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 unsigned bit_size
= glsl_get_bit_size(glsl_get_base_type(val
->type
));
42 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
44 unsigned elems
= glsl_get_length(val
->type
);
45 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
46 if (glsl_type_is_matrix(type
)) {
47 const struct glsl_type
*elem_type
=
48 glsl_vector_type(glsl_get_base_type(type
),
49 glsl_get_vector_elements(type
));
51 for (unsigned i
= 0; i
< elems
; i
++)
52 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
53 } else if (glsl_type_is_array(type
)) {
54 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
55 for (unsigned i
= 0; i
< elems
; i
++)
56 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
58 for (unsigned i
= 0; i
< elems
; i
++) {
59 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
60 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
68 static struct vtn_ssa_value
*
69 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
70 const struct glsl_type
*type
)
72 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
77 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
80 switch (glsl_get_base_type(type
)) {
85 case GLSL_TYPE_DOUBLE
:
86 if (glsl_type_is_vector_or_scalar(type
)) {
87 unsigned num_components
= glsl_get_vector_elements(val
->type
);
88 nir_load_const_instr
*load
=
89 nir_load_const_instr_create(b
->shader
, num_components
, 32);
91 for (unsigned i
= 0; i
< num_components
; i
++)
92 load
->value
.u32
[i
] = constant
->value
.u
[i
];
94 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
95 val
->def
= &load
->def
;
97 assert(glsl_type_is_matrix(type
));
98 unsigned rows
= glsl_get_vector_elements(val
->type
);
99 unsigned columns
= glsl_get_matrix_columns(val
->type
);
100 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
102 for (unsigned i
= 0; i
< columns
; i
++) {
103 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
104 col_val
->type
= glsl_get_column_type(val
->type
);
105 nir_load_const_instr
*load
=
106 nir_load_const_instr_create(b
->shader
, rows
, 32);
108 for (unsigned j
= 0; j
< rows
; j
++)
109 load
->value
.u32
[j
] = constant
->value
.u
[rows
* i
+ j
];
111 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
112 col_val
->def
= &load
->def
;
114 val
->elems
[i
] = col_val
;
119 case GLSL_TYPE_ARRAY
: {
120 unsigned elems
= glsl_get_length(val
->type
);
121 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
122 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
123 for (unsigned i
= 0; i
< elems
; i
++)
124 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
129 case GLSL_TYPE_STRUCT
: {
130 unsigned elems
= glsl_get_length(val
->type
);
131 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
132 for (unsigned i
= 0; i
< elems
; i
++) {
133 const struct glsl_type
*elem_type
=
134 glsl_get_struct_field(val
->type
, i
);
135 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
142 unreachable("bad constant type");
148 struct vtn_ssa_value
*
149 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
151 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
152 switch (val
->value_type
) {
153 case vtn_value_type_undef
:
154 return vtn_undef_ssa_value(b
, val
->type
->type
);
156 case vtn_value_type_constant
:
157 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
159 case vtn_value_type_ssa
:
162 case vtn_value_type_access_chain
:
163 /* This is needed for function parameters */
164 return vtn_variable_load(b
, val
->access_chain
);
167 unreachable("Invalid type for an SSA value");
172 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
173 unsigned word_count
, unsigned *words_used
)
175 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
177 /* Ammount of space taken by the string (including the null) */
178 unsigned len
= strlen(dup
) + 1;
179 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
185 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
186 const uint32_t *end
, vtn_instruction_handler handler
)
192 const uint32_t *w
= start
;
194 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
195 unsigned count
= w
[0] >> SpvWordCountShift
;
196 assert(count
>= 1 && w
+ count
<= end
);
200 break; /* Do nothing */
203 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
215 if (!handler(b
, opcode
, w
, count
))
227 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
228 const uint32_t *w
, unsigned count
)
231 case SpvOpExtInstImport
: {
232 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
233 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
234 val
->ext_handler
= vtn_handle_glsl450_instruction
;
236 assert(!"Unsupported extension");
242 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
243 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
250 unreachable("Unhandled opcode");
255 _foreach_decoration_helper(struct vtn_builder
*b
,
256 struct vtn_value
*base_value
,
258 struct vtn_value
*value
,
259 vtn_decoration_foreach_cb cb
, void *data
)
261 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
263 if (dec
->scope
== VTN_DEC_DECORATION
) {
264 member
= parent_member
;
265 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
266 assert(parent_member
== -1);
267 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
269 /* Not a decoration */
274 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
275 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
278 cb(b
, base_value
, member
, dec
, data
);
283 /** Iterates (recursively if needed) over all of the decorations on a value
285 * This function iterates over all of the decorations applied to a given
286 * value. If it encounters a decoration group, it recurses into the group
287 * and iterates over all of those decorations as well.
290 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
291 vtn_decoration_foreach_cb cb
, void *data
)
293 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
297 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
298 vtn_execution_mode_foreach_cb cb
, void *data
)
300 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
301 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
304 assert(dec
->group
== NULL
);
305 cb(b
, value
, dec
, data
);
310 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
311 const uint32_t *w
, unsigned count
)
313 const uint32_t *w_end
= w
+ count
;
314 const uint32_t target
= w
[1];
318 case SpvOpDecorationGroup
:
319 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
323 case SpvOpMemberDecorate
:
324 case SpvOpExecutionMode
: {
325 struct vtn_value
*val
= &b
->values
[target
];
327 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
330 dec
->scope
= VTN_DEC_DECORATION
;
332 case SpvOpMemberDecorate
:
333 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
335 case SpvOpExecutionMode
:
336 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
339 unreachable("Invalid decoration opcode");
341 dec
->decoration
= *(w
++);
344 /* Link into the list */
345 dec
->next
= val
->decoration
;
346 val
->decoration
= dec
;
350 case SpvOpGroupMemberDecorate
:
351 case SpvOpGroupDecorate
: {
352 struct vtn_value
*group
=
353 vtn_value(b
, target
, vtn_value_type_decoration_group
);
355 for (; w
< w_end
; w
++) {
356 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
357 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
360 if (opcode
== SpvOpGroupDecorate
) {
361 dec
->scope
= VTN_DEC_DECORATION
;
363 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
366 /* Link into the list */
367 dec
->next
= val
->decoration
;
368 val
->decoration
= dec
;
374 unreachable("Unhandled opcode");
378 struct member_decoration_ctx
{
380 struct glsl_struct_field
*fields
;
381 struct vtn_type
*type
;
384 /* does a shallow copy of a vtn_type */
386 static struct vtn_type
*
387 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
389 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
390 dest
->type
= src
->type
;
391 dest
->is_builtin
= src
->is_builtin
;
393 dest
->builtin
= src
->builtin
;
395 if (!glsl_type_is_scalar(src
->type
)) {
396 switch (glsl_get_base_type(src
->type
)) {
400 case GLSL_TYPE_FLOAT
:
401 case GLSL_TYPE_DOUBLE
:
402 case GLSL_TYPE_ARRAY
:
403 dest
->row_major
= src
->row_major
;
404 dest
->stride
= src
->stride
;
405 dest
->array_element
= src
->array_element
;
408 case GLSL_TYPE_STRUCT
: {
409 unsigned elems
= glsl_get_length(src
->type
);
411 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
412 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
414 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
415 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
420 unreachable("unhandled type");
427 static struct vtn_type
*
428 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
430 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
431 type
= type
->members
[member
];
433 /* We may have an array of matrices.... Oh, joy! */
434 while (glsl_type_is_array(type
->type
)) {
435 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
436 type
= type
->array_element
;
439 assert(glsl_type_is_matrix(type
->type
));
445 struct_member_decoration_cb(struct vtn_builder
*b
,
446 struct vtn_value
*val
, int member
,
447 const struct vtn_decoration
*dec
, void *void_ctx
)
449 struct member_decoration_ctx
*ctx
= void_ctx
;
454 assert(member
< ctx
->num_fields
);
456 switch (dec
->decoration
) {
457 case SpvDecorationNonWritable
:
458 case SpvDecorationRelaxedPrecision
:
459 break; /* FIXME: Do nothing with this for now. */
460 case SpvDecorationNoPerspective
:
461 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_NOPERSPECTIVE
;
463 case SpvDecorationFlat
:
464 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_FLAT
;
466 case SpvDecorationCentroid
:
467 ctx
->fields
[member
].centroid
= true;
469 case SpvDecorationSample
:
470 ctx
->fields
[member
].sample
= true;
472 case SpvDecorationLocation
:
473 ctx
->fields
[member
].location
= dec
->literals
[0];
475 case SpvDecorationBuiltIn
:
476 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
477 ctx
->type
->members
[member
]->is_builtin
= true;
478 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
479 ctx
->type
->builtin_block
= true;
481 case SpvDecorationOffset
:
482 ctx
->type
->offsets
[member
] = dec
->literals
[0];
484 case SpvDecorationMatrixStride
:
485 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
487 case SpvDecorationColMajor
:
488 break; /* Nothing to do here. Column-major is the default. */
489 case SpvDecorationRowMajor
:
490 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
493 unreachable("Unhandled member decoration");
498 type_decoration_cb(struct vtn_builder
*b
,
499 struct vtn_value
*val
, int member
,
500 const struct vtn_decoration
*dec
, void *ctx
)
502 struct vtn_type
*type
= val
->type
;
507 switch (dec
->decoration
) {
508 case SpvDecorationArrayStride
:
509 type
->stride
= dec
->literals
[0];
511 case SpvDecorationBlock
:
514 case SpvDecorationBufferBlock
:
515 type
->buffer_block
= true;
517 case SpvDecorationGLSLShared
:
518 case SpvDecorationGLSLPacked
:
519 /* Ignore these, since we get explicit offsets anyways */
522 case SpvDecorationStream
:
523 assert(dec
->literals
[0] == 0);
527 unreachable("Unhandled type decoration");
532 translate_image_format(SpvImageFormat format
)
535 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
536 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
537 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
538 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
539 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
540 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
541 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
542 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
543 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
544 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
545 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
546 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
547 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
548 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
549 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
550 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
551 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
552 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
553 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
554 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
555 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
556 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
557 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
558 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
559 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
560 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
561 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
562 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
563 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
564 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
565 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
566 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
567 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
568 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
569 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
570 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
571 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
572 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
573 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
574 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
576 assert(!"Invalid image format");
582 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
583 const uint32_t *w
, unsigned count
)
585 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
587 val
->type
= rzalloc(b
, struct vtn_type
);
588 val
->type
->is_builtin
= false;
589 val
->type
->val
= val
;
593 val
->type
->type
= glsl_void_type();
596 val
->type
->type
= glsl_bool_type();
599 const bool signedness
= w
[3];
600 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
604 val
->type
->type
= glsl_float_type();
607 case SpvOpTypeVector
: {
608 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
609 unsigned elems
= w
[3];
611 assert(glsl_type_is_scalar(base
->type
));
612 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
614 /* Vectors implicitly have sizeof(base_type) stride. For now, this
615 * is always 4 bytes. This will have to change if we want to start
616 * supporting doubles or half-floats.
618 val
->type
->stride
= 4;
619 val
->type
->array_element
= base
;
623 case SpvOpTypeMatrix
: {
624 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
625 unsigned columns
= w
[3];
627 assert(glsl_type_is_vector(base
->type
));
628 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
629 glsl_get_vector_elements(base
->type
),
631 assert(!glsl_type_is_error(val
->type
->type
));
632 val
->type
->array_element
= base
;
633 val
->type
->row_major
= false;
634 val
->type
->stride
= 0;
638 case SpvOpTypeRuntimeArray
:
639 case SpvOpTypeArray
: {
640 struct vtn_type
*array_element
=
641 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
644 if (opcode
== SpvOpTypeRuntimeArray
) {
645 /* A length of 0 is used to denote unsized arrays */
649 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->value
.u
[0];
652 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
653 val
->type
->array_element
= array_element
;
654 val
->type
->stride
= 0;
658 case SpvOpTypeStruct
: {
659 unsigned num_fields
= count
- 2;
660 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
661 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
663 NIR_VLA(struct glsl_struct_field
, fields
, count
);
664 for (unsigned i
= 0; i
< num_fields
; i
++) {
665 val
->type
->members
[i
] =
666 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
667 fields
[i
] = (struct glsl_struct_field
) {
668 .type
= val
->type
->members
[i
]->type
,
669 .name
= ralloc_asprintf(b
, "field%d", i
),
674 struct member_decoration_ctx ctx
= {
675 .num_fields
= num_fields
,
680 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
682 const char *name
= val
->name
? val
->name
: "struct";
684 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
688 case SpvOpTypeFunction
: {
689 const struct glsl_type
*return_type
=
690 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
691 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
692 for (unsigned i
= 0; i
< count
- 3; i
++) {
693 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
697 params
[i
].out
= true;
699 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
703 case SpvOpTypePointer
:
704 /* FIXME: For now, we'll just do the really lame thing and return
705 * the same type. The validator should ensure that the proper number
706 * of dereferences happen
708 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
711 case SpvOpTypeImage
: {
712 const struct glsl_type
*sampled_type
=
713 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
715 assert(glsl_type_is_vector_or_scalar(sampled_type
));
717 enum glsl_sampler_dim dim
;
718 switch ((SpvDim
)w
[3]) {
719 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
720 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
721 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
722 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
723 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
724 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
726 unreachable("Invalid SPIR-V Sampler dimension");
729 bool is_shadow
= w
[4];
730 bool is_array
= w
[5];
731 bool multisampled
= w
[6];
732 unsigned sampled
= w
[7];
733 SpvImageFormat format
= w
[8];
736 val
->type
->access_qualifier
= w
[9];
738 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
741 assert(dim
== GLSL_SAMPLER_DIM_2D
);
742 dim
= GLSL_SAMPLER_DIM_MS
;
745 val
->type
->image_format
= translate_image_format(format
);
748 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
749 glsl_get_base_type(sampled_type
));
750 } else if (sampled
== 2) {
753 val
->type
->type
= glsl_image_type(dim
, is_array
,
754 glsl_get_base_type(sampled_type
));
756 assert(!"We need to know if the image will be sampled");
761 case SpvOpTypeSampledImage
:
762 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
765 case SpvOpTypeSampler
:
766 /* The actual sampler type here doesn't really matter. It gets
767 * thrown away the moment you combine it with an image. What really
768 * matters is that it's a sampler type as opposed to an integer type
769 * so the backend knows what to do.
771 val
->type
->type
= glsl_bare_sampler_type();
774 case SpvOpTypeOpaque
:
776 case SpvOpTypeDeviceEvent
:
777 case SpvOpTypeReserveId
:
781 unreachable("Unhandled opcode");
784 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
787 static nir_constant
*
788 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
790 nir_constant
*c
= rzalloc(b
, nir_constant
);
792 switch (glsl_get_base_type(type
)) {
796 case GLSL_TYPE_FLOAT
:
797 case GLSL_TYPE_DOUBLE
:
798 /* Nothing to do here. It's already initialized to zero */
801 case GLSL_TYPE_ARRAY
:
802 assert(glsl_get_length(type
) > 0);
803 c
->num_elements
= glsl_get_length(type
);
804 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
806 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
807 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
808 c
->elements
[i
] = c
->elements
[0];
811 case GLSL_TYPE_STRUCT
:
812 c
->num_elements
= glsl_get_length(type
);
813 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
815 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
816 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
821 unreachable("Invalid type for null constant");
828 spec_constant_deocoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
829 int member
, const struct vtn_decoration
*dec
,
832 assert(member
== -1);
833 if (dec
->decoration
!= SpvDecorationSpecId
)
836 uint32_t *const_value
= data
;
838 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
839 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
840 *const_value
= b
->specializations
[i
].data
;
847 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
848 uint32_t const_value
)
850 vtn_foreach_decoration(b
, val
, spec_constant_deocoration_cb
, &const_value
);
855 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
856 const uint32_t *w
, unsigned count
)
858 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
859 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
860 val
->constant
= rzalloc(b
, nir_constant
);
862 case SpvOpConstantTrue
:
863 assert(val
->const_type
== glsl_bool_type());
864 val
->constant
->value
.u
[0] = NIR_TRUE
;
866 case SpvOpConstantFalse
:
867 assert(val
->const_type
== glsl_bool_type());
868 val
->constant
->value
.u
[0] = NIR_FALSE
;
871 case SpvOpSpecConstantTrue
:
872 case SpvOpSpecConstantFalse
: {
873 assert(val
->const_type
== glsl_bool_type());
875 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
876 val
->constant
->value
.u
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
881 assert(glsl_type_is_scalar(val
->const_type
));
882 val
->constant
->value
.u
[0] = w
[3];
884 case SpvOpSpecConstant
:
885 assert(glsl_type_is_scalar(val
->const_type
));
886 val
->constant
->value
.u
[0] = get_specialization(b
, val
, w
[3]);
888 case SpvOpSpecConstantComposite
:
889 case SpvOpConstantComposite
: {
890 unsigned elem_count
= count
- 3;
891 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
892 for (unsigned i
= 0; i
< elem_count
; i
++)
893 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
895 switch (glsl_get_base_type(val
->const_type
)) {
898 case GLSL_TYPE_FLOAT
:
900 if (glsl_type_is_matrix(val
->const_type
)) {
901 unsigned rows
= glsl_get_vector_elements(val
->const_type
);
902 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
903 for (unsigned i
= 0; i
< elem_count
; i
++)
904 for (unsigned j
= 0; j
< rows
; j
++)
905 val
->constant
->value
.u
[rows
* i
+ j
] = elems
[i
]->value
.u
[j
];
907 assert(glsl_type_is_vector(val
->const_type
));
908 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
909 for (unsigned i
= 0; i
< elem_count
; i
++)
910 val
->constant
->value
.u
[i
] = elems
[i
]->value
.u
[0];
915 case GLSL_TYPE_STRUCT
:
916 case GLSL_TYPE_ARRAY
:
917 ralloc_steal(val
->constant
, elems
);
918 val
->constant
->num_elements
= elem_count
;
919 val
->constant
->elements
= elems
;
923 unreachable("Unsupported type for constants");
928 case SpvOpSpecConstantOp
: {
929 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
931 case SpvOpVectorShuffle
: {
932 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
933 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
934 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
935 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
938 for (unsigned i
= 0; i
< len0
; i
++)
939 u
[i
] = v0
->constant
->value
.u
[i
];
940 for (unsigned i
= 0; i
< len1
; i
++)
941 u
[len0
+ i
] = v1
->constant
->value
.u
[i
];
943 for (unsigned i
= 0; i
< count
- 6; i
++) {
944 uint32_t comp
= w
[i
+ 6];
945 if (comp
== (uint32_t)-1) {
946 val
->constant
->value
.u
[i
] = 0xdeadbeef;
948 val
->constant
->value
.u
[i
] = u
[comp
];
954 case SpvOpCompositeExtract
:
955 case SpvOpCompositeInsert
: {
956 struct vtn_value
*comp
;
957 unsigned deref_start
;
958 struct nir_constant
**c
;
959 if (opcode
== SpvOpCompositeExtract
) {
960 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
964 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
966 val
->constant
= nir_constant_clone(comp
->constant
,
972 const struct glsl_type
*type
= comp
->const_type
;
973 for (unsigned i
= deref_start
; i
< count
; i
++) {
974 switch (glsl_get_base_type(type
)) {
977 case GLSL_TYPE_FLOAT
:
979 /* If we hit this granularity, we're picking off an element */
983 if (glsl_type_is_matrix(type
)) {
984 elem
+= w
[i
] * glsl_get_vector_elements(type
);
985 type
= glsl_get_column_type(type
);
987 assert(glsl_type_is_vector(type
));
989 type
= glsl_scalar_type(glsl_get_base_type(type
));
993 case GLSL_TYPE_ARRAY
:
994 c
= &(*c
)->elements
[w
[i
]];
995 type
= glsl_get_array_element(type
);
998 case GLSL_TYPE_STRUCT
:
999 c
= &(*c
)->elements
[w
[i
]];
1000 type
= glsl_get_struct_field(type
, w
[i
]);
1004 unreachable("Invalid constant type");
1008 if (opcode
== SpvOpCompositeExtract
) {
1012 unsigned num_components
= glsl_get_vector_elements(type
);
1013 for (unsigned i
= 0; i
< num_components
; i
++)
1014 val
->constant
->value
.u
[i
] = (*c
)->value
.u
[elem
+ i
];
1017 struct vtn_value
*insert
=
1018 vtn_value(b
, w
[4], vtn_value_type_constant
);
1019 assert(insert
->const_type
== type
);
1021 *c
= insert
->constant
;
1023 unsigned num_components
= glsl_get_vector_elements(type
);
1024 for (unsigned i
= 0; i
< num_components
; i
++)
1025 (*c
)->value
.u
[elem
+ i
] = insert
->constant
->value
.u
[i
];
1033 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
);
1035 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1037 glsl_get_bit_size(glsl_get_base_type(val
->const_type
));
1039 nir_const_value src
[3];
1041 for (unsigned i
= 0; i
< count
- 4; i
++) {
1043 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1045 unsigned j
= swap
? 1 - i
: i
;
1046 assert(bit_size
== 32);
1047 for (unsigned k
= 0; k
< num_components
; k
++)
1048 src
[j
].u32
[k
] = c
->value
.u
[k
];
1051 nir_const_value res
= nir_eval_const_opcode(op
, num_components
,
1054 for (unsigned k
= 0; k
< num_components
; k
++)
1055 val
->constant
->value
.u
[k
] = res
.u32
[k
];
1062 case SpvOpConstantNull
:
1063 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1066 case SpvOpConstantSampler
:
1067 assert(!"OpConstantSampler requires Kernel Capability");
1071 unreachable("Unhandled opcode");
1076 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1077 const uint32_t *w
, unsigned count
)
1079 struct nir_function
*callee
=
1080 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1082 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1083 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1084 unsigned arg_id
= w
[4 + i
];
1085 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1086 if (arg
->value_type
== vtn_value_type_access_chain
) {
1087 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1088 call
->params
[i
] = nir_deref_as_var(nir_copy_deref(call
, &d
->deref
));
1090 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1092 /* Make a temporary to store the argument in */
1094 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1095 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1097 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1101 nir_variable
*out_tmp
= NULL
;
1102 if (!glsl_type_is_void(callee
->return_type
)) {
1103 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1105 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1108 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1110 if (glsl_type_is_void(callee
->return_type
)) {
1111 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1113 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1114 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1118 struct vtn_ssa_value
*
1119 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1121 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1124 if (!glsl_type_is_vector_or_scalar(type
)) {
1125 unsigned elems
= glsl_get_length(type
);
1126 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1127 for (unsigned i
= 0; i
< elems
; i
++) {
1128 const struct glsl_type
*child_type
;
1130 switch (glsl_get_base_type(type
)) {
1132 case GLSL_TYPE_UINT
:
1133 case GLSL_TYPE_BOOL
:
1134 case GLSL_TYPE_FLOAT
:
1135 case GLSL_TYPE_DOUBLE
:
1136 child_type
= glsl_get_column_type(type
);
1138 case GLSL_TYPE_ARRAY
:
1139 child_type
= glsl_get_array_element(type
);
1141 case GLSL_TYPE_STRUCT
:
1142 child_type
= glsl_get_struct_field(type
, i
);
1145 unreachable("unkown base type");
1148 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1156 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1159 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1160 src
.src_type
= type
;
1165 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1166 const uint32_t *w
, unsigned count
)
1168 if (opcode
== SpvOpSampledImage
) {
1169 struct vtn_value
*val
=
1170 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1171 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1172 val
->sampled_image
->image
=
1173 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1174 val
->sampled_image
->sampler
=
1175 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1177 } else if (opcode
== SpvOpImage
) {
1178 struct vtn_value
*val
=
1179 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1180 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1181 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1182 val
->access_chain
= src_val
->sampled_image
->image
;
1184 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1185 val
->access_chain
= src_val
->access_chain
;
1190 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1191 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1193 struct vtn_sampled_image sampled
;
1194 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1195 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1196 sampled
= *sampled_val
->sampled_image
;
1198 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1199 sampled
.image
= NULL
;
1200 sampled
.sampler
= sampled_val
->access_chain
;
1203 const struct glsl_type
*image_type
;
1204 if (sampled
.image
) {
1205 image_type
= sampled
.image
->var
->var
->interface_type
;
1207 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1210 nir_tex_src srcs
[8]; /* 8 should be enough */
1211 nir_tex_src
*p
= srcs
;
1215 bool has_coord
= false;
1217 case SpvOpImageSampleImplicitLod
:
1218 case SpvOpImageSampleExplicitLod
:
1219 case SpvOpImageSampleDrefImplicitLod
:
1220 case SpvOpImageSampleDrefExplicitLod
:
1221 case SpvOpImageSampleProjImplicitLod
:
1222 case SpvOpImageSampleProjExplicitLod
:
1223 case SpvOpImageSampleProjDrefImplicitLod
:
1224 case SpvOpImageSampleProjDrefExplicitLod
:
1225 case SpvOpImageFetch
:
1226 case SpvOpImageGather
:
1227 case SpvOpImageDrefGather
:
1228 case SpvOpImageQueryLod
: {
1229 /* All these types have the coordinate as their first real argument */
1230 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, w
[idx
++]);
1232 p
->src
= nir_src_for_ssa(coord
->def
);
1233 p
->src_type
= nir_tex_src_coord
;
1242 /* These all have an explicit depth value as their next source */
1244 case SpvOpImageSampleDrefImplicitLod
:
1245 case SpvOpImageSampleDrefExplicitLod
:
1246 case SpvOpImageSampleProjDrefImplicitLod
:
1247 case SpvOpImageSampleProjDrefExplicitLod
:
1248 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparitor
);
1254 /* For OpImageQuerySizeLod, we always have an LOD */
1255 if (opcode
== SpvOpImageQuerySizeLod
)
1256 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1258 /* Figure out the base texture operation */
1261 case SpvOpImageSampleImplicitLod
:
1262 case SpvOpImageSampleDrefImplicitLod
:
1263 case SpvOpImageSampleProjImplicitLod
:
1264 case SpvOpImageSampleProjDrefImplicitLod
:
1265 texop
= nir_texop_tex
;
1268 case SpvOpImageSampleExplicitLod
:
1269 case SpvOpImageSampleDrefExplicitLod
:
1270 case SpvOpImageSampleProjExplicitLod
:
1271 case SpvOpImageSampleProjDrefExplicitLod
:
1272 texop
= nir_texop_txl
;
1275 case SpvOpImageFetch
:
1276 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1277 texop
= nir_texop_txf_ms
;
1279 texop
= nir_texop_txf
;
1283 case SpvOpImageGather
:
1284 case SpvOpImageDrefGather
:
1285 texop
= nir_texop_tg4
;
1288 case SpvOpImageQuerySizeLod
:
1289 case SpvOpImageQuerySize
:
1290 texop
= nir_texop_txs
;
1293 case SpvOpImageQueryLod
:
1294 texop
= nir_texop_lod
;
1297 case SpvOpImageQueryLevels
:
1298 texop
= nir_texop_query_levels
;
1301 case SpvOpImageQuerySamples
:
1303 unreachable("Unhandled opcode");
1306 /* Now we need to handle some number of optional arguments */
1308 uint32_t operands
= w
[idx
++];
1310 if (operands
& SpvImageOperandsBiasMask
) {
1311 assert(texop
== nir_texop_tex
);
1312 texop
= nir_texop_txb
;
1313 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1316 if (operands
& SpvImageOperandsLodMask
) {
1317 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1318 texop
== nir_texop_txf_ms
|| texop
== nir_texop_txs
);
1319 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1322 if (operands
& SpvImageOperandsGradMask
) {
1323 assert(texop
== nir_texop_tex
);
1324 texop
= nir_texop_txd
;
1325 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1326 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1329 if (operands
& SpvImageOperandsOffsetMask
||
1330 operands
& SpvImageOperandsConstOffsetMask
)
1331 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1333 if (operands
& SpvImageOperandsConstOffsetsMask
)
1334 assert(!"Constant offsets to texture gather not yet implemented");
1336 if (operands
& SpvImageOperandsSampleMask
) {
1337 assert(texop
== nir_texop_txf_ms
);
1338 texop
= nir_texop_txf_ms
;
1339 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1342 /* We should have now consumed exactly all of the arguments */
1343 assert(idx
== count
);
1345 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1348 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1350 instr
->sampler_dim
= glsl_get_sampler_dim(image_type
);
1351 instr
->is_array
= glsl_sampler_type_is_array(image_type
);
1352 instr
->is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1353 instr
->is_new_style_shadow
= instr
->is_shadow
;
1356 switch (instr
->sampler_dim
) {
1357 case GLSL_SAMPLER_DIM_1D
:
1358 case GLSL_SAMPLER_DIM_BUF
:
1359 instr
->coord_components
= 1;
1361 case GLSL_SAMPLER_DIM_2D
:
1362 case GLSL_SAMPLER_DIM_RECT
:
1363 case GLSL_SAMPLER_DIM_MS
:
1364 instr
->coord_components
= 2;
1366 case GLSL_SAMPLER_DIM_3D
:
1367 case GLSL_SAMPLER_DIM_CUBE
:
1368 instr
->coord_components
= 3;
1371 assert("Invalid sampler type");
1374 if (instr
->is_array
)
1375 instr
->coord_components
++;
1377 instr
->coord_components
= 0;
1380 switch (glsl_get_sampler_result_type(image_type
)) {
1381 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1382 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1383 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1384 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1386 unreachable("Invalid base type for sampler result");
1389 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1390 if (sampled
.image
) {
1391 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1392 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &image
->deref
));
1394 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1397 switch (instr
->op
) {
1402 /* These operations require a sampler */
1403 instr
->sampler
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1406 case nir_texop_txf_ms
:
1410 case nir_texop_query_levels
:
1411 case nir_texop_texture_samples
:
1412 case nir_texop_samples_identical
:
1414 instr
->sampler
= NULL
;
1418 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1419 nir_tex_instr_dest_size(instr
), 32, NULL
);
1421 assert(glsl_get_vector_elements(ret_type
->type
) ==
1422 nir_tex_instr_dest_size(instr
));
1424 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1425 val
->ssa
->def
= &instr
->dest
.ssa
;
1427 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1430 static nir_ssa_def
*
1431 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1433 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1435 /* The image_load_store intrinsics assume a 4-dim coordinate */
1436 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1437 unsigned swizzle
[4];
1438 for (unsigned i
= 0; i
< 4; i
++)
1439 swizzle
[i
] = MIN2(i
, dim
- 1);
1441 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1445 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1446 const uint32_t *w
, unsigned count
)
1448 /* Just get this one out of the way */
1449 if (opcode
== SpvOpImageTexelPointer
) {
1450 struct vtn_value
*val
=
1451 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1452 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1455 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1456 val
->image
->coord
= get_image_coord(b
, w
[4]);
1457 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1461 struct vtn_image_pointer image
;
1464 case SpvOpAtomicExchange
:
1465 case SpvOpAtomicCompareExchange
:
1466 case SpvOpAtomicCompareExchangeWeak
:
1467 case SpvOpAtomicIIncrement
:
1468 case SpvOpAtomicIDecrement
:
1469 case SpvOpAtomicIAdd
:
1470 case SpvOpAtomicISub
:
1471 case SpvOpAtomicSMin
:
1472 case SpvOpAtomicUMin
:
1473 case SpvOpAtomicSMax
:
1474 case SpvOpAtomicUMax
:
1475 case SpvOpAtomicAnd
:
1477 case SpvOpAtomicXor
:
1478 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1481 case SpvOpImageQuerySize
:
1483 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1485 image
.sample
= NULL
;
1488 case SpvOpImageRead
:
1490 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1491 image
.coord
= get_image_coord(b
, w
[4]);
1493 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1494 assert(w
[5] == SpvImageOperandsSampleMask
);
1495 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1497 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1501 case SpvOpImageWrite
:
1503 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1504 image
.coord
= get_image_coord(b
, w
[2]);
1508 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1509 assert(w
[4] == SpvImageOperandsSampleMask
);
1510 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1512 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1517 unreachable("Invalid image opcode");
1520 nir_intrinsic_op op
;
1522 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1523 OP(ImageQuerySize
, size
)
1525 OP(ImageWrite
, store
)
1526 OP(AtomicExchange
, atomic_exchange
)
1527 OP(AtomicCompareExchange
, atomic_comp_swap
)
1528 OP(AtomicIIncrement
, atomic_add
)
1529 OP(AtomicIDecrement
, atomic_add
)
1530 OP(AtomicIAdd
, atomic_add
)
1531 OP(AtomicISub
, atomic_add
)
1532 OP(AtomicSMin
, atomic_min
)
1533 OP(AtomicUMin
, atomic_min
)
1534 OP(AtomicSMax
, atomic_max
)
1535 OP(AtomicUMax
, atomic_max
)
1536 OP(AtomicAnd
, atomic_and
)
1537 OP(AtomicOr
, atomic_or
)
1538 OP(AtomicXor
, atomic_xor
)
1541 unreachable("Invalid image opcode");
1544 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1546 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1547 intrin
->variables
[0] =
1548 nir_deref_as_var(nir_copy_deref(&intrin
->instr
, &image_deref
->deref
));
1550 /* ImageQuerySize doesn't take any extra parameters */
1551 if (opcode
!= SpvOpImageQuerySize
) {
1552 /* The image coordinate is always 4 components but we may not have that
1553 * many. Swizzle to compensate.
1556 for (unsigned i
= 0; i
< 4; i
++)
1557 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1558 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1560 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1564 case SpvOpImageQuerySize
:
1565 case SpvOpImageRead
:
1567 case SpvOpImageWrite
:
1568 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1570 case SpvOpAtomicIIncrement
:
1571 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1573 case SpvOpAtomicIDecrement
:
1574 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1577 case SpvOpAtomicExchange
:
1578 case SpvOpAtomicIAdd
:
1579 case SpvOpAtomicSMin
:
1580 case SpvOpAtomicUMin
:
1581 case SpvOpAtomicSMax
:
1582 case SpvOpAtomicUMax
:
1583 case SpvOpAtomicAnd
:
1585 case SpvOpAtomicXor
:
1586 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1589 case SpvOpAtomicCompareExchange
:
1590 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1591 intrin
->src
[3] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1594 case SpvOpAtomicISub
:
1595 intrin
->src
[2] = nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1599 unreachable("Invalid image opcode");
1602 if (opcode
!= SpvOpImageWrite
) {
1603 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1604 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1605 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, 32, NULL
);
1607 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1609 /* The image intrinsics always return 4 channels but we may not want
1610 * that many. Emit a mov to trim it down.
1612 unsigned swiz
[4] = {0, 1, 2, 3};
1613 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1614 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1615 glsl_get_vector_elements(type
->type
), false);
1617 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1621 static nir_intrinsic_op
1622 get_ssbo_nir_atomic_op(SpvOp opcode
)
1625 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1626 OP(AtomicExchange
, atomic_exchange
)
1627 OP(AtomicCompareExchange
, atomic_comp_swap
)
1628 OP(AtomicIIncrement
, atomic_add
)
1629 OP(AtomicIDecrement
, atomic_add
)
1630 OP(AtomicIAdd
, atomic_add
)
1631 OP(AtomicISub
, atomic_add
)
1632 OP(AtomicSMin
, atomic_imin
)
1633 OP(AtomicUMin
, atomic_umin
)
1634 OP(AtomicSMax
, atomic_imax
)
1635 OP(AtomicUMax
, atomic_umax
)
1636 OP(AtomicAnd
, atomic_and
)
1637 OP(AtomicOr
, atomic_or
)
1638 OP(AtomicXor
, atomic_xor
)
1641 unreachable("Invalid SSBO atomic");
1645 static nir_intrinsic_op
1646 get_shared_nir_atomic_op(SpvOp opcode
)
1649 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
1650 OP(AtomicExchange
, atomic_exchange
)
1651 OP(AtomicCompareExchange
, atomic_comp_swap
)
1652 OP(AtomicIIncrement
, atomic_add
)
1653 OP(AtomicIDecrement
, atomic_add
)
1654 OP(AtomicIAdd
, atomic_add
)
1655 OP(AtomicISub
, atomic_add
)
1656 OP(AtomicSMin
, atomic_imin
)
1657 OP(AtomicUMin
, atomic_umin
)
1658 OP(AtomicSMax
, atomic_imax
)
1659 OP(AtomicUMax
, atomic_umax
)
1660 OP(AtomicAnd
, atomic_and
)
1661 OP(AtomicOr
, atomic_or
)
1662 OP(AtomicXor
, atomic_xor
)
1665 unreachable("Invalid shared atomic");
1670 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1671 const uint32_t *w
, nir_src
*src
)
1674 case SpvOpAtomicIIncrement
:
1675 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1678 case SpvOpAtomicIDecrement
:
1679 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1682 case SpvOpAtomicISub
:
1684 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1687 case SpvOpAtomicCompareExchange
:
1688 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1689 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1693 case SpvOpAtomicExchange
:
1694 case SpvOpAtomicIAdd
:
1695 case SpvOpAtomicSMin
:
1696 case SpvOpAtomicUMin
:
1697 case SpvOpAtomicSMax
:
1698 case SpvOpAtomicUMax
:
1699 case SpvOpAtomicAnd
:
1701 case SpvOpAtomicXor
:
1702 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1706 unreachable("Invalid SPIR-V atomic");
1711 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
1712 const uint32_t *w
, unsigned count
)
1714 struct vtn_access_chain
*chain
=
1715 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1716 nir_intrinsic_instr
*atomic
;
1719 SpvScope scope = w[4];
1720 SpvMemorySemanticsMask semantics = w[5];
1723 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
1724 nir_deref
*deref
= &vtn_access_chain_to_deref(b
, chain
)->deref
;
1725 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
1726 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1727 atomic
->variables
[0] = nir_deref_as_var(nir_copy_deref(atomic
, deref
));
1728 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
1730 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
1731 struct vtn_type
*type
;
1732 nir_ssa_def
*offset
, *index
;
1733 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
1735 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
1737 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1738 atomic
->src
[0] = nir_src_for_ssa(index
);
1739 atomic
->src
[1] = nir_src_for_ssa(offset
);
1740 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
1743 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
, 1, 32, NULL
);
1745 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1746 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1747 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
1748 val
->ssa
->def
= &atomic
->dest
.ssa
;
1749 val
->ssa
->type
= type
->type
;
1751 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
1754 static nir_alu_instr
*
1755 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
1758 switch (num_components
) {
1759 case 1: op
= nir_op_fmov
; break;
1760 case 2: op
= nir_op_vec2
; break;
1761 case 3: op
= nir_op_vec3
; break;
1762 case 4: op
= nir_op_vec4
; break;
1763 default: unreachable("bad vector size");
1766 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
1767 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
1769 vec
->dest
.write_mask
= (1 << num_components
) - 1;
1774 struct vtn_ssa_value
*
1775 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
1777 if (src
->transposed
)
1778 return src
->transposed
;
1780 struct vtn_ssa_value
*dest
=
1781 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
1783 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
1784 nir_alu_instr
*vec
= create_vec(b
->shader
,
1785 glsl_get_matrix_columns(src
->type
),
1786 glsl_get_bit_size(glsl_get_base_type(src
->type
)));
1787 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1788 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
1789 vec
->src
[0].swizzle
[0] = i
;
1791 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
1792 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
1793 vec
->src
[j
].swizzle
[0] = i
;
1796 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1797 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
1800 dest
->transposed
= src
;
1806 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
1808 unsigned swiz
[4] = { index
};
1809 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
1813 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
1816 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
1819 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
1821 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
1823 vec
->src
[i
].src
= nir_src_for_ssa(src
);
1824 vec
->src
[i
].swizzle
[0] = i
;
1828 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1830 return &vec
->dest
.dest
.ssa
;
1834 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1837 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
1838 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1839 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1840 vtn_vector_extract(b
, src
, i
), dest
);
1846 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1847 nir_ssa_def
*insert
, nir_ssa_def
*index
)
1849 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
1850 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1851 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1852 vtn_vector_insert(b
, src
, insert
, i
), dest
);
1857 static nir_ssa_def
*
1858 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
1859 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
1860 const uint32_t *indices
)
1862 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
1864 for (unsigned i
= 0; i
< num_components
; i
++) {
1865 uint32_t index
= indices
[i
];
1866 if (index
== 0xffffffff) {
1868 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
1869 } else if (index
< src0
->num_components
) {
1870 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
1871 vec
->src
[i
].swizzle
[0] = index
;
1873 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
1874 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
1878 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1880 return &vec
->dest
.dest
.ssa
;
1884 * Concatentates a number of vectors/scalars together to produce a vector
1886 static nir_ssa_def
*
1887 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
1888 unsigned num_srcs
, nir_ssa_def
**srcs
)
1890 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
1893 unsigned dest_idx
= 0;
1894 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1895 nir_ssa_def
*src
= srcs
[i
];
1896 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
1897 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
1898 vec
->src
[dest_idx
].swizzle
[0] = j
;
1903 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1905 return &vec
->dest
.dest
.ssa
;
1908 static struct vtn_ssa_value
*
1909 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
1911 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
1912 dest
->type
= src
->type
;
1914 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1915 dest
->def
= src
->def
;
1917 unsigned elems
= glsl_get_length(src
->type
);
1919 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
1920 for (unsigned i
= 0; i
< elems
; i
++)
1921 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
1927 static struct vtn_ssa_value
*
1928 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1929 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
1930 unsigned num_indices
)
1932 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
1934 struct vtn_ssa_value
*cur
= dest
;
1936 for (i
= 0; i
< num_indices
- 1; i
++) {
1937 cur
= cur
->elems
[indices
[i
]];
1940 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1941 /* According to the SPIR-V spec, OpCompositeInsert may work down to
1942 * the component granularity. In that case, the last index will be
1943 * the index to insert the scalar into the vector.
1946 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
1948 cur
->elems
[indices
[i
]] = insert
;
1954 static struct vtn_ssa_value
*
1955 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1956 const uint32_t *indices
, unsigned num_indices
)
1958 struct vtn_ssa_value
*cur
= src
;
1959 for (unsigned i
= 0; i
< num_indices
; i
++) {
1960 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1961 assert(i
== num_indices
- 1);
1962 /* According to the SPIR-V spec, OpCompositeExtract may work down to
1963 * the component granularity. The last index will be the index of the
1964 * vector to extract.
1967 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
1968 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
1969 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
1972 cur
= cur
->elems
[indices
[i
]];
1980 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
1981 const uint32_t *w
, unsigned count
)
1983 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1984 const struct glsl_type
*type
=
1985 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1986 val
->ssa
= vtn_create_ssa_value(b
, type
);
1989 case SpvOpVectorExtractDynamic
:
1990 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1991 vtn_ssa_value(b
, w
[4])->def
);
1994 case SpvOpVectorInsertDynamic
:
1995 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1996 vtn_ssa_value(b
, w
[4])->def
,
1997 vtn_ssa_value(b
, w
[5])->def
);
2000 case SpvOpVectorShuffle
:
2001 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2002 vtn_ssa_value(b
, w
[3])->def
,
2003 vtn_ssa_value(b
, w
[4])->def
,
2007 case SpvOpCompositeConstruct
: {
2008 unsigned elems
= count
- 3;
2009 if (glsl_type_is_vector_or_scalar(type
)) {
2010 nir_ssa_def
*srcs
[4];
2011 for (unsigned i
= 0; i
< elems
; i
++)
2012 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2014 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2017 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2018 for (unsigned i
= 0; i
< elems
; i
++)
2019 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2023 case SpvOpCompositeExtract
:
2024 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2028 case SpvOpCompositeInsert
:
2029 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2030 vtn_ssa_value(b
, w
[3]),
2034 case SpvOpCopyObject
:
2035 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2039 unreachable("unknown composite operation");
2044 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2045 const uint32_t *w
, unsigned count
)
2047 nir_intrinsic_op intrinsic_op
;
2049 case SpvOpEmitVertex
:
2050 case SpvOpEmitStreamVertex
:
2051 intrinsic_op
= nir_intrinsic_emit_vertex
;
2053 case SpvOpEndPrimitive
:
2054 case SpvOpEndStreamPrimitive
:
2055 intrinsic_op
= nir_intrinsic_end_primitive
;
2057 case SpvOpMemoryBarrier
:
2058 intrinsic_op
= nir_intrinsic_memory_barrier
;
2060 case SpvOpControlBarrier
:
2061 intrinsic_op
= nir_intrinsic_barrier
;
2064 unreachable("unknown barrier instruction");
2067 nir_intrinsic_instr
*intrin
=
2068 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2070 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2071 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2073 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2077 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2080 case SpvExecutionModeInputPoints
:
2081 case SpvExecutionModeOutputPoints
:
2082 return 0; /* GL_POINTS */
2083 case SpvExecutionModeInputLines
:
2084 return 1; /* GL_LINES */
2085 case SpvExecutionModeInputLinesAdjacency
:
2086 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2087 case SpvExecutionModeTriangles
:
2088 return 4; /* GL_TRIANGLES */
2089 case SpvExecutionModeInputTrianglesAdjacency
:
2090 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2091 case SpvExecutionModeQuads
:
2092 return 7; /* GL_QUADS */
2093 case SpvExecutionModeIsolines
:
2094 return 0x8E7A; /* GL_ISOLINES */
2095 case SpvExecutionModeOutputLineStrip
:
2096 return 3; /* GL_LINE_STRIP */
2097 case SpvExecutionModeOutputTriangleStrip
:
2098 return 5; /* GL_TRIANGLE_STRIP */
2100 assert(!"Invalid primitive type");
2106 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2109 case SpvExecutionModeInputPoints
:
2111 case SpvExecutionModeInputLines
:
2113 case SpvExecutionModeInputLinesAdjacency
:
2115 case SpvExecutionModeTriangles
:
2117 case SpvExecutionModeInputTrianglesAdjacency
:
2120 assert(!"Invalid GS input mode");
2125 static gl_shader_stage
2126 stage_for_execution_model(SpvExecutionModel model
)
2129 case SpvExecutionModelVertex
:
2130 return MESA_SHADER_VERTEX
;
2131 case SpvExecutionModelTessellationControl
:
2132 return MESA_SHADER_TESS_CTRL
;
2133 case SpvExecutionModelTessellationEvaluation
:
2134 return MESA_SHADER_TESS_EVAL
;
2135 case SpvExecutionModelGeometry
:
2136 return MESA_SHADER_GEOMETRY
;
2137 case SpvExecutionModelFragment
:
2138 return MESA_SHADER_FRAGMENT
;
2139 case SpvExecutionModelGLCompute
:
2140 return MESA_SHADER_COMPUTE
;
2142 unreachable("Unsupported execution model");
2147 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2148 const uint32_t *w
, unsigned count
)
2152 case SpvOpSourceExtension
:
2153 case SpvOpSourceContinued
:
2154 case SpvOpExtension
:
2155 /* Unhandled, but these are for debug so that's ok. */
2158 case SpvOpCapability
: {
2159 SpvCapability cap
= w
[1];
2161 case SpvCapabilityMatrix
:
2162 case SpvCapabilityShader
:
2163 case SpvCapabilityGeometry
:
2164 case SpvCapabilityTessellationPointSize
:
2165 case SpvCapabilityGeometryPointSize
:
2166 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2167 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2168 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2169 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2170 case SpvCapabilityImageRect
:
2171 case SpvCapabilitySampledRect
:
2172 case SpvCapabilitySampled1D
:
2173 case SpvCapabilityImage1D
:
2174 case SpvCapabilitySampledCubeArray
:
2175 case SpvCapabilitySampledBuffer
:
2176 case SpvCapabilityImageBuffer
:
2177 case SpvCapabilityImageQuery
:
2179 case SpvCapabilityClipDistance
:
2180 case SpvCapabilityCullDistance
:
2181 case SpvCapabilityGeometryStreams
:
2182 fprintf(stderr
, "WARNING: Unsupported SPIR-V Capability\n");
2185 assert(!"Unsupported capability");
2190 case SpvOpExtInstImport
:
2191 vtn_handle_extension(b
, opcode
, w
, count
);
2194 case SpvOpMemoryModel
:
2195 assert(w
[1] == SpvAddressingModelLogical
);
2196 assert(w
[2] == SpvMemoryModelGLSL450
);
2199 case SpvOpEntryPoint
: {
2200 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2201 /* Let this be a name label regardless */
2202 unsigned name_words
;
2203 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2205 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2206 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2209 assert(b
->entry_point
== NULL
);
2210 b
->entry_point
= entry_point
;
2215 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2216 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2220 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2223 case SpvOpMemberName
:
2227 case SpvOpExecutionMode
:
2228 case SpvOpDecorationGroup
:
2230 case SpvOpMemberDecorate
:
2231 case SpvOpGroupDecorate
:
2232 case SpvOpGroupMemberDecorate
:
2233 vtn_handle_decoration(b
, opcode
, w
, count
);
2237 return false; /* End of preamble */
2244 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2245 const struct vtn_decoration
*mode
, void *data
)
2247 assert(b
->entry_point
== entry_point
);
2249 switch(mode
->exec_mode
) {
2250 case SpvExecutionModeOriginUpperLeft
:
2251 case SpvExecutionModeOriginLowerLeft
:
2252 b
->origin_upper_left
=
2253 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2256 case SpvExecutionModeEarlyFragmentTests
:
2257 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2258 b
->shader
->info
.fs
.early_fragment_tests
= true;
2261 case SpvExecutionModeInvocations
:
2262 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2263 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2266 case SpvExecutionModeDepthReplacing
:
2267 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2268 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2270 case SpvExecutionModeDepthGreater
:
2271 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2272 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2274 case SpvExecutionModeDepthLess
:
2275 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2276 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2278 case SpvExecutionModeDepthUnchanged
:
2279 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2280 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2283 case SpvExecutionModeLocalSize
:
2284 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2285 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2286 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2287 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2289 case SpvExecutionModeLocalSizeHint
:
2290 break; /* Nothing do do with this */
2292 case SpvExecutionModeOutputVertices
:
2293 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2294 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2297 case SpvExecutionModeInputPoints
:
2298 case SpvExecutionModeInputLines
:
2299 case SpvExecutionModeInputLinesAdjacency
:
2300 case SpvExecutionModeTriangles
:
2301 case SpvExecutionModeInputTrianglesAdjacency
:
2302 case SpvExecutionModeQuads
:
2303 case SpvExecutionModeIsolines
:
2304 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2305 b
->shader
->info
.gs
.vertices_in
=
2306 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2308 assert(!"Tesselation shaders not yet supported");
2312 case SpvExecutionModeOutputPoints
:
2313 case SpvExecutionModeOutputLineStrip
:
2314 case SpvExecutionModeOutputTriangleStrip
:
2315 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2316 b
->shader
->info
.gs
.output_primitive
=
2317 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2320 case SpvExecutionModeSpacingEqual
:
2321 case SpvExecutionModeSpacingFractionalEven
:
2322 case SpvExecutionModeSpacingFractionalOdd
:
2323 case SpvExecutionModeVertexOrderCw
:
2324 case SpvExecutionModeVertexOrderCcw
:
2325 case SpvExecutionModePointMode
:
2326 assert(!"TODO: Add tessellation metadata");
2329 case SpvExecutionModePixelCenterInteger
:
2330 case SpvExecutionModeXfb
:
2331 assert(!"Unhandled execution mode");
2334 case SpvExecutionModeVecTypeHint
:
2335 case SpvExecutionModeContractionOff
:
2341 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2342 const uint32_t *w
, unsigned count
)
2346 case SpvOpSourceContinued
:
2347 case SpvOpSourceExtension
:
2348 case SpvOpExtension
:
2349 case SpvOpCapability
:
2350 case SpvOpExtInstImport
:
2351 case SpvOpMemoryModel
:
2352 case SpvOpEntryPoint
:
2353 case SpvOpExecutionMode
:
2356 case SpvOpMemberName
:
2357 case SpvOpDecorationGroup
:
2359 case SpvOpMemberDecorate
:
2360 case SpvOpGroupDecorate
:
2361 case SpvOpGroupMemberDecorate
:
2362 assert(!"Invalid opcode types and variables section");
2368 case SpvOpTypeFloat
:
2369 case SpvOpTypeVector
:
2370 case SpvOpTypeMatrix
:
2371 case SpvOpTypeImage
:
2372 case SpvOpTypeSampler
:
2373 case SpvOpTypeSampledImage
:
2374 case SpvOpTypeArray
:
2375 case SpvOpTypeRuntimeArray
:
2376 case SpvOpTypeStruct
:
2377 case SpvOpTypeOpaque
:
2378 case SpvOpTypePointer
:
2379 case SpvOpTypeFunction
:
2380 case SpvOpTypeEvent
:
2381 case SpvOpTypeDeviceEvent
:
2382 case SpvOpTypeReserveId
:
2383 case SpvOpTypeQueue
:
2385 vtn_handle_type(b
, opcode
, w
, count
);
2388 case SpvOpConstantTrue
:
2389 case SpvOpConstantFalse
:
2391 case SpvOpConstantComposite
:
2392 case SpvOpConstantSampler
:
2393 case SpvOpConstantNull
:
2394 case SpvOpSpecConstantTrue
:
2395 case SpvOpSpecConstantFalse
:
2396 case SpvOpSpecConstant
:
2397 case SpvOpSpecConstantComposite
:
2398 case SpvOpSpecConstantOp
:
2399 vtn_handle_constant(b
, opcode
, w
, count
);
2403 vtn_handle_variables(b
, opcode
, w
, count
);
2407 return false; /* End of preamble */
2414 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2415 const uint32_t *w
, unsigned count
)
2421 case SpvOpLoopMerge
:
2422 case SpvOpSelectionMerge
:
2423 /* This is handled by cfg pre-pass and walk_blocks */
2427 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2428 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2433 vtn_handle_extension(b
, opcode
, w
, count
);
2439 case SpvOpCopyMemory
:
2440 case SpvOpCopyMemorySized
:
2441 case SpvOpAccessChain
:
2442 case SpvOpInBoundsAccessChain
:
2443 case SpvOpArrayLength
:
2444 vtn_handle_variables(b
, opcode
, w
, count
);
2447 case SpvOpFunctionCall
:
2448 vtn_handle_function_call(b
, opcode
, w
, count
);
2451 case SpvOpSampledImage
:
2453 case SpvOpImageSampleImplicitLod
:
2454 case SpvOpImageSampleExplicitLod
:
2455 case SpvOpImageSampleDrefImplicitLod
:
2456 case SpvOpImageSampleDrefExplicitLod
:
2457 case SpvOpImageSampleProjImplicitLod
:
2458 case SpvOpImageSampleProjExplicitLod
:
2459 case SpvOpImageSampleProjDrefImplicitLod
:
2460 case SpvOpImageSampleProjDrefExplicitLod
:
2461 case SpvOpImageFetch
:
2462 case SpvOpImageGather
:
2463 case SpvOpImageDrefGather
:
2464 case SpvOpImageQuerySizeLod
:
2465 case SpvOpImageQueryLod
:
2466 case SpvOpImageQueryLevels
:
2467 case SpvOpImageQuerySamples
:
2468 vtn_handle_texture(b
, opcode
, w
, count
);
2471 case SpvOpImageRead
:
2472 case SpvOpImageWrite
:
2473 case SpvOpImageTexelPointer
:
2474 vtn_handle_image(b
, opcode
, w
, count
);
2477 case SpvOpImageQuerySize
: {
2478 struct vtn_access_chain
*image
=
2479 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2480 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2481 vtn_handle_image(b
, opcode
, w
, count
);
2483 vtn_handle_texture(b
, opcode
, w
, count
);
2488 case SpvOpAtomicExchange
:
2489 case SpvOpAtomicCompareExchange
:
2490 case SpvOpAtomicCompareExchangeWeak
:
2491 case SpvOpAtomicIIncrement
:
2492 case SpvOpAtomicIDecrement
:
2493 case SpvOpAtomicIAdd
:
2494 case SpvOpAtomicISub
:
2495 case SpvOpAtomicSMin
:
2496 case SpvOpAtomicUMin
:
2497 case SpvOpAtomicSMax
:
2498 case SpvOpAtomicUMax
:
2499 case SpvOpAtomicAnd
:
2501 case SpvOpAtomicXor
: {
2502 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2503 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2504 vtn_handle_image(b
, opcode
, w
, count
);
2506 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2507 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2517 case SpvOpConvertFToU
:
2518 case SpvOpConvertFToS
:
2519 case SpvOpConvertSToF
:
2520 case SpvOpConvertUToF
:
2524 case SpvOpQuantizeToF16
:
2525 case SpvOpConvertPtrToU
:
2526 case SpvOpConvertUToPtr
:
2527 case SpvOpPtrCastToGeneric
:
2528 case SpvOpGenericCastToPtr
:
2534 case SpvOpSignBitSet
:
2535 case SpvOpLessOrGreater
:
2537 case SpvOpUnordered
:
2552 case SpvOpVectorTimesScalar
:
2554 case SpvOpIAddCarry
:
2555 case SpvOpISubBorrow
:
2556 case SpvOpUMulExtended
:
2557 case SpvOpSMulExtended
:
2558 case SpvOpShiftRightLogical
:
2559 case SpvOpShiftRightArithmetic
:
2560 case SpvOpShiftLeftLogical
:
2561 case SpvOpLogicalEqual
:
2562 case SpvOpLogicalNotEqual
:
2563 case SpvOpLogicalOr
:
2564 case SpvOpLogicalAnd
:
2565 case SpvOpLogicalNot
:
2566 case SpvOpBitwiseOr
:
2567 case SpvOpBitwiseXor
:
2568 case SpvOpBitwiseAnd
:
2571 case SpvOpFOrdEqual
:
2572 case SpvOpFUnordEqual
:
2573 case SpvOpINotEqual
:
2574 case SpvOpFOrdNotEqual
:
2575 case SpvOpFUnordNotEqual
:
2576 case SpvOpULessThan
:
2577 case SpvOpSLessThan
:
2578 case SpvOpFOrdLessThan
:
2579 case SpvOpFUnordLessThan
:
2580 case SpvOpUGreaterThan
:
2581 case SpvOpSGreaterThan
:
2582 case SpvOpFOrdGreaterThan
:
2583 case SpvOpFUnordGreaterThan
:
2584 case SpvOpULessThanEqual
:
2585 case SpvOpSLessThanEqual
:
2586 case SpvOpFOrdLessThanEqual
:
2587 case SpvOpFUnordLessThanEqual
:
2588 case SpvOpUGreaterThanEqual
:
2589 case SpvOpSGreaterThanEqual
:
2590 case SpvOpFOrdGreaterThanEqual
:
2591 case SpvOpFUnordGreaterThanEqual
:
2597 case SpvOpFwidthFine
:
2598 case SpvOpDPdxCoarse
:
2599 case SpvOpDPdyCoarse
:
2600 case SpvOpFwidthCoarse
:
2601 case SpvOpBitFieldInsert
:
2602 case SpvOpBitFieldSExtract
:
2603 case SpvOpBitFieldUExtract
:
2604 case SpvOpBitReverse
:
2606 case SpvOpTranspose
:
2607 case SpvOpOuterProduct
:
2608 case SpvOpMatrixTimesScalar
:
2609 case SpvOpVectorTimesMatrix
:
2610 case SpvOpMatrixTimesVector
:
2611 case SpvOpMatrixTimesMatrix
:
2612 vtn_handle_alu(b
, opcode
, w
, count
);
2615 case SpvOpVectorExtractDynamic
:
2616 case SpvOpVectorInsertDynamic
:
2617 case SpvOpVectorShuffle
:
2618 case SpvOpCompositeConstruct
:
2619 case SpvOpCompositeExtract
:
2620 case SpvOpCompositeInsert
:
2621 case SpvOpCopyObject
:
2622 vtn_handle_composite(b
, opcode
, w
, count
);
2625 case SpvOpEmitVertex
:
2626 case SpvOpEndPrimitive
:
2627 case SpvOpEmitStreamVertex
:
2628 case SpvOpEndStreamPrimitive
:
2629 case SpvOpControlBarrier
:
2630 case SpvOpMemoryBarrier
:
2631 vtn_handle_barrier(b
, opcode
, w
, count
);
2635 unreachable("Unhandled opcode");
2642 spirv_to_nir(const uint32_t *words
, size_t word_count
,
2643 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
2644 gl_shader_stage stage
, const char *entry_point_name
,
2645 const nir_shader_compiler_options
*options
)
2647 const uint32_t *word_end
= words
+ word_count
;
2649 /* Handle the SPIR-V header (first 4 dwords) */
2650 assert(word_count
> 5);
2652 assert(words
[0] == SpvMagicNumber
);
2653 assert(words
[1] >= 0x10000);
2654 /* words[2] == generator magic */
2655 unsigned value_id_bound
= words
[3];
2656 assert(words
[4] == 0);
2660 /* Initialize the stn_builder object */
2661 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
2662 b
->value_id_bound
= value_id_bound
;
2663 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
2664 exec_list_make_empty(&b
->functions
);
2665 b
->entry_point_stage
= stage
;
2666 b
->entry_point_name
= entry_point_name
;
2668 /* Handle all the preamble instructions */
2669 words
= vtn_foreach_instruction(b
, words
, word_end
,
2670 vtn_handle_preamble_instruction
);
2672 if (b
->entry_point
== NULL
) {
2673 assert(!"Entry point not found");
2678 b
->shader
= nir_shader_create(NULL
, stage
, options
);
2680 /* Set shader info defaults */
2681 b
->shader
->info
.gs
.invocations
= 1;
2683 /* Parse execution modes */
2684 vtn_foreach_execution_mode(b
, b
->entry_point
,
2685 vtn_handle_execution_mode
, NULL
);
2687 b
->specializations
= spec
;
2688 b
->num_specializations
= num_spec
;
2690 /* Handle all variable, type, and constant instructions */
2691 words
= vtn_foreach_instruction(b
, words
, word_end
,
2692 vtn_handle_variable_or_type_instruction
);
2694 vtn_build_cfg(b
, words
, word_end
);
2696 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
2697 b
->impl
= func
->impl
;
2698 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
2699 _mesa_key_pointer_equal
);
2701 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
2704 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
2705 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
2706 assert(entry_point
);