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"
34 struct spec_constant_value
{
43 _vtn_warn(const char *file
, int line
, const char *msg
, ...)
49 formatted
= ralloc_vasprintf(NULL
, msg
, args
);
52 fprintf(stderr
, "%s:%d WARNING: %s\n", file
, line
, formatted
);
54 ralloc_free(formatted
);
57 static struct vtn_ssa_value
*
58 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
60 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
63 if (glsl_type_is_vector_or_scalar(type
)) {
64 unsigned num_components
= glsl_get_vector_elements(val
->type
);
65 unsigned bit_size
= glsl_get_bit_size(val
->type
);
66 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
68 unsigned elems
= glsl_get_length(val
->type
);
69 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
70 if (glsl_type_is_matrix(type
)) {
71 const struct glsl_type
*elem_type
=
72 glsl_vector_type(glsl_get_base_type(type
),
73 glsl_get_vector_elements(type
));
75 for (unsigned i
= 0; i
< elems
; i
++)
76 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
77 } else if (glsl_type_is_array(type
)) {
78 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
79 for (unsigned i
= 0; i
< elems
; i
++)
80 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
82 for (unsigned i
= 0; i
< elems
; i
++) {
83 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
84 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
92 static struct vtn_ssa_value
*
93 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
94 const struct glsl_type
*type
)
96 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
101 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
104 switch (glsl_get_base_type(type
)) {
108 case GLSL_TYPE_FLOAT
:
109 case GLSL_TYPE_DOUBLE
: {
110 int bit_size
= glsl_get_bit_size(type
);
111 if (glsl_type_is_vector_or_scalar(type
)) {
112 unsigned num_components
= glsl_get_vector_elements(val
->type
);
113 nir_load_const_instr
*load
=
114 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
116 load
->value
= constant
->values
[0];
118 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
119 val
->def
= &load
->def
;
121 assert(glsl_type_is_matrix(type
));
122 unsigned rows
= glsl_get_vector_elements(val
->type
);
123 unsigned columns
= glsl_get_matrix_columns(val
->type
);
124 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
126 for (unsigned i
= 0; i
< columns
; i
++) {
127 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
128 col_val
->type
= glsl_get_column_type(val
->type
);
129 nir_load_const_instr
*load
=
130 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
132 load
->value
= constant
->values
[i
];
134 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
135 col_val
->def
= &load
->def
;
137 val
->elems
[i
] = col_val
;
143 case GLSL_TYPE_ARRAY
: {
144 unsigned elems
= glsl_get_length(val
->type
);
145 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
146 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
147 for (unsigned i
= 0; i
< elems
; i
++)
148 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
153 case GLSL_TYPE_STRUCT
: {
154 unsigned elems
= glsl_get_length(val
->type
);
155 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
156 for (unsigned i
= 0; i
< elems
; i
++) {
157 const struct glsl_type
*elem_type
=
158 glsl_get_struct_field(val
->type
, i
);
159 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
166 unreachable("bad constant type");
172 struct vtn_ssa_value
*
173 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
175 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
176 switch (val
->value_type
) {
177 case vtn_value_type_undef
:
178 return vtn_undef_ssa_value(b
, val
->type
->type
);
180 case vtn_value_type_constant
:
181 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
183 case vtn_value_type_ssa
:
186 case vtn_value_type_access_chain
:
187 /* This is needed for function parameters */
188 return vtn_variable_load(b
, val
->access_chain
);
191 unreachable("Invalid type for an SSA value");
196 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
197 unsigned word_count
, unsigned *words_used
)
199 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
201 /* Ammount of space taken by the string (including the null) */
202 unsigned len
= strlen(dup
) + 1;
203 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
209 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
210 const uint32_t *end
, vtn_instruction_handler handler
)
216 const uint32_t *w
= start
;
218 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
219 unsigned count
= w
[0] >> SpvWordCountShift
;
220 assert(count
>= 1 && w
+ count
<= end
);
224 break; /* Do nothing */
227 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
239 if (!handler(b
, opcode
, w
, count
))
251 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
252 const uint32_t *w
, unsigned count
)
255 case SpvOpExtInstImport
: {
256 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
257 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
258 val
->ext_handler
= vtn_handle_glsl450_instruction
;
260 assert(!"Unsupported extension");
266 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
267 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
274 unreachable("Unhandled opcode");
279 _foreach_decoration_helper(struct vtn_builder
*b
,
280 struct vtn_value
*base_value
,
282 struct vtn_value
*value
,
283 vtn_decoration_foreach_cb cb
, void *data
)
285 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
287 if (dec
->scope
== VTN_DEC_DECORATION
) {
288 member
= parent_member
;
289 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
290 assert(parent_member
== -1);
291 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
293 /* Not a decoration */
298 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
299 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
302 cb(b
, base_value
, member
, dec
, data
);
307 /** Iterates (recursively if needed) over all of the decorations on a value
309 * This function iterates over all of the decorations applied to a given
310 * value. If it encounters a decoration group, it recurses into the group
311 * and iterates over all of those decorations as well.
314 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
315 vtn_decoration_foreach_cb cb
, void *data
)
317 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
321 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
322 vtn_execution_mode_foreach_cb cb
, void *data
)
324 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
325 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
328 assert(dec
->group
== NULL
);
329 cb(b
, value
, dec
, data
);
334 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
335 const uint32_t *w
, unsigned count
)
337 const uint32_t *w_end
= w
+ count
;
338 const uint32_t target
= w
[1];
342 case SpvOpDecorationGroup
:
343 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
347 case SpvOpMemberDecorate
:
348 case SpvOpExecutionMode
: {
349 struct vtn_value
*val
= &b
->values
[target
];
351 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
354 dec
->scope
= VTN_DEC_DECORATION
;
356 case SpvOpMemberDecorate
:
357 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
359 case SpvOpExecutionMode
:
360 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
363 unreachable("Invalid decoration opcode");
365 dec
->decoration
= *(w
++);
368 /* Link into the list */
369 dec
->next
= val
->decoration
;
370 val
->decoration
= dec
;
374 case SpvOpGroupMemberDecorate
:
375 case SpvOpGroupDecorate
: {
376 struct vtn_value
*group
=
377 vtn_value(b
, target
, vtn_value_type_decoration_group
);
379 for (; w
< w_end
; w
++) {
380 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
381 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
384 if (opcode
== SpvOpGroupDecorate
) {
385 dec
->scope
= VTN_DEC_DECORATION
;
387 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
390 /* Link into the list */
391 dec
->next
= val
->decoration
;
392 val
->decoration
= dec
;
398 unreachable("Unhandled opcode");
402 struct member_decoration_ctx
{
404 struct glsl_struct_field
*fields
;
405 struct vtn_type
*type
;
408 /* does a shallow copy of a vtn_type */
410 static struct vtn_type
*
411 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
413 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
414 dest
->type
= src
->type
;
415 dest
->is_builtin
= src
->is_builtin
;
417 dest
->builtin
= src
->builtin
;
419 if (!glsl_type_is_scalar(src
->type
)) {
420 switch (glsl_get_base_type(src
->type
)) {
424 case GLSL_TYPE_FLOAT
:
425 case GLSL_TYPE_DOUBLE
:
426 case GLSL_TYPE_ARRAY
:
427 dest
->row_major
= src
->row_major
;
428 dest
->stride
= src
->stride
;
429 dest
->array_element
= src
->array_element
;
432 case GLSL_TYPE_STRUCT
: {
433 unsigned elems
= glsl_get_length(src
->type
);
435 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
436 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
438 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
439 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
444 unreachable("unhandled type");
451 static struct vtn_type
*
452 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
454 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
455 type
= type
->members
[member
];
457 /* We may have an array of matrices.... Oh, joy! */
458 while (glsl_type_is_array(type
->type
)) {
459 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
460 type
= type
->array_element
;
463 assert(glsl_type_is_matrix(type
->type
));
469 struct_member_decoration_cb(struct vtn_builder
*b
,
470 struct vtn_value
*val
, int member
,
471 const struct vtn_decoration
*dec
, void *void_ctx
)
473 struct member_decoration_ctx
*ctx
= void_ctx
;
478 assert(member
< ctx
->num_fields
);
480 switch (dec
->decoration
) {
481 case SpvDecorationNonWritable
:
482 case SpvDecorationNonReadable
:
483 case SpvDecorationRelaxedPrecision
:
484 case SpvDecorationVolatile
:
485 case SpvDecorationCoherent
:
486 case SpvDecorationUniform
:
487 break; /* FIXME: Do nothing with this for now. */
488 case SpvDecorationNoPerspective
:
489 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
491 case SpvDecorationFlat
:
492 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
494 case SpvDecorationCentroid
:
495 ctx
->fields
[member
].centroid
= true;
497 case SpvDecorationSample
:
498 ctx
->fields
[member
].sample
= true;
500 case SpvDecorationStream
:
501 /* Vulkan only allows one GS stream */
502 assert(dec
->literals
[0] == 0);
504 case SpvDecorationLocation
:
505 ctx
->fields
[member
].location
= dec
->literals
[0];
507 case SpvDecorationComponent
:
508 break; /* FIXME: What should we do with these? */
509 case SpvDecorationBuiltIn
:
510 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
511 ctx
->type
->members
[member
]->is_builtin
= true;
512 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
513 ctx
->type
->builtin_block
= true;
515 case SpvDecorationOffset
:
516 ctx
->type
->offsets
[member
] = dec
->literals
[0];
518 case SpvDecorationMatrixStride
:
519 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
521 case SpvDecorationColMajor
:
522 break; /* Nothing to do here. Column-major is the default. */
523 case SpvDecorationRowMajor
:
524 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
527 case SpvDecorationPatch
:
528 vtn_warn("Tessellation not yet supported");
531 case SpvDecorationSpecId
:
532 case SpvDecorationBlock
:
533 case SpvDecorationBufferBlock
:
534 case SpvDecorationArrayStride
:
535 case SpvDecorationGLSLShared
:
536 case SpvDecorationGLSLPacked
:
537 case SpvDecorationInvariant
:
538 case SpvDecorationRestrict
:
539 case SpvDecorationAliased
:
540 case SpvDecorationConstant
:
541 case SpvDecorationIndex
:
542 case SpvDecorationBinding
:
543 case SpvDecorationDescriptorSet
:
544 case SpvDecorationLinkageAttributes
:
545 case SpvDecorationNoContraction
:
546 case SpvDecorationInputAttachmentIndex
:
547 vtn_warn("Decoration not allowed on struct members: %s",
548 spirv_decoration_to_string(dec
->decoration
));
551 case SpvDecorationXfbBuffer
:
552 case SpvDecorationXfbStride
:
553 vtn_warn("Vulkan does not have transform feedback");
556 case SpvDecorationCPacked
:
557 case SpvDecorationSaturatedConversion
:
558 case SpvDecorationFuncParamAttr
:
559 case SpvDecorationFPRoundingMode
:
560 case SpvDecorationFPFastMathMode
:
561 case SpvDecorationAlignment
:
562 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
563 spirv_decoration_to_string(dec
->decoration
));
569 type_decoration_cb(struct vtn_builder
*b
,
570 struct vtn_value
*val
, int member
,
571 const struct vtn_decoration
*dec
, void *ctx
)
573 struct vtn_type
*type
= val
->type
;
578 switch (dec
->decoration
) {
579 case SpvDecorationArrayStride
:
580 type
->stride
= dec
->literals
[0];
582 case SpvDecorationBlock
:
585 case SpvDecorationBufferBlock
:
586 type
->buffer_block
= true;
588 case SpvDecorationGLSLShared
:
589 case SpvDecorationGLSLPacked
:
590 /* Ignore these, since we get explicit offsets anyways */
593 case SpvDecorationRowMajor
:
594 case SpvDecorationColMajor
:
595 case SpvDecorationMatrixStride
:
596 case SpvDecorationBuiltIn
:
597 case SpvDecorationNoPerspective
:
598 case SpvDecorationFlat
:
599 case SpvDecorationPatch
:
600 case SpvDecorationCentroid
:
601 case SpvDecorationSample
:
602 case SpvDecorationVolatile
:
603 case SpvDecorationCoherent
:
604 case SpvDecorationNonWritable
:
605 case SpvDecorationNonReadable
:
606 case SpvDecorationUniform
:
607 case SpvDecorationStream
:
608 case SpvDecorationLocation
:
609 case SpvDecorationComponent
:
610 case SpvDecorationOffset
:
611 case SpvDecorationXfbBuffer
:
612 case SpvDecorationXfbStride
:
613 vtn_warn("Decoraiton only allowed for struct members: %s",
614 spirv_decoration_to_string(dec
->decoration
));
617 case SpvDecorationRelaxedPrecision
:
618 case SpvDecorationSpecId
:
619 case SpvDecorationInvariant
:
620 case SpvDecorationRestrict
:
621 case SpvDecorationAliased
:
622 case SpvDecorationConstant
:
623 case SpvDecorationIndex
:
624 case SpvDecorationBinding
:
625 case SpvDecorationDescriptorSet
:
626 case SpvDecorationLinkageAttributes
:
627 case SpvDecorationNoContraction
:
628 case SpvDecorationInputAttachmentIndex
:
629 vtn_warn("Decoraiton not allowed on types: %s",
630 spirv_decoration_to_string(dec
->decoration
));
633 case SpvDecorationCPacked
:
634 case SpvDecorationSaturatedConversion
:
635 case SpvDecorationFuncParamAttr
:
636 case SpvDecorationFPRoundingMode
:
637 case SpvDecorationFPFastMathMode
:
638 case SpvDecorationAlignment
:
639 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
640 spirv_decoration_to_string(dec
->decoration
));
646 translate_image_format(SpvImageFormat format
)
649 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
650 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
651 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
652 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
653 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
654 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
655 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
656 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
657 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
658 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
659 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
660 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
661 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
662 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
663 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
664 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
665 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
666 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
667 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
668 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
669 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
670 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
671 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
672 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
673 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
674 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
675 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
676 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
677 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
678 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
679 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
680 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
681 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
682 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
683 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
684 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
685 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
686 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
687 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
688 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
690 assert(!"Invalid image format");
696 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
697 const uint32_t *w
, unsigned count
)
699 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
701 val
->type
= rzalloc(b
, struct vtn_type
);
702 val
->type
->is_builtin
= false;
703 val
->type
->val
= val
;
707 val
->type
->type
= glsl_void_type();
710 val
->type
->type
= glsl_bool_type();
713 const bool signedness
= w
[3];
714 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
717 case SpvOpTypeFloat
: {
719 val
->type
->type
= bit_size
== 64 ? glsl_double_type() : glsl_float_type();
723 case SpvOpTypeVector
: {
724 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
725 unsigned elems
= w
[3];
727 assert(glsl_type_is_scalar(base
->type
));
728 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
730 /* Vectors implicitly have sizeof(base_type) stride. For now, this
731 * is always 4 bytes. This will have to change if we want to start
732 * supporting doubles or half-floats.
734 val
->type
->stride
= 4;
735 val
->type
->array_element
= base
;
739 case SpvOpTypeMatrix
: {
740 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
741 unsigned columns
= w
[3];
743 assert(glsl_type_is_vector(base
->type
));
744 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
745 glsl_get_vector_elements(base
->type
),
747 assert(!glsl_type_is_error(val
->type
->type
));
748 val
->type
->array_element
= base
;
749 val
->type
->row_major
= false;
750 val
->type
->stride
= 0;
754 case SpvOpTypeRuntimeArray
:
755 case SpvOpTypeArray
: {
756 struct vtn_type
*array_element
=
757 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
760 if (opcode
== SpvOpTypeRuntimeArray
) {
761 /* A length of 0 is used to denote unsized arrays */
765 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
768 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
769 val
->type
->array_element
= array_element
;
770 val
->type
->stride
= 0;
774 case SpvOpTypeStruct
: {
775 unsigned num_fields
= count
- 2;
776 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
777 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
779 NIR_VLA(struct glsl_struct_field
, fields
, count
);
780 for (unsigned i
= 0; i
< num_fields
; i
++) {
781 val
->type
->members
[i
] =
782 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
783 fields
[i
] = (struct glsl_struct_field
) {
784 .type
= val
->type
->members
[i
]->type
,
785 .name
= ralloc_asprintf(b
, "field%d", i
),
790 struct member_decoration_ctx ctx
= {
791 .num_fields
= num_fields
,
796 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
798 const char *name
= val
->name
? val
->name
: "struct";
800 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
804 case SpvOpTypeFunction
: {
805 const struct glsl_type
*return_type
=
806 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
807 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
808 for (unsigned i
= 0; i
< count
- 3; i
++) {
809 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
813 params
[i
].out
= true;
815 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
819 case SpvOpTypePointer
:
820 /* FIXME: For now, we'll just do the really lame thing and return
821 * the same type. The validator should ensure that the proper number
822 * of dereferences happen
824 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
827 case SpvOpTypeImage
: {
828 const struct glsl_type
*sampled_type
=
829 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
831 assert(glsl_type_is_vector_or_scalar(sampled_type
));
833 enum glsl_sampler_dim dim
;
834 switch ((SpvDim
)w
[3]) {
835 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
836 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
837 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
838 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
839 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
840 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
841 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
843 unreachable("Invalid SPIR-V Sampler dimension");
846 bool is_shadow
= w
[4];
847 bool is_array
= w
[5];
848 bool multisampled
= w
[6];
849 unsigned sampled
= w
[7];
850 SpvImageFormat format
= w
[8];
853 val
->type
->access_qualifier
= w
[9];
855 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
858 assert(dim
== GLSL_SAMPLER_DIM_2D
);
859 dim
= GLSL_SAMPLER_DIM_MS
;
862 val
->type
->image_format
= translate_image_format(format
);
865 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
866 glsl_get_base_type(sampled_type
));
867 } else if (sampled
== 2) {
868 assert((dim
== GLSL_SAMPLER_DIM_SUBPASS
) || format
);
870 val
->type
->type
= glsl_image_type(dim
, is_array
,
871 glsl_get_base_type(sampled_type
));
873 assert(!"We need to know if the image will be sampled");
878 case SpvOpTypeSampledImage
:
879 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
882 case SpvOpTypeSampler
:
883 /* The actual sampler type here doesn't really matter. It gets
884 * thrown away the moment you combine it with an image. What really
885 * matters is that it's a sampler type as opposed to an integer type
886 * so the backend knows what to do.
888 val
->type
->type
= glsl_bare_sampler_type();
891 case SpvOpTypeOpaque
:
893 case SpvOpTypeDeviceEvent
:
894 case SpvOpTypeReserveId
:
898 unreachable("Unhandled opcode");
901 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
904 static nir_constant
*
905 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
907 nir_constant
*c
= rzalloc(b
, nir_constant
);
909 switch (glsl_get_base_type(type
)) {
913 case GLSL_TYPE_FLOAT
:
914 case GLSL_TYPE_DOUBLE
:
915 /* Nothing to do here. It's already initialized to zero */
918 case GLSL_TYPE_ARRAY
:
919 assert(glsl_get_length(type
) > 0);
920 c
->num_elements
= glsl_get_length(type
);
921 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
923 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
924 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
925 c
->elements
[i
] = c
->elements
[0];
928 case GLSL_TYPE_STRUCT
:
929 c
->num_elements
= glsl_get_length(type
);
930 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
932 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
933 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
938 unreachable("Invalid type for null constant");
945 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
946 int member
, const struct vtn_decoration
*dec
,
949 assert(member
== -1);
950 if (dec
->decoration
!= SpvDecorationSpecId
)
953 struct spec_constant_value
*const_value
= data
;
955 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
956 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
957 if (const_value
->is_double
)
958 const_value
->data64
= b
->specializations
[i
].data64
;
960 const_value
->data32
= b
->specializations
[i
].data32
;
967 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
968 uint32_t const_value
)
970 struct spec_constant_value data
;
971 data
.is_double
= false;
972 data
.data32
= const_value
;
973 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
978 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
979 uint64_t const_value
)
981 struct spec_constant_value data
;
982 data
.is_double
= true;
983 data
.data64
= const_value
;
984 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
989 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
990 struct vtn_value
*val
,
992 const struct vtn_decoration
*dec
,
995 assert(member
== -1);
996 if (dec
->decoration
!= SpvDecorationBuiltIn
||
997 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1000 assert(val
->const_type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1002 b
->shader
->info
->cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1003 b
->shader
->info
->cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1004 b
->shader
->info
->cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1008 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1009 const uint32_t *w
, unsigned count
)
1011 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1012 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1013 val
->constant
= rzalloc(b
, nir_constant
);
1015 case SpvOpConstantTrue
:
1016 assert(val
->const_type
== glsl_bool_type());
1017 val
->constant
->values
[0].u32
[0] = NIR_TRUE
;
1019 case SpvOpConstantFalse
:
1020 assert(val
->const_type
== glsl_bool_type());
1021 val
->constant
->values
[0].u32
[0] = NIR_FALSE
;
1024 case SpvOpSpecConstantTrue
:
1025 case SpvOpSpecConstantFalse
: {
1026 assert(val
->const_type
== glsl_bool_type());
1028 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1029 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1033 case SpvOpConstant
: {
1034 assert(glsl_type_is_scalar(val
->const_type
));
1035 int bit_size
= glsl_get_bit_size(val
->const_type
);
1036 if (bit_size
== 64) {
1037 val
->constant
->values
->u32
[0] = w
[3];
1038 val
->constant
->values
->u32
[1] = w
[4];
1040 assert(bit_size
== 32);
1041 val
->constant
->values
->u32
[0] = w
[3];
1045 case SpvOpSpecConstant
: {
1046 assert(glsl_type_is_scalar(val
->const_type
));
1047 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1048 int bit_size
= glsl_get_bit_size(val
->const_type
);
1050 val
->constant
->values
[0].u64
[0] =
1051 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1053 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1056 case SpvOpSpecConstantComposite
:
1057 case SpvOpConstantComposite
: {
1058 unsigned elem_count
= count
- 3;
1059 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1060 for (unsigned i
= 0; i
< elem_count
; i
++)
1061 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1063 switch (glsl_get_base_type(val
->const_type
)) {
1064 case GLSL_TYPE_UINT
:
1066 case GLSL_TYPE_FLOAT
:
1067 case GLSL_TYPE_BOOL
:
1068 case GLSL_TYPE_DOUBLE
: {
1069 int bit_size
= glsl_get_bit_size(val
->const_type
);
1070 if (glsl_type_is_matrix(val
->const_type
)) {
1071 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
1072 for (unsigned i
= 0; i
< elem_count
; i
++)
1073 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1075 assert(glsl_type_is_vector(val
->const_type
));
1076 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
1077 for (unsigned i
= 0; i
< elem_count
; i
++) {
1078 if (bit_size
== 64) {
1079 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1081 assert(bit_size
== 32);
1082 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1089 case GLSL_TYPE_STRUCT
:
1090 case GLSL_TYPE_ARRAY
:
1091 ralloc_steal(val
->constant
, elems
);
1092 val
->constant
->num_elements
= elem_count
;
1093 val
->constant
->elements
= elems
;
1097 unreachable("Unsupported type for constants");
1102 case SpvOpSpecConstantOp
: {
1103 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1105 case SpvOpVectorShuffle
: {
1106 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1107 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1108 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
1109 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
1111 assert(len0
+ len1
< 16);
1113 unsigned bit_size
= glsl_get_bit_size(val
->const_type
);
1114 assert(bit_size
== glsl_get_bit_size(v0
->const_type
) &&
1115 bit_size
== glsl_get_bit_size(v1
->const_type
));
1117 if (bit_size
== 64) {
1119 for (unsigned i
= 0; i
< len0
; i
++)
1120 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1121 for (unsigned i
= 0; i
< len1
; i
++)
1122 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1124 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1125 uint32_t comp
= w
[i
+ 6];
1126 /* If component is not used, set the value to a known constant
1127 * to detect if it is wrongly used.
1129 if (comp
== (uint32_t)-1)
1130 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1132 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1136 for (unsigned i
= 0; i
< len0
; i
++)
1137 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1139 for (unsigned i
= 0; i
< len1
; i
++)
1140 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1142 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1143 uint32_t comp
= w
[i
+ 6];
1144 /* If component is not used, set the value to a known constant
1145 * to detect if it is wrongly used.
1147 if (comp
== (uint32_t)-1)
1148 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1150 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1156 case SpvOpCompositeExtract
:
1157 case SpvOpCompositeInsert
: {
1158 struct vtn_value
*comp
;
1159 unsigned deref_start
;
1160 struct nir_constant
**c
;
1161 if (opcode
== SpvOpCompositeExtract
) {
1162 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1164 c
= &comp
->constant
;
1166 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1168 val
->constant
= nir_constant_clone(comp
->constant
,
1175 const struct glsl_type
*type
= comp
->const_type
;
1176 for (unsigned i
= deref_start
; i
< count
; i
++) {
1177 switch (glsl_get_base_type(type
)) {
1178 case GLSL_TYPE_UINT
:
1180 case GLSL_TYPE_FLOAT
:
1181 case GLSL_TYPE_DOUBLE
:
1182 case GLSL_TYPE_BOOL
:
1183 /* If we hit this granularity, we're picking off an element */
1184 if (glsl_type_is_matrix(type
)) {
1185 assert(col
== 0 && elem
== -1);
1188 type
= glsl_get_column_type(type
);
1190 assert(elem
<= 0 && glsl_type_is_vector(type
));
1192 type
= glsl_scalar_type(glsl_get_base_type(type
));
1196 case GLSL_TYPE_ARRAY
:
1197 c
= &(*c
)->elements
[w
[i
]];
1198 type
= glsl_get_array_element(type
);
1201 case GLSL_TYPE_STRUCT
:
1202 c
= &(*c
)->elements
[w
[i
]];
1203 type
= glsl_get_struct_field(type
, w
[i
]);
1207 unreachable("Invalid constant type");
1211 if (opcode
== SpvOpCompositeExtract
) {
1215 unsigned num_components
= glsl_get_vector_elements(type
);
1216 unsigned bit_size
= glsl_get_bit_size(type
);
1217 for (unsigned i
= 0; i
< num_components
; i
++)
1218 if (bit_size
== 64) {
1219 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1221 assert(bit_size
== 32);
1222 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1226 struct vtn_value
*insert
=
1227 vtn_value(b
, w
[4], vtn_value_type_constant
);
1228 assert(insert
->const_type
== type
);
1230 *c
= insert
->constant
;
1232 unsigned num_components
= glsl_get_vector_elements(type
);
1233 unsigned bit_size
= glsl_get_bit_size(type
);
1234 for (unsigned i
= 0; i
< num_components
; i
++)
1235 if (bit_size
== 64) {
1236 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1238 assert(bit_size
== 32);
1239 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1248 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->const_type
);
1249 nir_alu_type src_alu_type
= dst_alu_type
;
1250 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
1252 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1254 glsl_get_bit_size(val
->const_type
);
1256 nir_const_value src
[4];
1258 for (unsigned i
= 0; i
< count
- 4; i
++) {
1260 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1262 unsigned j
= swap
? 1 - i
: i
;
1263 assert(bit_size
== 32);
1264 src
[j
] = c
->values
[0];
1267 val
->constant
->values
[0] =
1268 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1275 case SpvOpConstantNull
:
1276 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1279 case SpvOpConstantSampler
:
1280 assert(!"OpConstantSampler requires Kernel Capability");
1284 unreachable("Unhandled opcode");
1287 /* Now that we have the value, update the workgroup size if needed */
1288 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1292 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1293 const uint32_t *w
, unsigned count
)
1295 struct nir_function
*callee
=
1296 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1298 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1299 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1300 unsigned arg_id
= w
[4 + i
];
1301 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1302 if (arg
->value_type
== vtn_value_type_access_chain
) {
1303 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1304 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1306 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1308 /* Make a temporary to store the argument in */
1310 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1311 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1313 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1317 nir_variable
*out_tmp
= NULL
;
1318 if (!glsl_type_is_void(callee
->return_type
)) {
1319 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1321 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1324 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1326 if (glsl_type_is_void(callee
->return_type
)) {
1327 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1329 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1330 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1334 struct vtn_ssa_value
*
1335 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1337 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1340 if (!glsl_type_is_vector_or_scalar(type
)) {
1341 unsigned elems
= glsl_get_length(type
);
1342 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1343 for (unsigned i
= 0; i
< elems
; i
++) {
1344 const struct glsl_type
*child_type
;
1346 switch (glsl_get_base_type(type
)) {
1348 case GLSL_TYPE_UINT
:
1349 case GLSL_TYPE_BOOL
:
1350 case GLSL_TYPE_FLOAT
:
1351 case GLSL_TYPE_DOUBLE
:
1352 child_type
= glsl_get_column_type(type
);
1354 case GLSL_TYPE_ARRAY
:
1355 child_type
= glsl_get_array_element(type
);
1357 case GLSL_TYPE_STRUCT
:
1358 child_type
= glsl_get_struct_field(type
, i
);
1361 unreachable("unkown base type");
1364 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1372 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1375 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1376 src
.src_type
= type
;
1381 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1382 const uint32_t *w
, unsigned count
)
1384 if (opcode
== SpvOpSampledImage
) {
1385 struct vtn_value
*val
=
1386 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1387 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1388 val
->sampled_image
->image
=
1389 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1390 val
->sampled_image
->sampler
=
1391 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1393 } else if (opcode
== SpvOpImage
) {
1394 struct vtn_value
*val
=
1395 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1396 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1397 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1398 val
->access_chain
= src_val
->sampled_image
->image
;
1400 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1401 val
->access_chain
= src_val
->access_chain
;
1406 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1407 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1409 struct vtn_sampled_image sampled
;
1410 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1411 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1412 sampled
= *sampled_val
->sampled_image
;
1414 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1415 sampled
.image
= NULL
;
1416 sampled
.sampler
= sampled_val
->access_chain
;
1419 const struct glsl_type
*image_type
;
1420 if (sampled
.image
) {
1421 image_type
= sampled
.image
->var
->var
->interface_type
;
1423 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1425 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1426 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1427 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1429 /* Figure out the base texture operation */
1432 case SpvOpImageSampleImplicitLod
:
1433 case SpvOpImageSampleDrefImplicitLod
:
1434 case SpvOpImageSampleProjImplicitLod
:
1435 case SpvOpImageSampleProjDrefImplicitLod
:
1436 texop
= nir_texop_tex
;
1439 case SpvOpImageSampleExplicitLod
:
1440 case SpvOpImageSampleDrefExplicitLod
:
1441 case SpvOpImageSampleProjExplicitLod
:
1442 case SpvOpImageSampleProjDrefExplicitLod
:
1443 texop
= nir_texop_txl
;
1446 case SpvOpImageFetch
:
1447 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1448 texop
= nir_texop_txf_ms
;
1450 texop
= nir_texop_txf
;
1454 case SpvOpImageGather
:
1455 case SpvOpImageDrefGather
:
1456 texop
= nir_texop_tg4
;
1459 case SpvOpImageQuerySizeLod
:
1460 case SpvOpImageQuerySize
:
1461 texop
= nir_texop_txs
;
1464 case SpvOpImageQueryLod
:
1465 texop
= nir_texop_lod
;
1468 case SpvOpImageQueryLevels
:
1469 texop
= nir_texop_query_levels
;
1472 case SpvOpImageQuerySamples
:
1473 texop
= nir_texop_texture_samples
;
1477 unreachable("Unhandled opcode");
1480 nir_tex_src srcs
[8]; /* 8 should be enough */
1481 nir_tex_src
*p
= srcs
;
1485 struct nir_ssa_def
*coord
;
1486 unsigned coord_components
;
1488 case SpvOpImageSampleImplicitLod
:
1489 case SpvOpImageSampleExplicitLod
:
1490 case SpvOpImageSampleDrefImplicitLod
:
1491 case SpvOpImageSampleDrefExplicitLod
:
1492 case SpvOpImageSampleProjImplicitLod
:
1493 case SpvOpImageSampleProjExplicitLod
:
1494 case SpvOpImageSampleProjDrefImplicitLod
:
1495 case SpvOpImageSampleProjDrefExplicitLod
:
1496 case SpvOpImageFetch
:
1497 case SpvOpImageGather
:
1498 case SpvOpImageDrefGather
:
1499 case SpvOpImageQueryLod
: {
1500 /* All these types have the coordinate as their first real argument */
1501 switch (sampler_dim
) {
1502 case GLSL_SAMPLER_DIM_1D
:
1503 case GLSL_SAMPLER_DIM_BUF
:
1504 coord_components
= 1;
1506 case GLSL_SAMPLER_DIM_2D
:
1507 case GLSL_SAMPLER_DIM_RECT
:
1508 case GLSL_SAMPLER_DIM_MS
:
1509 coord_components
= 2;
1511 case GLSL_SAMPLER_DIM_3D
:
1512 case GLSL_SAMPLER_DIM_CUBE
:
1513 coord_components
= 3;
1516 unreachable("Invalid sampler type");
1519 if (is_array
&& texop
!= nir_texop_lod
)
1522 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1523 p
->src
= nir_src_for_ssa(coord
);
1524 p
->src_type
= nir_tex_src_coord
;
1531 coord_components
= 0;
1536 case SpvOpImageSampleProjImplicitLod
:
1537 case SpvOpImageSampleProjExplicitLod
:
1538 case SpvOpImageSampleProjDrefImplicitLod
:
1539 case SpvOpImageSampleProjDrefExplicitLod
:
1540 /* These have the projector as the last coordinate component */
1541 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1542 p
->src_type
= nir_tex_src_projector
;
1550 unsigned gather_component
= 0;
1552 case SpvOpImageSampleDrefImplicitLod
:
1553 case SpvOpImageSampleDrefExplicitLod
:
1554 case SpvOpImageSampleProjDrefImplicitLod
:
1555 case SpvOpImageSampleProjDrefExplicitLod
:
1556 case SpvOpImageDrefGather
:
1557 /* These all have an explicit depth value as their next source */
1558 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1561 case SpvOpImageGather
:
1562 /* This has a component as its next source */
1564 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1571 /* For OpImageQuerySizeLod, we always have an LOD */
1572 if (opcode
== SpvOpImageQuerySizeLod
)
1573 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1575 /* Now we need to handle some number of optional arguments */
1576 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1578 uint32_t operands
= w
[idx
++];
1580 if (operands
& SpvImageOperandsBiasMask
) {
1581 assert(texop
== nir_texop_tex
);
1582 texop
= nir_texop_txb
;
1583 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1586 if (operands
& SpvImageOperandsLodMask
) {
1587 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1588 texop
== nir_texop_txs
);
1589 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1592 if (operands
& SpvImageOperandsGradMask
) {
1593 assert(texop
== nir_texop_txl
);
1594 texop
= nir_texop_txd
;
1595 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1596 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1599 if (operands
& SpvImageOperandsOffsetMask
||
1600 operands
& SpvImageOperandsConstOffsetMask
)
1601 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1603 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1604 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1605 (*p
++) = (nir_tex_src
){};
1608 if (operands
& SpvImageOperandsSampleMask
) {
1609 assert(texop
== nir_texop_txf_ms
);
1610 texop
= nir_texop_txf_ms
;
1611 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1614 /* We should have now consumed exactly all of the arguments */
1615 assert(idx
== count
);
1617 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1620 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1622 instr
->coord_components
= coord_components
;
1623 instr
->sampler_dim
= sampler_dim
;
1624 instr
->is_array
= is_array
;
1625 instr
->is_shadow
= is_shadow
;
1626 instr
->is_new_style_shadow
=
1627 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1628 instr
->component
= gather_component
;
1630 switch (glsl_get_sampler_result_type(image_type
)) {
1631 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1632 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1633 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1634 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1636 unreachable("Invalid base type for sampler result");
1639 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1640 nir_deref_var
*texture
;
1641 if (sampled
.image
) {
1642 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1648 instr
->texture
= nir_deref_var_clone(texture
, instr
);
1650 switch (instr
->op
) {
1655 /* These operations require a sampler */
1656 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
1659 case nir_texop_txf_ms
:
1663 case nir_texop_query_levels
:
1664 case nir_texop_texture_samples
:
1665 case nir_texop_samples_identical
:
1667 instr
->sampler
= NULL
;
1669 case nir_texop_txf_ms_mcs
:
1670 unreachable("unexpected nir_texop_txf_ms_mcs");
1673 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1674 nir_tex_instr_dest_size(instr
), 32, NULL
);
1676 assert(glsl_get_vector_elements(ret_type
->type
) ==
1677 nir_tex_instr_dest_size(instr
));
1680 nir_instr
*instruction
;
1681 if (gather_offsets
) {
1682 assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
1683 assert(glsl_get_length(gather_offsets
->type
) == 4);
1684 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
1686 /* Copy the current instruction 4x */
1687 for (uint32_t i
= 1; i
< 4; i
++) {
1688 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
1689 instrs
[i
]->op
= instr
->op
;
1690 instrs
[i
]->coord_components
= instr
->coord_components
;
1691 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
1692 instrs
[i
]->is_array
= instr
->is_array
;
1693 instrs
[i
]->is_shadow
= instr
->is_shadow
;
1694 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
1695 instrs
[i
]->component
= instr
->component
;
1696 instrs
[i
]->dest_type
= instr
->dest_type
;
1697 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
1698 instrs
[i
]->sampler
= NULL
;
1700 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1702 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
1703 nir_tex_instr_dest_size(instr
), 32, NULL
);
1706 /* Fill in the last argument with the offset from the passed in offsets
1707 * and insert the instruction into the stream.
1709 for (uint32_t i
= 0; i
< 4; i
++) {
1711 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
1712 src
.src_type
= nir_tex_src_offset
;
1713 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
1714 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
1717 /* Combine the results of the 4 instructions by taking their .w
1720 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
1721 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
1722 vec4
->dest
.write_mask
= 0xf;
1723 for (uint32_t i
= 0; i
< 4; i
++) {
1724 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
1725 vec4
->src
[i
].swizzle
[0] = 3;
1727 def
= &vec4
->dest
.dest
.ssa
;
1728 instruction
= &vec4
->instr
;
1730 def
= &instr
->dest
.ssa
;
1731 instruction
= &instr
->instr
;
1734 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1735 val
->ssa
->def
= def
;
1737 nir_builder_instr_insert(&b
->nb
, instruction
);
1741 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1742 const uint32_t *w
, nir_src
*src
)
1745 case SpvOpAtomicIIncrement
:
1746 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1749 case SpvOpAtomicIDecrement
:
1750 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1753 case SpvOpAtomicISub
:
1755 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1758 case SpvOpAtomicCompareExchange
:
1759 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1760 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1763 case SpvOpAtomicExchange
:
1764 case SpvOpAtomicIAdd
:
1765 case SpvOpAtomicSMin
:
1766 case SpvOpAtomicUMin
:
1767 case SpvOpAtomicSMax
:
1768 case SpvOpAtomicUMax
:
1769 case SpvOpAtomicAnd
:
1771 case SpvOpAtomicXor
:
1772 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1776 unreachable("Invalid SPIR-V atomic");
1780 static nir_ssa_def
*
1781 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1783 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1785 /* The image_load_store intrinsics assume a 4-dim coordinate */
1786 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1787 unsigned swizzle
[4];
1788 for (unsigned i
= 0; i
< 4; i
++)
1789 swizzle
[i
] = MIN2(i
, dim
- 1);
1791 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1795 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1796 const uint32_t *w
, unsigned count
)
1798 /* Just get this one out of the way */
1799 if (opcode
== SpvOpImageTexelPointer
) {
1800 struct vtn_value
*val
=
1801 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1802 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1805 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1806 val
->image
->coord
= get_image_coord(b
, w
[4]);
1807 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1811 struct vtn_image_pointer image
;
1814 case SpvOpAtomicExchange
:
1815 case SpvOpAtomicCompareExchange
:
1816 case SpvOpAtomicCompareExchangeWeak
:
1817 case SpvOpAtomicIIncrement
:
1818 case SpvOpAtomicIDecrement
:
1819 case SpvOpAtomicIAdd
:
1820 case SpvOpAtomicISub
:
1821 case SpvOpAtomicLoad
:
1822 case SpvOpAtomicSMin
:
1823 case SpvOpAtomicUMin
:
1824 case SpvOpAtomicSMax
:
1825 case SpvOpAtomicUMax
:
1826 case SpvOpAtomicAnd
:
1828 case SpvOpAtomicXor
:
1829 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1832 case SpvOpAtomicStore
:
1833 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
1836 case SpvOpImageQuerySize
:
1838 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1840 image
.sample
= NULL
;
1843 case SpvOpImageRead
:
1845 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1846 image
.coord
= get_image_coord(b
, w
[4]);
1848 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1849 assert(w
[5] == SpvImageOperandsSampleMask
);
1850 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1852 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1856 case SpvOpImageWrite
:
1858 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1859 image
.coord
= get_image_coord(b
, w
[2]);
1863 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1864 assert(w
[4] == SpvImageOperandsSampleMask
);
1865 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1867 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1872 unreachable("Invalid image opcode");
1875 nir_intrinsic_op op
;
1877 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1878 OP(ImageQuerySize
, size
)
1880 OP(ImageWrite
, store
)
1881 OP(AtomicLoad
, load
)
1882 OP(AtomicStore
, store
)
1883 OP(AtomicExchange
, atomic_exchange
)
1884 OP(AtomicCompareExchange
, atomic_comp_swap
)
1885 OP(AtomicIIncrement
, atomic_add
)
1886 OP(AtomicIDecrement
, atomic_add
)
1887 OP(AtomicIAdd
, atomic_add
)
1888 OP(AtomicISub
, atomic_add
)
1889 OP(AtomicSMin
, atomic_min
)
1890 OP(AtomicUMin
, atomic_min
)
1891 OP(AtomicSMax
, atomic_max
)
1892 OP(AtomicUMax
, atomic_max
)
1893 OP(AtomicAnd
, atomic_and
)
1894 OP(AtomicOr
, atomic_or
)
1895 OP(AtomicXor
, atomic_xor
)
1898 unreachable("Invalid image opcode");
1901 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1903 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1904 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
1906 /* ImageQuerySize doesn't take any extra parameters */
1907 if (opcode
!= SpvOpImageQuerySize
) {
1908 /* The image coordinate is always 4 components but we may not have that
1909 * many. Swizzle to compensate.
1912 for (unsigned i
= 0; i
< 4; i
++)
1913 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1914 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1916 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1920 case SpvOpAtomicLoad
:
1921 case SpvOpImageQuerySize
:
1922 case SpvOpImageRead
:
1924 case SpvOpAtomicStore
:
1925 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
1927 case SpvOpImageWrite
:
1928 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1931 case SpvOpAtomicIIncrement
:
1932 case SpvOpAtomicIDecrement
:
1933 case SpvOpAtomicExchange
:
1934 case SpvOpAtomicIAdd
:
1935 case SpvOpAtomicSMin
:
1936 case SpvOpAtomicUMin
:
1937 case SpvOpAtomicSMax
:
1938 case SpvOpAtomicUMax
:
1939 case SpvOpAtomicAnd
:
1941 case SpvOpAtomicXor
:
1942 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
1946 unreachable("Invalid image opcode");
1949 if (opcode
!= SpvOpImageWrite
) {
1950 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1951 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1952 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, 32, NULL
);
1954 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1956 /* The image intrinsics always return 4 channels but we may not want
1957 * that many. Emit a mov to trim it down.
1959 unsigned swiz
[4] = {0, 1, 2, 3};
1960 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1961 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1962 glsl_get_vector_elements(type
->type
), false);
1964 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1968 static nir_intrinsic_op
1969 get_ssbo_nir_atomic_op(SpvOp opcode
)
1972 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
1973 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
1974 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1975 OP(AtomicExchange
, atomic_exchange
)
1976 OP(AtomicCompareExchange
, atomic_comp_swap
)
1977 OP(AtomicIIncrement
, atomic_add
)
1978 OP(AtomicIDecrement
, atomic_add
)
1979 OP(AtomicIAdd
, atomic_add
)
1980 OP(AtomicISub
, atomic_add
)
1981 OP(AtomicSMin
, atomic_imin
)
1982 OP(AtomicUMin
, atomic_umin
)
1983 OP(AtomicSMax
, atomic_imax
)
1984 OP(AtomicUMax
, atomic_umax
)
1985 OP(AtomicAnd
, atomic_and
)
1986 OP(AtomicOr
, atomic_or
)
1987 OP(AtomicXor
, atomic_xor
)
1990 unreachable("Invalid SSBO atomic");
1994 static nir_intrinsic_op
1995 get_shared_nir_atomic_op(SpvOp opcode
)
1998 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
1999 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2000 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2001 OP(AtomicExchange
, atomic_exchange
)
2002 OP(AtomicCompareExchange
, atomic_comp_swap
)
2003 OP(AtomicIIncrement
, atomic_add
)
2004 OP(AtomicIDecrement
, atomic_add
)
2005 OP(AtomicIAdd
, atomic_add
)
2006 OP(AtomicISub
, atomic_add
)
2007 OP(AtomicSMin
, atomic_imin
)
2008 OP(AtomicUMin
, atomic_umin
)
2009 OP(AtomicSMax
, atomic_imax
)
2010 OP(AtomicUMax
, atomic_umax
)
2011 OP(AtomicAnd
, atomic_and
)
2012 OP(AtomicOr
, atomic_or
)
2013 OP(AtomicXor
, atomic_xor
)
2016 unreachable("Invalid shared atomic");
2021 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2022 const uint32_t *w
, unsigned count
)
2024 struct vtn_access_chain
*chain
;
2025 nir_intrinsic_instr
*atomic
;
2028 case SpvOpAtomicLoad
:
2029 case SpvOpAtomicExchange
:
2030 case SpvOpAtomicCompareExchange
:
2031 case SpvOpAtomicCompareExchangeWeak
:
2032 case SpvOpAtomicIIncrement
:
2033 case SpvOpAtomicIDecrement
:
2034 case SpvOpAtomicIAdd
:
2035 case SpvOpAtomicISub
:
2036 case SpvOpAtomicSMin
:
2037 case SpvOpAtomicUMin
:
2038 case SpvOpAtomicSMax
:
2039 case SpvOpAtomicUMax
:
2040 case SpvOpAtomicAnd
:
2042 case SpvOpAtomicXor
:
2044 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2047 case SpvOpAtomicStore
:
2049 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
2053 unreachable("Invalid SPIR-V atomic");
2057 SpvScope scope = w[4];
2058 SpvMemorySemanticsMask semantics = w[5];
2061 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
2062 struct vtn_type
*type
= chain
->var
->type
;
2063 nir_deref_var
*deref
= vtn_access_chain_to_deref(b
, chain
);
2064 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
2065 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2066 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2069 case SpvOpAtomicLoad
:
2070 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2073 case SpvOpAtomicStore
:
2074 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2075 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2076 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2079 case SpvOpAtomicExchange
:
2080 case SpvOpAtomicCompareExchange
:
2081 case SpvOpAtomicCompareExchangeWeak
:
2082 case SpvOpAtomicIIncrement
:
2083 case SpvOpAtomicIDecrement
:
2084 case SpvOpAtomicIAdd
:
2085 case SpvOpAtomicISub
:
2086 case SpvOpAtomicSMin
:
2087 case SpvOpAtomicUMin
:
2088 case SpvOpAtomicSMax
:
2089 case SpvOpAtomicUMax
:
2090 case SpvOpAtomicAnd
:
2092 case SpvOpAtomicXor
:
2093 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2097 unreachable("Invalid SPIR-V atomic");
2101 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
2102 struct vtn_type
*type
;
2103 nir_ssa_def
*offset
, *index
;
2104 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
2106 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
2108 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2111 case SpvOpAtomicLoad
:
2112 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2113 atomic
->src
[0] = nir_src_for_ssa(index
);
2114 atomic
->src
[1] = nir_src_for_ssa(offset
);
2117 case SpvOpAtomicStore
:
2118 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2119 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2120 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2121 atomic
->src
[1] = nir_src_for_ssa(index
);
2122 atomic
->src
[2] = nir_src_for_ssa(offset
);
2125 case SpvOpAtomicExchange
:
2126 case SpvOpAtomicCompareExchange
:
2127 case SpvOpAtomicCompareExchangeWeak
:
2128 case SpvOpAtomicIIncrement
:
2129 case SpvOpAtomicIDecrement
:
2130 case SpvOpAtomicIAdd
:
2131 case SpvOpAtomicISub
:
2132 case SpvOpAtomicSMin
:
2133 case SpvOpAtomicUMin
:
2134 case SpvOpAtomicSMax
:
2135 case SpvOpAtomicUMax
:
2136 case SpvOpAtomicAnd
:
2138 case SpvOpAtomicXor
:
2139 atomic
->src
[0] = nir_src_for_ssa(index
);
2140 atomic
->src
[1] = nir_src_for_ssa(offset
);
2141 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
2145 unreachable("Invalid SPIR-V atomic");
2149 if (opcode
!= SpvOpAtomicStore
) {
2150 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2152 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2153 glsl_get_vector_elements(type
->type
),
2154 glsl_get_bit_size(type
->type
), NULL
);
2156 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2157 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2158 val
->ssa
->def
= &atomic
->dest
.ssa
;
2159 val
->ssa
->type
= type
->type
;
2162 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2165 static nir_alu_instr
*
2166 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
2169 switch (num_components
) {
2170 case 1: op
= nir_op_fmov
; break;
2171 case 2: op
= nir_op_vec2
; break;
2172 case 3: op
= nir_op_vec3
; break;
2173 case 4: op
= nir_op_vec4
; break;
2174 default: unreachable("bad vector size");
2177 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
2178 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2180 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2185 struct vtn_ssa_value
*
2186 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2188 if (src
->transposed
)
2189 return src
->transposed
;
2191 struct vtn_ssa_value
*dest
=
2192 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2194 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2195 nir_alu_instr
*vec
= create_vec(b
->shader
,
2196 glsl_get_matrix_columns(src
->type
),
2197 glsl_get_bit_size(src
->type
));
2198 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2199 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2200 vec
->src
[0].swizzle
[0] = i
;
2202 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2203 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2204 vec
->src
[j
].swizzle
[0] = i
;
2207 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2208 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2211 dest
->transposed
= src
;
2217 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2219 unsigned swiz
[4] = { index
};
2220 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2224 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2227 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
2230 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2232 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2234 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2235 vec
->src
[i
].swizzle
[0] = i
;
2239 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2241 return &vec
->dest
.dest
.ssa
;
2245 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2248 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2249 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2250 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2251 vtn_vector_extract(b
, src
, i
), dest
);
2257 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2258 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2260 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2261 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2262 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2263 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2268 static nir_ssa_def
*
2269 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2270 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2271 const uint32_t *indices
)
2273 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
2275 for (unsigned i
= 0; i
< num_components
; i
++) {
2276 uint32_t index
= indices
[i
];
2277 if (index
== 0xffffffff) {
2279 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2280 } else if (index
< src0
->num_components
) {
2281 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2282 vec
->src
[i
].swizzle
[0] = index
;
2284 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2285 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2289 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2291 return &vec
->dest
.dest
.ssa
;
2295 * Concatentates a number of vectors/scalars together to produce a vector
2297 static nir_ssa_def
*
2298 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2299 unsigned num_srcs
, nir_ssa_def
**srcs
)
2301 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
2304 unsigned dest_idx
= 0;
2305 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2306 nir_ssa_def
*src
= srcs
[i
];
2307 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2308 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2309 vec
->src
[dest_idx
].swizzle
[0] = j
;
2314 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2316 return &vec
->dest
.dest
.ssa
;
2319 static struct vtn_ssa_value
*
2320 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2322 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2323 dest
->type
= src
->type
;
2325 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2326 dest
->def
= src
->def
;
2328 unsigned elems
= glsl_get_length(src
->type
);
2330 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2331 for (unsigned i
= 0; i
< elems
; i
++)
2332 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2338 static struct vtn_ssa_value
*
2339 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2340 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2341 unsigned num_indices
)
2343 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2345 struct vtn_ssa_value
*cur
= dest
;
2347 for (i
= 0; i
< num_indices
- 1; i
++) {
2348 cur
= cur
->elems
[indices
[i
]];
2351 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2352 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2353 * the component granularity. In that case, the last index will be
2354 * the index to insert the scalar into the vector.
2357 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2359 cur
->elems
[indices
[i
]] = insert
;
2365 static struct vtn_ssa_value
*
2366 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2367 const uint32_t *indices
, unsigned num_indices
)
2369 struct vtn_ssa_value
*cur
= src
;
2370 for (unsigned i
= 0; i
< num_indices
; i
++) {
2371 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2372 assert(i
== num_indices
- 1);
2373 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2374 * the component granularity. The last index will be the index of the
2375 * vector to extract.
2378 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2379 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2380 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2383 cur
= cur
->elems
[indices
[i
]];
2391 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2392 const uint32_t *w
, unsigned count
)
2394 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2395 const struct glsl_type
*type
=
2396 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2397 val
->ssa
= vtn_create_ssa_value(b
, type
);
2400 case SpvOpVectorExtractDynamic
:
2401 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2402 vtn_ssa_value(b
, w
[4])->def
);
2405 case SpvOpVectorInsertDynamic
:
2406 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2407 vtn_ssa_value(b
, w
[4])->def
,
2408 vtn_ssa_value(b
, w
[5])->def
);
2411 case SpvOpVectorShuffle
:
2412 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2413 vtn_ssa_value(b
, w
[3])->def
,
2414 vtn_ssa_value(b
, w
[4])->def
,
2418 case SpvOpCompositeConstruct
: {
2419 unsigned elems
= count
- 3;
2420 if (glsl_type_is_vector_or_scalar(type
)) {
2421 nir_ssa_def
*srcs
[4];
2422 for (unsigned i
= 0; i
< elems
; i
++)
2423 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2425 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2428 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2429 for (unsigned i
= 0; i
< elems
; i
++)
2430 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2434 case SpvOpCompositeExtract
:
2435 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2439 case SpvOpCompositeInsert
:
2440 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2441 vtn_ssa_value(b
, w
[3]),
2445 case SpvOpCopyObject
:
2446 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2450 unreachable("unknown composite operation");
2455 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2456 const uint32_t *w
, unsigned count
)
2458 nir_intrinsic_op intrinsic_op
;
2460 case SpvOpEmitVertex
:
2461 case SpvOpEmitStreamVertex
:
2462 intrinsic_op
= nir_intrinsic_emit_vertex
;
2464 case SpvOpEndPrimitive
:
2465 case SpvOpEndStreamPrimitive
:
2466 intrinsic_op
= nir_intrinsic_end_primitive
;
2468 case SpvOpMemoryBarrier
:
2469 intrinsic_op
= nir_intrinsic_memory_barrier
;
2471 case SpvOpControlBarrier
:
2472 intrinsic_op
= nir_intrinsic_barrier
;
2475 unreachable("unknown barrier instruction");
2478 nir_intrinsic_instr
*intrin
=
2479 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2481 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2482 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2484 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2488 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2491 case SpvExecutionModeInputPoints
:
2492 case SpvExecutionModeOutputPoints
:
2493 return 0; /* GL_POINTS */
2494 case SpvExecutionModeInputLines
:
2495 return 1; /* GL_LINES */
2496 case SpvExecutionModeInputLinesAdjacency
:
2497 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2498 case SpvExecutionModeTriangles
:
2499 return 4; /* GL_TRIANGLES */
2500 case SpvExecutionModeInputTrianglesAdjacency
:
2501 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2502 case SpvExecutionModeQuads
:
2503 return 7; /* GL_QUADS */
2504 case SpvExecutionModeIsolines
:
2505 return 0x8E7A; /* GL_ISOLINES */
2506 case SpvExecutionModeOutputLineStrip
:
2507 return 3; /* GL_LINE_STRIP */
2508 case SpvExecutionModeOutputTriangleStrip
:
2509 return 5; /* GL_TRIANGLE_STRIP */
2511 assert(!"Invalid primitive type");
2517 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2520 case SpvExecutionModeInputPoints
:
2522 case SpvExecutionModeInputLines
:
2524 case SpvExecutionModeInputLinesAdjacency
:
2526 case SpvExecutionModeTriangles
:
2528 case SpvExecutionModeInputTrianglesAdjacency
:
2531 assert(!"Invalid GS input mode");
2536 static gl_shader_stage
2537 stage_for_execution_model(SpvExecutionModel model
)
2540 case SpvExecutionModelVertex
:
2541 return MESA_SHADER_VERTEX
;
2542 case SpvExecutionModelTessellationControl
:
2543 return MESA_SHADER_TESS_CTRL
;
2544 case SpvExecutionModelTessellationEvaluation
:
2545 return MESA_SHADER_TESS_EVAL
;
2546 case SpvExecutionModelGeometry
:
2547 return MESA_SHADER_GEOMETRY
;
2548 case SpvExecutionModelFragment
:
2549 return MESA_SHADER_FRAGMENT
;
2550 case SpvExecutionModelGLCompute
:
2551 return MESA_SHADER_COMPUTE
;
2553 unreachable("Unsupported execution model");
2557 #define spv_check_supported(name, cap) do { \
2558 if (!(b->ext && b->ext->name)) \
2559 vtn_warn("Unsupported SPIR-V capability: %s", \
2560 spirv_capability_to_string(cap)); \
2564 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2565 const uint32_t *w
, unsigned count
)
2569 case SpvOpSourceExtension
:
2570 case SpvOpSourceContinued
:
2571 case SpvOpExtension
:
2572 /* Unhandled, but these are for debug so that's ok. */
2575 case SpvOpCapability
: {
2576 SpvCapability cap
= w
[1];
2578 case SpvCapabilityMatrix
:
2579 case SpvCapabilityShader
:
2580 case SpvCapabilityGeometry
:
2581 case SpvCapabilityGeometryPointSize
:
2582 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2583 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2584 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2585 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2586 case SpvCapabilityImageRect
:
2587 case SpvCapabilitySampledRect
:
2588 case SpvCapabilitySampled1D
:
2589 case SpvCapabilityImage1D
:
2590 case SpvCapabilitySampledCubeArray
:
2591 case SpvCapabilitySampledBuffer
:
2592 case SpvCapabilityImageBuffer
:
2593 case SpvCapabilityImageQuery
:
2594 case SpvCapabilityDerivativeControl
:
2595 case SpvCapabilityInterpolationFunction
:
2596 case SpvCapabilityMultiViewport
:
2597 case SpvCapabilitySampleRateShading
:
2598 case SpvCapabilityClipDistance
:
2599 case SpvCapabilityCullDistance
:
2600 case SpvCapabilityInputAttachment
:
2601 case SpvCapabilityImageGatherExtended
:
2602 case SpvCapabilityStorageImageExtendedFormats
:
2605 case SpvCapabilityGeometryStreams
:
2606 case SpvCapabilityTessellation
:
2607 case SpvCapabilityTessellationPointSize
:
2608 case SpvCapabilityLinkage
:
2609 case SpvCapabilityVector16
:
2610 case SpvCapabilityFloat16Buffer
:
2611 case SpvCapabilityFloat16
:
2612 case SpvCapabilityInt64
:
2613 case SpvCapabilityInt64Atomics
:
2614 case SpvCapabilityAtomicStorage
:
2615 case SpvCapabilityInt16
:
2616 case SpvCapabilityStorageImageMultisample
:
2617 case SpvCapabilityImageCubeArray
:
2618 case SpvCapabilityInt8
:
2619 case SpvCapabilitySparseResidency
:
2620 case SpvCapabilityMinLod
:
2621 case SpvCapabilityTransformFeedback
:
2622 case SpvCapabilityStorageImageReadWithoutFormat
:
2623 case SpvCapabilityStorageImageWriteWithoutFormat
:
2624 vtn_warn("Unsupported SPIR-V capability: %s",
2625 spirv_capability_to_string(cap
));
2628 case SpvCapabilityFloat64
:
2629 spv_check_supported(float64
, cap
);
2632 case SpvCapabilityAddresses
:
2633 case SpvCapabilityKernel
:
2634 case SpvCapabilityImageBasic
:
2635 case SpvCapabilityImageReadWrite
:
2636 case SpvCapabilityImageMipmap
:
2637 case SpvCapabilityPipes
:
2638 case SpvCapabilityGroups
:
2639 case SpvCapabilityDeviceEnqueue
:
2640 case SpvCapabilityLiteralSampler
:
2641 case SpvCapabilityGenericPointer
:
2642 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2643 spirv_capability_to_string(cap
));
2646 case SpvCapabilityImageMSArray
:
2647 spv_check_supported(image_ms_array
, cap
);
2653 case SpvOpExtInstImport
:
2654 vtn_handle_extension(b
, opcode
, w
, count
);
2657 case SpvOpMemoryModel
:
2658 assert(w
[1] == SpvAddressingModelLogical
);
2659 assert(w
[2] == SpvMemoryModelGLSL450
);
2662 case SpvOpEntryPoint
: {
2663 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2664 /* Let this be a name label regardless */
2665 unsigned name_words
;
2666 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2668 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2669 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2672 assert(b
->entry_point
== NULL
);
2673 b
->entry_point
= entry_point
;
2678 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2679 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2683 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2686 case SpvOpMemberName
:
2690 case SpvOpExecutionMode
:
2691 case SpvOpDecorationGroup
:
2693 case SpvOpMemberDecorate
:
2694 case SpvOpGroupDecorate
:
2695 case SpvOpGroupMemberDecorate
:
2696 vtn_handle_decoration(b
, opcode
, w
, count
);
2700 return false; /* End of preamble */
2707 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2708 const struct vtn_decoration
*mode
, void *data
)
2710 assert(b
->entry_point
== entry_point
);
2712 switch(mode
->exec_mode
) {
2713 case SpvExecutionModeOriginUpperLeft
:
2714 case SpvExecutionModeOriginLowerLeft
:
2715 b
->origin_upper_left
=
2716 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2719 case SpvExecutionModeEarlyFragmentTests
:
2720 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2721 b
->shader
->info
->fs
.early_fragment_tests
= true;
2724 case SpvExecutionModeInvocations
:
2725 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2726 b
->shader
->info
->gs
.invocations
= MAX2(1, mode
->literals
[0]);
2729 case SpvExecutionModeDepthReplacing
:
2730 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2731 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2733 case SpvExecutionModeDepthGreater
:
2734 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2735 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2737 case SpvExecutionModeDepthLess
:
2738 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2739 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2741 case SpvExecutionModeDepthUnchanged
:
2742 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2743 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2746 case SpvExecutionModeLocalSize
:
2747 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2748 b
->shader
->info
->cs
.local_size
[0] = mode
->literals
[0];
2749 b
->shader
->info
->cs
.local_size
[1] = mode
->literals
[1];
2750 b
->shader
->info
->cs
.local_size
[2] = mode
->literals
[2];
2752 case SpvExecutionModeLocalSizeHint
:
2753 break; /* Nothing to do with this */
2755 case SpvExecutionModeOutputVertices
:
2756 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2757 b
->shader
->info
->gs
.vertices_out
= mode
->literals
[0];
2760 case SpvExecutionModeInputPoints
:
2761 case SpvExecutionModeInputLines
:
2762 case SpvExecutionModeInputLinesAdjacency
:
2763 case SpvExecutionModeTriangles
:
2764 case SpvExecutionModeInputTrianglesAdjacency
:
2765 case SpvExecutionModeQuads
:
2766 case SpvExecutionModeIsolines
:
2767 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2768 b
->shader
->info
->gs
.vertices_in
=
2769 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2771 assert(!"Tesselation shaders not yet supported");
2775 case SpvExecutionModeOutputPoints
:
2776 case SpvExecutionModeOutputLineStrip
:
2777 case SpvExecutionModeOutputTriangleStrip
:
2778 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2779 b
->shader
->info
->gs
.output_primitive
=
2780 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2783 case SpvExecutionModeSpacingEqual
:
2784 case SpvExecutionModeSpacingFractionalEven
:
2785 case SpvExecutionModeSpacingFractionalOdd
:
2786 case SpvExecutionModeVertexOrderCw
:
2787 case SpvExecutionModeVertexOrderCcw
:
2788 case SpvExecutionModePointMode
:
2789 assert(!"TODO: Add tessellation metadata");
2792 case SpvExecutionModePixelCenterInteger
:
2793 b
->pixel_center_integer
= true;
2796 case SpvExecutionModeXfb
:
2797 assert(!"Unhandled execution mode");
2800 case SpvExecutionModeVecTypeHint
:
2801 case SpvExecutionModeContractionOff
:
2807 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2808 const uint32_t *w
, unsigned count
)
2812 case SpvOpSourceContinued
:
2813 case SpvOpSourceExtension
:
2814 case SpvOpExtension
:
2815 case SpvOpCapability
:
2816 case SpvOpExtInstImport
:
2817 case SpvOpMemoryModel
:
2818 case SpvOpEntryPoint
:
2819 case SpvOpExecutionMode
:
2822 case SpvOpMemberName
:
2823 case SpvOpDecorationGroup
:
2825 case SpvOpMemberDecorate
:
2826 case SpvOpGroupDecorate
:
2827 case SpvOpGroupMemberDecorate
:
2828 assert(!"Invalid opcode types and variables section");
2834 case SpvOpTypeFloat
:
2835 case SpvOpTypeVector
:
2836 case SpvOpTypeMatrix
:
2837 case SpvOpTypeImage
:
2838 case SpvOpTypeSampler
:
2839 case SpvOpTypeSampledImage
:
2840 case SpvOpTypeArray
:
2841 case SpvOpTypeRuntimeArray
:
2842 case SpvOpTypeStruct
:
2843 case SpvOpTypeOpaque
:
2844 case SpvOpTypePointer
:
2845 case SpvOpTypeFunction
:
2846 case SpvOpTypeEvent
:
2847 case SpvOpTypeDeviceEvent
:
2848 case SpvOpTypeReserveId
:
2849 case SpvOpTypeQueue
:
2851 vtn_handle_type(b
, opcode
, w
, count
);
2854 case SpvOpConstantTrue
:
2855 case SpvOpConstantFalse
:
2857 case SpvOpConstantComposite
:
2858 case SpvOpConstantSampler
:
2859 case SpvOpConstantNull
:
2860 case SpvOpSpecConstantTrue
:
2861 case SpvOpSpecConstantFalse
:
2862 case SpvOpSpecConstant
:
2863 case SpvOpSpecConstantComposite
:
2864 case SpvOpSpecConstantOp
:
2865 vtn_handle_constant(b
, opcode
, w
, count
);
2869 vtn_handle_variables(b
, opcode
, w
, count
);
2873 return false; /* End of preamble */
2880 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2881 const uint32_t *w
, unsigned count
)
2887 case SpvOpLoopMerge
:
2888 case SpvOpSelectionMerge
:
2889 /* This is handled by cfg pre-pass and walk_blocks */
2893 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2894 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2899 vtn_handle_extension(b
, opcode
, w
, count
);
2905 case SpvOpCopyMemory
:
2906 case SpvOpCopyMemorySized
:
2907 case SpvOpAccessChain
:
2908 case SpvOpInBoundsAccessChain
:
2909 case SpvOpArrayLength
:
2910 vtn_handle_variables(b
, opcode
, w
, count
);
2913 case SpvOpFunctionCall
:
2914 vtn_handle_function_call(b
, opcode
, w
, count
);
2917 case SpvOpSampledImage
:
2919 case SpvOpImageSampleImplicitLod
:
2920 case SpvOpImageSampleExplicitLod
:
2921 case SpvOpImageSampleDrefImplicitLod
:
2922 case SpvOpImageSampleDrefExplicitLod
:
2923 case SpvOpImageSampleProjImplicitLod
:
2924 case SpvOpImageSampleProjExplicitLod
:
2925 case SpvOpImageSampleProjDrefImplicitLod
:
2926 case SpvOpImageSampleProjDrefExplicitLod
:
2927 case SpvOpImageFetch
:
2928 case SpvOpImageGather
:
2929 case SpvOpImageDrefGather
:
2930 case SpvOpImageQuerySizeLod
:
2931 case SpvOpImageQueryLod
:
2932 case SpvOpImageQueryLevels
:
2933 case SpvOpImageQuerySamples
:
2934 vtn_handle_texture(b
, opcode
, w
, count
);
2937 case SpvOpImageRead
:
2938 case SpvOpImageWrite
:
2939 case SpvOpImageTexelPointer
:
2940 vtn_handle_image(b
, opcode
, w
, count
);
2943 case SpvOpImageQuerySize
: {
2944 struct vtn_access_chain
*image
=
2945 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2946 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2947 vtn_handle_image(b
, opcode
, w
, count
);
2949 vtn_handle_texture(b
, opcode
, w
, count
);
2954 case SpvOpAtomicLoad
:
2955 case SpvOpAtomicExchange
:
2956 case SpvOpAtomicCompareExchange
:
2957 case SpvOpAtomicCompareExchangeWeak
:
2958 case SpvOpAtomicIIncrement
:
2959 case SpvOpAtomicIDecrement
:
2960 case SpvOpAtomicIAdd
:
2961 case SpvOpAtomicISub
:
2962 case SpvOpAtomicSMin
:
2963 case SpvOpAtomicUMin
:
2964 case SpvOpAtomicSMax
:
2965 case SpvOpAtomicUMax
:
2966 case SpvOpAtomicAnd
:
2968 case SpvOpAtomicXor
: {
2969 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2970 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2971 vtn_handle_image(b
, opcode
, w
, count
);
2973 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2974 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2979 case SpvOpAtomicStore
: {
2980 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
2981 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2982 vtn_handle_image(b
, opcode
, w
, count
);
2984 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2985 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2995 case SpvOpConvertFToU
:
2996 case SpvOpConvertFToS
:
2997 case SpvOpConvertSToF
:
2998 case SpvOpConvertUToF
:
3002 case SpvOpQuantizeToF16
:
3003 case SpvOpConvertPtrToU
:
3004 case SpvOpConvertUToPtr
:
3005 case SpvOpPtrCastToGeneric
:
3006 case SpvOpGenericCastToPtr
:
3012 case SpvOpSignBitSet
:
3013 case SpvOpLessOrGreater
:
3015 case SpvOpUnordered
:
3030 case SpvOpVectorTimesScalar
:
3032 case SpvOpIAddCarry
:
3033 case SpvOpISubBorrow
:
3034 case SpvOpUMulExtended
:
3035 case SpvOpSMulExtended
:
3036 case SpvOpShiftRightLogical
:
3037 case SpvOpShiftRightArithmetic
:
3038 case SpvOpShiftLeftLogical
:
3039 case SpvOpLogicalEqual
:
3040 case SpvOpLogicalNotEqual
:
3041 case SpvOpLogicalOr
:
3042 case SpvOpLogicalAnd
:
3043 case SpvOpLogicalNot
:
3044 case SpvOpBitwiseOr
:
3045 case SpvOpBitwiseXor
:
3046 case SpvOpBitwiseAnd
:
3049 case SpvOpFOrdEqual
:
3050 case SpvOpFUnordEqual
:
3051 case SpvOpINotEqual
:
3052 case SpvOpFOrdNotEqual
:
3053 case SpvOpFUnordNotEqual
:
3054 case SpvOpULessThan
:
3055 case SpvOpSLessThan
:
3056 case SpvOpFOrdLessThan
:
3057 case SpvOpFUnordLessThan
:
3058 case SpvOpUGreaterThan
:
3059 case SpvOpSGreaterThan
:
3060 case SpvOpFOrdGreaterThan
:
3061 case SpvOpFUnordGreaterThan
:
3062 case SpvOpULessThanEqual
:
3063 case SpvOpSLessThanEqual
:
3064 case SpvOpFOrdLessThanEqual
:
3065 case SpvOpFUnordLessThanEqual
:
3066 case SpvOpUGreaterThanEqual
:
3067 case SpvOpSGreaterThanEqual
:
3068 case SpvOpFOrdGreaterThanEqual
:
3069 case SpvOpFUnordGreaterThanEqual
:
3075 case SpvOpFwidthFine
:
3076 case SpvOpDPdxCoarse
:
3077 case SpvOpDPdyCoarse
:
3078 case SpvOpFwidthCoarse
:
3079 case SpvOpBitFieldInsert
:
3080 case SpvOpBitFieldSExtract
:
3081 case SpvOpBitFieldUExtract
:
3082 case SpvOpBitReverse
:
3084 case SpvOpTranspose
:
3085 case SpvOpOuterProduct
:
3086 case SpvOpMatrixTimesScalar
:
3087 case SpvOpVectorTimesMatrix
:
3088 case SpvOpMatrixTimesVector
:
3089 case SpvOpMatrixTimesMatrix
:
3090 vtn_handle_alu(b
, opcode
, w
, count
);
3093 case SpvOpVectorExtractDynamic
:
3094 case SpvOpVectorInsertDynamic
:
3095 case SpvOpVectorShuffle
:
3096 case SpvOpCompositeConstruct
:
3097 case SpvOpCompositeExtract
:
3098 case SpvOpCompositeInsert
:
3099 case SpvOpCopyObject
:
3100 vtn_handle_composite(b
, opcode
, w
, count
);
3103 case SpvOpEmitVertex
:
3104 case SpvOpEndPrimitive
:
3105 case SpvOpEmitStreamVertex
:
3106 case SpvOpEndStreamPrimitive
:
3107 case SpvOpControlBarrier
:
3108 case SpvOpMemoryBarrier
:
3109 vtn_handle_barrier(b
, opcode
, w
, count
);
3113 unreachable("Unhandled opcode");
3120 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3121 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3122 gl_shader_stage stage
, const char *entry_point_name
,
3123 const struct nir_spirv_supported_extensions
*ext
,
3124 const nir_shader_compiler_options
*options
)
3126 const uint32_t *word_end
= words
+ word_count
;
3128 /* Handle the SPIR-V header (first 4 dwords) */
3129 assert(word_count
> 5);
3131 assert(words
[0] == SpvMagicNumber
);
3132 assert(words
[1] >= 0x10000);
3133 /* words[2] == generator magic */
3134 unsigned value_id_bound
= words
[3];
3135 assert(words
[4] == 0);
3139 /* Initialize the stn_builder object */
3140 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3141 b
->value_id_bound
= value_id_bound
;
3142 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3143 exec_list_make_empty(&b
->functions
);
3144 b
->entry_point_stage
= stage
;
3145 b
->entry_point_name
= entry_point_name
;
3148 /* Handle all the preamble instructions */
3149 words
= vtn_foreach_instruction(b
, words
, word_end
,
3150 vtn_handle_preamble_instruction
);
3152 if (b
->entry_point
== NULL
) {
3153 assert(!"Entry point not found");
3158 b
->shader
= nir_shader_create(NULL
, stage
, options
, NULL
);
3160 /* Set shader info defaults */
3161 b
->shader
->info
->gs
.invocations
= 1;
3163 /* Parse execution modes */
3164 vtn_foreach_execution_mode(b
, b
->entry_point
,
3165 vtn_handle_execution_mode
, NULL
);
3167 b
->specializations
= spec
;
3168 b
->num_specializations
= num_spec
;
3170 /* Handle all variable, type, and constant instructions */
3171 words
= vtn_foreach_instruction(b
, words
, word_end
,
3172 vtn_handle_variable_or_type_instruction
);
3174 vtn_build_cfg(b
, words
, word_end
);
3176 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3177 b
->impl
= func
->impl
;
3178 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3179 _mesa_key_pointer_equal
);
3181 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3184 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3185 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3186 assert(entry_point
);