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"
33 static struct vtn_ssa_value
*
34 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
36 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
39 if (glsl_type_is_vector_or_scalar(type
)) {
40 unsigned num_components
= glsl_get_vector_elements(val
->type
);
41 nir_ssa_undef_instr
*undef
=
42 nir_ssa_undef_instr_create(b
->shader
, num_components
);
44 nir_instr_insert_before_cf_list(&b
->impl
->body
, &undef
->instr
);
45 val
->def
= &undef
->def
;
47 unsigned elems
= glsl_get_length(val
->type
);
48 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
49 if (glsl_type_is_matrix(type
)) {
50 const struct glsl_type
*elem_type
=
51 glsl_vector_type(glsl_get_base_type(type
),
52 glsl_get_vector_elements(type
));
54 for (unsigned i
= 0; i
< elems
; i
++)
55 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
56 } else if (glsl_type_is_array(type
)) {
57 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
58 for (unsigned i
= 0; i
< elems
; i
++)
59 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
61 for (unsigned i
= 0; i
< elems
; i
++) {
62 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
63 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
71 static struct vtn_ssa_value
*
72 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
73 const struct glsl_type
*type
)
75 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
80 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
83 switch (glsl_get_base_type(type
)) {
88 case GLSL_TYPE_DOUBLE
:
89 if (glsl_type_is_vector_or_scalar(type
)) {
90 unsigned num_components
= glsl_get_vector_elements(val
->type
);
91 nir_load_const_instr
*load
=
92 nir_load_const_instr_create(b
->shader
, num_components
);
94 for (unsigned i
= 0; i
< num_components
; i
++)
95 load
->value
.u
[i
] = constant
->value
.u
[i
];
97 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
98 val
->def
= &load
->def
;
100 assert(glsl_type_is_matrix(type
));
101 unsigned rows
= glsl_get_vector_elements(val
->type
);
102 unsigned columns
= glsl_get_matrix_columns(val
->type
);
103 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
105 for (unsigned i
= 0; i
< columns
; i
++) {
106 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
107 col_val
->type
= glsl_get_column_type(val
->type
);
108 nir_load_const_instr
*load
=
109 nir_load_const_instr_create(b
->shader
, rows
);
111 for (unsigned j
= 0; j
< rows
; j
++)
112 load
->value
.u
[j
] = constant
->value
.u
[rows
* i
+ j
];
114 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
115 col_val
->def
= &load
->def
;
117 val
->elems
[i
] = col_val
;
122 case GLSL_TYPE_ARRAY
: {
123 unsigned elems
= glsl_get_length(val
->type
);
124 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
125 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
126 for (unsigned i
= 0; i
< elems
; i
++)
127 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
132 case GLSL_TYPE_STRUCT
: {
133 unsigned elems
= glsl_get_length(val
->type
);
134 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
135 for (unsigned i
= 0; i
< elems
; i
++) {
136 const struct glsl_type
*elem_type
=
137 glsl_get_struct_field(val
->type
, i
);
138 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
145 unreachable("bad constant type");
151 struct vtn_ssa_value
*
152 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
154 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
155 switch (val
->value_type
) {
156 case vtn_value_type_undef
:
157 return vtn_undef_ssa_value(b
, val
->type
->type
);
159 case vtn_value_type_constant
:
160 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
162 case vtn_value_type_ssa
:
165 case vtn_value_type_access_chain
:
166 /* This is needed for function parameters */
167 return vtn_variable_load(b
, val
->access_chain
);
170 unreachable("Invalid type for an SSA value");
175 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
176 unsigned word_count
, unsigned *words_used
)
178 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
180 /* Ammount of space taken by the string (including the null) */
181 unsigned len
= strlen(dup
) + 1;
182 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
188 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
189 const uint32_t *end
, vtn_instruction_handler handler
)
195 const uint32_t *w
= start
;
197 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
198 unsigned count
= w
[0] >> SpvWordCountShift
;
199 assert(count
>= 1 && w
+ count
<= end
);
203 break; /* Do nothing */
206 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
218 if (!handler(b
, opcode
, w
, count
))
230 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
231 const uint32_t *w
, unsigned count
)
234 case SpvOpExtInstImport
: {
235 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
236 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
237 val
->ext_handler
= vtn_handle_glsl450_instruction
;
239 assert(!"Unsupported extension");
245 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
246 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
253 unreachable("Unhandled opcode");
258 _foreach_decoration_helper(struct vtn_builder
*b
,
259 struct vtn_value
*base_value
,
261 struct vtn_value
*value
,
262 vtn_decoration_foreach_cb cb
, void *data
)
264 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
266 if (dec
->scope
== VTN_DEC_DECORATION
) {
267 member
= parent_member
;
268 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
269 assert(parent_member
== -1);
270 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
272 /* Not a decoration */
277 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
278 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
281 cb(b
, base_value
, member
, dec
, data
);
286 /** Iterates (recursively if needed) over all of the decorations on a value
288 * This function iterates over all of the decorations applied to a given
289 * value. If it encounters a decoration group, it recurses into the group
290 * and iterates over all of those decorations as well.
293 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
294 vtn_decoration_foreach_cb cb
, void *data
)
296 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
300 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
301 vtn_execution_mode_foreach_cb cb
, void *data
)
303 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
304 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
307 assert(dec
->group
== NULL
);
308 cb(b
, value
, dec
, data
);
313 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
314 const uint32_t *w
, unsigned count
)
316 const uint32_t *w_end
= w
+ count
;
317 const uint32_t target
= w
[1];
321 case SpvOpDecorationGroup
:
322 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
326 case SpvOpMemberDecorate
:
327 case SpvOpExecutionMode
: {
328 struct vtn_value
*val
= &b
->values
[target
];
330 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
333 dec
->scope
= VTN_DEC_DECORATION
;
335 case SpvOpMemberDecorate
:
336 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
338 case SpvOpExecutionMode
:
339 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
342 unreachable("Invalid decoration opcode");
344 dec
->decoration
= *(w
++);
347 /* Link into the list */
348 dec
->next
= val
->decoration
;
349 val
->decoration
= dec
;
353 case SpvOpGroupMemberDecorate
:
354 case SpvOpGroupDecorate
: {
355 struct vtn_value
*group
=
356 vtn_value(b
, target
, vtn_value_type_decoration_group
);
358 for (; w
< w_end
; w
++) {
359 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
360 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
363 if (opcode
== SpvOpGroupDecorate
) {
364 dec
->scope
= VTN_DEC_DECORATION
;
366 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
369 /* Link into the list */
370 dec
->next
= val
->decoration
;
371 val
->decoration
= dec
;
377 unreachable("Unhandled opcode");
381 struct member_decoration_ctx
{
382 struct glsl_struct_field
*fields
;
383 struct vtn_type
*type
;
386 /* does a shallow copy of a vtn_type */
388 static struct vtn_type
*
389 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
391 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
392 dest
->type
= src
->type
;
393 dest
->is_builtin
= src
->is_builtin
;
395 dest
->builtin
= src
->builtin
;
397 if (!glsl_type_is_scalar(src
->type
)) {
398 switch (glsl_get_base_type(src
->type
)) {
402 case GLSL_TYPE_FLOAT
:
403 case GLSL_TYPE_DOUBLE
:
404 case GLSL_TYPE_ARRAY
:
405 dest
->row_major
= src
->row_major
;
406 dest
->stride
= src
->stride
;
407 dest
->array_element
= src
->array_element
;
410 case GLSL_TYPE_STRUCT
: {
411 unsigned elems
= glsl_get_length(src
->type
);
413 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
414 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
416 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
417 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
422 unreachable("unhandled type");
429 static struct vtn_type
*
430 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
432 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
433 type
= type
->members
[member
];
435 /* We may have an array of matrices.... Oh, joy! */
436 while (glsl_type_is_array(type
->type
)) {
437 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
438 type
= type
->array_element
;
441 assert(glsl_type_is_matrix(type
->type
));
447 struct_member_decoration_cb(struct vtn_builder
*b
,
448 struct vtn_value
*val
, int member
,
449 const struct vtn_decoration
*dec
, void *void_ctx
)
451 struct member_decoration_ctx
*ctx
= void_ctx
;
456 switch (dec
->decoration
) {
457 case SpvDecorationRelaxedPrecision
:
458 break; /* FIXME: Do nothing with this for now. */
459 case SpvDecorationNoPerspective
:
460 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_NOPERSPECTIVE
;
462 case SpvDecorationFlat
:
463 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_FLAT
;
465 case SpvDecorationCentroid
:
466 ctx
->fields
[member
].centroid
= true;
468 case SpvDecorationSample
:
469 ctx
->fields
[member
].sample
= true;
471 case SpvDecorationLocation
:
472 ctx
->fields
[member
].location
= dec
->literals
[0];
474 case SpvDecorationBuiltIn
:
475 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
476 ctx
->type
->members
[member
]->is_builtin
= true;
477 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
478 ctx
->type
->builtin_block
= true;
480 case SpvDecorationOffset
:
481 ctx
->type
->offsets
[member
] = dec
->literals
[0];
483 case SpvDecorationMatrixStride
:
484 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
486 case SpvDecorationColMajor
:
487 break; /* Nothing to do here. Column-major is the default. */
488 case SpvDecorationRowMajor
:
489 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
492 unreachable("Unhandled member decoration");
497 type_decoration_cb(struct vtn_builder
*b
,
498 struct vtn_value
*val
, int member
,
499 const struct vtn_decoration
*dec
, void *ctx
)
501 struct vtn_type
*type
= val
->type
;
506 switch (dec
->decoration
) {
507 case SpvDecorationArrayStride
:
508 type
->stride
= dec
->literals
[0];
510 case SpvDecorationBlock
:
513 case SpvDecorationBufferBlock
:
514 type
->buffer_block
= true;
516 case SpvDecorationGLSLShared
:
517 case SpvDecorationGLSLPacked
:
518 /* Ignore these, since we get explicit offsets anyways */
521 case SpvDecorationStream
:
522 assert(dec
->literals
[0] == 0);
526 unreachable("Unhandled type decoration");
531 translate_image_format(SpvImageFormat format
)
534 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
535 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
536 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
537 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
538 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
539 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
540 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
541 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
542 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
543 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
544 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
545 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
546 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
547 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
548 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
549 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
550 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
551 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
552 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
553 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
554 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
555 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
556 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
557 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
558 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
559 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
560 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
561 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
562 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
563 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
564 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
565 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
566 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
567 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
568 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
569 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
570 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
571 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
572 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
573 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
575 assert(!"Invalid image format");
581 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
582 const uint32_t *w
, unsigned count
)
584 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
586 val
->type
= rzalloc(b
, struct vtn_type
);
587 val
->type
->is_builtin
= false;
588 val
->type
->val
= val
;
592 val
->type
->type
= glsl_void_type();
595 val
->type
->type
= glsl_bool_type();
598 val
->type
->type
= glsl_int_type();
601 val
->type
->type
= glsl_float_type();
604 case SpvOpTypeVector
: {
605 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
606 unsigned elems
= w
[3];
608 assert(glsl_type_is_scalar(base
->type
));
609 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
611 /* Vectors implicitly have sizeof(base_type) stride. For now, this
612 * is always 4 bytes. This will have to change if we want to start
613 * supporting doubles or half-floats.
615 val
->type
->stride
= 4;
616 val
->type
->array_element
= base
;
620 case SpvOpTypeMatrix
: {
621 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
622 unsigned columns
= w
[3];
624 assert(glsl_type_is_vector(base
->type
));
625 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
626 glsl_get_vector_elements(base
->type
),
628 assert(!glsl_type_is_error(val
->type
->type
));
629 val
->type
->array_element
= base
;
630 val
->type
->row_major
= false;
631 val
->type
->stride
= 0;
635 case SpvOpTypeRuntimeArray
:
636 case SpvOpTypeArray
: {
637 struct vtn_type
*array_element
=
638 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
641 if (opcode
== SpvOpTypeRuntimeArray
) {
642 /* A length of 0 is used to denote unsized arrays */
646 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->value
.u
[0];
649 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
650 val
->type
->array_element
= array_element
;
651 val
->type
->stride
= 0;
655 case SpvOpTypeStruct
: {
656 unsigned num_fields
= count
- 2;
657 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
658 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
660 NIR_VLA(struct glsl_struct_field
, fields
, count
);
661 for (unsigned i
= 0; i
< num_fields
; i
++) {
662 val
->type
->members
[i
] =
663 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
664 fields
[i
] = (struct glsl_struct_field
) {
665 .type
= val
->type
->members
[i
]->type
,
666 .name
= ralloc_asprintf(b
, "field%d", i
),
671 struct member_decoration_ctx ctx
= {
676 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
678 const char *name
= val
->name
? val
->name
: "struct";
680 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
684 case SpvOpTypeFunction
: {
685 const struct glsl_type
*return_type
=
686 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
687 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
688 for (unsigned i
= 0; i
< count
- 3; i
++) {
689 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
693 params
[i
].out
= true;
695 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
699 case SpvOpTypePointer
:
700 /* FIXME: For now, we'll just do the really lame thing and return
701 * the same type. The validator should ensure that the proper number
702 * of dereferences happen
704 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
707 case SpvOpTypeImage
: {
708 const struct glsl_type
*sampled_type
=
709 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
711 assert(glsl_type_is_vector_or_scalar(sampled_type
));
713 enum glsl_sampler_dim dim
;
714 switch ((SpvDim
)w
[3]) {
715 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
716 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
717 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
718 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
719 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
720 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
722 unreachable("Invalid SPIR-V Sampler dimension");
725 bool is_shadow
= w
[4];
726 bool is_array
= w
[5];
727 bool multisampled
= w
[6];
728 unsigned sampled
= w
[7];
729 SpvImageFormat format
= w
[8];
732 val
->type
->access_qualifier
= w
[9];
734 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
736 assert(!multisampled
&& "FIXME: Handl multi-sampled textures");
738 val
->type
->image_format
= translate_image_format(format
);
741 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
742 glsl_get_base_type(sampled_type
));
743 } else if (sampled
== 2) {
746 val
->type
->type
= glsl_image_type(dim
, is_array
,
747 glsl_get_base_type(sampled_type
));
749 assert(!"We need to know if the image will be sampled");
754 case SpvOpTypeSampledImage
:
755 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
758 case SpvOpTypeSampler
:
759 /* The actual sampler type here doesn't really matter. It gets
760 * thrown away the moment you combine it with an image. What really
761 * matters is that it's a sampler type as opposed to an integer type
762 * so the backend knows what to do.
764 * TODO: Eventually we should consider adding a "bare sampler" type
767 val
->type
->type
= glsl_sampler_type(GLSL_SAMPLER_DIM_2D
, false, false,
771 case SpvOpTypeOpaque
:
773 case SpvOpTypeDeviceEvent
:
774 case SpvOpTypeReserveId
:
778 unreachable("Unhandled opcode");
781 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
784 static nir_constant
*
785 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
787 nir_constant
*c
= rzalloc(b
, nir_constant
);
789 switch (glsl_get_base_type(type
)) {
793 case GLSL_TYPE_FLOAT
:
794 case GLSL_TYPE_DOUBLE
:
795 /* Nothing to do here. It's already initialized to zero */
798 case GLSL_TYPE_ARRAY
:
799 assert(glsl_get_length(type
) > 0);
800 c
->num_elements
= glsl_get_length(type
);
801 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
803 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
804 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
805 c
->elements
[i
] = c
->elements
[0];
808 case GLSL_TYPE_STRUCT
:
809 c
->num_elements
= glsl_get_length(type
);
810 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
812 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
813 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
818 unreachable("Invalid type for null constant");
825 spec_constant_deocoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
826 int member
, const struct vtn_decoration
*dec
,
829 assert(member
== -1);
830 if (dec
->decoration
!= SpvDecorationSpecId
)
833 uint32_t *const_value
= data
;
835 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
836 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
837 *const_value
= b
->specializations
[i
].data
;
844 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
845 uint32_t const_value
)
847 vtn_foreach_decoration(b
, val
, spec_constant_deocoration_cb
, &const_value
);
852 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
853 const uint32_t *w
, unsigned count
)
855 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
856 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
857 val
->constant
= rzalloc(b
, nir_constant
);
859 case SpvOpConstantTrue
:
860 assert(val
->const_type
== glsl_bool_type());
861 val
->constant
->value
.u
[0] = NIR_TRUE
;
863 case SpvOpConstantFalse
:
864 assert(val
->const_type
== glsl_bool_type());
865 val
->constant
->value
.u
[0] = NIR_FALSE
;
868 case SpvOpSpecConstantTrue
:
869 case SpvOpSpecConstantFalse
: {
870 assert(val
->const_type
== glsl_bool_type());
872 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
873 val
->constant
->value
.u
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
878 assert(glsl_type_is_scalar(val
->const_type
));
879 val
->constant
->value
.u
[0] = w
[3];
881 case SpvOpSpecConstant
:
882 assert(glsl_type_is_scalar(val
->const_type
));
883 val
->constant
->value
.u
[0] = get_specialization(b
, val
, w
[3]);
885 case SpvOpSpecConstantComposite
:
886 case SpvOpConstantComposite
: {
887 unsigned elem_count
= count
- 3;
888 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
889 for (unsigned i
= 0; i
< elem_count
; i
++)
890 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
892 switch (glsl_get_base_type(val
->const_type
)) {
895 case GLSL_TYPE_FLOAT
:
897 if (glsl_type_is_matrix(val
->const_type
)) {
898 unsigned rows
= glsl_get_vector_elements(val
->const_type
);
899 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
900 for (unsigned i
= 0; i
< elem_count
; i
++)
901 for (unsigned j
= 0; j
< rows
; j
++)
902 val
->constant
->value
.u
[rows
* i
+ j
] = elems
[i
]->value
.u
[j
];
904 assert(glsl_type_is_vector(val
->const_type
));
905 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
906 for (unsigned i
= 0; i
< elem_count
; i
++)
907 val
->constant
->value
.u
[i
] = elems
[i
]->value
.u
[0];
912 case GLSL_TYPE_STRUCT
:
913 case GLSL_TYPE_ARRAY
:
914 ralloc_steal(val
->constant
, elems
);
915 val
->constant
->num_elements
= elem_count
;
916 val
->constant
->elements
= elems
;
920 unreachable("Unsupported type for constants");
925 case SpvOpSpecConstantOp
: {
926 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
928 case SpvOpVectorShuffle
: {
929 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
930 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
931 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
932 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
935 for (unsigned i
= 0; i
< len0
; i
++)
936 u
[i
] = v0
->constant
->value
.u
[i
];
937 for (unsigned i
= 0; i
< len1
; i
++)
938 u
[len0
+ i
] = v1
->constant
->value
.u
[i
];
940 for (unsigned i
= 0; i
< count
- 6; i
++) {
941 uint32_t comp
= w
[i
+ 6];
942 if (comp
== (uint32_t)-1) {
943 val
->constant
->value
.u
[i
] = 0xdeadbeef;
945 val
->constant
->value
.u
[i
] = u
[comp
];
951 case SpvOpCompositeExtract
:
952 case SpvOpCompositeInsert
: {
953 struct vtn_value
*comp
;
954 unsigned deref_start
;
955 struct nir_constant
**c
;
956 if (opcode
== SpvOpCompositeExtract
) {
957 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
961 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
963 val
->constant
= nir_constant_clone(comp
->constant
,
969 const struct glsl_type
*type
= comp
->const_type
;
970 for (unsigned i
= deref_start
; i
< count
; i
++) {
971 switch (glsl_get_base_type(type
)) {
974 case GLSL_TYPE_FLOAT
:
976 /* If we hit this granularity, we're picking off an element */
980 if (glsl_type_is_matrix(type
)) {
981 elem
+= w
[i
] * glsl_get_vector_elements(type
);
982 type
= glsl_get_column_type(type
);
984 assert(glsl_type_is_vector(type
));
986 type
= glsl_scalar_type(glsl_get_base_type(type
));
990 case GLSL_TYPE_ARRAY
:
991 c
= &(*c
)->elements
[w
[i
]];
992 type
= glsl_get_array_element(type
);
995 case GLSL_TYPE_STRUCT
:
996 c
= &(*c
)->elements
[w
[i
]];
997 type
= glsl_get_struct_field(type
, w
[i
]);
1001 unreachable("Invalid constant type");
1005 if (opcode
== SpvOpCompositeExtract
) {
1009 unsigned num_components
= glsl_get_vector_elements(type
);
1010 for (unsigned i
= 0; i
< num_components
; i
++)
1011 val
->constant
->value
.u
[i
] = (*c
)->value
.u
[elem
+ i
];
1014 struct vtn_value
*insert
=
1015 vtn_value(b
, w
[4], vtn_value_type_constant
);
1016 assert(insert
->const_type
== type
);
1018 *c
= insert
->constant
;
1020 unsigned num_components
= glsl_get_vector_elements(type
);
1021 for (unsigned i
= 0; i
< num_components
; i
++)
1022 (*c
)->value
.u
[elem
+ i
] = insert
->constant
->value
.u
[i
];
1030 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
);
1032 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1034 nir_const_value src
[3];
1036 for (unsigned i
= 0; i
< count
- 4; i
++) {
1038 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1040 unsigned j
= swap
? 1 - i
: i
;
1041 for (unsigned k
= 0; k
< num_components
; k
++)
1042 src
[j
].u
[k
] = c
->value
.u
[k
];
1045 nir_const_value res
= nir_eval_const_opcode(op
, num_components
, src
);
1047 for (unsigned k
= 0; k
< num_components
; k
++)
1048 val
->constant
->value
.u
[k
] = res
.u
[k
];
1055 case SpvOpConstantNull
:
1056 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1059 case SpvOpConstantSampler
:
1060 assert(!"OpConstantSampler requires Kernel Capability");
1064 unreachable("Unhandled opcode");
1069 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1070 const uint32_t *w
, unsigned count
)
1072 struct nir_function
*callee
=
1073 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1075 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1076 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1077 unsigned arg_id
= w
[4 + i
];
1078 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1079 if (arg
->value_type
== vtn_value_type_access_chain
) {
1080 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1081 call
->params
[i
] = nir_deref_as_var(nir_copy_deref(call
, &d
->deref
));
1083 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1085 /* Make a temporary to store the argument in */
1087 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1088 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1090 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1094 nir_variable
*out_tmp
= NULL
;
1095 if (!glsl_type_is_void(callee
->return_type
)) {
1096 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1098 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1101 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1103 if (glsl_type_is_void(callee
->return_type
)) {
1104 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1106 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1107 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1111 struct vtn_ssa_value
*
1112 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1114 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1117 if (!glsl_type_is_vector_or_scalar(type
)) {
1118 unsigned elems
= glsl_get_length(type
);
1119 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1120 for (unsigned i
= 0; i
< elems
; i
++) {
1121 const struct glsl_type
*child_type
;
1123 switch (glsl_get_base_type(type
)) {
1125 case GLSL_TYPE_UINT
:
1126 case GLSL_TYPE_BOOL
:
1127 case GLSL_TYPE_FLOAT
:
1128 case GLSL_TYPE_DOUBLE
:
1129 child_type
= glsl_get_column_type(type
);
1131 case GLSL_TYPE_ARRAY
:
1132 child_type
= glsl_get_array_element(type
);
1134 case GLSL_TYPE_STRUCT
:
1135 child_type
= glsl_get_struct_field(type
, i
);
1138 unreachable("unkown base type");
1141 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1149 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1152 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1153 src
.src_type
= type
;
1158 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1159 const uint32_t *w
, unsigned count
)
1161 if (opcode
== SpvOpSampledImage
) {
1162 struct vtn_value
*val
=
1163 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1164 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1165 val
->sampled_image
->image
=
1166 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1167 val
->sampled_image
->sampler
=
1168 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1170 } else if (opcode
== SpvOpImage
) {
1171 struct vtn_value
*val
=
1172 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1173 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1174 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1175 val
->access_chain
= src_val
->sampled_image
->image
;
1177 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1178 val
->access_chain
= src_val
->access_chain
;
1183 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1184 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1186 struct vtn_sampled_image sampled
;
1187 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1188 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1189 sampled
= *sampled_val
->sampled_image
;
1191 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1192 sampled
.image
= NULL
;
1193 sampled
.sampler
= sampled_val
->access_chain
;
1196 nir_tex_src srcs
[8]; /* 8 should be enough */
1197 nir_tex_src
*p
= srcs
;
1201 bool has_coord
= false;
1203 case SpvOpImageSampleImplicitLod
:
1204 case SpvOpImageSampleExplicitLod
:
1205 case SpvOpImageSampleDrefImplicitLod
:
1206 case SpvOpImageSampleDrefExplicitLod
:
1207 case SpvOpImageSampleProjImplicitLod
:
1208 case SpvOpImageSampleProjExplicitLod
:
1209 case SpvOpImageSampleProjDrefImplicitLod
:
1210 case SpvOpImageSampleProjDrefExplicitLod
:
1211 case SpvOpImageFetch
:
1212 case SpvOpImageGather
:
1213 case SpvOpImageDrefGather
:
1214 case SpvOpImageQueryLod
: {
1215 /* All these types have the coordinate as their first real argument */
1216 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, w
[idx
++]);
1218 p
->src
= nir_src_for_ssa(coord
->def
);
1219 p
->src_type
= nir_tex_src_coord
;
1228 /* These all have an explicit depth value as their next source */
1230 case SpvOpImageSampleDrefImplicitLod
:
1231 case SpvOpImageSampleDrefExplicitLod
:
1232 case SpvOpImageSampleProjDrefImplicitLod
:
1233 case SpvOpImageSampleProjDrefExplicitLod
:
1234 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparitor
);
1240 /* For OpImageQuerySizeLod, we always have an LOD */
1241 if (opcode
== SpvOpImageQuerySizeLod
)
1242 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1244 /* Figure out the base texture operation */
1247 case SpvOpImageSampleImplicitLod
:
1248 case SpvOpImageSampleDrefImplicitLod
:
1249 case SpvOpImageSampleProjImplicitLod
:
1250 case SpvOpImageSampleProjDrefImplicitLod
:
1251 texop
= nir_texop_tex
;
1254 case SpvOpImageSampleExplicitLod
:
1255 case SpvOpImageSampleDrefExplicitLod
:
1256 case SpvOpImageSampleProjExplicitLod
:
1257 case SpvOpImageSampleProjDrefExplicitLod
:
1258 texop
= nir_texop_txl
;
1261 case SpvOpImageFetch
:
1262 texop
= nir_texop_txf
;
1265 case SpvOpImageGather
:
1266 case SpvOpImageDrefGather
:
1267 texop
= nir_texop_tg4
;
1270 case SpvOpImageQuerySizeLod
:
1271 case SpvOpImageQuerySize
:
1272 texop
= nir_texop_txs
;
1275 case SpvOpImageQueryLod
:
1276 texop
= nir_texop_lod
;
1279 case SpvOpImageQueryLevels
:
1280 texop
= nir_texop_query_levels
;
1283 case SpvOpImageQuerySamples
:
1285 unreachable("Unhandled opcode");
1288 /* Now we need to handle some number of optional arguments */
1290 uint32_t operands
= w
[idx
++];
1292 if (operands
& SpvImageOperandsBiasMask
) {
1293 assert(texop
== nir_texop_tex
);
1294 texop
= nir_texop_txb
;
1295 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1298 if (operands
& SpvImageOperandsLodMask
) {
1299 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1300 texop
== nir_texop_txs
);
1301 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1304 if (operands
& SpvImageOperandsGradMask
) {
1305 assert(texop
== nir_texop_tex
);
1306 texop
= nir_texop_txd
;
1307 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1308 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1311 if (operands
& SpvImageOperandsOffsetMask
||
1312 operands
& SpvImageOperandsConstOffsetMask
)
1313 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1315 if (operands
& SpvImageOperandsConstOffsetsMask
)
1316 assert(!"Constant offsets to texture gather not yet implemented");
1318 if (operands
& SpvImageOperandsSampleMask
) {
1319 assert(texop
== nir_texop_txf
);
1320 texop
= nir_texop_txf_ms
;
1321 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1324 /* We should have now consumed exactly all of the arguments */
1325 assert(idx
== count
);
1327 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1330 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
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 instr
->sampler_dim
= glsl_get_sampler_dim(image_type
);
1340 instr
->is_array
= glsl_sampler_type_is_array(image_type
);
1341 instr
->is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1342 instr
->is_new_style_shadow
= instr
->is_shadow
;
1345 switch (instr
->sampler_dim
) {
1346 case GLSL_SAMPLER_DIM_1D
:
1347 case GLSL_SAMPLER_DIM_BUF
:
1348 instr
->coord_components
= 1;
1350 case GLSL_SAMPLER_DIM_2D
:
1351 case GLSL_SAMPLER_DIM_RECT
:
1352 instr
->coord_components
= 2;
1354 case GLSL_SAMPLER_DIM_3D
:
1355 case GLSL_SAMPLER_DIM_CUBE
:
1356 case GLSL_SAMPLER_DIM_MS
:
1357 instr
->coord_components
= 3;
1360 assert("Invalid sampler type");
1363 if (instr
->is_array
)
1364 instr
->coord_components
++;
1366 instr
->coord_components
= 0;
1369 switch (glsl_get_sampler_result_type(image_type
)) {
1370 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1371 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1372 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1373 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1375 unreachable("Invalid base type for sampler result");
1378 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1379 instr
->sampler
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1380 if (sampled
.image
) {
1381 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1382 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &image
->deref
));
1384 instr
->texture
= NULL
;
1387 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1388 nir_tex_instr_dest_size(instr
), NULL
);
1390 assert(glsl_get_vector_elements(ret_type
->type
) ==
1391 nir_tex_instr_dest_size(instr
));
1393 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1394 val
->ssa
->def
= &instr
->dest
.ssa
;
1396 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1399 static nir_ssa_def
*
1400 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1402 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1404 /* The image_load_store intrinsics assume a 4-dim coordinate */
1405 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1406 unsigned swizzle
[4];
1407 for (unsigned i
= 0; i
< 4; i
++)
1408 swizzle
[i
] = MIN2(i
, dim
- 1);
1410 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1414 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1415 const uint32_t *w
, unsigned count
)
1417 /* Just get this one out of the way */
1418 if (opcode
== SpvOpImageTexelPointer
) {
1419 struct vtn_value
*val
=
1420 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1421 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1424 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1425 val
->image
->coord
= get_image_coord(b
, w
[4]);
1426 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1430 struct vtn_image_pointer image
;
1433 case SpvOpAtomicExchange
:
1434 case SpvOpAtomicCompareExchange
:
1435 case SpvOpAtomicCompareExchangeWeak
:
1436 case SpvOpAtomicIIncrement
:
1437 case SpvOpAtomicIDecrement
:
1438 case SpvOpAtomicIAdd
:
1439 case SpvOpAtomicISub
:
1440 case SpvOpAtomicSMin
:
1441 case SpvOpAtomicUMin
:
1442 case SpvOpAtomicSMax
:
1443 case SpvOpAtomicUMax
:
1444 case SpvOpAtomicAnd
:
1446 case SpvOpAtomicXor
:
1447 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1450 case SpvOpImageQuerySize
:
1452 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1454 image
.sample
= NULL
;
1457 case SpvOpImageRead
:
1459 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1460 image
.coord
= get_image_coord(b
, w
[4]);
1462 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1463 assert(w
[5] == SpvImageOperandsSampleMask
);
1464 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1466 image
.sample
= nir_ssa_undef(&b
->nb
, 1);
1470 case SpvOpImageWrite
:
1472 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1473 image
.coord
= get_image_coord(b
, w
[2]);
1477 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1478 assert(w
[4] == SpvImageOperandsSampleMask
);
1479 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1481 image
.sample
= nir_ssa_undef(&b
->nb
, 1);
1486 unreachable("Invalid image opcode");
1489 nir_intrinsic_op op
;
1491 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1492 OP(ImageQuerySize
, size
)
1494 OP(ImageWrite
, store
)
1495 OP(AtomicExchange
, atomic_exchange
)
1496 OP(AtomicCompareExchange
, atomic_comp_swap
)
1497 OP(AtomicIIncrement
, atomic_add
)
1498 OP(AtomicIDecrement
, atomic_add
)
1499 OP(AtomicIAdd
, atomic_add
)
1500 OP(AtomicISub
, atomic_add
)
1501 OP(AtomicSMin
, atomic_min
)
1502 OP(AtomicUMin
, atomic_min
)
1503 OP(AtomicSMax
, atomic_max
)
1504 OP(AtomicUMax
, atomic_max
)
1505 OP(AtomicAnd
, atomic_and
)
1506 OP(AtomicOr
, atomic_or
)
1507 OP(AtomicXor
, atomic_xor
)
1510 unreachable("Invalid image opcode");
1513 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1515 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1516 intrin
->variables
[0] =
1517 nir_deref_as_var(nir_copy_deref(&intrin
->instr
, &image_deref
->deref
));
1519 /* ImageQuerySize doesn't take any extra parameters */
1520 if (opcode
!= SpvOpImageQuerySize
) {
1521 /* The image coordinate is always 4 components but we may not have that
1522 * many. Swizzle to compensate.
1525 for (unsigned i
= 0; i
< 4; i
++)
1526 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1527 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1529 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1533 case SpvOpImageQuerySize
:
1534 case SpvOpImageRead
:
1536 case SpvOpImageWrite
:
1537 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1539 case SpvOpAtomicIIncrement
:
1540 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1542 case SpvOpAtomicIDecrement
:
1543 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1546 case SpvOpAtomicExchange
:
1547 case SpvOpAtomicIAdd
:
1548 case SpvOpAtomicSMin
:
1549 case SpvOpAtomicUMin
:
1550 case SpvOpAtomicSMax
:
1551 case SpvOpAtomicUMax
:
1552 case SpvOpAtomicAnd
:
1554 case SpvOpAtomicXor
:
1555 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1558 case SpvOpAtomicCompareExchange
:
1559 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1560 intrin
->src
[3] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1563 case SpvOpAtomicISub
:
1564 intrin
->src
[2] = nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1568 unreachable("Invalid image opcode");
1571 if (opcode
!= SpvOpImageWrite
) {
1572 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1573 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1574 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, NULL
);
1576 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1578 /* The image intrinsics always return 4 channels but we may not want
1579 * that many. Emit a mov to trim it down.
1581 unsigned swiz
[4] = {0, 1, 2, 3};
1582 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1583 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1584 glsl_get_vector_elements(type
->type
), false);
1586 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1590 static nir_intrinsic_op
1591 get_ssbo_nir_atomic_op(SpvOp opcode
)
1594 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1595 OP(AtomicExchange
, atomic_exchange
)
1596 OP(AtomicCompareExchange
, atomic_comp_swap
)
1597 OP(AtomicIIncrement
, atomic_add
)
1598 OP(AtomicIDecrement
, atomic_add
)
1599 OP(AtomicIAdd
, atomic_add
)
1600 OP(AtomicISub
, atomic_add
)
1601 OP(AtomicSMin
, atomic_imin
)
1602 OP(AtomicUMin
, atomic_umin
)
1603 OP(AtomicSMax
, atomic_imax
)
1604 OP(AtomicUMax
, atomic_umax
)
1605 OP(AtomicAnd
, atomic_and
)
1606 OP(AtomicOr
, atomic_or
)
1607 OP(AtomicXor
, atomic_xor
)
1610 unreachable("Invalid SSBO atomic");
1614 static nir_intrinsic_op
1615 get_shared_nir_atomic_op(SpvOp opcode
)
1618 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
1619 OP(AtomicExchange
, atomic_exchange
)
1620 OP(AtomicCompareExchange
, atomic_comp_swap
)
1621 OP(AtomicIIncrement
, atomic_add
)
1622 OP(AtomicIDecrement
, atomic_add
)
1623 OP(AtomicIAdd
, atomic_add
)
1624 OP(AtomicISub
, atomic_add
)
1625 OP(AtomicSMin
, atomic_imin
)
1626 OP(AtomicUMin
, atomic_umin
)
1627 OP(AtomicSMax
, atomic_imax
)
1628 OP(AtomicUMax
, atomic_umax
)
1629 OP(AtomicAnd
, atomic_and
)
1630 OP(AtomicOr
, atomic_or
)
1631 OP(AtomicXor
, atomic_xor
)
1634 unreachable("Invalid shared atomic");
1639 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1640 const uint32_t *w
, nir_src
*src
)
1643 case SpvOpAtomicIIncrement
:
1644 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1647 case SpvOpAtomicIDecrement
:
1648 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1651 case SpvOpAtomicISub
:
1653 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1656 case SpvOpAtomicCompareExchange
:
1657 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1658 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1662 case SpvOpAtomicExchange
:
1663 case SpvOpAtomicIAdd
:
1664 case SpvOpAtomicSMin
:
1665 case SpvOpAtomicUMin
:
1666 case SpvOpAtomicSMax
:
1667 case SpvOpAtomicUMax
:
1668 case SpvOpAtomicAnd
:
1670 case SpvOpAtomicXor
:
1671 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1675 unreachable("Invalid SPIR-V atomic");
1680 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
1681 const uint32_t *w
, unsigned count
)
1683 struct vtn_access_chain
*chain
=
1684 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1685 nir_intrinsic_instr
*atomic
;
1688 SpvScope scope = w[4];
1689 SpvMemorySemanticsMask semantics = w[5];
1692 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
1693 nir_deref
*deref
= &vtn_access_chain_to_deref(b
, chain
)->deref
;
1694 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
1695 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1696 atomic
->variables
[0] = nir_deref_as_var(nir_copy_deref(atomic
, deref
));
1697 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
1699 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
1700 struct vtn_type
*type
;
1701 nir_ssa_def
*offset
, *index
;
1702 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
1704 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
1706 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1707 atomic
->src
[0] = nir_src_for_ssa(index
);
1708 atomic
->src
[1] = nir_src_for_ssa(offset
);
1709 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
1712 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
, 1, NULL
);
1714 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1715 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1716 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
1717 val
->ssa
->def
= &atomic
->dest
.ssa
;
1718 val
->ssa
->type
= type
->type
;
1720 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
1723 static nir_alu_instr
*
1724 create_vec(nir_shader
*shader
, unsigned num_components
)
1727 switch (num_components
) {
1728 case 1: op
= nir_op_fmov
; break;
1729 case 2: op
= nir_op_vec2
; break;
1730 case 3: op
= nir_op_vec3
; break;
1731 case 4: op
= nir_op_vec4
; break;
1732 default: unreachable("bad vector size");
1735 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
1736 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
, NULL
);
1737 vec
->dest
.write_mask
= (1 << num_components
) - 1;
1742 struct vtn_ssa_value
*
1743 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
1745 if (src
->transposed
)
1746 return src
->transposed
;
1748 struct vtn_ssa_value
*dest
=
1749 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
1751 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
1752 nir_alu_instr
*vec
= create_vec(b
->shader
,
1753 glsl_get_matrix_columns(src
->type
));
1754 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1755 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
1756 vec
->src
[0].swizzle
[0] = i
;
1758 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
1759 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
1760 vec
->src
[j
].swizzle
[0] = i
;
1763 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1764 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
1767 dest
->transposed
= src
;
1773 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
1775 unsigned swiz
[4] = { index
};
1776 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
1780 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
1783 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
);
1785 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
1787 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
1789 vec
->src
[i
].src
= nir_src_for_ssa(src
);
1790 vec
->src
[i
].swizzle
[0] = i
;
1794 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1796 return &vec
->dest
.dest
.ssa
;
1800 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1803 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
1804 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1805 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1806 vtn_vector_extract(b
, src
, i
), dest
);
1812 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1813 nir_ssa_def
*insert
, nir_ssa_def
*index
)
1815 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
1816 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1817 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1818 vtn_vector_insert(b
, src
, insert
, i
), dest
);
1823 static nir_ssa_def
*
1824 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
1825 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
1826 const uint32_t *indices
)
1828 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
);
1830 nir_ssa_undef_instr
*undef
= nir_ssa_undef_instr_create(b
->shader
, 1);
1831 nir_builder_instr_insert(&b
->nb
, &undef
->instr
);
1833 for (unsigned i
= 0; i
< num_components
; i
++) {
1834 uint32_t index
= indices
[i
];
1835 if (index
== 0xffffffff) {
1836 vec
->src
[i
].src
= nir_src_for_ssa(&undef
->def
);
1837 } else if (index
< src0
->num_components
) {
1838 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
1839 vec
->src
[i
].swizzle
[0] = index
;
1841 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
1842 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
1846 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1848 return &vec
->dest
.dest
.ssa
;
1852 * Concatentates a number of vectors/scalars together to produce a vector
1854 static nir_ssa_def
*
1855 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
1856 unsigned num_srcs
, nir_ssa_def
**srcs
)
1858 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
);
1860 unsigned dest_idx
= 0;
1861 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1862 nir_ssa_def
*src
= srcs
[i
];
1863 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
1864 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
1865 vec
->src
[dest_idx
].swizzle
[0] = j
;
1870 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1872 return &vec
->dest
.dest
.ssa
;
1875 static struct vtn_ssa_value
*
1876 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
1878 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
1879 dest
->type
= src
->type
;
1881 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1882 dest
->def
= src
->def
;
1884 unsigned elems
= glsl_get_length(src
->type
);
1886 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
1887 for (unsigned i
= 0; i
< elems
; i
++)
1888 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
1894 static struct vtn_ssa_value
*
1895 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1896 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
1897 unsigned num_indices
)
1899 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
1901 struct vtn_ssa_value
*cur
= dest
;
1903 for (i
= 0; i
< num_indices
- 1; i
++) {
1904 cur
= cur
->elems
[indices
[i
]];
1907 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1908 /* According to the SPIR-V spec, OpCompositeInsert may work down to
1909 * the component granularity. In that case, the last index will be
1910 * the index to insert the scalar into the vector.
1913 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
1915 cur
->elems
[indices
[i
]] = insert
;
1921 static struct vtn_ssa_value
*
1922 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1923 const uint32_t *indices
, unsigned num_indices
)
1925 struct vtn_ssa_value
*cur
= src
;
1926 for (unsigned i
= 0; i
< num_indices
; i
++) {
1927 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1928 assert(i
== num_indices
- 1);
1929 /* According to the SPIR-V spec, OpCompositeExtract may work down to
1930 * the component granularity. The last index will be the index of the
1931 * vector to extract.
1934 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
1935 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
1936 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
1939 cur
= cur
->elems
[indices
[i
]];
1947 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
1948 const uint32_t *w
, unsigned count
)
1950 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1951 const struct glsl_type
*type
=
1952 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1953 val
->ssa
= vtn_create_ssa_value(b
, type
);
1956 case SpvOpVectorExtractDynamic
:
1957 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1958 vtn_ssa_value(b
, w
[4])->def
);
1961 case SpvOpVectorInsertDynamic
:
1962 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1963 vtn_ssa_value(b
, w
[4])->def
,
1964 vtn_ssa_value(b
, w
[5])->def
);
1967 case SpvOpVectorShuffle
:
1968 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
1969 vtn_ssa_value(b
, w
[3])->def
,
1970 vtn_ssa_value(b
, w
[4])->def
,
1974 case SpvOpCompositeConstruct
: {
1975 unsigned elems
= count
- 3;
1976 if (glsl_type_is_vector_or_scalar(type
)) {
1977 nir_ssa_def
*srcs
[4];
1978 for (unsigned i
= 0; i
< elems
; i
++)
1979 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
1981 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
1984 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1985 for (unsigned i
= 0; i
< elems
; i
++)
1986 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
1990 case SpvOpCompositeExtract
:
1991 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
1995 case SpvOpCompositeInsert
:
1996 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
1997 vtn_ssa_value(b
, w
[3]),
2001 case SpvOpCopyObject
:
2002 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2006 unreachable("unknown composite operation");
2011 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2012 const uint32_t *w
, unsigned count
)
2014 nir_intrinsic_op intrinsic_op
;
2016 case SpvOpEmitVertex
:
2017 case SpvOpEmitStreamVertex
:
2018 intrinsic_op
= nir_intrinsic_emit_vertex
;
2020 case SpvOpEndPrimitive
:
2021 case SpvOpEndStreamPrimitive
:
2022 intrinsic_op
= nir_intrinsic_end_primitive
;
2024 case SpvOpMemoryBarrier
:
2025 intrinsic_op
= nir_intrinsic_memory_barrier
;
2027 case SpvOpControlBarrier
:
2028 intrinsic_op
= nir_intrinsic_barrier
;
2031 unreachable("unknown barrier instruction");
2034 nir_intrinsic_instr
*intrin
=
2035 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2037 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2038 intrin
->const_index
[0] = w
[1];
2040 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2044 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2047 case SpvExecutionModeInputPoints
:
2048 case SpvExecutionModeOutputPoints
:
2049 return 0; /* GL_POINTS */
2050 case SpvExecutionModeInputLines
:
2051 return 1; /* GL_LINES */
2052 case SpvExecutionModeInputLinesAdjacency
:
2053 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2054 case SpvExecutionModeTriangles
:
2055 return 4; /* GL_TRIANGLES */
2056 case SpvExecutionModeInputTrianglesAdjacency
:
2057 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2058 case SpvExecutionModeQuads
:
2059 return 7; /* GL_QUADS */
2060 case SpvExecutionModeIsolines
:
2061 return 0x8E7A; /* GL_ISOLINES */
2062 case SpvExecutionModeOutputLineStrip
:
2063 return 3; /* GL_LINE_STRIP */
2064 case SpvExecutionModeOutputTriangleStrip
:
2065 return 5; /* GL_TRIANGLE_STRIP */
2067 assert(!"Invalid primitive type");
2073 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2076 case SpvExecutionModeInputPoints
:
2078 case SpvExecutionModeInputLines
:
2080 case SpvExecutionModeInputLinesAdjacency
:
2082 case SpvExecutionModeTriangles
:
2084 case SpvExecutionModeInputTrianglesAdjacency
:
2087 assert(!"Invalid GS input mode");
2092 static gl_shader_stage
2093 stage_for_execution_model(SpvExecutionModel model
)
2096 case SpvExecutionModelVertex
:
2097 return MESA_SHADER_VERTEX
;
2098 case SpvExecutionModelTessellationControl
:
2099 return MESA_SHADER_TESS_CTRL
;
2100 case SpvExecutionModelTessellationEvaluation
:
2101 return MESA_SHADER_TESS_EVAL
;
2102 case SpvExecutionModelGeometry
:
2103 return MESA_SHADER_GEOMETRY
;
2104 case SpvExecutionModelFragment
:
2105 return MESA_SHADER_FRAGMENT
;
2106 case SpvExecutionModelGLCompute
:
2107 return MESA_SHADER_COMPUTE
;
2109 unreachable("Unsupported execution model");
2114 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2115 const uint32_t *w
, unsigned count
)
2119 case SpvOpSourceExtension
:
2120 case SpvOpSourceContinued
:
2121 case SpvOpExtension
:
2122 /* Unhandled, but these are for debug so that's ok. */
2125 case SpvOpCapability
:
2126 switch ((SpvCapability
)w
[1]) {
2127 case SpvCapabilityMatrix
:
2128 case SpvCapabilityShader
:
2129 case SpvCapabilityGeometry
:
2132 assert(!"Unsupported capability");
2136 case SpvOpExtInstImport
:
2137 vtn_handle_extension(b
, opcode
, w
, count
);
2140 case SpvOpMemoryModel
:
2141 assert(w
[1] == SpvAddressingModelLogical
);
2142 assert(w
[2] == SpvMemoryModelGLSL450
);
2145 case SpvOpEntryPoint
: {
2146 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2147 /* Let this be a name label regardless */
2148 unsigned name_words
;
2149 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2151 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2152 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2155 assert(b
->entry_point
== NULL
);
2156 b
->entry_point
= entry_point
;
2161 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2162 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2166 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2169 case SpvOpMemberName
:
2173 case SpvOpExecutionMode
:
2174 case SpvOpDecorationGroup
:
2176 case SpvOpMemberDecorate
:
2177 case SpvOpGroupDecorate
:
2178 case SpvOpGroupMemberDecorate
:
2179 vtn_handle_decoration(b
, opcode
, w
, count
);
2183 return false; /* End of preamble */
2190 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2191 const struct vtn_decoration
*mode
, void *data
)
2193 assert(b
->entry_point
== entry_point
);
2195 switch(mode
->exec_mode
) {
2196 case SpvExecutionModeOriginUpperLeft
:
2197 case SpvExecutionModeOriginLowerLeft
:
2198 b
->origin_upper_left
=
2199 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2202 case SpvExecutionModeEarlyFragmentTests
:
2203 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2204 b
->shader
->info
.fs
.early_fragment_tests
= true;
2207 case SpvExecutionModeInvocations
:
2208 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2209 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2212 case SpvExecutionModeDepthReplacing
:
2213 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2214 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2216 case SpvExecutionModeDepthGreater
:
2217 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2218 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2220 case SpvExecutionModeDepthLess
:
2221 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2222 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2224 case SpvExecutionModeDepthUnchanged
:
2225 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2226 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2229 case SpvExecutionModeLocalSize
:
2230 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2231 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2232 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2233 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2235 case SpvExecutionModeLocalSizeHint
:
2236 break; /* Nothing do do with this */
2238 case SpvExecutionModeOutputVertices
:
2239 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2240 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2243 case SpvExecutionModeInputPoints
:
2244 case SpvExecutionModeInputLines
:
2245 case SpvExecutionModeInputLinesAdjacency
:
2246 case SpvExecutionModeTriangles
:
2247 case SpvExecutionModeInputTrianglesAdjacency
:
2248 case SpvExecutionModeQuads
:
2249 case SpvExecutionModeIsolines
:
2250 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2251 b
->shader
->info
.gs
.vertices_in
=
2252 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2254 assert(!"Tesselation shaders not yet supported");
2258 case SpvExecutionModeOutputPoints
:
2259 case SpvExecutionModeOutputLineStrip
:
2260 case SpvExecutionModeOutputTriangleStrip
:
2261 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2262 b
->shader
->info
.gs
.output_primitive
=
2263 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2266 case SpvExecutionModeSpacingEqual
:
2267 case SpvExecutionModeSpacingFractionalEven
:
2268 case SpvExecutionModeSpacingFractionalOdd
:
2269 case SpvExecutionModeVertexOrderCw
:
2270 case SpvExecutionModeVertexOrderCcw
:
2271 case SpvExecutionModePointMode
:
2272 assert(!"TODO: Add tessellation metadata");
2275 case SpvExecutionModePixelCenterInteger
:
2276 case SpvExecutionModeXfb
:
2277 assert(!"Unhandled execution mode");
2280 case SpvExecutionModeVecTypeHint
:
2281 case SpvExecutionModeContractionOff
:
2287 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2288 const uint32_t *w
, unsigned count
)
2292 case SpvOpSourceContinued
:
2293 case SpvOpSourceExtension
:
2294 case SpvOpExtension
:
2295 case SpvOpCapability
:
2296 case SpvOpExtInstImport
:
2297 case SpvOpMemoryModel
:
2298 case SpvOpEntryPoint
:
2299 case SpvOpExecutionMode
:
2302 case SpvOpMemberName
:
2303 case SpvOpDecorationGroup
:
2305 case SpvOpMemberDecorate
:
2306 case SpvOpGroupDecorate
:
2307 case SpvOpGroupMemberDecorate
:
2308 assert(!"Invalid opcode types and variables section");
2314 case SpvOpTypeFloat
:
2315 case SpvOpTypeVector
:
2316 case SpvOpTypeMatrix
:
2317 case SpvOpTypeImage
:
2318 case SpvOpTypeSampler
:
2319 case SpvOpTypeSampledImage
:
2320 case SpvOpTypeArray
:
2321 case SpvOpTypeRuntimeArray
:
2322 case SpvOpTypeStruct
:
2323 case SpvOpTypeOpaque
:
2324 case SpvOpTypePointer
:
2325 case SpvOpTypeFunction
:
2326 case SpvOpTypeEvent
:
2327 case SpvOpTypeDeviceEvent
:
2328 case SpvOpTypeReserveId
:
2329 case SpvOpTypeQueue
:
2331 vtn_handle_type(b
, opcode
, w
, count
);
2334 case SpvOpConstantTrue
:
2335 case SpvOpConstantFalse
:
2337 case SpvOpConstantComposite
:
2338 case SpvOpConstantSampler
:
2339 case SpvOpConstantNull
:
2340 case SpvOpSpecConstantTrue
:
2341 case SpvOpSpecConstantFalse
:
2342 case SpvOpSpecConstant
:
2343 case SpvOpSpecConstantComposite
:
2344 case SpvOpSpecConstantOp
:
2345 vtn_handle_constant(b
, opcode
, w
, count
);
2349 vtn_handle_variables(b
, opcode
, w
, count
);
2353 return false; /* End of preamble */
2360 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2361 const uint32_t *w
, unsigned count
)
2367 case SpvOpLoopMerge
:
2368 case SpvOpSelectionMerge
:
2369 /* This is handled by cfg pre-pass and walk_blocks */
2373 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2374 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2379 vtn_handle_extension(b
, opcode
, w
, count
);
2385 case SpvOpCopyMemory
:
2386 case SpvOpCopyMemorySized
:
2387 case SpvOpAccessChain
:
2388 case SpvOpInBoundsAccessChain
:
2389 case SpvOpArrayLength
:
2390 vtn_handle_variables(b
, opcode
, w
, count
);
2393 case SpvOpFunctionCall
:
2394 vtn_handle_function_call(b
, opcode
, w
, count
);
2397 case SpvOpSampledImage
:
2399 case SpvOpImageSampleImplicitLod
:
2400 case SpvOpImageSampleExplicitLod
:
2401 case SpvOpImageSampleDrefImplicitLod
:
2402 case SpvOpImageSampleDrefExplicitLod
:
2403 case SpvOpImageSampleProjImplicitLod
:
2404 case SpvOpImageSampleProjExplicitLod
:
2405 case SpvOpImageSampleProjDrefImplicitLod
:
2406 case SpvOpImageSampleProjDrefExplicitLod
:
2407 case SpvOpImageFetch
:
2408 case SpvOpImageGather
:
2409 case SpvOpImageDrefGather
:
2410 case SpvOpImageQuerySizeLod
:
2411 case SpvOpImageQueryLod
:
2412 case SpvOpImageQueryLevels
:
2413 case SpvOpImageQuerySamples
:
2414 vtn_handle_texture(b
, opcode
, w
, count
);
2417 case SpvOpImageRead
:
2418 case SpvOpImageWrite
:
2419 case SpvOpImageTexelPointer
:
2420 vtn_handle_image(b
, opcode
, w
, count
);
2423 case SpvOpImageQuerySize
: {
2424 struct vtn_access_chain
*image
=
2425 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2426 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2427 vtn_handle_image(b
, opcode
, w
, count
);
2429 vtn_handle_texture(b
, opcode
, w
, count
);
2434 case SpvOpAtomicExchange
:
2435 case SpvOpAtomicCompareExchange
:
2436 case SpvOpAtomicCompareExchangeWeak
:
2437 case SpvOpAtomicIIncrement
:
2438 case SpvOpAtomicIDecrement
:
2439 case SpvOpAtomicIAdd
:
2440 case SpvOpAtomicISub
:
2441 case SpvOpAtomicSMin
:
2442 case SpvOpAtomicUMin
:
2443 case SpvOpAtomicSMax
:
2444 case SpvOpAtomicUMax
:
2445 case SpvOpAtomicAnd
:
2447 case SpvOpAtomicXor
: {
2448 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2449 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2450 vtn_handle_image(b
, opcode
, w
, count
);
2452 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2453 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2463 case SpvOpConvertFToU
:
2464 case SpvOpConvertFToS
:
2465 case SpvOpConvertSToF
:
2466 case SpvOpConvertUToF
:
2470 case SpvOpQuantizeToF16
:
2471 case SpvOpConvertPtrToU
:
2472 case SpvOpConvertUToPtr
:
2473 case SpvOpPtrCastToGeneric
:
2474 case SpvOpGenericCastToPtr
:
2480 case SpvOpSignBitSet
:
2481 case SpvOpLessOrGreater
:
2483 case SpvOpUnordered
:
2498 case SpvOpVectorTimesScalar
:
2500 case SpvOpIAddCarry
:
2501 case SpvOpISubBorrow
:
2502 case SpvOpUMulExtended
:
2503 case SpvOpSMulExtended
:
2504 case SpvOpShiftRightLogical
:
2505 case SpvOpShiftRightArithmetic
:
2506 case SpvOpShiftLeftLogical
:
2507 case SpvOpLogicalEqual
:
2508 case SpvOpLogicalNotEqual
:
2509 case SpvOpLogicalOr
:
2510 case SpvOpLogicalAnd
:
2511 case SpvOpLogicalNot
:
2512 case SpvOpBitwiseOr
:
2513 case SpvOpBitwiseXor
:
2514 case SpvOpBitwiseAnd
:
2517 case SpvOpFOrdEqual
:
2518 case SpvOpFUnordEqual
:
2519 case SpvOpINotEqual
:
2520 case SpvOpFOrdNotEqual
:
2521 case SpvOpFUnordNotEqual
:
2522 case SpvOpULessThan
:
2523 case SpvOpSLessThan
:
2524 case SpvOpFOrdLessThan
:
2525 case SpvOpFUnordLessThan
:
2526 case SpvOpUGreaterThan
:
2527 case SpvOpSGreaterThan
:
2528 case SpvOpFOrdGreaterThan
:
2529 case SpvOpFUnordGreaterThan
:
2530 case SpvOpULessThanEqual
:
2531 case SpvOpSLessThanEqual
:
2532 case SpvOpFOrdLessThanEqual
:
2533 case SpvOpFUnordLessThanEqual
:
2534 case SpvOpUGreaterThanEqual
:
2535 case SpvOpSGreaterThanEqual
:
2536 case SpvOpFOrdGreaterThanEqual
:
2537 case SpvOpFUnordGreaterThanEqual
:
2543 case SpvOpFwidthFine
:
2544 case SpvOpDPdxCoarse
:
2545 case SpvOpDPdyCoarse
:
2546 case SpvOpFwidthCoarse
:
2547 case SpvOpBitFieldInsert
:
2548 case SpvOpBitFieldSExtract
:
2549 case SpvOpBitFieldUExtract
:
2550 case SpvOpBitReverse
:
2552 case SpvOpTranspose
:
2553 case SpvOpOuterProduct
:
2554 case SpvOpMatrixTimesScalar
:
2555 case SpvOpVectorTimesMatrix
:
2556 case SpvOpMatrixTimesVector
:
2557 case SpvOpMatrixTimesMatrix
:
2558 vtn_handle_alu(b
, opcode
, w
, count
);
2561 case SpvOpVectorExtractDynamic
:
2562 case SpvOpVectorInsertDynamic
:
2563 case SpvOpVectorShuffle
:
2564 case SpvOpCompositeConstruct
:
2565 case SpvOpCompositeExtract
:
2566 case SpvOpCompositeInsert
:
2567 case SpvOpCopyObject
:
2568 vtn_handle_composite(b
, opcode
, w
, count
);
2571 case SpvOpEmitVertex
:
2572 case SpvOpEndPrimitive
:
2573 case SpvOpEmitStreamVertex
:
2574 case SpvOpEndStreamPrimitive
:
2575 case SpvOpControlBarrier
:
2576 case SpvOpMemoryBarrier
:
2577 vtn_handle_barrier(b
, opcode
, w
, count
);
2581 unreachable("Unhandled opcode");
2588 spirv_to_nir(const uint32_t *words
, size_t word_count
,
2589 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
2590 gl_shader_stage stage
, const char *entry_point_name
,
2591 const nir_shader_compiler_options
*options
)
2593 const uint32_t *word_end
= words
+ word_count
;
2595 /* Handle the SPIR-V header (first 4 dwords) */
2596 assert(word_count
> 5);
2598 assert(words
[0] == SpvMagicNumber
);
2599 assert(words
[1] >= 0x10000);
2600 /* words[2] == generator magic */
2601 unsigned value_id_bound
= words
[3];
2602 assert(words
[4] == 0);
2606 /* Initialize the stn_builder object */
2607 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
2608 b
->value_id_bound
= value_id_bound
;
2609 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
2610 exec_list_make_empty(&b
->functions
);
2611 b
->entry_point_stage
= stage
;
2612 b
->entry_point_name
= entry_point_name
;
2614 /* Handle all the preamble instructions */
2615 words
= vtn_foreach_instruction(b
, words
, word_end
,
2616 vtn_handle_preamble_instruction
);
2618 if (b
->entry_point
== NULL
) {
2619 assert(!"Entry point not found");
2624 b
->shader
= nir_shader_create(NULL
, stage
, options
);
2626 /* Parse execution modes */
2627 vtn_foreach_execution_mode(b
, b
->entry_point
,
2628 vtn_handle_execution_mode
, NULL
);
2630 b
->specializations
= spec
;
2631 b
->num_specializations
= num_spec
;
2633 /* Handle all variable, type, and constant instructions */
2634 words
= vtn_foreach_instruction(b
, words
, word_end
,
2635 vtn_handle_variable_or_type_instruction
);
2637 vtn_build_cfg(b
, words
, word_end
);
2639 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
2640 b
->impl
= func
->impl
;
2641 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
2642 _mesa_key_pointer_equal
);
2644 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
2647 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
2648 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
2649 assert(entry_point
);