniformIMul:
4147 case SpvOpGroupNonUniformFMul:
4148 case SpvOpGroupNonUniformSMin:
4149 case SpvOpGroupNonUniformUMin:
4150 case SpvOpGroupNonUniformFMin:
4151 case SpvOpGroupNonUniformSMax:
4152 case SpvOpGroupNonUniformUMax:
4153 case SpvOpGroupNonUniformFMax:
4154 case SpvOpGroupNonUniformBitwiseAnd:
4155 case SpvOpGroupNonUniformBitwiseOr:
4156 case SpvOpGroupNonUniformBitwiseXor:
4157 case SpvOpGroupNonUniformLogicalAnd:
4158 case SpvOpGroupNonUniformLogicalOr:
4159 case SpvOpGroupNonUniformLogicalXor:
4160 case SpvOpGroupNonUniformQuadBroadcast:
4161 case SpvOpGroupNonUniformQuadSwap:
4162 vtn_handle_subgroup(b, opcode, w, count);
4166 vtn_fail("Unhandled opcode");
4173 vtn_create_builder(const uint32_t *words, size_t word_count,
4174 gl_shader_stage stage, const char *entry_point_name,
4175 const struct spirv_to_nir_options *options)
4177 /* Initialize the vtn_builder object */
4178 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
4180 b->spirv_word_count = word_count;
4184 exec_list_make_empty(&b->functions);
4185 b->entry_point_stage = stage;
4186 b->entry_point_name = entry_point_name;
4187 b->options = options;
4190 * Handle the SPIR-V header (first 5 dwords).
4191 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4193 if (word_count <= 5)
4196 if (words[0] != SpvMagicNumber) {
4197 vtn_err("words[0] was 0x%x, want 0x%x", words[0], SpvMagicNumber);
4200 if (words[1] < 0x10000) {
4201 vtn_err("words[1] was 0x%x, want >= 0x10000", words[1]);
4205 /* words[2] == generator magic */
4206 unsigned value_id_bound = words[3];
4207 if (words[4] != 0) {
4208 vtn_err("words[4] was %u, want 0", words[4]);
4212 b->value_id_bound = value_id_bound;
4213 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
4222 spirv_to_nir(const uint32_t *words, size_t word_count,
4223 struct nir_spirv_specialization *spec, unsigned num_spec,
4224 gl_shader_stage stage, const char *entry_point_name,
4225 const struct spirv_to_nir_options *options,
4226 const nir_shader_compiler_options *nir_options)
4229 const uint32_t *word_end = words + word_count;
4231 struct vtn_builder *b = vtn_create_builder(words, word_count,
4232 stage, entry_point_name,
4238 /* See also _vtn_fail() */
4239 if (setjmp(b->fail_jump)) {
4244 /* Skip the SPIR-V header, handled at vtn_create_builder */
4247 /* Handle all the preamble instructions */
4248 words = vtn_foreach_instruction(b, words, word_end,
4249 vtn_handle_preamble_instruction);
4251 if (b->entry_point == NULL) {
4252 vtn_fail("Entry point not found");
4257 b->shader = nir_shader_create(b, stage, nir_options, NULL);
4259 /* Set shader info defaults */
4260 b->shader->info.gs.invocations = 1;
4262 /* Parse execution modes */
4263 vtn_foreach_execution_mode(b, b->entry_point,
4264 vtn_handle_execution_mode, NULL);
4266 b->specializations = spec;
4267 b->num_specializations = num_spec;
4269 /* Handle all variable, type, and constant instructions */
4270 words = vtn_foreach_instruction(b, words, word_end,
4271 vtn_handle_variable_or_type_instruction);
4273 /* Set types on all vtn_values */
4274 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
4276 vtn_build_cfg(b, words, word_end);
4278 assert(b->entry_point->value_type == vtn_value_type_function);
4279 b->entry_point->func->referenced = true;
4284 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
4285 if (func->referenced && !func->emitted) {
4286 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
4287 _mesa_key_pointer_equal);
4289 vtn_function_emit(b, func, vtn_handle_body_instruction);
4295 /* We sometimes generate bogus derefs that, while never used, give the
4296 * validator a bit of heartburn. Run dead code to get rid of them.
4298 nir_opt_dce(b->shader);
4300 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
4301 nir_function *entry_point = b->entry_point->func->impl->function;
4302 vtn_assert(entry_point);
4304 /* Unparent the shader from the vtn_builder before we delete the builder */
4305 ralloc_steal(NULL, b->shader);