2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
35 _vtn_warn(const char *file
, int line
, const char *msg
, ...)
41 formatted
= ralloc_vasprintf(NULL
, msg
, args
);
44 fprintf(stderr
, "%s:%d WARNING: %s\n", file
, line
, formatted
);
46 ralloc_free(formatted
);
49 static struct vtn_ssa_value
*
50 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
52 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
55 if (glsl_type_is_vector_or_scalar(type
)) {
56 unsigned num_components
= glsl_get_vector_elements(val
->type
);
57 unsigned bit_size
= glsl_get_bit_size(val
->type
);
58 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
60 unsigned elems
= glsl_get_length(val
->type
);
61 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
62 if (glsl_type_is_matrix(type
)) {
63 const struct glsl_type
*elem_type
=
64 glsl_vector_type(glsl_get_base_type(type
),
65 glsl_get_vector_elements(type
));
67 for (unsigned i
= 0; i
< elems
; i
++)
68 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
69 } else if (glsl_type_is_array(type
)) {
70 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
71 for (unsigned i
= 0; i
< elems
; i
++)
72 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
74 for (unsigned i
= 0; i
< elems
; i
++) {
75 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
76 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
84 static struct vtn_ssa_value
*
85 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
86 const struct glsl_type
*type
)
88 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
93 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
96 switch (glsl_get_base_type(type
)) {
100 case GLSL_TYPE_FLOAT
:
101 case GLSL_TYPE_DOUBLE
:
102 if (glsl_type_is_vector_or_scalar(type
)) {
103 unsigned num_components
= glsl_get_vector_elements(val
->type
);
104 nir_load_const_instr
*load
=
105 nir_load_const_instr_create(b
->shader
, num_components
, 32);
107 for (unsigned i
= 0; i
< num_components
; i
++)
108 load
->value
.u32
[i
] = constant
->value
.u
[i
];
110 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
111 val
->def
= &load
->def
;
113 assert(glsl_type_is_matrix(type
));
114 unsigned rows
= glsl_get_vector_elements(val
->type
);
115 unsigned columns
= glsl_get_matrix_columns(val
->type
);
116 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
118 for (unsigned i
= 0; i
< columns
; i
++) {
119 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
120 col_val
->type
= glsl_get_column_type(val
->type
);
121 nir_load_const_instr
*load
=
122 nir_load_const_instr_create(b
->shader
, rows
, 32);
124 for (unsigned j
= 0; j
< rows
; j
++)
125 load
->value
.u32
[j
] = constant
->value
.u
[rows
* i
+ j
];
127 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
128 col_val
->def
= &load
->def
;
130 val
->elems
[i
] = col_val
;
135 case GLSL_TYPE_ARRAY
: {
136 unsigned elems
= glsl_get_length(val
->type
);
137 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
138 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
139 for (unsigned i
= 0; i
< elems
; i
++)
140 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
145 case GLSL_TYPE_STRUCT
: {
146 unsigned elems
= glsl_get_length(val
->type
);
147 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
148 for (unsigned i
= 0; i
< elems
; i
++) {
149 const struct glsl_type
*elem_type
=
150 glsl_get_struct_field(val
->type
, i
);
151 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
158 unreachable("bad constant type");
164 struct vtn_ssa_value
*
165 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
167 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
168 switch (val
->value_type
) {
169 case vtn_value_type_undef
:
170 return vtn_undef_ssa_value(b
, val
->type
->type
);
172 case vtn_value_type_constant
:
173 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
175 case vtn_value_type_ssa
:
178 case vtn_value_type_access_chain
:
179 /* This is needed for function parameters */
180 return vtn_variable_load(b
, val
->access_chain
);
183 unreachable("Invalid type for an SSA value");
188 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
189 unsigned word_count
, unsigned *words_used
)
191 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
193 /* Ammount of space taken by the string (including the null) */
194 unsigned len
= strlen(dup
) + 1;
195 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
201 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
202 const uint32_t *end
, vtn_instruction_handler handler
)
208 const uint32_t *w
= start
;
210 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
211 unsigned count
= w
[0] >> SpvWordCountShift
;
212 assert(count
>= 1 && w
+ count
<= end
);
216 break; /* Do nothing */
219 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
231 if (!handler(b
, opcode
, w
, count
))
243 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
244 const uint32_t *w
, unsigned count
)
247 case SpvOpExtInstImport
: {
248 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
249 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
250 val
->ext_handler
= vtn_handle_glsl450_instruction
;
252 assert(!"Unsupported extension");
258 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
259 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
266 unreachable("Unhandled opcode");
271 _foreach_decoration_helper(struct vtn_builder
*b
,
272 struct vtn_value
*base_value
,
274 struct vtn_value
*value
,
275 vtn_decoration_foreach_cb cb
, void *data
)
277 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
279 if (dec
->scope
== VTN_DEC_DECORATION
) {
280 member
= parent_member
;
281 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
282 assert(parent_member
== -1);
283 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
285 /* Not a decoration */
290 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
291 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
294 cb(b
, base_value
, member
, dec
, data
);
299 /** Iterates (recursively if needed) over all of the decorations on a value
301 * This function iterates over all of the decorations applied to a given
302 * value. If it encounters a decoration group, it recurses into the group
303 * and iterates over all of those decorations as well.
306 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
307 vtn_decoration_foreach_cb cb
, void *data
)
309 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
313 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
314 vtn_execution_mode_foreach_cb cb
, void *data
)
316 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
317 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
320 assert(dec
->group
== NULL
);
321 cb(b
, value
, dec
, data
);
326 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
327 const uint32_t *w
, unsigned count
)
329 const uint32_t *w_end
= w
+ count
;
330 const uint32_t target
= w
[1];
334 case SpvOpDecorationGroup
:
335 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
339 case SpvOpMemberDecorate
:
340 case SpvOpExecutionMode
: {
341 struct vtn_value
*val
= &b
->values
[target
];
343 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
346 dec
->scope
= VTN_DEC_DECORATION
;
348 case SpvOpMemberDecorate
:
349 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
351 case SpvOpExecutionMode
:
352 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
355 unreachable("Invalid decoration opcode");
357 dec
->decoration
= *(w
++);
360 /* Link into the list */
361 dec
->next
= val
->decoration
;
362 val
->decoration
= dec
;
366 case SpvOpGroupMemberDecorate
:
367 case SpvOpGroupDecorate
: {
368 struct vtn_value
*group
=
369 vtn_value(b
, target
, vtn_value_type_decoration_group
);
371 for (; w
< w_end
; w
++) {
372 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
373 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
376 if (opcode
== SpvOpGroupDecorate
) {
377 dec
->scope
= VTN_DEC_DECORATION
;
379 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
382 /* Link into the list */
383 dec
->next
= val
->decoration
;
384 val
->decoration
= dec
;
390 unreachable("Unhandled opcode");
394 struct member_decoration_ctx
{
396 struct glsl_struct_field
*fields
;
397 struct vtn_type
*type
;
400 /* does a shallow copy of a vtn_type */
402 static struct vtn_type
*
403 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
405 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
406 dest
->type
= src
->type
;
407 dest
->is_builtin
= src
->is_builtin
;
409 dest
->builtin
= src
->builtin
;
411 if (!glsl_type_is_scalar(src
->type
)) {
412 switch (glsl_get_base_type(src
->type
)) {
416 case GLSL_TYPE_FLOAT
:
417 case GLSL_TYPE_DOUBLE
:
418 case GLSL_TYPE_ARRAY
:
419 dest
->row_major
= src
->row_major
;
420 dest
->stride
= src
->stride
;
421 dest
->array_element
= src
->array_element
;
424 case GLSL_TYPE_STRUCT
: {
425 unsigned elems
= glsl_get_length(src
->type
);
427 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
428 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
430 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
431 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
436 unreachable("unhandled type");
443 static struct vtn_type
*
444 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
446 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
447 type
= type
->members
[member
];
449 /* We may have an array of matrices.... Oh, joy! */
450 while (glsl_type_is_array(type
->type
)) {
451 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
452 type
= type
->array_element
;
455 assert(glsl_type_is_matrix(type
->type
));
461 struct_member_decoration_cb(struct vtn_builder
*b
,
462 struct vtn_value
*val
, int member
,
463 const struct vtn_decoration
*dec
, void *void_ctx
)
465 struct member_decoration_ctx
*ctx
= void_ctx
;
470 assert(member
< ctx
->num_fields
);
472 switch (dec
->decoration
) {
473 case SpvDecorationNonWritable
:
474 case SpvDecorationNonReadable
:
475 case SpvDecorationRelaxedPrecision
:
476 case SpvDecorationVolatile
:
477 case SpvDecorationCoherent
:
478 case SpvDecorationUniform
:
479 break; /* FIXME: Do nothing with this for now. */
480 case SpvDecorationNoPerspective
:
481 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_NOPERSPECTIVE
;
483 case SpvDecorationFlat
:
484 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_FLAT
;
486 case SpvDecorationCentroid
:
487 ctx
->fields
[member
].centroid
= true;
489 case SpvDecorationSample
:
490 ctx
->fields
[member
].sample
= true;
492 case SpvDecorationStream
:
493 /* Vulkan only allows one GS stream */
494 assert(dec
->literals
[0] == 0);
496 case SpvDecorationLocation
:
497 ctx
->fields
[member
].location
= dec
->literals
[0];
499 case SpvDecorationComponent
:
500 break; /* FIXME: What should we do with these? */
501 case SpvDecorationBuiltIn
:
502 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
503 ctx
->type
->members
[member
]->is_builtin
= true;
504 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
505 ctx
->type
->builtin_block
= true;
507 case SpvDecorationOffset
:
508 ctx
->type
->offsets
[member
] = dec
->literals
[0];
510 case SpvDecorationMatrixStride
:
511 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
513 case SpvDecorationColMajor
:
514 break; /* Nothing to do here. Column-major is the default. */
515 case SpvDecorationRowMajor
:
516 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
519 case SpvDecorationPatch
:
520 vtn_warn("Tessellation not yet supported");
523 case SpvDecorationSpecId
:
524 case SpvDecorationBlock
:
525 case SpvDecorationBufferBlock
:
526 case SpvDecorationArrayStride
:
527 case SpvDecorationGLSLShared
:
528 case SpvDecorationGLSLPacked
:
529 case SpvDecorationInvariant
:
530 case SpvDecorationRestrict
:
531 case SpvDecorationAliased
:
532 case SpvDecorationConstant
:
533 case SpvDecorationIndex
:
534 case SpvDecorationBinding
:
535 case SpvDecorationDescriptorSet
:
536 case SpvDecorationLinkageAttributes
:
537 case SpvDecorationNoContraction
:
538 case SpvDecorationInputAttachmentIndex
:
539 vtn_warn("Decoration not allowed on struct members: %s",
540 spirv_decoration_to_string(dec
->decoration
));
543 case SpvDecorationXfbBuffer
:
544 case SpvDecorationXfbStride
:
545 vtn_warn("Vulkan does not have transform feedback");
548 case SpvDecorationCPacked
:
549 case SpvDecorationSaturatedConversion
:
550 case SpvDecorationFuncParamAttr
:
551 case SpvDecorationFPRoundingMode
:
552 case SpvDecorationFPFastMathMode
:
553 case SpvDecorationAlignment
:
554 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
555 spirv_decoration_to_string(dec
->decoration
));
561 type_decoration_cb(struct vtn_builder
*b
,
562 struct vtn_value
*val
, int member
,
563 const struct vtn_decoration
*dec
, void *ctx
)
565 struct vtn_type
*type
= val
->type
;
570 switch (dec
->decoration
) {
571 case SpvDecorationArrayStride
:
572 type
->stride
= dec
->literals
[0];
574 case SpvDecorationBlock
:
577 case SpvDecorationBufferBlock
:
578 type
->buffer_block
= true;
580 case SpvDecorationGLSLShared
:
581 case SpvDecorationGLSLPacked
:
582 /* Ignore these, since we get explicit offsets anyways */
585 case SpvDecorationRowMajor
:
586 case SpvDecorationColMajor
:
587 case SpvDecorationMatrixStride
:
588 case SpvDecorationBuiltIn
:
589 case SpvDecorationNoPerspective
:
590 case SpvDecorationFlat
:
591 case SpvDecorationPatch
:
592 case SpvDecorationCentroid
:
593 case SpvDecorationSample
:
594 case SpvDecorationVolatile
:
595 case SpvDecorationCoherent
:
596 case SpvDecorationNonWritable
:
597 case SpvDecorationNonReadable
:
598 case SpvDecorationUniform
:
599 case SpvDecorationStream
:
600 case SpvDecorationLocation
:
601 case SpvDecorationComponent
:
602 case SpvDecorationOffset
:
603 case SpvDecorationXfbBuffer
:
604 case SpvDecorationXfbStride
:
605 vtn_warn("Decoraiton only allowed for struct members: %s",
606 spirv_decoration_to_string(dec
->decoration
));
609 case SpvDecorationRelaxedPrecision
:
610 case SpvDecorationSpecId
:
611 case SpvDecorationInvariant
:
612 case SpvDecorationRestrict
:
613 case SpvDecorationAliased
:
614 case SpvDecorationConstant
:
615 case SpvDecorationIndex
:
616 case SpvDecorationBinding
:
617 case SpvDecorationDescriptorSet
:
618 case SpvDecorationLinkageAttributes
:
619 case SpvDecorationNoContraction
:
620 case SpvDecorationInputAttachmentIndex
:
621 vtn_warn("Decoraiton not allowed on types: %s",
622 spirv_decoration_to_string(dec
->decoration
));
625 case SpvDecorationCPacked
:
626 case SpvDecorationSaturatedConversion
:
627 case SpvDecorationFuncParamAttr
:
628 case SpvDecorationFPRoundingMode
:
629 case SpvDecorationFPFastMathMode
:
630 case SpvDecorationAlignment
:
631 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
632 spirv_decoration_to_string(dec
->decoration
));
638 translate_image_format(SpvImageFormat format
)
641 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
642 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
643 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
644 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
645 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
646 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
647 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
648 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
649 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
650 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
651 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
652 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
653 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
654 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
655 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
656 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
657 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
658 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
659 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
660 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
661 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
662 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
663 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
664 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
665 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
666 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
667 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
668 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
669 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
670 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
671 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
672 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
673 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
674 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
675 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
676 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
677 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
678 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
679 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
680 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
682 assert(!"Invalid image format");
688 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
689 const uint32_t *w
, unsigned count
)
691 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
693 val
->type
= rzalloc(b
, struct vtn_type
);
694 val
->type
->is_builtin
= false;
695 val
->type
->val
= val
;
699 val
->type
->type
= glsl_void_type();
702 val
->type
->type
= glsl_bool_type();
705 const bool signedness
= w
[3];
706 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
710 val
->type
->type
= glsl_float_type();
713 case SpvOpTypeVector
: {
714 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
715 unsigned elems
= w
[3];
717 assert(glsl_type_is_scalar(base
->type
));
718 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
720 /* Vectors implicitly have sizeof(base_type) stride. For now, this
721 * is always 4 bytes. This will have to change if we want to start
722 * supporting doubles or half-floats.
724 val
->type
->stride
= 4;
725 val
->type
->array_element
= base
;
729 case SpvOpTypeMatrix
: {
730 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
731 unsigned columns
= w
[3];
733 assert(glsl_type_is_vector(base
->type
));
734 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
735 glsl_get_vector_elements(base
->type
),
737 assert(!glsl_type_is_error(val
->type
->type
));
738 val
->type
->array_element
= base
;
739 val
->type
->row_major
= false;
740 val
->type
->stride
= 0;
744 case SpvOpTypeRuntimeArray
:
745 case SpvOpTypeArray
: {
746 struct vtn_type
*array_element
=
747 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
750 if (opcode
== SpvOpTypeRuntimeArray
) {
751 /* A length of 0 is used to denote unsized arrays */
755 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->value
.u
[0];
758 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
759 val
->type
->array_element
= array_element
;
760 val
->type
->stride
= 0;
764 case SpvOpTypeStruct
: {
765 unsigned num_fields
= count
- 2;
766 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
767 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
769 NIR_VLA(struct glsl_struct_field
, fields
, count
);
770 for (unsigned i
= 0; i
< num_fields
; i
++) {
771 val
->type
->members
[i
] =
772 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
773 fields
[i
] = (struct glsl_struct_field
) {
774 .type
= val
->type
->members
[i
]->type
,
775 .name
= ralloc_asprintf(b
, "field%d", i
),
780 struct member_decoration_ctx ctx
= {
781 .num_fields
= num_fields
,
786 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
788 const char *name
= val
->name
? val
->name
: "struct";
790 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
794 case SpvOpTypeFunction
: {
795 const struct glsl_type
*return_type
=
796 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
797 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
798 for (unsigned i
= 0; i
< count
- 3; i
++) {
799 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
803 params
[i
].out
= true;
805 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
809 case SpvOpTypePointer
:
810 /* FIXME: For now, we'll just do the really lame thing and return
811 * the same type. The validator should ensure that the proper number
812 * of dereferences happen
814 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
817 case SpvOpTypeImage
: {
818 const struct glsl_type
*sampled_type
=
819 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
821 assert(glsl_type_is_vector_or_scalar(sampled_type
));
823 enum glsl_sampler_dim dim
;
824 switch ((SpvDim
)w
[3]) {
825 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
826 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
827 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
828 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
829 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
830 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
832 unreachable("Invalid SPIR-V Sampler dimension");
835 bool is_shadow
= w
[4];
836 bool is_array
= w
[5];
837 bool multisampled
= w
[6];
838 unsigned sampled
= w
[7];
839 SpvImageFormat format
= w
[8];
842 val
->type
->access_qualifier
= w
[9];
844 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
847 assert(dim
== GLSL_SAMPLER_DIM_2D
);
848 dim
= GLSL_SAMPLER_DIM_MS
;
851 val
->type
->image_format
= translate_image_format(format
);
854 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
855 glsl_get_base_type(sampled_type
));
856 } else if (sampled
== 2) {
859 val
->type
->type
= glsl_image_type(dim
, is_array
,
860 glsl_get_base_type(sampled_type
));
862 assert(!"We need to know if the image will be sampled");
867 case SpvOpTypeSampledImage
:
868 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
871 case SpvOpTypeSampler
:
872 /* The actual sampler type here doesn't really matter. It gets
873 * thrown away the moment you combine it with an image. What really
874 * matters is that it's a sampler type as opposed to an integer type
875 * so the backend knows what to do.
877 val
->type
->type
= glsl_bare_sampler_type();
880 case SpvOpTypeOpaque
:
882 case SpvOpTypeDeviceEvent
:
883 case SpvOpTypeReserveId
:
887 unreachable("Unhandled opcode");
890 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
893 static nir_constant
*
894 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
896 nir_constant
*c
= rzalloc(b
, nir_constant
);
898 switch (glsl_get_base_type(type
)) {
902 case GLSL_TYPE_FLOAT
:
903 case GLSL_TYPE_DOUBLE
:
904 /* Nothing to do here. It's already initialized to zero */
907 case GLSL_TYPE_ARRAY
:
908 assert(glsl_get_length(type
) > 0);
909 c
->num_elements
= glsl_get_length(type
);
910 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
912 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
913 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
914 c
->elements
[i
] = c
->elements
[0];
917 case GLSL_TYPE_STRUCT
:
918 c
->num_elements
= glsl_get_length(type
);
919 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
921 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
922 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
927 unreachable("Invalid type for null constant");
934 spec_constant_deocoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
935 int member
, const struct vtn_decoration
*dec
,
938 assert(member
== -1);
939 if (dec
->decoration
!= SpvDecorationSpecId
)
942 uint32_t *const_value
= data
;
944 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
945 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
946 *const_value
= b
->specializations
[i
].data
;
953 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
954 uint32_t const_value
)
956 vtn_foreach_decoration(b
, val
, spec_constant_deocoration_cb
, &const_value
);
961 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
962 struct vtn_value
*val
,
964 const struct vtn_decoration
*dec
,
967 assert(member
== -1);
968 if (dec
->decoration
!= SpvDecorationBuiltIn
||
969 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
972 assert(val
->const_type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
974 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->value
.u
[0];
975 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->value
.u
[1];
976 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->value
.u
[2];
980 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
981 const uint32_t *w
, unsigned count
)
983 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
984 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
985 val
->constant
= rzalloc(b
, nir_constant
);
987 case SpvOpConstantTrue
:
988 assert(val
->const_type
== glsl_bool_type());
989 val
->constant
->value
.u
[0] = NIR_TRUE
;
991 case SpvOpConstantFalse
:
992 assert(val
->const_type
== glsl_bool_type());
993 val
->constant
->value
.u
[0] = NIR_FALSE
;
996 case SpvOpSpecConstantTrue
:
997 case SpvOpSpecConstantFalse
: {
998 assert(val
->const_type
== glsl_bool_type());
1000 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1001 val
->constant
->value
.u
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1006 assert(glsl_type_is_scalar(val
->const_type
));
1007 val
->constant
->value
.u
[0] = w
[3];
1009 case SpvOpSpecConstant
:
1010 assert(glsl_type_is_scalar(val
->const_type
));
1011 val
->constant
->value
.u
[0] = get_specialization(b
, val
, w
[3]);
1013 case SpvOpSpecConstantComposite
:
1014 case SpvOpConstantComposite
: {
1015 unsigned elem_count
= count
- 3;
1016 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1017 for (unsigned i
= 0; i
< elem_count
; i
++)
1018 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1020 switch (glsl_get_base_type(val
->const_type
)) {
1021 case GLSL_TYPE_UINT
:
1023 case GLSL_TYPE_FLOAT
:
1024 case GLSL_TYPE_BOOL
:
1025 if (glsl_type_is_matrix(val
->const_type
)) {
1026 unsigned rows
= glsl_get_vector_elements(val
->const_type
);
1027 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
1028 for (unsigned i
= 0; i
< elem_count
; i
++)
1029 for (unsigned j
= 0; j
< rows
; j
++)
1030 val
->constant
->value
.u
[rows
* i
+ j
] = elems
[i
]->value
.u
[j
];
1032 assert(glsl_type_is_vector(val
->const_type
));
1033 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
1034 for (unsigned i
= 0; i
< elem_count
; i
++)
1035 val
->constant
->value
.u
[i
] = elems
[i
]->value
.u
[0];
1040 case GLSL_TYPE_STRUCT
:
1041 case GLSL_TYPE_ARRAY
:
1042 ralloc_steal(val
->constant
, elems
);
1043 val
->constant
->num_elements
= elem_count
;
1044 val
->constant
->elements
= elems
;
1048 unreachable("Unsupported type for constants");
1053 case SpvOpSpecConstantOp
: {
1054 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1056 case SpvOpVectorShuffle
: {
1057 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1058 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1059 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
1060 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
1063 for (unsigned i
= 0; i
< len0
; i
++)
1064 u
[i
] = v0
->constant
->value
.u
[i
];
1065 for (unsigned i
= 0; i
< len1
; i
++)
1066 u
[len0
+ i
] = v1
->constant
->value
.u
[i
];
1068 for (unsigned i
= 0; i
< count
- 6; i
++) {
1069 uint32_t comp
= w
[i
+ 6];
1070 if (comp
== (uint32_t)-1) {
1071 val
->constant
->value
.u
[i
] = 0xdeadbeef;
1073 val
->constant
->value
.u
[i
] = u
[comp
];
1079 case SpvOpCompositeExtract
:
1080 case SpvOpCompositeInsert
: {
1081 struct vtn_value
*comp
;
1082 unsigned deref_start
;
1083 struct nir_constant
**c
;
1084 if (opcode
== SpvOpCompositeExtract
) {
1085 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1087 c
= &comp
->constant
;
1089 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1091 val
->constant
= nir_constant_clone(comp
->constant
,
1097 const struct glsl_type
*type
= comp
->const_type
;
1098 for (unsigned i
= deref_start
; i
< count
; i
++) {
1099 switch (glsl_get_base_type(type
)) {
1100 case GLSL_TYPE_UINT
:
1102 case GLSL_TYPE_FLOAT
:
1103 case GLSL_TYPE_BOOL
:
1104 /* If we hit this granularity, we're picking off an element */
1108 if (glsl_type_is_matrix(type
)) {
1109 elem
+= w
[i
] * glsl_get_vector_elements(type
);
1110 type
= glsl_get_column_type(type
);
1112 assert(glsl_type_is_vector(type
));
1114 type
= glsl_scalar_type(glsl_get_base_type(type
));
1118 case GLSL_TYPE_ARRAY
:
1119 c
= &(*c
)->elements
[w
[i
]];
1120 type
= glsl_get_array_element(type
);
1123 case GLSL_TYPE_STRUCT
:
1124 c
= &(*c
)->elements
[w
[i
]];
1125 type
= glsl_get_struct_field(type
, w
[i
]);
1129 unreachable("Invalid constant type");
1133 if (opcode
== SpvOpCompositeExtract
) {
1137 unsigned num_components
= glsl_get_vector_elements(type
);
1138 for (unsigned i
= 0; i
< num_components
; i
++)
1139 val
->constant
->value
.u
[i
] = (*c
)->value
.u
[elem
+ i
];
1142 struct vtn_value
*insert
=
1143 vtn_value(b
, w
[4], vtn_value_type_constant
);
1144 assert(insert
->const_type
== type
);
1146 *c
= insert
->constant
;
1148 unsigned num_components
= glsl_get_vector_elements(type
);
1149 for (unsigned i
= 0; i
< num_components
; i
++)
1150 (*c
)->value
.u
[elem
+ i
] = insert
->constant
->value
.u
[i
];
1158 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
);
1160 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1162 glsl_get_bit_size(val
->const_type
);
1164 nir_const_value src
[4];
1166 for (unsigned i
= 0; i
< count
- 4; i
++) {
1168 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1170 unsigned j
= swap
? 1 - i
: i
;
1171 assert(bit_size
== 32);
1172 for (unsigned k
= 0; k
< num_components
; k
++)
1173 src
[j
].u32
[k
] = c
->value
.u
[k
];
1176 nir_const_value res
= nir_eval_const_opcode(op
, num_components
,
1179 for (unsigned k
= 0; k
< num_components
; k
++)
1180 val
->constant
->value
.u
[k
] = res
.u32
[k
];
1188 case SpvOpConstantNull
:
1189 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1192 case SpvOpConstantSampler
:
1193 assert(!"OpConstantSampler requires Kernel Capability");
1197 unreachable("Unhandled opcode");
1200 /* Now that we have the value, update the workgroup size if needed */
1201 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1205 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1206 const uint32_t *w
, unsigned count
)
1208 struct nir_function
*callee
=
1209 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1211 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1212 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1213 unsigned arg_id
= w
[4 + i
];
1214 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1215 if (arg
->value_type
== vtn_value_type_access_chain
) {
1216 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1217 call
->params
[i
] = nir_deref_as_var(nir_copy_deref(call
, &d
->deref
));
1219 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1221 /* Make a temporary to store the argument in */
1223 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1224 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1226 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1230 nir_variable
*out_tmp
= NULL
;
1231 if (!glsl_type_is_void(callee
->return_type
)) {
1232 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1234 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1237 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1239 if (glsl_type_is_void(callee
->return_type
)) {
1240 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1242 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1243 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1247 struct vtn_ssa_value
*
1248 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1250 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1253 if (!glsl_type_is_vector_or_scalar(type
)) {
1254 unsigned elems
= glsl_get_length(type
);
1255 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1256 for (unsigned i
= 0; i
< elems
; i
++) {
1257 const struct glsl_type
*child_type
;
1259 switch (glsl_get_base_type(type
)) {
1261 case GLSL_TYPE_UINT
:
1262 case GLSL_TYPE_BOOL
:
1263 case GLSL_TYPE_FLOAT
:
1264 case GLSL_TYPE_DOUBLE
:
1265 child_type
= glsl_get_column_type(type
);
1267 case GLSL_TYPE_ARRAY
:
1268 child_type
= glsl_get_array_element(type
);
1270 case GLSL_TYPE_STRUCT
:
1271 child_type
= glsl_get_struct_field(type
, i
);
1274 unreachable("unkown base type");
1277 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1285 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1288 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1289 src
.src_type
= type
;
1294 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1295 const uint32_t *w
, unsigned count
)
1297 if (opcode
== SpvOpSampledImage
) {
1298 struct vtn_value
*val
=
1299 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1300 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1301 val
->sampled_image
->image
=
1302 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1303 val
->sampled_image
->sampler
=
1304 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1306 } else if (opcode
== SpvOpImage
) {
1307 struct vtn_value
*val
=
1308 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1309 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1310 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1311 val
->access_chain
= src_val
->sampled_image
->image
;
1313 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1314 val
->access_chain
= src_val
->access_chain
;
1319 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1320 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1322 struct vtn_sampled_image sampled
;
1323 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1324 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1325 sampled
= *sampled_val
->sampled_image
;
1327 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1328 sampled
.image
= NULL
;
1329 sampled
.sampler
= sampled_val
->access_chain
;
1332 const struct glsl_type
*image_type
;
1333 if (sampled
.image
) {
1334 image_type
= sampled
.image
->var
->var
->interface_type
;
1336 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1339 nir_tex_src srcs
[8]; /* 8 should be enough */
1340 nir_tex_src
*p
= srcs
;
1344 bool has_coord
= false;
1346 case SpvOpImageSampleImplicitLod
:
1347 case SpvOpImageSampleExplicitLod
:
1348 case SpvOpImageSampleDrefImplicitLod
:
1349 case SpvOpImageSampleDrefExplicitLod
:
1350 case SpvOpImageSampleProjImplicitLod
:
1351 case SpvOpImageSampleProjExplicitLod
:
1352 case SpvOpImageSampleProjDrefImplicitLod
:
1353 case SpvOpImageSampleProjDrefExplicitLod
:
1354 case SpvOpImageFetch
:
1355 case SpvOpImageGather
:
1356 case SpvOpImageDrefGather
:
1357 case SpvOpImageQueryLod
: {
1358 /* All these types have the coordinate as their first real argument */
1359 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, w
[idx
++]);
1361 p
->src
= nir_src_for_ssa(coord
->def
);
1362 p
->src_type
= nir_tex_src_coord
;
1371 /* These all have an explicit depth value as their next source */
1373 case SpvOpImageSampleDrefImplicitLod
:
1374 case SpvOpImageSampleDrefExplicitLod
:
1375 case SpvOpImageSampleProjDrefImplicitLod
:
1376 case SpvOpImageSampleProjDrefExplicitLod
:
1377 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparitor
);
1383 /* For OpImageQuerySizeLod, we always have an LOD */
1384 if (opcode
== SpvOpImageQuerySizeLod
)
1385 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1387 /* Figure out the base texture operation */
1390 case SpvOpImageSampleImplicitLod
:
1391 case SpvOpImageSampleDrefImplicitLod
:
1392 case SpvOpImageSampleProjImplicitLod
:
1393 case SpvOpImageSampleProjDrefImplicitLod
:
1394 texop
= nir_texop_tex
;
1397 case SpvOpImageSampleExplicitLod
:
1398 case SpvOpImageSampleDrefExplicitLod
:
1399 case SpvOpImageSampleProjExplicitLod
:
1400 case SpvOpImageSampleProjDrefExplicitLod
:
1401 texop
= nir_texop_txl
;
1404 case SpvOpImageFetch
:
1405 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1406 texop
= nir_texop_txf_ms
;
1408 texop
= nir_texop_txf
;
1412 case SpvOpImageGather
:
1413 case SpvOpImageDrefGather
:
1414 texop
= nir_texop_tg4
;
1417 case SpvOpImageQuerySizeLod
:
1418 case SpvOpImageQuerySize
:
1419 texop
= nir_texop_txs
;
1422 case SpvOpImageQueryLod
:
1423 texop
= nir_texop_lod
;
1426 case SpvOpImageQueryLevels
:
1427 texop
= nir_texop_query_levels
;
1430 case SpvOpImageQuerySamples
:
1432 unreachable("Unhandled opcode");
1435 /* Now we need to handle some number of optional arguments */
1437 uint32_t operands
= w
[idx
++];
1439 if (operands
& SpvImageOperandsBiasMask
) {
1440 assert(texop
== nir_texop_tex
);
1441 texop
= nir_texop_txb
;
1442 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1445 if (operands
& SpvImageOperandsLodMask
) {
1446 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1447 texop
== nir_texop_txf_ms
|| texop
== nir_texop_txs
);
1448 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1451 if (operands
& SpvImageOperandsGradMask
) {
1452 assert(texop
== nir_texop_tex
);
1453 texop
= nir_texop_txd
;
1454 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1455 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1458 if (operands
& SpvImageOperandsOffsetMask
||
1459 operands
& SpvImageOperandsConstOffsetMask
)
1460 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1462 if (operands
& SpvImageOperandsConstOffsetsMask
)
1463 assert(!"Constant offsets to texture gather not yet implemented");
1465 if (operands
& SpvImageOperandsSampleMask
) {
1466 assert(texop
== nir_texop_txf_ms
);
1467 texop
= nir_texop_txf_ms
;
1468 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1471 /* We should have now consumed exactly all of the arguments */
1472 assert(idx
== count
);
1474 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1477 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1479 instr
->sampler_dim
= glsl_get_sampler_dim(image_type
);
1480 instr
->is_array
= glsl_sampler_type_is_array(image_type
);
1481 instr
->is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1482 instr
->is_new_style_shadow
= instr
->is_shadow
;
1485 switch (instr
->sampler_dim
) {
1486 case GLSL_SAMPLER_DIM_1D
:
1487 case GLSL_SAMPLER_DIM_BUF
:
1488 instr
->coord_components
= 1;
1490 case GLSL_SAMPLER_DIM_2D
:
1491 case GLSL_SAMPLER_DIM_RECT
:
1492 case GLSL_SAMPLER_DIM_MS
:
1493 instr
->coord_components
= 2;
1495 case GLSL_SAMPLER_DIM_3D
:
1496 case GLSL_SAMPLER_DIM_CUBE
:
1497 instr
->coord_components
= 3;
1500 assert("Invalid sampler type");
1503 if (instr
->is_array
)
1504 instr
->coord_components
++;
1506 instr
->coord_components
= 0;
1509 switch (glsl_get_sampler_result_type(image_type
)) {
1510 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1511 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1512 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1513 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1515 unreachable("Invalid base type for sampler result");
1518 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1519 if (sampled
.image
) {
1520 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1521 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &image
->deref
));
1523 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1526 switch (instr
->op
) {
1531 /* These operations require a sampler */
1532 instr
->sampler
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1535 case nir_texop_txf_ms
:
1539 case nir_texop_query_levels
:
1540 case nir_texop_texture_samples
:
1541 case nir_texop_samples_identical
:
1543 instr
->sampler
= NULL
;
1545 case nir_texop_txf_ms_mcs
:
1546 unreachable("unexpected nir_texop_txf_ms_mcs");
1549 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1550 nir_tex_instr_dest_size(instr
), 32, NULL
);
1552 assert(glsl_get_vector_elements(ret_type
->type
) ==
1553 nir_tex_instr_dest_size(instr
));
1555 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1556 val
->ssa
->def
= &instr
->dest
.ssa
;
1558 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1561 static nir_ssa_def
*
1562 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1564 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1566 /* The image_load_store intrinsics assume a 4-dim coordinate */
1567 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1568 unsigned swizzle
[4];
1569 for (unsigned i
= 0; i
< 4; i
++)
1570 swizzle
[i
] = MIN2(i
, dim
- 1);
1572 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1576 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1577 const uint32_t *w
, unsigned count
)
1579 /* Just get this one out of the way */
1580 if (opcode
== SpvOpImageTexelPointer
) {
1581 struct vtn_value
*val
=
1582 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1583 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1586 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1587 val
->image
->coord
= get_image_coord(b
, w
[4]);
1588 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1592 struct vtn_image_pointer image
;
1595 case SpvOpAtomicExchange
:
1596 case SpvOpAtomicCompareExchange
:
1597 case SpvOpAtomicCompareExchangeWeak
:
1598 case SpvOpAtomicIIncrement
:
1599 case SpvOpAtomicIDecrement
:
1600 case SpvOpAtomicIAdd
:
1601 case SpvOpAtomicISub
:
1602 case SpvOpAtomicSMin
:
1603 case SpvOpAtomicUMin
:
1604 case SpvOpAtomicSMax
:
1605 case SpvOpAtomicUMax
:
1606 case SpvOpAtomicAnd
:
1608 case SpvOpAtomicXor
:
1609 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1612 case SpvOpImageQuerySize
:
1614 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1616 image
.sample
= NULL
;
1619 case SpvOpImageRead
:
1621 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1622 image
.coord
= get_image_coord(b
, w
[4]);
1624 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1625 assert(w
[5] == SpvImageOperandsSampleMask
);
1626 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1628 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1632 case SpvOpImageWrite
:
1634 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1635 image
.coord
= get_image_coord(b
, w
[2]);
1639 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1640 assert(w
[4] == SpvImageOperandsSampleMask
);
1641 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1643 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1648 unreachable("Invalid image opcode");
1651 nir_intrinsic_op op
;
1653 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1654 OP(ImageQuerySize
, size
)
1656 OP(ImageWrite
, store
)
1657 OP(AtomicExchange
, atomic_exchange
)
1658 OP(AtomicCompareExchange
, atomic_comp_swap
)
1659 OP(AtomicIIncrement
, atomic_add
)
1660 OP(AtomicIDecrement
, atomic_add
)
1661 OP(AtomicIAdd
, atomic_add
)
1662 OP(AtomicISub
, atomic_add
)
1663 OP(AtomicSMin
, atomic_min
)
1664 OP(AtomicUMin
, atomic_min
)
1665 OP(AtomicSMax
, atomic_max
)
1666 OP(AtomicUMax
, atomic_max
)
1667 OP(AtomicAnd
, atomic_and
)
1668 OP(AtomicOr
, atomic_or
)
1669 OP(AtomicXor
, atomic_xor
)
1672 unreachable("Invalid image opcode");
1675 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1677 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1678 intrin
->variables
[0] =
1679 nir_deref_as_var(nir_copy_deref(&intrin
->instr
, &image_deref
->deref
));
1681 /* ImageQuerySize doesn't take any extra parameters */
1682 if (opcode
!= SpvOpImageQuerySize
) {
1683 /* The image coordinate is always 4 components but we may not have that
1684 * many. Swizzle to compensate.
1687 for (unsigned i
= 0; i
< 4; i
++)
1688 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1689 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1691 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1695 case SpvOpImageQuerySize
:
1696 case SpvOpImageRead
:
1698 case SpvOpImageWrite
:
1699 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1701 case SpvOpAtomicIIncrement
:
1702 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1704 case SpvOpAtomicIDecrement
:
1705 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1708 case SpvOpAtomicExchange
:
1709 case SpvOpAtomicIAdd
:
1710 case SpvOpAtomicSMin
:
1711 case SpvOpAtomicUMin
:
1712 case SpvOpAtomicSMax
:
1713 case SpvOpAtomicUMax
:
1714 case SpvOpAtomicAnd
:
1716 case SpvOpAtomicXor
:
1717 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1720 case SpvOpAtomicCompareExchange
:
1721 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1722 intrin
->src
[3] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1725 case SpvOpAtomicISub
:
1726 intrin
->src
[2] = nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1730 unreachable("Invalid image opcode");
1733 if (opcode
!= SpvOpImageWrite
) {
1734 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1735 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1736 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, 32, NULL
);
1738 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1740 /* The image intrinsics always return 4 channels but we may not want
1741 * that many. Emit a mov to trim it down.
1743 unsigned swiz
[4] = {0, 1, 2, 3};
1744 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1745 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1746 glsl_get_vector_elements(type
->type
), false);
1748 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1752 static nir_intrinsic_op
1753 get_ssbo_nir_atomic_op(SpvOp opcode
)
1756 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1757 OP(AtomicExchange
, atomic_exchange
)
1758 OP(AtomicCompareExchange
, atomic_comp_swap
)
1759 OP(AtomicIIncrement
, atomic_add
)
1760 OP(AtomicIDecrement
, atomic_add
)
1761 OP(AtomicIAdd
, atomic_add
)
1762 OP(AtomicISub
, atomic_add
)
1763 OP(AtomicSMin
, atomic_imin
)
1764 OP(AtomicUMin
, atomic_umin
)
1765 OP(AtomicSMax
, atomic_imax
)
1766 OP(AtomicUMax
, atomic_umax
)
1767 OP(AtomicAnd
, atomic_and
)
1768 OP(AtomicOr
, atomic_or
)
1769 OP(AtomicXor
, atomic_xor
)
1772 unreachable("Invalid SSBO atomic");
1776 static nir_intrinsic_op
1777 get_shared_nir_atomic_op(SpvOp opcode
)
1780 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
1781 OP(AtomicExchange
, atomic_exchange
)
1782 OP(AtomicCompareExchange
, atomic_comp_swap
)
1783 OP(AtomicIIncrement
, atomic_add
)
1784 OP(AtomicIDecrement
, atomic_add
)
1785 OP(AtomicIAdd
, atomic_add
)
1786 OP(AtomicISub
, atomic_add
)
1787 OP(AtomicSMin
, atomic_imin
)
1788 OP(AtomicUMin
, atomic_umin
)
1789 OP(AtomicSMax
, atomic_imax
)
1790 OP(AtomicUMax
, atomic_umax
)
1791 OP(AtomicAnd
, atomic_and
)
1792 OP(AtomicOr
, atomic_or
)
1793 OP(AtomicXor
, atomic_xor
)
1796 unreachable("Invalid shared atomic");
1801 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1802 const uint32_t *w
, nir_src
*src
)
1805 case SpvOpAtomicIIncrement
:
1806 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1809 case SpvOpAtomicIDecrement
:
1810 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1813 case SpvOpAtomicISub
:
1815 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1818 case SpvOpAtomicCompareExchange
:
1819 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1820 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1824 case SpvOpAtomicExchange
:
1825 case SpvOpAtomicIAdd
:
1826 case SpvOpAtomicSMin
:
1827 case SpvOpAtomicUMin
:
1828 case SpvOpAtomicSMax
:
1829 case SpvOpAtomicUMax
:
1830 case SpvOpAtomicAnd
:
1832 case SpvOpAtomicXor
:
1833 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1837 unreachable("Invalid SPIR-V atomic");
1842 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
1843 const uint32_t *w
, unsigned count
)
1845 struct vtn_access_chain
*chain
=
1846 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1847 nir_intrinsic_instr
*atomic
;
1850 SpvScope scope = w[4];
1851 SpvMemorySemanticsMask semantics = w[5];
1854 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
1855 nir_deref
*deref
= &vtn_access_chain_to_deref(b
, chain
)->deref
;
1856 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
1857 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1858 atomic
->variables
[0] = nir_deref_as_var(nir_copy_deref(atomic
, deref
));
1859 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
1861 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
1862 struct vtn_type
*type
;
1863 nir_ssa_def
*offset
, *index
;
1864 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
1866 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
1868 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1869 atomic
->src
[0] = nir_src_for_ssa(index
);
1870 atomic
->src
[1] = nir_src_for_ssa(offset
);
1871 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
1874 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
, 1, 32, NULL
);
1876 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1877 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1878 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
1879 val
->ssa
->def
= &atomic
->dest
.ssa
;
1880 val
->ssa
->type
= type
->type
;
1882 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
1885 static nir_alu_instr
*
1886 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
1889 switch (num_components
) {
1890 case 1: op
= nir_op_fmov
; break;
1891 case 2: op
= nir_op_vec2
; break;
1892 case 3: op
= nir_op_vec3
; break;
1893 case 4: op
= nir_op_vec4
; break;
1894 default: unreachable("bad vector size");
1897 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
1898 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
1900 vec
->dest
.write_mask
= (1 << num_components
) - 1;
1905 struct vtn_ssa_value
*
1906 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
1908 if (src
->transposed
)
1909 return src
->transposed
;
1911 struct vtn_ssa_value
*dest
=
1912 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
1914 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
1915 nir_alu_instr
*vec
= create_vec(b
->shader
,
1916 glsl_get_matrix_columns(src
->type
),
1917 glsl_get_bit_size(src
->type
));
1918 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1919 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
1920 vec
->src
[0].swizzle
[0] = i
;
1922 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
1923 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
1924 vec
->src
[j
].swizzle
[0] = i
;
1927 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1928 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
1931 dest
->transposed
= src
;
1937 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
1939 unsigned swiz
[4] = { index
};
1940 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
1944 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
1947 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
1950 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
1952 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
1954 vec
->src
[i
].src
= nir_src_for_ssa(src
);
1955 vec
->src
[i
].swizzle
[0] = i
;
1959 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1961 return &vec
->dest
.dest
.ssa
;
1965 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1968 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
1969 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1970 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1971 vtn_vector_extract(b
, src
, i
), dest
);
1977 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1978 nir_ssa_def
*insert
, nir_ssa_def
*index
)
1980 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
1981 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1982 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1983 vtn_vector_insert(b
, src
, insert
, i
), dest
);
1988 static nir_ssa_def
*
1989 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
1990 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
1991 const uint32_t *indices
)
1993 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
1995 for (unsigned i
= 0; i
< num_components
; i
++) {
1996 uint32_t index
= indices
[i
];
1997 if (index
== 0xffffffff) {
1999 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2000 } else if (index
< src0
->num_components
) {
2001 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2002 vec
->src
[i
].swizzle
[0] = index
;
2004 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2005 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2009 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2011 return &vec
->dest
.dest
.ssa
;
2015 * Concatentates a number of vectors/scalars together to produce a vector
2017 static nir_ssa_def
*
2018 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2019 unsigned num_srcs
, nir_ssa_def
**srcs
)
2021 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
2024 unsigned dest_idx
= 0;
2025 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2026 nir_ssa_def
*src
= srcs
[i
];
2027 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2028 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2029 vec
->src
[dest_idx
].swizzle
[0] = j
;
2034 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2036 return &vec
->dest
.dest
.ssa
;
2039 static struct vtn_ssa_value
*
2040 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2042 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2043 dest
->type
= src
->type
;
2045 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2046 dest
->def
= src
->def
;
2048 unsigned elems
= glsl_get_length(src
->type
);
2050 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2051 for (unsigned i
= 0; i
< elems
; i
++)
2052 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2058 static struct vtn_ssa_value
*
2059 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2060 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2061 unsigned num_indices
)
2063 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2065 struct vtn_ssa_value
*cur
= dest
;
2067 for (i
= 0; i
< num_indices
- 1; i
++) {
2068 cur
= cur
->elems
[indices
[i
]];
2071 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2072 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2073 * the component granularity. In that case, the last index will be
2074 * the index to insert the scalar into the vector.
2077 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2079 cur
->elems
[indices
[i
]] = insert
;
2085 static struct vtn_ssa_value
*
2086 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2087 const uint32_t *indices
, unsigned num_indices
)
2089 struct vtn_ssa_value
*cur
= src
;
2090 for (unsigned i
= 0; i
< num_indices
; i
++) {
2091 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2092 assert(i
== num_indices
- 1);
2093 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2094 * the component granularity. The last index will be the index of the
2095 * vector to extract.
2098 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2099 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2100 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2103 cur
= cur
->elems
[indices
[i
]];
2111 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2112 const uint32_t *w
, unsigned count
)
2114 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2115 const struct glsl_type
*type
=
2116 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2117 val
->ssa
= vtn_create_ssa_value(b
, type
);
2120 case SpvOpVectorExtractDynamic
:
2121 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2122 vtn_ssa_value(b
, w
[4])->def
);
2125 case SpvOpVectorInsertDynamic
:
2126 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2127 vtn_ssa_value(b
, w
[4])->def
,
2128 vtn_ssa_value(b
, w
[5])->def
);
2131 case SpvOpVectorShuffle
:
2132 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2133 vtn_ssa_value(b
, w
[3])->def
,
2134 vtn_ssa_value(b
, w
[4])->def
,
2138 case SpvOpCompositeConstruct
: {
2139 unsigned elems
= count
- 3;
2140 if (glsl_type_is_vector_or_scalar(type
)) {
2141 nir_ssa_def
*srcs
[4];
2142 for (unsigned i
= 0; i
< elems
; i
++)
2143 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2145 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2148 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2149 for (unsigned i
= 0; i
< elems
; i
++)
2150 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2154 case SpvOpCompositeExtract
:
2155 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2159 case SpvOpCompositeInsert
:
2160 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2161 vtn_ssa_value(b
, w
[3]),
2165 case SpvOpCopyObject
:
2166 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2170 unreachable("unknown composite operation");
2175 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2176 const uint32_t *w
, unsigned count
)
2178 nir_intrinsic_op intrinsic_op
;
2180 case SpvOpEmitVertex
:
2181 case SpvOpEmitStreamVertex
:
2182 intrinsic_op
= nir_intrinsic_emit_vertex
;
2184 case SpvOpEndPrimitive
:
2185 case SpvOpEndStreamPrimitive
:
2186 intrinsic_op
= nir_intrinsic_end_primitive
;
2188 case SpvOpMemoryBarrier
:
2189 intrinsic_op
= nir_intrinsic_memory_barrier
;
2191 case SpvOpControlBarrier
:
2192 intrinsic_op
= nir_intrinsic_barrier
;
2195 unreachable("unknown barrier instruction");
2198 nir_intrinsic_instr
*intrin
=
2199 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2201 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2202 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2204 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2208 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2211 case SpvExecutionModeInputPoints
:
2212 case SpvExecutionModeOutputPoints
:
2213 return 0; /* GL_POINTS */
2214 case SpvExecutionModeInputLines
:
2215 return 1; /* GL_LINES */
2216 case SpvExecutionModeInputLinesAdjacency
:
2217 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2218 case SpvExecutionModeTriangles
:
2219 return 4; /* GL_TRIANGLES */
2220 case SpvExecutionModeInputTrianglesAdjacency
:
2221 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2222 case SpvExecutionModeQuads
:
2223 return 7; /* GL_QUADS */
2224 case SpvExecutionModeIsolines
:
2225 return 0x8E7A; /* GL_ISOLINES */
2226 case SpvExecutionModeOutputLineStrip
:
2227 return 3; /* GL_LINE_STRIP */
2228 case SpvExecutionModeOutputTriangleStrip
:
2229 return 5; /* GL_TRIANGLE_STRIP */
2231 assert(!"Invalid primitive type");
2237 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2240 case SpvExecutionModeInputPoints
:
2242 case SpvExecutionModeInputLines
:
2244 case SpvExecutionModeInputLinesAdjacency
:
2246 case SpvExecutionModeTriangles
:
2248 case SpvExecutionModeInputTrianglesAdjacency
:
2251 assert(!"Invalid GS input mode");
2256 static gl_shader_stage
2257 stage_for_execution_model(SpvExecutionModel model
)
2260 case SpvExecutionModelVertex
:
2261 return MESA_SHADER_VERTEX
;
2262 case SpvExecutionModelTessellationControl
:
2263 return MESA_SHADER_TESS_CTRL
;
2264 case SpvExecutionModelTessellationEvaluation
:
2265 return MESA_SHADER_TESS_EVAL
;
2266 case SpvExecutionModelGeometry
:
2267 return MESA_SHADER_GEOMETRY
;
2268 case SpvExecutionModelFragment
:
2269 return MESA_SHADER_FRAGMENT
;
2270 case SpvExecutionModelGLCompute
:
2271 return MESA_SHADER_COMPUTE
;
2273 unreachable("Unsupported execution model");
2278 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2279 const uint32_t *w
, unsigned count
)
2283 case SpvOpSourceExtension
:
2284 case SpvOpSourceContinued
:
2285 case SpvOpExtension
:
2286 /* Unhandled, but these are for debug so that's ok. */
2289 case SpvOpCapability
: {
2290 SpvCapability cap
= w
[1];
2292 case SpvCapabilityMatrix
:
2293 case SpvCapabilityShader
:
2294 case SpvCapabilityGeometry
:
2295 case SpvCapabilityGeometryPointSize
:
2296 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2297 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2298 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2299 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2300 case SpvCapabilityImageRect
:
2301 case SpvCapabilitySampledRect
:
2302 case SpvCapabilitySampled1D
:
2303 case SpvCapabilityImage1D
:
2304 case SpvCapabilitySampledCubeArray
:
2305 case SpvCapabilitySampledBuffer
:
2306 case SpvCapabilityImageBuffer
:
2307 case SpvCapabilityImageQuery
:
2308 case SpvCapabilityDerivativeControl
:
2309 case SpvCapabilityInterpolationFunction
:
2310 case SpvCapabilityMultiViewport
:
2313 case SpvCapabilityClipDistance
:
2314 case SpvCapabilityCullDistance
:
2315 case SpvCapabilityGeometryStreams
:
2316 case SpvCapabilityTessellation
:
2317 case SpvCapabilityTessellationPointSize
:
2318 case SpvCapabilityLinkage
:
2319 case SpvCapabilityVector16
:
2320 case SpvCapabilityFloat16Buffer
:
2321 case SpvCapabilityFloat16
:
2322 case SpvCapabilityFloat64
:
2323 case SpvCapabilityInt64
:
2324 case SpvCapabilityInt64Atomics
:
2325 case SpvCapabilityAtomicStorage
:
2326 case SpvCapabilityInt16
:
2327 case SpvCapabilityImageGatherExtended
:
2328 case SpvCapabilityStorageImageMultisample
:
2329 case SpvCapabilityImageCubeArray
:
2330 case SpvCapabilitySampleRateShading
:
2331 case SpvCapabilityInt8
:
2332 case SpvCapabilityInputAttachment
:
2333 case SpvCapabilitySparseResidency
:
2334 case SpvCapabilityMinLod
:
2335 case SpvCapabilityImageMSArray
:
2336 case SpvCapabilityStorageImageExtendedFormats
:
2337 case SpvCapabilityTransformFeedback
:
2338 case SpvCapabilityStorageImageReadWithoutFormat
:
2339 case SpvCapabilityStorageImageWriteWithoutFormat
:
2340 vtn_warn("Unsupported SPIR-V capability: %s",
2341 spirv_capability_to_string(cap
));
2344 case SpvCapabilityAddresses
:
2345 case SpvCapabilityKernel
:
2346 case SpvCapabilityImageBasic
:
2347 case SpvCapabilityImageReadWrite
:
2348 case SpvCapabilityImageMipmap
:
2349 case SpvCapabilityPipes
:
2350 case SpvCapabilityGroups
:
2351 case SpvCapabilityDeviceEnqueue
:
2352 case SpvCapabilityLiteralSampler
:
2353 case SpvCapabilityGenericPointer
:
2354 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2355 spirv_capability_to_string(cap
));
2361 case SpvOpExtInstImport
:
2362 vtn_handle_extension(b
, opcode
, w
, count
);
2365 case SpvOpMemoryModel
:
2366 assert(w
[1] == SpvAddressingModelLogical
);
2367 assert(w
[2] == SpvMemoryModelGLSL450
);
2370 case SpvOpEntryPoint
: {
2371 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2372 /* Let this be a name label regardless */
2373 unsigned name_words
;
2374 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2376 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2377 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2380 assert(b
->entry_point
== NULL
);
2381 b
->entry_point
= entry_point
;
2386 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2387 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2391 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2394 case SpvOpMemberName
:
2398 case SpvOpExecutionMode
:
2399 case SpvOpDecorationGroup
:
2401 case SpvOpMemberDecorate
:
2402 case SpvOpGroupDecorate
:
2403 case SpvOpGroupMemberDecorate
:
2404 vtn_handle_decoration(b
, opcode
, w
, count
);
2408 return false; /* End of preamble */
2415 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2416 const struct vtn_decoration
*mode
, void *data
)
2418 assert(b
->entry_point
== entry_point
);
2420 switch(mode
->exec_mode
) {
2421 case SpvExecutionModeOriginUpperLeft
:
2422 case SpvExecutionModeOriginLowerLeft
:
2423 b
->origin_upper_left
=
2424 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2427 case SpvExecutionModeEarlyFragmentTests
:
2428 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2429 b
->shader
->info
.fs
.early_fragment_tests
= true;
2432 case SpvExecutionModeInvocations
:
2433 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2434 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2437 case SpvExecutionModeDepthReplacing
:
2438 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2439 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2441 case SpvExecutionModeDepthGreater
:
2442 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2443 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2445 case SpvExecutionModeDepthLess
:
2446 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2447 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2449 case SpvExecutionModeDepthUnchanged
:
2450 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2451 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2454 case SpvExecutionModeLocalSize
:
2455 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2456 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2457 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2458 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2460 case SpvExecutionModeLocalSizeHint
:
2461 break; /* Nothing do do with this */
2463 case SpvExecutionModeOutputVertices
:
2464 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2465 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2468 case SpvExecutionModeInputPoints
:
2469 case SpvExecutionModeInputLines
:
2470 case SpvExecutionModeInputLinesAdjacency
:
2471 case SpvExecutionModeTriangles
:
2472 case SpvExecutionModeInputTrianglesAdjacency
:
2473 case SpvExecutionModeQuads
:
2474 case SpvExecutionModeIsolines
:
2475 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2476 b
->shader
->info
.gs
.vertices_in
=
2477 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2479 assert(!"Tesselation shaders not yet supported");
2483 case SpvExecutionModeOutputPoints
:
2484 case SpvExecutionModeOutputLineStrip
:
2485 case SpvExecutionModeOutputTriangleStrip
:
2486 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2487 b
->shader
->info
.gs
.output_primitive
=
2488 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2491 case SpvExecutionModeSpacingEqual
:
2492 case SpvExecutionModeSpacingFractionalEven
:
2493 case SpvExecutionModeSpacingFractionalOdd
:
2494 case SpvExecutionModeVertexOrderCw
:
2495 case SpvExecutionModeVertexOrderCcw
:
2496 case SpvExecutionModePointMode
:
2497 assert(!"TODO: Add tessellation metadata");
2500 case SpvExecutionModePixelCenterInteger
:
2501 b
->pixel_center_integer
= true;
2504 case SpvExecutionModeXfb
:
2505 assert(!"Unhandled execution mode");
2508 case SpvExecutionModeVecTypeHint
:
2509 case SpvExecutionModeContractionOff
:
2515 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2516 const uint32_t *w
, unsigned count
)
2520 case SpvOpSourceContinued
:
2521 case SpvOpSourceExtension
:
2522 case SpvOpExtension
:
2523 case SpvOpCapability
:
2524 case SpvOpExtInstImport
:
2525 case SpvOpMemoryModel
:
2526 case SpvOpEntryPoint
:
2527 case SpvOpExecutionMode
:
2530 case SpvOpMemberName
:
2531 case SpvOpDecorationGroup
:
2533 case SpvOpMemberDecorate
:
2534 case SpvOpGroupDecorate
:
2535 case SpvOpGroupMemberDecorate
:
2536 assert(!"Invalid opcode types and variables section");
2542 case SpvOpTypeFloat
:
2543 case SpvOpTypeVector
:
2544 case SpvOpTypeMatrix
:
2545 case SpvOpTypeImage
:
2546 case SpvOpTypeSampler
:
2547 case SpvOpTypeSampledImage
:
2548 case SpvOpTypeArray
:
2549 case SpvOpTypeRuntimeArray
:
2550 case SpvOpTypeStruct
:
2551 case SpvOpTypeOpaque
:
2552 case SpvOpTypePointer
:
2553 case SpvOpTypeFunction
:
2554 case SpvOpTypeEvent
:
2555 case SpvOpTypeDeviceEvent
:
2556 case SpvOpTypeReserveId
:
2557 case SpvOpTypeQueue
:
2559 vtn_handle_type(b
, opcode
, w
, count
);
2562 case SpvOpConstantTrue
:
2563 case SpvOpConstantFalse
:
2565 case SpvOpConstantComposite
:
2566 case SpvOpConstantSampler
:
2567 case SpvOpConstantNull
:
2568 case SpvOpSpecConstantTrue
:
2569 case SpvOpSpecConstantFalse
:
2570 case SpvOpSpecConstant
:
2571 case SpvOpSpecConstantComposite
:
2572 case SpvOpSpecConstantOp
:
2573 vtn_handle_constant(b
, opcode
, w
, count
);
2577 vtn_handle_variables(b
, opcode
, w
, count
);
2581 return false; /* End of preamble */
2588 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2589 const uint32_t *w
, unsigned count
)
2595 case SpvOpLoopMerge
:
2596 case SpvOpSelectionMerge
:
2597 /* This is handled by cfg pre-pass and walk_blocks */
2601 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2602 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2607 vtn_handle_extension(b
, opcode
, w
, count
);
2613 case SpvOpCopyMemory
:
2614 case SpvOpCopyMemorySized
:
2615 case SpvOpAccessChain
:
2616 case SpvOpInBoundsAccessChain
:
2617 case SpvOpArrayLength
:
2618 vtn_handle_variables(b
, opcode
, w
, count
);
2621 case SpvOpFunctionCall
:
2622 vtn_handle_function_call(b
, opcode
, w
, count
);
2625 case SpvOpSampledImage
:
2627 case SpvOpImageSampleImplicitLod
:
2628 case SpvOpImageSampleExplicitLod
:
2629 case SpvOpImageSampleDrefImplicitLod
:
2630 case SpvOpImageSampleDrefExplicitLod
:
2631 case SpvOpImageSampleProjImplicitLod
:
2632 case SpvOpImageSampleProjExplicitLod
:
2633 case SpvOpImageSampleProjDrefImplicitLod
:
2634 case SpvOpImageSampleProjDrefExplicitLod
:
2635 case SpvOpImageFetch
:
2636 case SpvOpImageGather
:
2637 case SpvOpImageDrefGather
:
2638 case SpvOpImageQuerySizeLod
:
2639 case SpvOpImageQueryLod
:
2640 case SpvOpImageQueryLevels
:
2641 case SpvOpImageQuerySamples
:
2642 vtn_handle_texture(b
, opcode
, w
, count
);
2645 case SpvOpImageRead
:
2646 case SpvOpImageWrite
:
2647 case SpvOpImageTexelPointer
:
2648 vtn_handle_image(b
, opcode
, w
, count
);
2651 case SpvOpImageQuerySize
: {
2652 struct vtn_access_chain
*image
=
2653 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2654 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2655 vtn_handle_image(b
, opcode
, w
, count
);
2657 vtn_handle_texture(b
, opcode
, w
, count
);
2662 case SpvOpAtomicExchange
:
2663 case SpvOpAtomicCompareExchange
:
2664 case SpvOpAtomicCompareExchangeWeak
:
2665 case SpvOpAtomicIIncrement
:
2666 case SpvOpAtomicIDecrement
:
2667 case SpvOpAtomicIAdd
:
2668 case SpvOpAtomicISub
:
2669 case SpvOpAtomicSMin
:
2670 case SpvOpAtomicUMin
:
2671 case SpvOpAtomicSMax
:
2672 case SpvOpAtomicUMax
:
2673 case SpvOpAtomicAnd
:
2675 case SpvOpAtomicXor
: {
2676 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2677 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2678 vtn_handle_image(b
, opcode
, w
, count
);
2680 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2681 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2691 case SpvOpConvertFToU
:
2692 case SpvOpConvertFToS
:
2693 case SpvOpConvertSToF
:
2694 case SpvOpConvertUToF
:
2698 case SpvOpQuantizeToF16
:
2699 case SpvOpConvertPtrToU
:
2700 case SpvOpConvertUToPtr
:
2701 case SpvOpPtrCastToGeneric
:
2702 case SpvOpGenericCastToPtr
:
2708 case SpvOpSignBitSet
:
2709 case SpvOpLessOrGreater
:
2711 case SpvOpUnordered
:
2726 case SpvOpVectorTimesScalar
:
2728 case SpvOpIAddCarry
:
2729 case SpvOpISubBorrow
:
2730 case SpvOpUMulExtended
:
2731 case SpvOpSMulExtended
:
2732 case SpvOpShiftRightLogical
:
2733 case SpvOpShiftRightArithmetic
:
2734 case SpvOpShiftLeftLogical
:
2735 case SpvOpLogicalEqual
:
2736 case SpvOpLogicalNotEqual
:
2737 case SpvOpLogicalOr
:
2738 case SpvOpLogicalAnd
:
2739 case SpvOpLogicalNot
:
2740 case SpvOpBitwiseOr
:
2741 case SpvOpBitwiseXor
:
2742 case SpvOpBitwiseAnd
:
2745 case SpvOpFOrdEqual
:
2746 case SpvOpFUnordEqual
:
2747 case SpvOpINotEqual
:
2748 case SpvOpFOrdNotEqual
:
2749 case SpvOpFUnordNotEqual
:
2750 case SpvOpULessThan
:
2751 case SpvOpSLessThan
:
2752 case SpvOpFOrdLessThan
:
2753 case SpvOpFUnordLessThan
:
2754 case SpvOpUGreaterThan
:
2755 case SpvOpSGreaterThan
:
2756 case SpvOpFOrdGreaterThan
:
2757 case SpvOpFUnordGreaterThan
:
2758 case SpvOpULessThanEqual
:
2759 case SpvOpSLessThanEqual
:
2760 case SpvOpFOrdLessThanEqual
:
2761 case SpvOpFUnordLessThanEqual
:
2762 case SpvOpUGreaterThanEqual
:
2763 case SpvOpSGreaterThanEqual
:
2764 case SpvOpFOrdGreaterThanEqual
:
2765 case SpvOpFUnordGreaterThanEqual
:
2771 case SpvOpFwidthFine
:
2772 case SpvOpDPdxCoarse
:
2773 case SpvOpDPdyCoarse
:
2774 case SpvOpFwidthCoarse
:
2775 case SpvOpBitFieldInsert
:
2776 case SpvOpBitFieldSExtract
:
2777 case SpvOpBitFieldUExtract
:
2778 case SpvOpBitReverse
:
2780 case SpvOpTranspose
:
2781 case SpvOpOuterProduct
:
2782 case SpvOpMatrixTimesScalar
:
2783 case SpvOpVectorTimesMatrix
:
2784 case SpvOpMatrixTimesVector
:
2785 case SpvOpMatrixTimesMatrix
:
2786 vtn_handle_alu(b
, opcode
, w
, count
);
2789 case SpvOpVectorExtractDynamic
:
2790 case SpvOpVectorInsertDynamic
:
2791 case SpvOpVectorShuffle
:
2792 case SpvOpCompositeConstruct
:
2793 case SpvOpCompositeExtract
:
2794 case SpvOpCompositeInsert
:
2795 case SpvOpCopyObject
:
2796 vtn_handle_composite(b
, opcode
, w
, count
);
2799 case SpvOpEmitVertex
:
2800 case SpvOpEndPrimitive
:
2801 case SpvOpEmitStreamVertex
:
2802 case SpvOpEndStreamPrimitive
:
2803 case SpvOpControlBarrier
:
2804 case SpvOpMemoryBarrier
:
2805 vtn_handle_barrier(b
, opcode
, w
, count
);
2809 unreachable("Unhandled opcode");
2816 spirv_to_nir(const uint32_t *words
, size_t word_count
,
2817 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
2818 gl_shader_stage stage
, const char *entry_point_name
,
2819 const nir_shader_compiler_options
*options
)
2821 const uint32_t *word_end
= words
+ word_count
;
2823 /* Handle the SPIR-V header (first 4 dwords) */
2824 assert(word_count
> 5);
2826 assert(words
[0] == SpvMagicNumber
);
2827 assert(words
[1] >= 0x10000);
2828 /* words[2] == generator magic */
2829 unsigned value_id_bound
= words
[3];
2830 assert(words
[4] == 0);
2834 /* Initialize the stn_builder object */
2835 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
2836 b
->value_id_bound
= value_id_bound
;
2837 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
2838 exec_list_make_empty(&b
->functions
);
2839 b
->entry_point_stage
= stage
;
2840 b
->entry_point_name
= entry_point_name
;
2842 /* Handle all the preamble instructions */
2843 words
= vtn_foreach_instruction(b
, words
, word_end
,
2844 vtn_handle_preamble_instruction
);
2846 if (b
->entry_point
== NULL
) {
2847 assert(!"Entry point not found");
2852 b
->shader
= nir_shader_create(NULL
, stage
, options
);
2854 /* Set shader info defaults */
2855 b
->shader
->info
.gs
.invocations
= 1;
2857 /* Parse execution modes */
2858 vtn_foreach_execution_mode(b
, b
->entry_point
,
2859 vtn_handle_execution_mode
, NULL
);
2861 b
->specializations
= spec
;
2862 b
->num_specializations
= num_spec
;
2864 /* Handle all variable, type, and constant instructions */
2865 words
= vtn_foreach_instruction(b
, words
, word_end
,
2866 vtn_handle_variable_or_type_instruction
);
2868 vtn_build_cfg(b
, words
, word_end
);
2870 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
2871 b
->impl
= func
->impl
;
2872 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
2873 _mesa_key_pointer_equal
);
2875 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
2878 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
2879 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
2880 assert(entry_point
);