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)
387 case SpvOpExtInstImport: {
388 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
389 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
390 val->ext_handler = vtn_handle_glsl450_instruction;
391 } else if ((strcmp((const char *)&w[2], "SPV_AMD_gcn_shader") == 0)
392 && (b->options && b->options->caps.gcn_shader)) {
393 val->ext_handler = vtn_handle_amd_gcn_shader_instruction;
394 } else if ((strcmp((const char *)&w[2], "SPV_AMD_shader_trinary_minmax") == 0)
395 && (b->options && b->options->caps.trinary_minmax)) {
396 val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
398 vtn_fail("Unsupported extension");
404 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
405 bool handled = val->ext_handler(b, w[4], w, count);
411 vtn_fail("Unhandled opcode");
416 _foreach_decoration_helper(struct vtn_builder *b,
417 struct vtn_value *base_value,
419 struct vtn_value *value,
420 vtn_decoration_foreach_cb cb, void *data)
422 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
424 if (dec->scope == VTN_DEC_DECORATION) {
425 member = parent_member;
426 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
427 vtn_fail_if(value->value_type != vtn_value_type_type ||
428 value->type->base_type != vtn_base_type_struct,
429 "OpMemberDecorate and OpGroupMemberDecorate are only "
430 "allowed on OpTypeStruct");
431 /* This means we haven't recursed yet */
432 assert(value == base_value);
434 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
436 vtn_fail_if(member >= base_value->type->length,
437 "OpMemberDecorate specifies member %d but the "
438 "OpTypeStruct has only %u members",
439 member, base_value->type->length);
441 /* Not a decoration */
442 assert(dec->scope == VTN_DEC_EXECUTION_MODE);
447 assert(dec->group->value_type == vtn_value_type_decoration_group);
448 _foreach_decoration_helper(b, base_value, member, dec->group,
451 cb(b, base_value, member, dec, data);
456 /** Iterates (recursively if needed) over all of the decorations on a value
458 * This function iterates over all of the decorations applied to a given
459 * value. If it encounters a decoration group, it recurses into the group
460 * and iterates over all of those decorations as well.
463 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
464 vtn_decoration_foreach_cb cb, void *data)
466 _foreach_decoration_helper(b, value, -1, value, cb, data);
470 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
471 vtn_execution_mode_foreach_cb cb, void *data)
473 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
474 if (dec->scope != VTN_DEC_EXECUTION_MODE)
477 assert(dec->group == NULL);
478 cb(b, value, dec, data);
483 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
484 const uint32_t *w, unsigned count)
486 const uint32_t *w_end = w + count;
487 const uint32_t target = w[1];
491 case SpvOpDecorationGroup:
492 vtn_push_value(b, target, vtn_value_type_decoration_group);
496 case SpvOpMemberDecorate:
497 case SpvOpExecutionMode: {
498 struct vtn_value *val = vtn_untyped_value(b, target);
500 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
503 dec->scope = VTN_DEC_DECORATION;
505 case SpvOpMemberDecorate:
506 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
507 vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
508 "Member argument of OpMemberDecorate too large");
510 case SpvOpExecutionMode:
511 dec->scope = VTN_DEC_EXECUTION_MODE;
514 unreachable("Invalid decoration opcode");
516 dec->decoration = *(w++);
519 /* Link into the list */
520 dec->next = val->decoration;
521 val->decoration = dec;
525 case SpvOpGroupMemberDecorate:
526 case SpvOpGroupDecorate: {
527 struct vtn_value *group =
528 vtn_value(b, target, vtn_value_type_decoration_group);
530 for (; w < w_end; w++) {
531 struct vtn_value *val = vtn_untyped_value(b, *w);
532 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
535 if (opcode == SpvOpGroupDecorate) {
536 dec->scope = VTN_DEC_DECORATION;
538 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
539 vtn_fail_if(dec->scope < 0, /* Check for overflow */
540 "Member argument of OpGroupMemberDecorate too large");
543 /* Link into the list */
544 dec->next = val->decoration;
545 val->decoration = dec;
551 unreachable("Unhandled opcode");
555 struct member_decoration_ctx {
557 struct glsl_struct_field *fields;
558 struct vtn_type *type;
561 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
562 * OpStore, or OpCopyMemory between them without breaking anything.
563 * Technically, the SPIR-V rules require the exact same type ID but this lets
564 * us internally be a bit looser.
567 vtn_types_compatible(struct vtn_builder *b,
568 struct vtn_type *t1, struct vtn_type *t2)
570 if (t1->id == t2->id)
573 if (t1->base_type != t2->base_type)
576 switch (t1->base_type) {
577 case vtn_base_type_void:
578 case vtn_base_type_scalar:
579 case vtn_base_type_vector:
580 case vtn_base_type_matrix:
581 case vtn_base_type_image:
582 case vtn_base_type_sampler:
583 case vtn_base_type_sampled_image:
584 return t1->type == t2->type;
586 case vtn_base_type_array:
587 return t1->length == t2->length &&
588 vtn_types_compatible(b, t1->array_element, t2->array_element);
590 case vtn_base_type_pointer:
591 return vtn_types_compatible(b, t1->deref, t2->deref);
593 case vtn_base_type_struct:
594 if (t1->length != t2->length)
597 for (unsigned i = 0; i < t1->length; i++) {
598 if (!vtn_types_compatible(b, t1->members[i], t2->members[i]))
603 case vtn_base_type_function:
604 /* This case shouldn't get hit since you can't copy around function
605 * types. Just require them to be identical.
610 vtn_fail("Invalid base type");
613 /* does a shallow copy of a vtn_type */
615 static struct vtn_type *
616 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
618 struct vtn_type *dest = ralloc(b, struct vtn_type);
621 switch (src->base_type) {
622 case vtn_base_type_void:
623 case vtn_base_type_scalar:
624 case vtn_base_type_vector:
625 case vtn_base_type_matrix:
626 case vtn_base_type_array:
627 case vtn_base_type_pointer:
628 case vtn_base_type_image:
629 case vtn_base_type_sampler:
630 case vtn_base_type_sampled_image:
631 /* Nothing more to do */
634 case vtn_base_type_struct:
635 dest->members = ralloc_array(b, struct vtn_type *, src->length);
636 memcpy(dest->members, src->members,
637 src->length * sizeof(src->members[0]));
639 dest->offsets = ralloc_array(b, unsigned, src->length);
640 memcpy(dest->offsets, src->offsets,
641 src->length * sizeof(src->offsets[0]));
644 case vtn_base_type_function:
645 dest->params = ralloc_array(b, struct vtn_type *, src->length);
646 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
653 static struct vtn_type *
654 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
656 type->members[member] = vtn_type_copy(b, type->members[member]);
657 type = type->members[member];
659 /* We may have an array of matrices.... Oh, joy! */
660 while (glsl_type_is_array(type->type)) {
661 type->array_element = vtn_type_copy(b, type->array_element);
662 type = type->array_element;
665 vtn_assert(glsl_type_is_matrix(type->type));
671 struct_member_decoration_cb(struct vtn_builder *b,
672 struct vtn_value *val, int member,
673 const struct vtn_decoration *dec, void *void_ctx)
675 struct member_decoration_ctx *ctx = void_ctx;
680 assert(member < ctx->num_fields);
682 switch (dec->decoration) {
683 case SpvDecorationNonWritable:
684 case SpvDecorationNonReadable:
685 case SpvDecorationRelaxedPrecision:
686 case SpvDecorationVolatile:
687 case SpvDecorationCoherent:
688 case SpvDecorationUniform:
689 break; /* FIXME: Do nothing with this for now. */
690 case SpvDecorationNoPerspective:
691 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
693 case SpvDecorationFlat:
694 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
696 case SpvDecorationCentroid:
697 ctx->fields[member].centroid = true;
699 case SpvDecorationSample:
700 ctx->fields[member].sample = true;
702 case SpvDecorationStream:
703 /* Vulkan only allows one GS stream */
704 vtn_assert(dec->literals[0] == 0);
706 case SpvDecorationLocation:
707 ctx->fields[member].location = dec->literals[0];
709 case SpvDecorationComponent:
710 break; /* FIXME: What should we do with these? */
711 case SpvDecorationBuiltIn:
712 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
713 ctx->type->members[member]->is_builtin = true;
714 ctx->type->members[member]->builtin = dec->literals[0];
715 ctx->type->builtin_block = true;
717 case SpvDecorationOffset:
718 ctx->type->offsets[member] = dec->literals[0];
720 case SpvDecorationMatrixStride:
721 /* Handled as a second pass */
723 case SpvDecorationColMajor:
724 break; /* Nothing to do here. Column-major is the default. */
725 case SpvDecorationRowMajor:
726 mutable_matrix_member(b, ctx->type, member)->row_major = true;
729 case SpvDecorationPatch:
732 case SpvDecorationSpecId:
733 case SpvDecorationBlock:
734 case SpvDecorationBufferBlock:
735 case SpvDecorationArrayStride:
736 case SpvDecorationGLSLShared:
737 case SpvDecorationGLSLPacked:
738 case SpvDecorationInvariant:
739 case SpvDecorationRestrict:
740 case SpvDecorationAliased:
741 case SpvDecorationConstant:
742 case SpvDecorationIndex:
743 case SpvDecorationBinding:
744 case SpvDecorationDescriptorSet:
745 case SpvDecorationLinkageAttributes:
746 case SpvDecorationNoContraction:
747 case SpvDecorationInputAttachmentIndex:
748 vtn_warn("Decoration not allowed on struct members: %s",
749 spirv_decoration_to_string(dec->decoration));
752 case SpvDecorationXfbBuffer:
753 case SpvDecorationXfbStride:
754 vtn_warn("Vulkan does not have transform feedback");
757 case SpvDecorationCPacked:
758 case SpvDecorationSaturatedConversion:
759 case SpvDecorationFuncParamAttr:
760 case SpvDecorationFPRoundingMode:
761 case SpvDecorationFPFastMathMode:
762 case SpvDecorationAlignment:
763 vtn_warn("Decoration only allowed for CL-style kernels: %s",
764 spirv_decoration_to_string(dec->decoration));
768 vtn_fail("Unhandled decoration");
772 /* Matrix strides are handled as a separate pass because we need to know
773 * whether the matrix is row-major or not first.
776 struct_member_matrix_stride_cb(struct vtn_builder *b,
777 struct vtn_value *val, int member,
778 const struct vtn_decoration *dec,
781 if (dec->decoration != SpvDecorationMatrixStride)
784 vtn_fail_if(member < 0,
785 "The MatrixStride decoration is only allowed on members "
788 struct member_decoration_ctx *ctx = void_ctx;
790 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
791 if (mat_type->row_major) {
792 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
793 mat_type->stride = mat_type->array_element->stride;
794 mat_type->array_element->stride = dec->literals[0];
796 vtn_assert(mat_type->array_element->stride > 0);
797 mat_type->stride = dec->literals[0];
802 type_decoration_cb(struct vtn_builder *b,
803 struct vtn_value *val, int member,
804 const struct vtn_decoration *dec, void *ctx)
806 struct vtn_type *type = val->type;
809 /* This should have been handled by OpTypeStruct */
810 assert(val->type->base_type == vtn_base_type_struct);
811 assert(member >= 0 && member < val->type->length);
815 switch (dec->decoration) {
816 case SpvDecorationArrayStride:
817 vtn_assert(type->base_type == vtn_base_type_matrix ||
818 type->base_type == vtn_base_type_array ||
819 type->base_type == vtn_base_type_pointer);
820 type->stride = dec->literals[0];
822 case SpvDecorationBlock:
823 vtn_assert(type->base_type == vtn_base_type_struct);
826 case SpvDecorationBufferBlock:
827 vtn_assert(type->base_type == vtn_base_type_struct);
828 type->buffer_block = true;
830 case SpvDecorationGLSLShared:
831 case SpvDecorationGLSLPacked:
832 /* Ignore these, since we get explicit offsets anyways */
835 case SpvDecorationRowMajor:
836 case SpvDecorationColMajor:
837 case SpvDecorationMatrixStride:
838 case SpvDecorationBuiltIn:
839 case SpvDecorationNoPerspective:
840 case SpvDecorationFlat:
841 case SpvDecorationPatch:
842 case SpvDecorationCentroid:
843 case SpvDecorationSample:
844 case SpvDecorationVolatile:
845 case SpvDecorationCoherent:
846 case SpvDecorationNonWritable:
847 case SpvDecorationNonReadable:
848 case SpvDecorationUniform:
849 case SpvDecorationStream:
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 SpvDecorationRelaxedPrecision:
860 case SpvDecorationSpecId:
861 case SpvDecorationInvariant:
862 case SpvDecorationRestrict:
863 case SpvDecorationAliased:
864 case SpvDecorationConstant:
865 case SpvDecorationIndex:
866 case SpvDecorationBinding:
867 case SpvDecorationDescriptorSet:
868 case SpvDecorationLinkageAttributes:
869 case SpvDecorationNoContraction:
870 case SpvDecorationInputAttachmentIndex:
871 vtn_warn("Decoration not allowed on types: %s",
872 spirv_decoration_to_string(dec->decoration));
875 case SpvDecorationCPacked:
876 case SpvDecorationSaturatedConversion:
877 case SpvDecorationFuncParamAttr:
878 case SpvDecorationFPRoundingMode:
879 case SpvDecorationFPFastMathMode:
880 case SpvDecorationAlignment:
881 vtn_warn("Decoration only allowed for CL-style kernels: %s",
882 spirv_decoration_to_string(dec->decoration));
886 vtn_fail("Unhandled decoration");
891 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
894 case SpvImageFormatUnknown: return 0; /* GL_NONE */
895 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
896 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
897 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
898 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
899 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
900 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
901 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
902 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
903 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
904 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
905 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
906 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
907 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
908 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
909 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
910 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
911 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
912 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
913 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
914 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
915 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
916 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
917 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
918 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
919 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
920 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
921 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
922 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
923 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
924 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
925 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
926 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
927 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
928 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
929 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
930 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
931 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
932 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
933 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
935 vtn_fail("Invalid image format");
939 static struct vtn_type *
940 vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
941 uint32_t *size_out, uint32_t *align_out)
943 switch (type->base_type) {
944 case vtn_base_type_scalar: {
945 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
946 *size_out = comp_size;
947 *align_out = comp_size;
951 case vtn_base_type_vector: {
952 uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
953 unsigned align_comps = type->length == 3 ? 4 : type->length;
954 *size_out = comp_size * type->length,
955 *align_out = comp_size * align_comps;
959 case vtn_base_type_matrix:
960 case vtn_base_type_array: {
961 /* We're going to add an array stride */
962 type = vtn_type_copy(b, type);
963 uint32_t elem_size, elem_align;
964 type->array_element = vtn_type_layout_std430(b, type->array_element,
965 &elem_size, &elem_align);
966 type->stride = vtn_align_u32(elem_size, elem_align);
967 *size_out = type->stride * type->length;
968 *align_out = elem_align;
972 case vtn_base_type_struct: {
973 /* We're going to add member offsets */
974 type = vtn_type_copy(b, type);
977 for (unsigned i = 0; i < type->length; i++) {
978 uint32_t mem_size, mem_align;
979 type->members[i] = vtn_type_layout_std430(b, type->members[i],
980 &mem_size, &mem_align);
981 offset = vtn_align_u32(offset, mem_align);
982 type->offsets[i] = offset;
984 align = MAX2(align, mem_align);
992 unreachable("Invalid SPIR-V type for std430");
997 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
998 const uint32_t *w, unsigned count)
1000 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
1002 val->type = rzalloc(b, struct vtn_type);
1003 val->type->id = w[1];
1007 val->type->base_type = vtn_base_type_void;
1008 val->type->type = glsl_void_type();
1011 val->type->base_type = vtn_base_type_scalar;
1012 val->type->type = glsl_bool_type();
1013 val->type->length = 1;
1015 case SpvOpTypeInt: {
1016 int bit_size = w[2];
1017 const bool signedness = w[3];
1018 val->type->base_type = vtn_base_type_scalar;
1021 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
1024 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
1027 val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
1030 val->type->type = (signedness ? glsl_int8_t_type() : glsl_uint8_t_type());
1033 vtn_fail("Invalid int bit size");
1035 val->type->length = 1;
1039 case SpvOpTypeFloat: {
1040 int bit_size = w[2];
1041 val->type->base_type = vtn_base_type_scalar;
1044 val->type->type = glsl_float16_t_type();
1047 val->type->type = glsl_float_type();
1050 val->type->type = glsl_double_type();
1053 vtn_fail("Invalid float bit size");
1055 val->type->length = 1;
1059 case SpvOpTypeVector: {
1060 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1061 unsigned elems = w[3];
1063 vtn_fail_if(base->base_type != vtn_base_type_scalar,
1064 "Base type for OpTypeVector must be a scalar");
1065 vtn_fail_if((elems < 2 || elems > 4) && (elems != 8) && (elems != 16),
1066 "Invalid component count for OpTypeVector");
1068 val->type->base_type = vtn_base_type_vector;
1069 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
1070 val->type->length = elems;
1071 val->type->stride = glsl_get_bit_size(base->type) / 8;
1072 val->type->array_element = base;
1076 case SpvOpTypeMatrix: {
1077 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
1078 unsigned columns = w[3];
1080 vtn_fail_if(base->base_type != vtn_base_type_vector,
1081 "Base type for OpTypeMatrix must be a vector");
1082 vtn_fail_if(columns < 2 || columns > 4,
1083 "Invalid column count for OpTypeMatrix");
1085 val->type->base_type = vtn_base_type_matrix;
1086 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
1087 glsl_get_vector_elements(base->type),
1089 vtn_fail_if(glsl_type_is_error(val->type->type),
1090 "Unsupported base type for OpTypeMatrix");
1091 assert(!glsl_type_is_error(val->type->type));
1092 val->type->length = columns;
1093 val->type->array_element = base;
1094 val->type->row_major = false;
1095 val->type->stride = 0;
1099 case SpvOpTypeRuntimeArray:
1100 case SpvOpTypeArray: {
1101 struct vtn_type *array_element =
1102 vtn_value(b, w[2], vtn_value_type_type)->type;
1104 if (opcode == SpvOpTypeRuntimeArray) {
1105 /* A length of 0 is used to denote unsized arrays */
1106 val->type->length = 0;
1109 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
1112 val->type->base_type = vtn_base_type_array;
1113 val->type->type = glsl_array_type(array_element->type, val->type->length);
1114 val->type->array_element = array_element;
1115 val->type->stride = 0;
1119 case SpvOpTypeStruct: {
1120 unsigned num_fields = count - 2;
1121 val->type->base_type = vtn_base_type_struct;
1122 val->type->length = num_fields;
1123 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
1124 val->type->offsets = ralloc_array(b, unsigned, num_fields);
1126 NIR_VLA(struct glsl_struct_field, fields, count);
1127 for (unsigned i = 0; i < num_fields; i++) {
1128 val->type->members[i] =
1129 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
1130 fields[i] = (struct glsl_struct_field) {
1131 .type = val->type->members[i]->type,
1132 .name = ralloc_asprintf(b, "field%d", i),
1137 struct member_decoration_ctx ctx = {
1138 .num_fields = num_fields,
1143 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
1144 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
1146 const char *name = val->name ? val->name : "struct";
1148 val->type->type = glsl_struct_type(fields, num_fields, name);
1152 case SpvOpTypeFunction: {
1153 val->type->base_type = vtn_base_type_function;
1154 val->type->type = NULL;
1156 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
1158 const unsigned num_params = count - 3;
1159 val->type->length = num_params;
1160 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
1161 for (unsigned i = 0; i < count - 3; i++) {
1162 val->type->params[i] =
1163 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
1168 case SpvOpTypePointer: {
1169 SpvStorageClass storage_class = w[2];
1170 struct vtn_type *deref_type =
1171 vtn_value(b, w[3], vtn_value_type_type)->type;
1173 val->type->base_type = vtn_base_type_pointer;
1174 val->type->storage_class = storage_class;
1175 val->type->deref = deref_type;
1177 if (storage_class == SpvStorageClassUniform ||
1178 storage_class == SpvStorageClassStorageBuffer) {
1179 /* These can actually be stored to nir_variables and used as SSA
1180 * values so they need a real glsl_type.
1182 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
1185 if (storage_class == SpvStorageClassPushConstant) {
1186 /* These can actually be stored to nir_variables and used as SSA
1187 * values so they need a real glsl_type.
1189 val->type->type = glsl_uint_type();
1192 if (storage_class == SpvStorageClassWorkgroup &&
1193 b->options->lower_workgroup_access_to_offsets) {
1194 uint32_t size, align;
1195 val->type->deref = vtn_type_layout_std430(b, val->type->deref,
1197 val->type->length = size;
1198 val->type->align = align;
1199 /* These can actually be stored to nir_variables and used as SSA
1200 * values so they need a real glsl_type.
1202 val->type->type = glsl_uint_type();
1207 case SpvOpTypeImage: {
1208 val->type->base_type = vtn_base_type_image;
1210 const struct vtn_type *sampled_type =
1211 vtn_value(b, w[2], vtn_value_type_type)->type;
1213 vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar ||
1214 glsl_get_bit_size(sampled_type->type) != 32,
1215 "Sampled type of OpTypeImage must be a 32-bit scalar");
1217 enum glsl_sampler_dim dim;
1218 switch ((SpvDim)w[3]) {
1219 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
1220 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
1221 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
1222 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
1223 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
1224 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
1225 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
1227 vtn_fail("Invalid SPIR-V image dimensionality");
1230 bool is_shadow = w[4];
1231 bool is_array = w[5];
1232 bool multisampled = w[6];
1233 unsigned sampled = w[7];
1234 SpvImageFormat format = w[8];
1237 val->type->access_qualifier = w[9];
1239 val->type->access_qualifier = SpvAccessQualifierReadWrite;
1242 if (dim == GLSL_SAMPLER_DIM_2D)
1243 dim = GLSL_SAMPLER_DIM_MS;
1244 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1245 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1247 vtn_fail("Unsupported multisampled image type");
1250 val->type->image_format = translate_image_format(b, format);
1252 enum glsl_base_type sampled_base_type =
1253 glsl_get_base_type(sampled_type->type);
1255 val->type->sampled = true;
1256 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1258 } else if (sampled == 2) {
1259 vtn_assert(!is_shadow);
1260 val->type->sampled = false;
1261 val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
1263 vtn_fail("We need to know if the image will be sampled");
1268 case SpvOpTypeSampledImage:
1269 val->type->base_type = vtn_base_type_sampled_image;
1270 val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type;
1271 val->type->type = val->type->image->type;
1274 case SpvOpTypeSampler:
1275 /* The actual sampler type here doesn't really matter. It gets
1276 * thrown away the moment you combine it with an image. What really
1277 * matters is that it's a sampler type as opposed to an integer type
1278 * so the backend knows what to do.
1280 val->type->base_type = vtn_base_type_sampler;
1281 val->type->type = glsl_bare_sampler_type();
1284 case SpvOpTypeOpaque:
1285 case SpvOpTypeEvent:
1286 case SpvOpTypeDeviceEvent:
1287 case SpvOpTypeReserveId:
1288 case SpvOpTypeQueue:
1291 vtn_fail("Unhandled opcode");
1294 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1297 static nir_constant *
1298 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1300 nir_constant *c = rzalloc(b, nir_constant);
1302 /* For pointers and other typeless things, we have to return something but
1303 * it doesn't matter what.
1308 switch (glsl_get_base_type(type)) {
1310 case GLSL_TYPE_UINT:
1311 case GLSL_TYPE_INT16:
1312 case GLSL_TYPE_UINT16:
1313 case GLSL_TYPE_UINT8:
1314 case GLSL_TYPE_INT8:
1315 case GLSL_TYPE_INT64:
1316 case GLSL_TYPE_UINT64:
1317 case GLSL_TYPE_BOOL:
1318 case GLSL_TYPE_FLOAT:
1319 case GLSL_TYPE_FLOAT16:
1320 case GLSL_TYPE_DOUBLE:
1321 /* Nothing to do here. It's already initialized to zero */
1324 case GLSL_TYPE_ARRAY:
1325 vtn_assert(glsl_get_length(type) > 0);
1326 c->num_elements = glsl_get_length(type);
1327 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1329 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1330 for (unsigned i = 1; i < c->num_elements; i++)
1331 c->elements[i] = c->elements[0];
1334 case GLSL_TYPE_STRUCT:
1335 c->num_elements = glsl_get_length(type);
1336 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1338 for (unsigned i = 0; i < c->num_elements; i++) {
1339 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1344 vtn_fail("Invalid type for null constant");
1351 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1352 int member, const struct vtn_decoration *dec,
1355 vtn_assert(member == -1);
1356 if (dec->decoration != SpvDecorationSpecId)
1359 struct spec_constant_value *const_value = data;
1361 for (unsigned i = 0; i < b->num_specializations; i++) {
1362 if (b->specializations[i].id == dec->literals[0]) {
1363 if (const_value->is_double)
1364 const_value->data64 = b->specializations[i].data64;
1366 const_value->data32 = b->specializations[i].data32;
1373 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1374 uint32_t const_value)
1376 struct spec_constant_value data;
1377 data.is_double = false;
1378 data.data32 = const_value;
1379 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1384 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1385 uint64_t const_value)
1387 struct spec_constant_value data;
1388 data.is_double = true;
1389 data.data64 = const_value;
1390 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1395 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1396 struct vtn_value *val,
1398 const struct vtn_decoration *dec,
1401 vtn_assert(member == -1);
1402 if (dec->decoration != SpvDecorationBuiltIn ||
1403 dec->literals[0] != SpvBuiltInWorkgroupSize)
1406 vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1408 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1409 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1410 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1414 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1415 const uint32_t *w, unsigned count)
1417 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1418 val->constant = rzalloc(b, nir_constant);
1420 case SpvOpConstantTrue:
1421 case SpvOpConstantFalse:
1422 case SpvOpSpecConstantTrue:
1423 case SpvOpSpecConstantFalse: {
1424 vtn_fail_if(val->type->type != glsl_bool_type(),
1425 "Result type of %s must be OpTypeBool",
1426 spirv_op_to_string(opcode));
1428 uint32_t int_val = (opcode == SpvOpConstantTrue ||
1429 opcode == SpvOpSpecConstantTrue);
1431 if (opcode == SpvOpSpecConstantTrue ||
1432 opcode == SpvOpSpecConstantFalse)
1433 int_val = get_specialization(b, val, int_val);
1435 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1439 case SpvOpConstant: {
1440 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1441 "Result type of %s must be a scalar",
1442 spirv_op_to_string(opcode));
1443 int bit_size = glsl_get_bit_size(val->type->type);
1446 val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
1449 val->constant->values->u32[0] = w[3];
1452 val->constant->values->u16[0] = w[3];
1455 val->constant->values->u8[0] = w[3];
1458 vtn_fail("Unsupported SpvOpConstant bit size");
1463 case SpvOpSpecConstant: {
1464 vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
1465 "Result type of %s must be a scalar",
1466 spirv_op_to_string(opcode));
1467 int bit_size = glsl_get_bit_size(val->type->type);
1470 val->constant->values[0].u64[0] =
1471 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1474 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1477 val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
1480 val->constant->values[0].u8[0] = get_specialization(b, val, w[3]);
1483 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1488 case SpvOpSpecConstantComposite:
1489 case SpvOpConstantComposite: {
1490 unsigned elem_count = count - 3;
1491 vtn_fail_if(elem_count != val->type->length,
1492 "%s has %u constituents, expected %u",
1493 spirv_op_to_string(opcode), elem_count, val->type->length);
1495 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1496 for (unsigned i = 0; i < elem_count; i++)
1497 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1499 switch (val->type->base_type) {
1500 case vtn_base_type_vector: {
1501 assert(glsl_type_is_vector(val->type->type));
1502 int bit_size = glsl_get_bit_size(val->type->type);
1503 for (unsigned i = 0; i < elem_count; i++) {
1506 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1509 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1512 val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
1515 val->constant->values[0].u8[i] = elems[i]->values[0].u8[0];
1518 vtn_fail("Invalid SpvOpConstantComposite bit size");
1524 case vtn_base_type_matrix:
1525 assert(glsl_type_is_matrix(val->type->type));
1526 for (unsigned i = 0; i < elem_count; i++)
1527 val->constant->values[i] = elems[i]->values[0];
1530 case vtn_base_type_struct:
1531 case vtn_base_type_array:
1532 ralloc_steal(val->constant, elems);
1533 val->constant->num_elements = elem_count;
1534 val->constant->elements = elems;
1538 vtn_fail("Result type of %s must be a composite type",
1539 spirv_op_to_string(opcode));
1544 case SpvOpSpecConstantOp: {
1545 SpvOp opcode = get_specialization(b, val, w[3]);
1547 case SpvOpVectorShuffle: {
1548 struct vtn_value *v0 = &b->values[w[4]];
1549 struct vtn_value *v1 = &b->values[w[5]];
1551 vtn_assert(v0->value_type == vtn_value_type_constant ||
1552 v0->value_type == vtn_value_type_undef);
1553 vtn_assert(v1->value_type == vtn_value_type_constant ||
1554 v1->value_type == vtn_value_type_undef);
1556 unsigned len0 = glsl_get_vector_elements(v0->type->type);
1557 unsigned len1 = glsl_get_vector_elements(v1->type->type);
1559 vtn_assert(len0 + len1 < 16);
1561 unsigned bit_size = glsl_get_bit_size(val->type->type);
1562 unsigned bit_size0 = glsl_get_bit_size(v0->type->type);
1563 unsigned bit_size1 = glsl_get_bit_size(v1->type->type);
1565 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1566 (void)bit_size0; (void)bit_size1;
1568 if (bit_size == 64) {
1570 if (v0->value_type == vtn_value_type_constant) {
1571 for (unsigned i = 0; i < len0; i++)
1572 u64[i] = v0->constant->values[0].u64[i];
1574 if (v1->value_type == vtn_value_type_constant) {
1575 for (unsigned i = 0; i < len1; i++)
1576 u64[len0 + i] = v1->constant->values[0].u64[i];
1579 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1580 uint32_t comp = w[i + 6];
1581 /* If component is not used, set the value to a known constant
1582 * to detect if it is wrongly used.
1584 if (comp == (uint32_t)-1)
1585 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1587 val->constant->values[0].u64[j] = u64[comp];
1590 /* This is for both 32-bit and 16-bit values */
1592 if (v0->value_type == vtn_value_type_constant) {
1593 for (unsigned i = 0; i < len0; i++)
1594 u32[i] = v0->constant->values[0].u32[i];
1596 if (v1->value_type == vtn_value_type_constant) {
1597 for (unsigned i = 0; i < len1; i++)
1598 u32[len0 + i] = v1->constant->values[0].u32[i];
1601 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1602 uint32_t comp = w[i + 6];
1603 /* If component is not used, set the value to a known constant
1604 * to detect if it is wrongly used.
1606 if (comp == (uint32_t)-1)
1607 val->constant->values[0].u32[j] = 0xdeadbeef;
1609 val->constant->values[0].u32[j] = u32[comp];
1615 case SpvOpCompositeExtract:
1616 case SpvOpCompositeInsert: {
1617 struct vtn_value *comp;
1618 unsigned deref_start;
1619 struct nir_constant **c;
1620 if (opcode == SpvOpCompositeExtract) {
1621 comp = vtn_value(b, w[4], vtn_value_type_constant);
1623 c = &comp->constant;
1625 comp = vtn_value(b, w[5], vtn_value_type_constant);
1627 val->constant = nir_constant_clone(comp->constant,
1634 const struct vtn_type *type = comp->type;
1635 for (unsigned i = deref_start; i < count; i++) {
1636 vtn_fail_if(w[i] > type->length,
1637 "%uth index of %s is %u but the type has only "
1638 "%u elements", i - deref_start,
1639 spirv_op_to_string(opcode), w[i], type->length);
1641 switch (type->base_type) {
1642 case vtn_base_type_vector:
1644 type = type->array_element;
1647 case vtn_base_type_matrix:
1648 assert(col == 0 && elem == -1);
1651 type = type->array_element;
1654 case vtn_base_type_array:
1655 c = &(*c)->elements[w[i]];
1656 type = type->array_element;
1659 case vtn_base_type_struct:
1660 c = &(*c)->elements[w[i]];
1661 type = type->members[w[i]];
1665 vtn_fail("%s must only index into composite types",
1666 spirv_op_to_string(opcode));
1670 if (opcode == SpvOpCompositeExtract) {
1674 unsigned num_components = type->length;
1675 unsigned bit_size = glsl_get_bit_size(type->type);
1676 for (unsigned i = 0; i < num_components; i++)
1679 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1682 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1685 val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
1688 val->constant->values[0].u8[i] = (*c)->values[col].u8[elem + i];
1691 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1695 struct vtn_value *insert =
1696 vtn_value(b, w[4], vtn_value_type_constant);
1697 vtn_assert(insert->type == type);
1699 *c = insert->constant;
1701 unsigned num_components = type->length;
1702 unsigned bit_size = glsl_get_bit_size(type->type);
1703 for (unsigned i = 0; i < num_components; i++)
1706 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1709 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1712 (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
1715 (*c)->values[col].u8[elem + i] = insert->constant->values[0].u8[i];
1718 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1727 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->type->type);
1728 nir_alu_type src_alu_type = dst_alu_type;
1729 unsigned num_components = glsl_get_vector_elements(val->type->type);
1732 vtn_assert(count <= 7);
1737 /* We have a source in a conversion */
1739 nir_get_nir_type_for_glsl_type(
1740 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1741 /* We use the bitsize of the conversion source to evaluate the opcode later */
1742 bit_size = glsl_get_bit_size(
1743 vtn_value(b, w[4], vtn_value_type_constant)->type->type);
1746 bit_size = glsl_get_bit_size(val->type->type);
1749 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1750 nir_alu_type_get_type_size(src_alu_type),
1751 nir_alu_type_get_type_size(dst_alu_type));
1752 nir_const_value src[4];
1754 for (unsigned i = 0; i < count - 4; i++) {
1756 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1758 unsigned j = swap ? 1 - i : i;
1759 src[j] = c->values[0];
1762 val->constant->values[0] =
1763 nir_eval_const_opcode(op, num_components, bit_size, src);
1770 case SpvOpConstantNull:
1771 val->constant = vtn_null_constant(b, val->type->type);
1774 case SpvOpConstantSampler:
1775 vtn_fail("OpConstantSampler requires Kernel Capability");
1779 vtn_fail("Unhandled opcode");
1782 /* Now that we have the value, update the workgroup size if needed */
1783 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1787 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1788 const uint32_t *w, unsigned count)
1790 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1791 struct vtn_function *vtn_callee =
1792 vtn_value(b, w[3], vtn_value_type_function)->func;
1793 struct nir_function *callee = vtn_callee->impl->function;
1795 vtn_callee->referenced = true;
1797 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1799 unsigned param_idx = 0;
1801 nir_deref_instr *ret_deref = NULL;
1802 struct vtn_type *ret_type = vtn_callee->type->return_type;
1803 if (ret_type->base_type != vtn_base_type_void) {
1804 nir_variable *ret_tmp =
1805 nir_local_variable_create(b->nb.impl, ret_type->type, "return_tmp");
1806 ret_deref = nir_build_deref_var(&b->nb, ret_tmp);
1807 call->params[param_idx++] = nir_src_for_ssa(&ret_deref->dest.ssa);
1810 for (unsigned i = 0; i < vtn_callee->type->length; i++) {
1811 struct vtn_type *arg_type = vtn_callee->type->params[i];
1812 unsigned arg_id = w[4 + i];
1814 if (arg_type->base_type == vtn_base_type_sampled_image) {
1815 struct vtn_sampled_image *sampled_image =
1816 vtn_value(b, arg_id, vtn_value_type_sampled_image)->sampled_image;
1818 call->params[param_idx++] =
1819 nir_src_for_ssa(&sampled_image->image->deref->dest.ssa);
1820 call->params[param_idx++] =
1821 nir_src_for_ssa(&sampled_image->sampler->deref->dest.ssa);
1822 } else if (arg_type->base_type == vtn_base_type_pointer ||
1823 arg_type->base_type == vtn_base_type_image ||
1824 arg_type->base_type == vtn_base_type_sampler) {
1825 struct vtn_pointer *pointer =
1826 vtn_value(b, arg_id, vtn_value_type_pointer)->pointer;
1827 call->params[param_idx++] =
1828 nir_src_for_ssa(vtn_pointer_to_ssa(b, pointer));
1830 /* This is a regular SSA value and we need a temporary */
1832 nir_local_variable_create(b->nb.impl, arg_type->type, "arg_tmp");
1833 nir_deref_instr *tmp_deref = nir_build_deref_var(&b->nb, tmp);
1834 vtn_local_store(b, vtn_ssa_value(b, arg_id), tmp_deref);
1835 call->params[param_idx++] = nir_src_for_ssa(&tmp_deref->dest.ssa);
1838 assert(param_idx == call->num_params);
1840 nir_builder_instr_insert(&b->nb, &call->instr);
1842 if (ret_type->base_type == vtn_base_type_void) {
1843 vtn_push_value(b, w[2], vtn_value_type_undef);
1845 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, ret_deref));
1849 struct vtn_ssa_value *
1850 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1852 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1855 if (!glsl_type_is_vector_or_scalar(type)) {
1856 unsigned elems = glsl_get_length(type);
1857 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1858 for (unsigned i = 0; i < elems; i++) {
1859 const struct glsl_type *child_type;
1861 switch (glsl_get_base_type(type)) {
1863 case GLSL_TYPE_UINT:
1864 case GLSL_TYPE_INT16:
1865 case GLSL_TYPE_UINT16:
1866 case GLSL_TYPE_UINT8:
1867 case GLSL_TYPE_INT8:
1868 case GLSL_TYPE_INT64:
1869 case GLSL_TYPE_UINT64:
1870 case GLSL_TYPE_BOOL:
1871 case GLSL_TYPE_FLOAT:
1872 case GLSL_TYPE_FLOAT16:
1873 case GLSL_TYPE_DOUBLE:
1874 child_type = glsl_get_column_type(type);
1876 case GLSL_TYPE_ARRAY:
1877 child_type = glsl_get_array_element(type);
1879 case GLSL_TYPE_STRUCT:
1880 child_type = glsl_get_struct_field(type, i);
1883 vtn_fail("unkown base type");
1886 val->elems[i] = vtn_create_ssa_value(b, child_type);
1894 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1897 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1898 src.src_type = type;
1903 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1904 const uint32_t *w, unsigned count)
1906 if (opcode == SpvOpSampledImage) {
1907 struct vtn_value *val =
1908 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1909 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1910 val->sampled_image->type =
1911 vtn_value(b, w[1], vtn_value_type_type)->type;
1912 val->sampled_image->image =
1913 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1914 val->sampled_image->sampler =
1915 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1917 } else if (opcode == SpvOpImage) {
1918 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1919 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1920 if (src_val->value_type == vtn_value_type_sampled_image) {
1921 val->pointer = src_val->sampled_image->image;
1923 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1924 val->pointer = src_val->pointer;
1929 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1930 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1932 struct vtn_sampled_image sampled;
1933 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1934 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1935 sampled = *sampled_val->sampled_image;
1937 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1938 sampled.type = sampled_val->pointer->type;
1939 sampled.image = NULL;
1940 sampled.sampler = sampled_val->pointer;
1943 const struct glsl_type *image_type = sampled.type->type;
1944 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1945 const bool is_array = glsl_sampler_type_is_array(image_type);
1947 /* Figure out the base texture operation */
1950 case SpvOpImageSampleImplicitLod:
1951 case SpvOpImageSampleDrefImplicitLod:
1952 case SpvOpImageSampleProjImplicitLod:
1953 case SpvOpImageSampleProjDrefImplicitLod:
1954 texop = nir_texop_tex;
1957 case SpvOpImageSampleExplicitLod:
1958 case SpvOpImageSampleDrefExplicitLod:
1959 case SpvOpImageSampleProjExplicitLod:
1960 case SpvOpImageSampleProjDrefExplicitLod:
1961 texop = nir_texop_txl;
1964 case SpvOpImageFetch:
1965 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1966 texop = nir_texop_txf_ms;
1968 texop = nir_texop_txf;
1972 case SpvOpImageGather:
1973 case SpvOpImageDrefGather:
1974 texop = nir_texop_tg4;
1977 case SpvOpImageQuerySizeLod:
1978 case SpvOpImageQuerySize:
1979 texop = nir_texop_txs;
1982 case SpvOpImageQueryLod:
1983 texop = nir_texop_lod;
1986 case SpvOpImageQueryLevels:
1987 texop = nir_texop_query_levels;
1990 case SpvOpImageQuerySamples:
1991 texop = nir_texop_texture_samples;
1995 vtn_fail("Unhandled opcode");
1998 nir_tex_src srcs[10]; /* 10 should be enough */
1999 nir_tex_src *p = srcs;
2001 nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
2002 nir_deref_instr *texture =
2003 sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
2005 p->src = nir_src_for_ssa(&texture->dest.ssa);
2006 p->src_type = nir_tex_src_texture_deref;
2015 /* These operations require a sampler */
2016 p->src = nir_src_for_ssa(&sampler->dest.ssa);
2017 p->src_type = nir_tex_src_sampler_deref;
2021 case nir_texop_txf_ms:
2024 case nir_texop_query_levels:
2025 case nir_texop_texture_samples:
2026 case nir_texop_samples_identical:
2029 case nir_texop_txf_ms_mcs:
2030 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2035 struct nir_ssa_def *coord;
2036 unsigned coord_components;
2038 case SpvOpImageSampleImplicitLod:
2039 case SpvOpImageSampleExplicitLod:
2040 case SpvOpImageSampleDrefImplicitLod:
2041 case SpvOpImageSampleDrefExplicitLod:
2042 case SpvOpImageSampleProjImplicitLod:
2043 case SpvOpImageSampleProjExplicitLod:
2044 case SpvOpImageSampleProjDrefImplicitLod:
2045 case SpvOpImageSampleProjDrefExplicitLod:
2046 case SpvOpImageFetch:
2047 case SpvOpImageGather:
2048 case SpvOpImageDrefGather:
2049 case SpvOpImageQueryLod: {
2050 /* All these types have the coordinate as their first real argument */
2051 switch (sampler_dim) {
2052 case GLSL_SAMPLER_DIM_1D:
2053 case GLSL_SAMPLER_DIM_BUF:
2054 coord_components = 1;
2056 case GLSL_SAMPLER_DIM_2D:
2057 case GLSL_SAMPLER_DIM_RECT:
2058 case GLSL_SAMPLER_DIM_MS:
2059 coord_components = 2;
2061 case GLSL_SAMPLER_DIM_3D:
2062 case GLSL_SAMPLER_DIM_CUBE:
2063 coord_components = 3;
2066 vtn_fail("Invalid sampler type");
2069 if (is_array && texop != nir_texop_lod)
2072 coord = vtn_ssa_value(b, w[idx++])->def;
2073 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
2074 (1 << coord_components) - 1));
2075 p->src_type = nir_tex_src_coord;
2082 coord_components = 0;
2087 case SpvOpImageSampleProjImplicitLod:
2088 case SpvOpImageSampleProjExplicitLod:
2089 case SpvOpImageSampleProjDrefImplicitLod:
2090 case SpvOpImageSampleProjDrefExplicitLod:
2091 /* These have the projector as the last coordinate component */
2092 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
2093 p->src_type = nir_tex_src_projector;
2101 bool is_shadow = false;
2102 unsigned gather_component = 0;
2104 case SpvOpImageSampleDrefImplicitLod:
2105 case SpvOpImageSampleDrefExplicitLod:
2106 case SpvOpImageSampleProjDrefImplicitLod:
2107 case SpvOpImageSampleProjDrefExplicitLod:
2108 case SpvOpImageDrefGather:
2109 /* These all have an explicit depth value as their next source */
2111 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
2114 case SpvOpImageGather:
2115 /* This has a component as its next source */
2117 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
2124 /* For OpImageQuerySizeLod, we always have an LOD */
2125 if (opcode == SpvOpImageQuerySizeLod)
2126 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2128 /* Now we need to handle some number of optional arguments */
2129 const struct vtn_ssa_value *gather_offsets = NULL;
2131 uint32_t operands = w[idx++];
2133 if (operands & SpvImageOperandsBiasMask) {
2134 vtn_assert(texop == nir_texop_tex);
2135 texop = nir_texop_txb;
2136 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
2139 if (operands & SpvImageOperandsLodMask) {
2140 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
2141 texop == nir_texop_txs);
2142 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
2145 if (operands & SpvImageOperandsGradMask) {
2146 vtn_assert(texop == nir_texop_txl);
2147 texop = nir_texop_txd;
2148 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
2149 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
2152 if (operands & SpvImageOperandsOffsetMask ||
2153 operands & SpvImageOperandsConstOffsetMask)
2154 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
2156 if (operands & SpvImageOperandsConstOffsetsMask) {
2157 nir_tex_src none = {0};
2158 gather_offsets = vtn_ssa_value(b, w[idx++]);
2162 if (operands & SpvImageOperandsSampleMask) {
2163 vtn_assert(texop == nir_texop_txf_ms);
2164 texop = nir_texop_txf_ms;
2165 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
2168 /* We should have now consumed exactly all of the arguments */
2169 vtn_assert(idx == count);
2171 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
2174 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
2176 instr->coord_components = coord_components;
2177 instr->sampler_dim = sampler_dim;
2178 instr->is_array = is_array;
2179 instr->is_shadow = is_shadow;
2180 instr->is_new_style_shadow =
2181 is_shadow && glsl_get_components(ret_type->type) == 1;
2182 instr->component = gather_component;
2184 switch (glsl_get_sampler_result_type(image_type)) {
2185 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
2186 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
2187 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
2188 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
2190 vtn_fail("Invalid base type for sampler result");
2193 nir_ssa_dest_init(&instr->instr, &instr->dest,
2194 nir_tex_instr_dest_size(instr), 32, NULL);
2196 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
2197 nir_tex_instr_dest_size(instr));
2200 nir_instr *instruction;
2201 if (gather_offsets) {
2202 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
2203 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
2204 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
2206 /* Copy the current instruction 4x */
2207 for (uint32_t i = 1; i < 4; i++) {
2208 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
2209 instrs[i]->op = instr->op;
2210 instrs[i]->coord_components = instr->coord_components;
2211 instrs[i]->sampler_dim = instr->sampler_dim;
2212 instrs[i]->is_array = instr->is_array;
2213 instrs[i]->is_shadow = instr->is_shadow;
2214 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
2215 instrs[i]->component = instr->component;
2216 instrs[i]->dest_type = instr->dest_type;
2218 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
2220 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
2221 nir_tex_instr_dest_size(instr), 32, NULL);
2224 /* Fill in the last argument with the offset from the passed in offsets
2225 * and insert the instruction into the stream.
2227 for (uint32_t i = 0; i < 4; i++) {
2229 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
2230 src.src_type = nir_tex_src_offset;
2231 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
2232 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
2235 /* Combine the results of the 4 instructions by taking their .w
2238 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
2239 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
2240 vec4->dest.write_mask = 0xf;
2241 for (uint32_t i = 0; i < 4; i++) {
2242 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
2243 vec4->src[i].swizzle[0] = 3;
2245 def = &vec4->dest.dest.ssa;
2246 instruction = &vec4->instr;
2248 def = &instr->dest.ssa;
2249 instruction = &instr->instr;
2252 val->ssa = vtn_create_ssa_value(b, ret_type->type);
2253 val->ssa->def = def;
2255 nir_builder_instr_insert(&b->nb, instruction);
2259 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
2260 const uint32_t *w, nir_src *src)
2263 case SpvOpAtomicIIncrement:
2264 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
2267 case SpvOpAtomicIDecrement:
2268 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
2271 case SpvOpAtomicISub:
2273 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
2276 case SpvOpAtomicCompareExchange:
2277 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
2278 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
2281 case SpvOpAtomicExchange:
2282 case SpvOpAtomicIAdd:
2283 case SpvOpAtomicSMin:
2284 case SpvOpAtomicUMin:
2285 case SpvOpAtomicSMax:
2286 case SpvOpAtomicUMax:
2287 case SpvOpAtomicAnd:
2289 case SpvOpAtomicXor:
2290 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
2294 vtn_fail("Invalid SPIR-V atomic");
2298 static nir_ssa_def *
2299 get_image_coord(struct vtn_builder *b, uint32_t value)
2301 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
2303 /* The image_load_store intrinsics assume a 4-dim coordinate */
2304 unsigned dim = glsl_get_vector_elements(coord->type);
2305 unsigned swizzle[4];
2306 for (unsigned i = 0; i < 4; i++)
2307 swizzle[i] = MIN2(i, dim - 1);
2309 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
2313 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
2314 const uint32_t *w, unsigned count)
2316 /* Just get this one out of the way */
2317 if (opcode == SpvOpImageTexelPointer) {
2318 struct vtn_value *val =
2319 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
2320 val->image = ralloc(b, struct vtn_image_pointer);
2322 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2323 val->image->coord = get_image_coord(b, w[4]);
2324 val->image->sample = vtn_ssa_value(b, w[5])->def;
2328 struct vtn_image_pointer image;
2331 case SpvOpAtomicExchange:
2332 case SpvOpAtomicCompareExchange:
2333 case SpvOpAtomicCompareExchangeWeak:
2334 case SpvOpAtomicIIncrement:
2335 case SpvOpAtomicIDecrement:
2336 case SpvOpAtomicIAdd:
2337 case SpvOpAtomicISub:
2338 case SpvOpAtomicLoad:
2339 case SpvOpAtomicSMin:
2340 case SpvOpAtomicUMin:
2341 case SpvOpAtomicSMax:
2342 case SpvOpAtomicUMax:
2343 case SpvOpAtomicAnd:
2345 case SpvOpAtomicXor:
2346 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2349 case SpvOpAtomicStore:
2350 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2353 case SpvOpImageQuerySize:
2354 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2356 image.sample = NULL;
2359 case SpvOpImageRead:
2360 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2361 image.coord = get_image_coord(b, w[4]);
2363 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2364 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2365 image.sample = vtn_ssa_value(b, w[6])->def;
2367 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2371 case SpvOpImageWrite:
2372 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2373 image.coord = get_image_coord(b, w[2]);
2377 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2378 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2379 image.sample = vtn_ssa_value(b, w[5])->def;
2381 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2386 vtn_fail("Invalid image opcode");
2389 nir_intrinsic_op op;
2391 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2392 OP(ImageQuerySize, size)
2394 OP(ImageWrite, store)
2395 OP(AtomicLoad, load)
2396 OP(AtomicStore, store)
2397 OP(AtomicExchange, atomic_exchange)
2398 OP(AtomicCompareExchange, atomic_comp_swap)
2399 OP(AtomicIIncrement, atomic_add)
2400 OP(AtomicIDecrement, atomic_add)
2401 OP(AtomicIAdd, atomic_add)
2402 OP(AtomicISub, atomic_add)
2403 OP(AtomicSMin, atomic_min)
2404 OP(AtomicUMin, atomic_min)
2405 OP(AtomicSMax, atomic_max)
2406 OP(AtomicUMax, atomic_max)
2407 OP(AtomicAnd, atomic_and)
2408 OP(AtomicOr, atomic_or)
2409 OP(AtomicXor, atomic_xor)
2412 vtn_fail("Invalid image opcode");
2415 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2417 nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image.image);
2418 intrin->src[0] = nir_src_for_ssa(&image_deref->dest.ssa);
2420 /* ImageQuerySize doesn't take any extra parameters */
2421 if (opcode != SpvOpImageQuerySize) {
2422 /* The image coordinate is always 4 components but we may not have that
2423 * many. Swizzle to compensate.
2426 for (unsigned i = 0; i < 4; i++)
2427 swiz[i] = i < image.coord->num_components ? i : 0;
2428 intrin->src[1] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2430 intrin->src[2] = nir_src_for_ssa(image.sample);
2434 case SpvOpAtomicLoad:
2435 case SpvOpImageQuerySize:
2436 case SpvOpImageRead:
2438 case SpvOpAtomicStore:
2439 intrin->src[3] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2441 case SpvOpImageWrite:
2442 intrin->src[3] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2445 case SpvOpAtomicCompareExchange:
2446 case SpvOpAtomicIIncrement:
2447 case SpvOpAtomicIDecrement:
2448 case SpvOpAtomicExchange:
2449 case SpvOpAtomicIAdd:
2450 case SpvOpAtomicISub:
2451 case SpvOpAtomicSMin:
2452 case SpvOpAtomicUMin:
2453 case SpvOpAtomicSMax:
2454 case SpvOpAtomicUMax:
2455 case SpvOpAtomicAnd:
2457 case SpvOpAtomicXor:
2458 fill_common_atomic_sources(b, opcode, w, &intrin->src[3]);
2462 vtn_fail("Invalid image opcode");
2465 if (opcode != SpvOpImageWrite) {
2466 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2467 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2469 unsigned dest_components = nir_intrinsic_dest_components(intrin);
2470 if (intrin->intrinsic == nir_intrinsic_image_deref_size) {
2471 dest_components = intrin->num_components =
2472 glsl_get_vector_elements(type->type);
2475 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2476 dest_components, 32, NULL);
2478 nir_builder_instr_insert(&b->nb, &intrin->instr);
2480 val->ssa = vtn_create_ssa_value(b, type->type);
2481 val->ssa->def = &intrin->dest.ssa;
2483 nir_builder_instr_insert(&b->nb, &intrin->instr);
2487 static nir_intrinsic_op
2488 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2491 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2492 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2493 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2494 OP(AtomicExchange, atomic_exchange)
2495 OP(AtomicCompareExchange, atomic_comp_swap)
2496 OP(AtomicIIncrement, atomic_add)
2497 OP(AtomicIDecrement, atomic_add)
2498 OP(AtomicIAdd, atomic_add)
2499 OP(AtomicISub, atomic_add)
2500 OP(AtomicSMin, atomic_imin)
2501 OP(AtomicUMin, atomic_umin)
2502 OP(AtomicSMax, atomic_imax)
2503 OP(AtomicUMax, atomic_umax)
2504 OP(AtomicAnd, atomic_and)
2505 OP(AtomicOr, atomic_or)
2506 OP(AtomicXor, atomic_xor)
2509 vtn_fail("Invalid SSBO atomic");
2513 static nir_intrinsic_op
2514 get_uniform_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2517 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2518 OP(AtomicLoad, read_deref)
2519 OP(AtomicExchange, exchange)
2520 OP(AtomicCompareExchange, comp_swap)
2521 OP(AtomicIIncrement, inc_deref)
2522 OP(AtomicIDecrement, post_dec_deref)
2523 OP(AtomicIAdd, add_deref)
2524 OP(AtomicISub, add_deref)
2525 OP(AtomicUMin, min_deref)
2526 OP(AtomicUMax, max_deref)
2527 OP(AtomicAnd, and_deref)
2528 OP(AtomicOr, or_deref)
2529 OP(AtomicXor, xor_deref)
2532 /* We left the following out: AtomicStore, AtomicSMin and
2533 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2534 * moment Atomic Counter support is needed for ARB_spirv support, so is
2535 * only need to support GLSL Atomic Counters that are uints and don't
2536 * allow direct storage.
2538 unreachable("Invalid uniform atomic");
2542 static nir_intrinsic_op
2543 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2546 case SpvOpAtomicLoad: return nir_intrinsic_load_shared;
2547 case SpvOpAtomicStore: return nir_intrinsic_store_shared;
2548 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2549 OP(AtomicExchange, atomic_exchange)
2550 OP(AtomicCompareExchange, atomic_comp_swap)
2551 OP(AtomicIIncrement, atomic_add)
2552 OP(AtomicIDecrement, atomic_add)
2553 OP(AtomicIAdd, atomic_add)
2554 OP(AtomicISub, atomic_add)
2555 OP(AtomicSMin, atomic_imin)
2556 OP(AtomicUMin, atomic_umin)
2557 OP(AtomicSMax, atomic_imax)
2558 OP(AtomicUMax, atomic_umax)
2559 OP(AtomicAnd, atomic_and)
2560 OP(AtomicOr, atomic_or)
2561 OP(AtomicXor, atomic_xor)
2564 vtn_fail("Invalid shared atomic");
2568 static nir_intrinsic_op
2569 get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2572 case SpvOpAtomicLoad: return nir_intrinsic_load_deref;
2573 case SpvOpAtomicStore: return nir_intrinsic_store_deref;
2574 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2575 OP(AtomicExchange, atomic_exchange)
2576 OP(AtomicCompareExchange, atomic_comp_swap)
2577 OP(AtomicIIncrement, atomic_add)
2578 OP(AtomicIDecrement, atomic_add)
2579 OP(AtomicIAdd, atomic_add)
2580 OP(AtomicISub, atomic_add)
2581 OP(AtomicSMin, atomic_imin)
2582 OP(AtomicUMin, atomic_umin)
2583 OP(AtomicSMax, atomic_imax)
2584 OP(AtomicUMax, atomic_umax)
2585 OP(AtomicAnd, atomic_and)
2586 OP(AtomicOr, atomic_or)
2587 OP(AtomicXor, atomic_xor)
2590 vtn_fail("Invalid shared atomic");
2595 * Handles shared atomics, ssbo atomics and atomic counters.
2598 vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
2599 const uint32_t *w, unsigned count)
2601 struct vtn_pointer *ptr;
2602 nir_intrinsic_instr *atomic;
2605 case SpvOpAtomicLoad:
2606 case SpvOpAtomicExchange:
2607 case SpvOpAtomicCompareExchange:
2608 case SpvOpAtomicCompareExchangeWeak:
2609 case SpvOpAtomicIIncrement:
2610 case SpvOpAtomicIDecrement:
2611 case SpvOpAtomicIAdd:
2612 case SpvOpAtomicISub:
2613 case SpvOpAtomicSMin:
2614 case SpvOpAtomicUMin:
2615 case SpvOpAtomicSMax:
2616 case SpvOpAtomicUMax:
2617 case SpvOpAtomicAnd:
2619 case SpvOpAtomicXor:
2620 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2623 case SpvOpAtomicStore:
2624 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2628 vtn_fail("Invalid SPIR-V atomic");
2632 SpvScope scope = w[4];
2633 SpvMemorySemanticsMask semantics = w[5];
2636 /* uniform as "atomic counter uniform" */
2637 if (ptr->mode == vtn_variable_mode_uniform) {
2638 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
2639 const struct glsl_type *deref_type = deref->type;
2640 nir_intrinsic_op op = get_uniform_nir_atomic_op(b, opcode);
2641 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2642 atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
2644 /* SSBO needs to initialize index/offset. In this case we don't need to,
2645 * as that info is already stored on the ptr->var->var nir_variable (see
2646 * vtn_create_variable)
2650 case SpvOpAtomicLoad:
2651 atomic->num_components = glsl_get_vector_elements(deref_type);
2654 case SpvOpAtomicStore:
2655 atomic->num_components = glsl_get_vector_elements(deref_type);
2656 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2659 case SpvOpAtomicExchange:
2660 case SpvOpAtomicCompareExchange:
2661 case SpvOpAtomicCompareExchangeWeak:
2662 case SpvOpAtomicIIncrement:
2663 case SpvOpAtomicIDecrement:
2664 case SpvOpAtomicIAdd:
2665 case SpvOpAtomicISub:
2666 case SpvOpAtomicSMin:
2667 case SpvOpAtomicUMin:
2668 case SpvOpAtomicSMax:
2669 case SpvOpAtomicUMax:
2670 case SpvOpAtomicAnd:
2672 case SpvOpAtomicXor:
2673 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2674 * atomic counter uniforms doesn't have sources
2679 unreachable("Invalid SPIR-V atomic");
2682 } else if (ptr->mode == vtn_variable_mode_workgroup &&
2683 !b->options->lower_workgroup_access_to_offsets) {
2684 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
2685 const struct glsl_type *deref_type = deref->type;
2686 nir_intrinsic_op op = get_deref_nir_atomic_op(b, opcode);
2687 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2688 atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
2691 case SpvOpAtomicLoad:
2692 atomic->num_components = glsl_get_vector_elements(deref_type);
2695 case SpvOpAtomicStore:
2696 atomic->num_components = glsl_get_vector_elements(deref_type);
2697 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2698 atomic->src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2701 case SpvOpAtomicExchange:
2702 case SpvOpAtomicCompareExchange:
2703 case SpvOpAtomicCompareExchangeWeak:
2704 case SpvOpAtomicIIncrement:
2705 case SpvOpAtomicIDecrement:
2706 case SpvOpAtomicIAdd:
2707 case SpvOpAtomicISub:
2708 case SpvOpAtomicSMin:
2709 case SpvOpAtomicUMin:
2710 case SpvOpAtomicSMax:
2711 case SpvOpAtomicUMax:
2712 case SpvOpAtomicAnd:
2714 case SpvOpAtomicXor:
2715 fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
2719 vtn_fail("Invalid SPIR-V atomic");
2723 nir_ssa_def *offset, *index;
2724 offset = vtn_pointer_to_offset(b, ptr, &index);
2726 nir_intrinsic_op op;
2727 if (ptr->mode == vtn_variable_mode_ssbo) {
2728 op = get_ssbo_nir_atomic_op(b, opcode);
2730 vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
2731 b->options->lower_workgroup_access_to_offsets);
2732 op = get_shared_nir_atomic_op(b, opcode);
2735 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2739 case SpvOpAtomicLoad:
2740 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2741 if (ptr->mode == vtn_variable_mode_ssbo)
2742 atomic->src[src++] = nir_src_for_ssa(index);
2743 atomic->src[src++] = nir_src_for_ssa(offset);
2746 case SpvOpAtomicStore:
2747 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2748 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2749 atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2750 if (ptr->mode == vtn_variable_mode_ssbo)
2751 atomic->src[src++] = nir_src_for_ssa(index);
2752 atomic->src[src++] = nir_src_for_ssa(offset);
2755 case SpvOpAtomicExchange:
2756 case SpvOpAtomicCompareExchange:
2757 case SpvOpAtomicCompareExchangeWeak:
2758 case SpvOpAtomicIIncrement:
2759 case SpvOpAtomicIDecrement:
2760 case SpvOpAtomicIAdd:
2761 case SpvOpAtomicISub:
2762 case SpvOpAtomicSMin:
2763 case SpvOpAtomicUMin:
2764 case SpvOpAtomicSMax:
2765 case SpvOpAtomicUMax:
2766 case SpvOpAtomicAnd:
2768 case SpvOpAtomicXor:
2769 if (ptr->mode == vtn_variable_mode_ssbo)
2770 atomic->src[src++] = nir_src_for_ssa(index);
2771 atomic->src[src++] = nir_src_for_ssa(offset);
2772 fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
2776 vtn_fail("Invalid SPIR-V atomic");
2780 if (opcode != SpvOpAtomicStore) {
2781 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2783 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2784 glsl_get_vector_elements(type->type),
2785 glsl_get_bit_size(type->type), NULL);
2787 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2788 val->ssa = rzalloc(b, struct vtn_ssa_value);
2789 val->ssa->def = &atomic->dest.ssa;
2790 val->ssa->type = type->type;
2793 nir_builder_instr_insert(&b->nb, &atomic->instr);
2796 static nir_alu_instr *
2797 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2800 switch (num_components) {
2801 case 1: op = nir_op_fmov; break;
2802 case 2: op = nir_op_vec2; break;
2803 case 3: op = nir_op_vec3; break;
2804 case 4: op = nir_op_vec4; break;
2805 default: vtn_fail("bad vector size");
2808 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2809 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2811 vec->dest.write_mask = (1 << num_components) - 1;
2816 struct vtn_ssa_value *
2817 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2819 if (src->transposed)
2820 return src->transposed;
2822 struct vtn_ssa_value *dest =
2823 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2825 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2826 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2827 glsl_get_bit_size(src->type));
2828 if (glsl_type_is_vector_or_scalar(src->type)) {
2829 vec->src[0].src = nir_src_for_ssa(src->def);
2830 vec->src[0].swizzle[0] = i;
2832 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2833 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2834 vec->src[j].swizzle[0] = i;
2837 nir_builder_instr_insert(&b->nb, &vec->instr);
2838 dest->elems[i]->def = &vec->dest.dest.ssa;
2841 dest->transposed = src;
2847 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2849 unsigned swiz[4] = { index };
2850 return nir_swizzle(&b->nb, src, swiz, 1, true);
2854 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2857 nir_alu_instr *vec = create_vec(b, src->num_components,
2860 for (unsigned i = 0; i < src->num_components; i++) {
2862 vec->src[i].src = nir_src_for_ssa(insert);
2864 vec->src[i].src = nir_src_for_ssa(src);
2865 vec->src[i].swizzle[0] = i;
2869 nir_builder_instr_insert(&b->nb, &vec->instr);
2871 return &vec->dest.dest.ssa;
2875 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2878 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2879 for (unsigned i = 1; i < src->num_components; i++)
2880 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2881 vtn_vector_extract(b, src, i), dest);
2887 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2888 nir_ssa_def *insert, nir_ssa_def *index)
2890 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2891 for (unsigned i = 1; i < src->num_components; i++)
2892 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2893 vtn_vector_insert(b, src, insert, i), dest);
2898 static nir_ssa_def *
2899 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2900 nir_ssa_def *src0, nir_ssa_def *src1,
2901 const uint32_t *indices)
2903 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2905 for (unsigned i = 0; i < num_components; i++) {
2906 uint32_t index = indices[i];
2907 if (index == 0xffffffff) {
2909 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2910 } else if (index < src0->num_components) {
2911 vec->src[i].src = nir_src_for_ssa(src0);
2912 vec->src[i].swizzle[0] = index;
2914 vec->src[i].src = nir_src_for_ssa(src1);
2915 vec->src[i].swizzle[0] = index - src0->num_components;
2919 nir_builder_instr_insert(&b->nb, &vec->instr);
2921 return &vec->dest.dest.ssa;
2925 * Concatentates a number of vectors/scalars together to produce a vector
2927 static nir_ssa_def *
2928 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2929 unsigned num_srcs, nir_ssa_def **srcs)
2931 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2933 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2935 * "When constructing a vector, there must be at least two Constituent
2938 vtn_assert(num_srcs >= 2);
2940 unsigned dest_idx = 0;
2941 for (unsigned i = 0; i < num_srcs; i++) {
2942 nir_ssa_def *src = srcs[i];
2943 vtn_assert(dest_idx + src->num_components <= num_components);
2944 for (unsigned j = 0; j < src->num_components; j++) {
2945 vec->src[dest_idx].src = nir_src_for_ssa(src);
2946 vec->src[dest_idx].swizzle[0] = j;
2951 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2953 * "When constructing a vector, the total number of components in all
2954 * the operands must equal the number of components in Result Type."
2956 vtn_assert(dest_idx == num_components);
2958 nir_builder_instr_insert(&b->nb, &vec->instr);
2960 return &vec->dest.dest.ssa;
2963 static struct vtn_ssa_value *
2964 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2966 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2967 dest->type = src->type;
2969 if (glsl_type_is_vector_or_scalar(src->type)) {
2970 dest->def = src->def;
2972 unsigned elems = glsl_get_length(src->type);
2974 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2975 for (unsigned i = 0; i < elems; i++)
2976 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2982 static struct vtn_ssa_value *
2983 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2984 struct vtn_ssa_value *insert, const uint32_t *indices,
2985 unsigned num_indices)
2987 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2989 struct vtn_ssa_value *cur = dest;
2991 for (i = 0; i < num_indices - 1; i++) {
2992 cur = cur->elems[indices[i]];
2995 if (glsl_type_is_vector_or_scalar(cur->type)) {
2996 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2997 * the component granularity. In that case, the last index will be
2998 * the index to insert the scalar into the vector.
3001 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
3003 cur->elems[indices[i]] = insert;
3009 static struct vtn_ssa_value *
3010 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
3011 const uint32_t *indices, unsigned num_indices)
3013 struct vtn_ssa_value *cur = src;
3014 for (unsigned i = 0; i < num_indices; i++) {
3015 if (glsl_type_is_vector_or_scalar(cur->type)) {
3016 vtn_assert(i == num_indices - 1);
3017 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3018 * the component granularity. The last index will be the index of the
3019 * vector to extract.
3022 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
3023 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
3024 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
3027 cur = cur->elems[indices[i]];
3035 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
3036 const uint32_t *w, unsigned count)
3038 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
3039 const struct glsl_type *type =
3040 vtn_value(b, w[1], vtn_value_type_type)->type->type;
3041 val->ssa = vtn_create_ssa_value(b, type);
3044 case SpvOpVectorExtractDynamic:
3045 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
3046 vtn_ssa_value(b, w[4])->def);
3049 case SpvOpVectorInsertDynamic:
3050 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
3051 vtn_ssa_value(b, w[4])->def,
3052 vtn_ssa_value(b, w[5])->def);
3055 case SpvOpVectorShuffle:
3056 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
3057 vtn_ssa_value(b, w[3])->def,
3058 vtn_ssa_value(b, w[4])->def,
3062 case SpvOpCompositeConstruct: {
3063 unsigned elems = count - 3;
3065 if (glsl_type_is_vector_or_scalar(type)) {
3066 nir_ssa_def *srcs[4];
3067 for (unsigned i = 0; i < elems; i++)
3068 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
3070 vtn_vector_construct(b, glsl_get_vector_elements(type),
3073 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
3074 for (unsigned i = 0; i < elems; i++)
3075 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
3079 case SpvOpCompositeExtract:
3080 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
3084 case SpvOpCompositeInsert:
3085 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
3086 vtn_ssa_value(b, w[3]),
3090 case SpvOpCopyObject:
3091 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
3095 vtn_fail("unknown composite operation");
3100 vtn_emit_barrier(struct vtn_builder *b, nir_intrinsic_op op)
3102 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
3103 nir_builder_instr_insert(&b->nb, &intrin->instr);
3107 vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
3108 SpvMemorySemanticsMask semantics)
3110 static const SpvMemorySemanticsMask all_memory_semantics =
3111 SpvMemorySemanticsUniformMemoryMask |
3112 SpvMemorySemanticsWorkgroupMemoryMask |
3113 SpvMemorySemanticsAtomicCounterMemoryMask |
3114 SpvMemorySemanticsImageMemoryMask;
3116 /* If we're not actually doing a memory barrier, bail */
3117 if (!(semantics & all_memory_semantics))
3120 /* GL and Vulkan don't have these */
3121 vtn_assert(scope != SpvScopeCrossDevice);
3123 if (scope == SpvScopeSubgroup)
3124 return; /* Nothing to do here */
3126 if (scope == SpvScopeWorkgroup) {
3127 vtn_emit_barrier(b, nir_intrinsic_group_memory_barrier);
3131 /* There's only two scopes thing left */
3132 vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice);
3134 if ((semantics & all_memory_semantics) == all_memory_semantics) {
3135 vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
3139 /* Issue a bunch of more specific barriers */
3140 uint32_t bits = semantics;
3142 SpvMemorySemanticsMask semantic = 1 << u_bit_scan(&bits);
3144 case SpvMemorySemanticsUniformMemoryMask:
3145 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer);
3147 case SpvMemorySemanticsWorkgroupMemoryMask:
3148 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared);
3150 case SpvMemorySemanticsAtomicCounterMemoryMask:
3151 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter);
3153 case SpvMemorySemanticsImageMemoryMask:
3154 vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image);
3163 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
3164 const uint32_t *w, unsigned count)
3167 case SpvOpEmitVertex:
3168 case SpvOpEmitStreamVertex:
3169 case SpvOpEndPrimitive:
3170 case SpvOpEndStreamPrimitive: {
3171 nir_intrinsic_op intrinsic_op;
3173 case SpvOpEmitVertex:
3174 case SpvOpEmitStreamVertex:
3175 intrinsic_op = nir_intrinsic_emit_vertex;
3177 case SpvOpEndPrimitive:
3178 case SpvOpEndStreamPrimitive:
3179 intrinsic_op = nir_intrinsic_end_primitive;
3182 unreachable("Invalid opcode");
3185 nir_intrinsic_instr *intrin =
3186 nir_intrinsic_instr_create(b->shader, intrinsic_op);
3189 case SpvOpEmitStreamVertex:
3190 case SpvOpEndStreamPrimitive:
3191 nir_intrinsic_set_stream_id(intrin, w[1]);
3197 nir_builder_instr_insert(&b->nb, &intrin->instr);
3201 case SpvOpMemoryBarrier: {
3202 SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
3203 SpvMemorySemanticsMask semantics =
3204 vtn_constant_value(b, w[2])->values[0].u32[0];
3205 vtn_emit_memory_barrier(b, scope, semantics);
3209 case SpvOpControlBarrier: {
3210 SpvScope execution_scope =
3211 vtn_constant_value(b, w[1])->values[0].u32[0];
3212 if (execution_scope == SpvScopeWorkgroup)
3213 vtn_emit_barrier(b, nir_intrinsic_barrier);
3215 SpvScope memory_scope =
3216 vtn_constant_value(b, w[2])->values[0].u32[0];
3217 SpvMemorySemanticsMask memory_semantics =
3218 vtn_constant_value(b, w[3])->values[0].u32[0];
3219 vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
3224 unreachable("unknown barrier instruction");
3229 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
3230 SpvExecutionMode mode)
3233 case SpvExecutionModeInputPoints:
3234 case SpvExecutionModeOutputPoints:
3235 return 0; /* GL_POINTS */
3236 case SpvExecutionModeInputLines:
3237 return 1; /* GL_LINES */
3238 case SpvExecutionModeInputLinesAdjacency:
3239 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3240 case SpvExecutionModeTriangles:
3241 return 4; /* GL_TRIANGLES */
3242 case SpvExecutionModeInputTrianglesAdjacency:
3243 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3244 case SpvExecutionModeQuads:
3245 return 7; /* GL_QUADS */
3246 case SpvExecutionModeIsolines:
3247 return 0x8E7A; /* GL_ISOLINES */
3248 case SpvExecutionModeOutputLineStrip:
3249 return 3; /* GL_LINE_STRIP */
3250 case SpvExecutionModeOutputTriangleStrip:
3251 return 5; /* GL_TRIANGLE_STRIP */
3253 vtn_fail("Invalid primitive type");
3258 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
3259 SpvExecutionMode mode)
3262 case SpvExecutionModeInputPoints:
3264 case SpvExecutionModeInputLines:
3266 case SpvExecutionModeInputLinesAdjacency:
3268 case SpvExecutionModeTriangles:
3270 case SpvExecutionModeInputTrianglesAdjacency:
3273 vtn_fail("Invalid GS input mode");
3277 static gl_shader_stage
3278 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
3281 case SpvExecutionModelVertex:
3282 return MESA_SHADER_VERTEX;
3283 case SpvExecutionModelTessellationControl:
3284 return MESA_SHADER_TESS_CTRL;
3285 case SpvExecutionModelTessellationEvaluation:
3286 return MESA_SHADER_TESS_EVAL;
3287 case SpvExecutionModelGeometry:
3288 return MESA_SHADER_GEOMETRY;
3289 case SpvExecutionModelFragment:
3290 return MESA_SHADER_FRAGMENT;
3291 case SpvExecutionModelGLCompute:
3292 return MESA_SHADER_COMPUTE;
3294 vtn_fail("Unsupported execution model");
3298 #define spv_check_supported(name, cap) do { \
3299 if (!(b->options && b->options->caps.name)) \
3300 vtn_warn("Unsupported SPIR-V capability: %s", \
3301 spirv_capability_to_string(cap)); \
3306 vtn_handle_entry_point(struct vtn_builder *b, const uint32_t *w,
3309 struct vtn_value *entry_point = &b->values[w[2]];
3310 /* Let this be a name label regardless */
3311 unsigned name_words;
3312 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
3314 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
3315 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
3318 vtn_assert(b->entry_point == NULL);
3319 b->entry_point = entry_point;
3323 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
3324 const uint32_t *w, unsigned count)
3331 case SpvSourceLanguageUnknown: lang = "unknown"; break;
3332 case SpvSourceLanguageESSL: lang = "ESSL"; break;
3333 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
3334 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
3335 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
3336 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
3339 uint32_t version = w[2];
3342 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
3344 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
3348 case SpvOpSourceExtension:
3349 case SpvOpSourceContinued:
3350 case SpvOpExtension:
3351 case SpvOpModuleProcessed:
3352 /* Unhandled, but these are for debug so that's ok. */
3355 case SpvOpCapability: {
3356 SpvCapability cap = w[1];
3358 case SpvCapabilityMatrix:
3359 case SpvCapabilityShader:
3360 case SpvCapabilityGeometry:
3361 case SpvCapabilityGeometryPointSize:
3362 case SpvCapabilityUniformBufferArrayDynamicIndexing:
3363 case SpvCapabilitySampledImageArrayDynamicIndexing:
3364 case SpvCapabilityStorageBufferArrayDynamicIndexing:
3365 case SpvCapabilityStorageImageArrayDynamicIndexing:
3366 case SpvCapabilityImageRect:
3367 case SpvCapabilitySampledRect:
3368 case SpvCapabilitySampled1D:
3369 case SpvCapabilityImage1D:
3370 case SpvCapabilitySampledCubeArray:
3371 case SpvCapabilityImageCubeArray:
3372 case SpvCapabilitySampledBuffer:
3373 case SpvCapabilityImageBuffer:
3374 case SpvCapabilityImageQuery:
3375 case SpvCapabilityDerivativeControl:
3376 case SpvCapabilityInterpolationFunction:
3377 case SpvCapabilityMultiViewport:
3378 case SpvCapabilitySampleRateShading:
3379 case SpvCapabilityClipDistance:
3380 case SpvCapabilityCullDistance:
3381 case SpvCapabilityInputAttachment:
3382 case SpvCapabilityImageGatherExtended:
3383 case SpvCapabilityStorageImageExtendedFormats:
3386 case SpvCapabilityGeometryStreams:
3387 case SpvCapabilityLinkage:
3388 case SpvCapabilityVector16:
3389 case SpvCapabilityFloat16Buffer:
3390 case SpvCapabilityFloat16:
3391 case SpvCapabilityInt64Atomics:
3392 case SpvCapabilityStorageImageMultisample:
3393 case SpvCapabilityInt8:
3394 case SpvCapabilitySparseResidency:
3395 case SpvCapabilityMinLod:
3396 case SpvCapabilityTransformFeedback:
3397 vtn_warn("Unsupported SPIR-V capability: %s",
3398 spirv_capability_to_string(cap));
3401 case SpvCapabilityAtomicStorage:
3402 spv_check_supported(atomic_storage, cap);
3405 case SpvCapabilityFloat64:
3406 spv_check_supported(float64, cap);
3408 case SpvCapabilityInt64:
3409 spv_check_supported(int64, cap);
3411 case SpvCapabilityInt16:
3412 spv_check_supported(int16, cap);
3415 case SpvCapabilityAddresses:
3416 case SpvCapabilityKernel:
3417 case SpvCapabilityImageBasic:
3418 case SpvCapabilityImageReadWrite:
3419 case SpvCapabilityImageMipmap:
3420 case SpvCapabilityPipes:
3421 case SpvCapabilityGroups:
3422 case SpvCapabilityDeviceEnqueue:
3423 case SpvCapabilityLiteralSampler:
3424 case SpvCapabilityGenericPointer:
3425 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3426 spirv_capability_to_string(cap));
3429 case SpvCapabilityImageMSArray:
3430 spv_check_supported(image_ms_array, cap);
3433 case SpvCapabilityTessellation:
3434 case SpvCapabilityTessellationPointSize:
3435 spv_check_supported(tessellation, cap);
3438 case SpvCapabilityDrawParameters:
3439 spv_check_supported(draw_parameters, cap);
3442 case SpvCapabilityStorageImageReadWithoutFormat:
3443 spv_check_supported(image_read_without_format, cap);
3446 case SpvCapabilityStorageImageWriteWithoutFormat:
3447 spv_check_supported(image_write_without_format, cap);
3450 case SpvCapabilityDeviceGroup:
3451 spv_check_supported(device_group, cap);
3454 case SpvCapabilityMultiView:
3455 spv_check_supported(multiview, cap);
3458 case SpvCapabilityGroupNonUniform:
3459 spv_check_supported(subgroup_basic, cap);
3462 case SpvCapabilityGroupNonUniformVote:
3463 spv_check_supported(subgroup_vote, cap);
3466 case SpvCapabilitySubgroupBallotKHR:
3467 case SpvCapabilityGroupNonUniformBallot:
3468 spv_check_supported(subgroup_ballot, cap);
3471 case SpvCapabilityGroupNonUniformShuffle:
3472 case SpvCapabilityGroupNonUniformShuffleRelative:
3473 spv_check_supported(subgroup_shuffle, cap);
3476 case SpvCapabilityGroupNonUniformQuad:
3477 spv_check_supported(subgroup_quad, cap);
3480 case SpvCapabilityGroupNonUniformArithmetic:
3481 case SpvCapabilityGroupNonUniformClustered:
3482 spv_check_supported(subgroup_arithmetic, cap);
3485 case SpvCapabilityVariablePointersStorageBuffer:
3486 case SpvCapabilityVariablePointers:
3487 spv_check_supported(variable_pointers, cap);
3490 case SpvCapabilityStorageUniformBufferBlock16:
3491 case SpvCapabilityStorageUniform16:
3492 case SpvCapabilityStoragePushConstant16:
3493 case SpvCapabilityStorageInputOutput16:
3494 spv_check_supported(storage_16bit, cap);
3497 case SpvCapabilityShaderViewportIndexLayerEXT:
3498 spv_check_supported(shader_viewport_index_layer, cap);
3501 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT:
3502 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT:
3503 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT:
3504 spv_check_supported(descriptor_array_dynamic_indexing, cap);
3507 case SpvCapabilityRuntimeDescriptorArrayEXT:
3508 spv_check_supported(runtime_descriptor_array, cap);
3511 case SpvCapabilityStencilExportEXT:
3512 spv_check_supported(stencil_export, cap);
3516 vtn_fail("Unhandled capability");
3521 case SpvOpExtInstImport:
3522 vtn_handle_extension(b, opcode, w, count);
3525 case SpvOpMemoryModel:
3526 vtn_assert(w[1] == SpvAddressingModelLogical);
3527 vtn_assert(w[2] == SpvMemoryModelSimple ||
3528 w[2] == SpvMemoryModelGLSL450);
3531 case SpvOpEntryPoint:
3532 vtn_handle_entry_point(b, w, count);
3536 vtn_push_value(b, w[1], vtn_value_type_string)->str =
3537 vtn_string_literal(b, &w[2], count - 2, NULL);
3541 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
3544 case SpvOpMemberName:
3548 case SpvOpExecutionMode:
3549 case SpvOpDecorationGroup:
3551 case SpvOpMemberDecorate:
3552 case SpvOpGroupDecorate:
3553 case SpvOpGroupMemberDecorate:
3554 vtn_handle_decoration(b, opcode, w, count);
3558 return false; /* End of preamble */
3565 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
3566 const struct vtn_decoration *mode, void *data)
3568 vtn_assert(b->entry_point == entry_point);
3570 switch(mode->exec_mode) {
3571 case SpvExecutionModeOriginUpperLeft:
3572 case SpvExecutionModeOriginLowerLeft:
3573 b->origin_upper_left =
3574 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
3577 case SpvExecutionModeEarlyFragmentTests:
3578 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3579 b->shader->info.fs.early_fragment_tests = true;
3582 case SpvExecutionModeInvocations:
3583 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3584 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
3587 case SpvExecutionModeDepthReplacing:
3588 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3589 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
3591 case SpvExecutionModeDepthGreater:
3592 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3593 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
3595 case SpvExecutionModeDepthLess:
3596 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3597 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
3599 case SpvExecutionModeDepthUnchanged:
3600 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3601 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
3604 case SpvExecutionModeLocalSize:
3605 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
3606 b->shader->info.cs.local_size[0] = mode->literals[0];
3607 b->shader->info.cs.local_size[1] = mode->literals[1];
3608 b->shader->info.cs.local_size[2] = mode->literals[2];
3610 case SpvExecutionModeLocalSizeHint:
3611 break; /* Nothing to do with this */
3613 case SpvExecutionModeOutputVertices:
3614 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3615 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3616 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3618 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3619 b->shader->info.gs.vertices_out = mode->literals[0];
3623 case SpvExecutionModeInputPoints:
3624 case SpvExecutionModeInputLines:
3625 case SpvExecutionModeInputLinesAdjacency:
3626 case SpvExecutionModeTriangles:
3627 case SpvExecutionModeInputTrianglesAdjacency:
3628 case SpvExecutionModeQuads:
3629 case SpvExecutionModeIsolines:
3630 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3631 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3632 b->shader->info.tess.primitive_mode =
3633 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3635 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3636 b->shader->info.gs.vertices_in =
3637 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3641 case SpvExecutionModeOutputPoints:
3642 case SpvExecutionModeOutputLineStrip:
3643 case SpvExecutionModeOutputTriangleStrip:
3644 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3645 b->shader->info.gs.output_primitive =
3646 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3649 case SpvExecutionModeSpacingEqual:
3650 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3651 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3652 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3654 case SpvExecutionModeSpacingFractionalEven:
3655 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3656 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3657 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3659 case SpvExecutionModeSpacingFractionalOdd:
3660 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3661 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3662 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3664 case SpvExecutionModeVertexOrderCw:
3665 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3666 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3667 b->shader->info.tess.ccw = false;
3669 case SpvExecutionModeVertexOrderCcw:
3670 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3671 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3672 b->shader->info.tess.ccw = true;
3674 case SpvExecutionModePointMode:
3675 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3676 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3677 b->shader->info.tess.point_mode = true;
3680 case SpvExecutionModePixelCenterInteger:
3681 b->pixel_center_integer = true;
3684 case SpvExecutionModeXfb:
3685 vtn_fail("Unhandled execution mode");
3688 case SpvExecutionModeVecTypeHint:
3689 case SpvExecutionModeContractionOff:
3692 case SpvExecutionModeStencilRefReplacingEXT:
3693 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
3697 vtn_fail("Unhandled execution mode");
3702 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3703 const uint32_t *w, unsigned count)
3705 vtn_set_instruction_result_type(b, opcode, w, count);
3709 case SpvOpSourceContinued:
3710 case SpvOpSourceExtension:
3711 case SpvOpExtension:
3712 case SpvOpCapability:
3713 case SpvOpExtInstImport:
3714 case SpvOpMemoryModel:
3715 case SpvOpEntryPoint:
3716 case SpvOpExecutionMode:
3719 case SpvOpMemberName:
3720 case SpvOpDecorationGroup:
3722 case SpvOpMemberDecorate:
3723 case SpvOpGroupDecorate:
3724 case SpvOpGroupMemberDecorate:
3725 vtn_fail("Invalid opcode types and variables section");
3731 case SpvOpTypeFloat:
3732 case SpvOpTypeVector:
3733 case SpvOpTypeMatrix:
3734 case SpvOpTypeImage:
3735 case SpvOpTypeSampler:
3736 case SpvOpTypeSampledImage:
3737 case SpvOpTypeArray:
3738 case SpvOpTypeRuntimeArray:
3739 case SpvOpTypeStruct:
3740 case SpvOpTypeOpaque:
3741 case SpvOpTypePointer:
3742 case SpvOpTypeFunction:
3743 case SpvOpTypeEvent:
3744 case SpvOpTypeDeviceEvent:
3745 case SpvOpTypeReserveId:
3746 case SpvOpTypeQueue:
3748 vtn_handle_type(b, opcode, w, count);
3751 case SpvOpConstantTrue:
3752 case SpvOpConstantFalse:
3754 case SpvOpConstantComposite:
3755 case SpvOpConstantSampler:
3756 case SpvOpConstantNull:
3757 case SpvOpSpecConstantTrue:
3758 case SpvOpSpecConstantFalse:
3759 case SpvOpSpecConstant:
3760 case SpvOpSpecConstantComposite:
3761 case SpvOpSpecConstantOp:
3762 vtn_handle_constant(b, opcode, w, count);
3767 vtn_handle_variables(b, opcode, w, count);
3771 return false; /* End of preamble */
3778 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3779 const uint32_t *w, unsigned count)
3785 case SpvOpLoopMerge:
3786 case SpvOpSelectionMerge:
3787 /* This is handled by cfg pre-pass and walk_blocks */
3791 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3792 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3797 vtn_handle_extension(b, opcode, w, count);
3803 case SpvOpCopyMemory:
3804 case SpvOpCopyMemorySized:
3805 case SpvOpAccessChain:
3806 case SpvOpPtrAccessChain:
3807 case SpvOpInBoundsAccessChain:
3808 case SpvOpArrayLength:
3809 vtn_handle_variables(b, opcode, w, count);
3812 case SpvOpFunctionCall:
3813 vtn_handle_function_call(b, opcode, w, count);
3816 case SpvOpSampledImage:
3818 case SpvOpImageSampleImplicitLod:
3819 case SpvOpImageSampleExplicitLod:
3820 case SpvOpImageSampleDrefImplicitLod:
3821 case SpvOpImageSampleDrefExplicitLod:
3822 case SpvOpImageSampleProjImplicitLod:
3823 case SpvOpImageSampleProjExplicitLod:
3824 case SpvOpImageSampleProjDrefImplicitLod:
3825 case SpvOpImageSampleProjDrefExplicitLod:
3826 case SpvOpImageFetch:
3827 case SpvOpImageGather:
3828 case SpvOpImageDrefGather:
3829 case SpvOpImageQuerySizeLod:
3830 case SpvOpImageQueryLod:
3831 case SpvOpImageQueryLevels:
3832 case SpvOpImageQuerySamples:
3833 vtn_handle_texture(b, opcode, w, count);
3836 case SpvOpImageRead:
3837 case SpvOpImageWrite:
3838 case SpvOpImageTexelPointer:
3839 vtn_handle_image(b, opcode, w, count);
3842 case SpvOpImageQuerySize: {
3843 struct vtn_pointer *image =
3844 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3845 if (glsl_type_is_image(image->type->type)) {
3846 vtn_handle_image(b, opcode, w, count);
3848 vtn_assert(glsl_type_is_sampler(image->type->type));
3849 vtn_handle_texture(b, opcode, w, count);
3854 case SpvOpAtomicLoad:
3855 case SpvOpAtomicExchange:
3856 case SpvOpAtomicCompareExchange:
3857 case SpvOpAtomicCompareExchangeWeak:
3858 case SpvOpAtomicIIncrement:
3859 case SpvOpAtomicIDecrement:
3860 case SpvOpAtomicIAdd:
3861 case SpvOpAtomicISub:
3862 case SpvOpAtomicSMin:
3863 case SpvOpAtomicUMin:
3864 case SpvOpAtomicSMax:
3865 case SpvOpAtomicUMax:
3866 case SpvOpAtomicAnd:
3868 case SpvOpAtomicXor: {
3869 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3870 if (pointer->value_type == vtn_value_type_image_pointer) {
3871 vtn_handle_image(b, opcode, w, count);
3873 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3874 vtn_handle_atomics(b, opcode, w, count);
3879 case SpvOpAtomicStore: {
3880 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3881 if (pointer->value_type == vtn_value_type_image_pointer) {
3882 vtn_handle_image(b, opcode, w, count);
3884 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3885 vtn_handle_atomics(b, opcode, w, count);
3891 /* Handle OpSelect up-front here because it needs to be able to handle
3892 * pointers and not just regular vectors and scalars.
3894 struct vtn_value *res_val = vtn_untyped_value(b, w[2]);
3895 struct vtn_value *sel_val = vtn_untyped_value(b, w[3]);
3896 struct vtn_value *obj1_val = vtn_untyped_value(b, w[4]);
3897 struct vtn_value *obj2_val = vtn_untyped_value(b, w[5]);
3899 const struct glsl_type *sel_type;
3900 switch (res_val->type->base_type) {
3901 case vtn_base_type_scalar:
3902 sel_type = glsl_bool_type();
3904 case vtn_base_type_vector:
3905 sel_type = glsl_vector_type(GLSL_TYPE_BOOL, res_val->type->length);
3907 case vtn_base_type_pointer:
3908 /* We need to have actual storage for pointer types */
3909 vtn_fail_if(res_val->type->type == NULL,
3910 "Invalid pointer result type for OpSelect");
3911 sel_type = glsl_bool_type();
3914 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3917 if (unlikely(sel_val->type->type != sel_type)) {
3918 if (sel_val->type->type == glsl_bool_type()) {
3919 /* This case is illegal but some older versions of GLSLang produce
3920 * it. The GLSLang issue was fixed on March 30, 2017:
3922 * https://github.com/KhronosGroup/glslang/issues/809
3924 * Unfortunately, there are applications in the wild which are
3925 * shipping with this bug so it isn't nice to fail on them so we
3926 * throw a warning instead. It's not actually a problem for us as
3927 * nir_builder will just splat the condition out which is most
3928 * likely what the client wanted anyway.
3930 vtn_warn("Condition type of OpSelect must have the same number "
3931 "of components as Result Type");
3933 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3934 "of Boolean type. It must have the same number of "
3935 "components as Result Type");
3939 vtn_fail_if(obj1_val->type != res_val->type ||
3940 obj2_val->type != res_val->type,
3941 "Object types must match the result type in OpSelect");
3943 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3944 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3945 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3946 vtn_ssa_value(b, w[4])->def,
3947 vtn_ssa_value(b, w[5])->def);
3948 vtn_push_ssa(b, w[2], res_type, ssa);
3957 case SpvOpConvertFToU:
3958 case SpvOpConvertFToS:
3959 case SpvOpConvertSToF:
3960 case SpvOpConvertUToF:
3964 case SpvOpQuantizeToF16:
3965 case SpvOpConvertPtrToU:
3966 case SpvOpConvertUToPtr:
3967 case SpvOpPtrCastToGeneric:
3968 case SpvOpGenericCastToPtr:
3974 case SpvOpSignBitSet:
3975 case SpvOpLessOrGreater:
3977 case SpvOpUnordered:
3992 case SpvOpVectorTimesScalar:
3994 case SpvOpIAddCarry:
3995 case SpvOpISubBorrow:
3996 case SpvOpUMulExtended:
3997 case SpvOpSMulExtended:
3998 case SpvOpShiftRightLogical:
3999 case SpvOpShiftRightArithmetic:
4000 case SpvOpShiftLeftLogical:
4001 case SpvOpLogicalEqual:
4002 case SpvOpLogicalNotEqual:
4003 case SpvOpLogicalOr:
4004 case SpvOpLogicalAnd:
4005 case SpvOpLogicalNot:
4006 case SpvOpBitwiseOr:
4007 case SpvOpBitwiseXor:
4008 case SpvOpBitwiseAnd:
4010 case SpvOpFOrdEqual:
4011 case SpvOpFUnordEqual:
4012 case SpvOpINotEqual:
4013 case SpvOpFOrdNotEqual:
4014 case SpvOpFUnordNotEqual:
4015 case SpvOpULessThan:
4016 case SpvOpSLessThan:
4017 case SpvOpFOrdLessThan:
4018 case SpvOpFUnordLessThan:
4019 case SpvOpUGreaterThan:
4020 case SpvOpSGreaterThan:
4021 case SpvOpFOrdGreaterThan:
4022 case SpvOpFUnordGreaterThan:
4023 case SpvOpULessThanEqual:
4024 case SpvOpSLessThanEqual:
4025 case SpvOpFOrdLessThanEqual:
4026 case SpvOpFUnordLessThanEqual:
4027 case SpvOpUGreaterThanEqual:
4028 case SpvOpSGreaterThanEqual:
4029 case SpvOpFOrdGreaterThanEqual:
4030 case SpvOpFUnordGreaterThanEqual:
4036 case SpvOpFwidthFine:
4037 case SpvOpDPdxCoarse:
4038 case SpvOpDPdyCoarse:
4039 case SpvOpFwidthCoarse:
4040 case SpvOpBitFieldInsert:
4041 case SpvOpBitFieldSExtract:
4042 case SpvOpBitFieldUExtract:
4043 case SpvOpBitReverse:
4045 case SpvOpTranspose:
4046 case SpvOpOuterProduct:
4047 case SpvOpMatrixTimesScalar:
4048 case SpvOpVectorTimesMatrix:
4049 case SpvOpMatrixTimesVector:
4050 case SpvOpMatrixTimesMatrix:
4051 vtn_handle_alu(b, opcode, w, count);
4054 case SpvOpVectorExtractDynamic:
4055 case SpvOpVectorInsertDynamic:
4056 case SpvOpVectorShuffle:
4057 case SpvOpCompositeConstruct:
4058 case SpvOpCompositeExtract:
4059 case SpvOpCompositeInsert:
4060 case SpvOpCopyObject:
4061 vtn_handle_composite(b, opcode, w, count);
4064 case SpvOpEmitVertex:
4065 case SpvOpEndPrimitive:
4066 case SpvOpEmitStreamVertex:
4067 case SpvOpEndStreamPrimitive:
4068 case SpvOpControlBarrier:
4069 case SpvOpMemoryBarrier:
4070 vtn_handle_barrier(b, opcode, w, count);
4073 case SpvOpGroupNonUniformElect:
4074 case SpvOpGroupNonUniformAll:
4075 case SpvOpGroupNonUniformAny:
4076 case SpvOpGroupNonUniformAllEqual:
4077 case SpvOpGroupNonUniformBroadcast:
4078 case SpvOpGroupNonUniformBroadcastFirst:
4079 case SpvOpGroupNonUniformBallot:
4080 case SpvOpGroupNonUniformInverseBallot:
4081 case SpvOpGroupNonUniformBallotBitExtract:
4082 case SpvOpGroupNonUniformBallotBitCount:
4083 case SpvOpGroupNonUniformBallotFindLSB:
4084 case SpvOpGroupNonUniformBallotFindMSB:
4085 case SpvOpGroupNonUniformShuffle:
4086 case SpvOpGroupNonUniformShuffleXor:
4087 case SpvOpGroupNonUniformShuffleUp:
4088 case SpvOpGroupNonUniformShuffleDown:
4089 case SpvOpGroupNonUniformIAdd:
4090 case SpvOpGroupNonUniformFAdd:
4091 case SpvOpGroupNonUniformIMul:
4092 case SpvOpGroupNonUniformFMul:
4093 case SpvOpGroupNonUniformSMin:
4094 case SpvOpGroupNonUniformUMin:
4095 case SpvOpGroupNonUniformFMin:
4096 case SpvOpGroupNonUniformSMax:
4097 case SpvOpGroupNonUniformUMax:
4098 case SpvOpGroupNonUniformFMax:
4099 case SpvOpGroupNonUniformBitwiseAnd:
4100 case SpvOpGroupNonUniformBitwiseOr:
4101 case SpvOpGroupNonUniformBitwiseXor:
4102 case SpvOpGroupNonUniformLogicalAnd:
4103 case SpvOpGroupNonUniformLogicalOr:
4104 case SpvOpGroupNonUniformLogicalXor:
4105 case SpvOpGroupNonUniformQuadBroadcast:
4106 case SpvOpGroupNonUniformQuadSwap:
4107 vtn_handle_subgroup(b, opcode, w, count);
4111 vtn_fail("Unhandled opcode");
4118 vtn_create_builder(const uint32_t *words, size_t word_count,
4119 gl_shader_stage stage, const char *entry_point_name,
4120 const struct spirv_to_nir_options *options)
4122 /* Initialize the vtn_builder object */
4123 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
4125 b->spirv_word_count = word_count;
4129 exec_list_make_empty(&b->functions);
4130 b->entry_point_stage = stage;
4131 b->entry_point_name = entry_point_name;
4132 b->options = options;
4135 * Handle the SPIR-V header (first 5 dwords).
4136 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4138 if (word_count <= 5)
4141 if (words[0] != SpvMagicNumber) {
4142 vtn_err("words[0] was 0x%x, want 0x%x", words[0], SpvMagicNumber);
4145 if (words[1] < 0x10000) {
4146 vtn_err("words[1] was 0x%x, want >= 0x10000", words[1]);
4150 /* words[2] == generator magic */
4151 unsigned value_id_bound = words[3];
4152 if (words[4] != 0) {
4153 vtn_err("words[4] was %u, want 0", words[4]);
4157 b->value_id_bound = value_id_bound;
4158 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
4167 spirv_to_nir(const uint32_t *words, size_t word_count,
4168 struct nir_spirv_specialization *spec, unsigned num_spec,
4169 gl_shader_stage stage, const char *entry_point_name,
4170 const struct spirv_to_nir_options *options,
4171 const nir_shader_compiler_options *nir_options)
4174 const uint32_t *word_end = words + word_count;
4176 struct vtn_builder *b = vtn_create_builder(words, word_count,
4177 stage, entry_point_name,
4183 /* See also _vtn_fail() */
4184 if (setjmp(b->fail_jump)) {
4189 /* Skip the SPIR-V header, handled at vtn_create_builder */
4192 /* Handle all the preamble instructions */
4193 words = vtn_foreach_instruction(b, words, word_end,
4194 vtn_handle_preamble_instruction);
4196 if (b->entry_point == NULL) {
4197 vtn_fail("Entry point not found");
4202 b->shader = nir_shader_create(b, stage, nir_options, NULL);
4204 /* Set shader info defaults */
4205 b->shader->info.gs.invocations = 1;
4207 /* Parse execution modes */
4208 vtn_foreach_execution_mode(b, b->entry_point,
4209 vtn_handle_execution_mode, NULL);
4211 b->specializations = spec;
4212 b->num_specializations = num_spec;
4214 /* Handle all variable, type, and constant instructions */
4215 words = vtn_foreach_instruction(b, words, word_end,
4216 vtn_handle_variable_or_type_instruction);
4218 /* Set types on all vtn_values */
4219 vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
4221 vtn_build_cfg(b, words, word_end);
4223 assert(b->entry_point->value_type == vtn_value_type_function);
4224 b->entry_point->func->referenced = true;
4229 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
4230 if (func->referenced && !func->emitted) {
4231 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
4232 _mesa_key_pointer_equal);
4234 vtn_function_emit(b, func, vtn_handle_body_instruction);
4240 /* We sometimes generate bogus derefs that, while never used, give the
4241 * validator a bit of heartburn. Run dead code to get rid of them.
4243 nir_opt_dce(b->shader);
4245 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
4246 nir_function *entry_point = b->entry_point->func->impl->function;
4247 vtn_assert(entry_point);
4249 /* Unparent the shader from the vtn_builder before we delete the builder */
4250 ralloc_steal(NULL, b->shader);