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 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 const bool signedness
= w
[3];
599 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
603 val
->type
->type
= glsl_float_type();
606 case SpvOpTypeVector
: {
607 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
608 unsigned elems
= w
[3];
610 assert(glsl_type_is_scalar(base
->type
));
611 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
613 /* Vectors implicitly have sizeof(base_type) stride. For now, this
614 * is always 4 bytes. This will have to change if we want to start
615 * supporting doubles or half-floats.
617 val
->type
->stride
= 4;
618 val
->type
->array_element
= base
;
622 case SpvOpTypeMatrix
: {
623 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
624 unsigned columns
= w
[3];
626 assert(glsl_type_is_vector(base
->type
));
627 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
628 glsl_get_vector_elements(base
->type
),
630 assert(!glsl_type_is_error(val
->type
->type
));
631 val
->type
->array_element
= base
;
632 val
->type
->row_major
= false;
633 val
->type
->stride
= 0;
637 case SpvOpTypeRuntimeArray
:
638 case SpvOpTypeArray
: {
639 struct vtn_type
*array_element
=
640 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
643 if (opcode
== SpvOpTypeRuntimeArray
) {
644 /* A length of 0 is used to denote unsized arrays */
648 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->value
.u
[0];
651 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
652 val
->type
->array_element
= array_element
;
653 val
->type
->stride
= 0;
657 case SpvOpTypeStruct
: {
658 unsigned num_fields
= count
- 2;
659 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
660 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
662 NIR_VLA(struct glsl_struct_field
, fields
, count
);
663 for (unsigned i
= 0; i
< num_fields
; i
++) {
664 val
->type
->members
[i
] =
665 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
666 fields
[i
] = (struct glsl_struct_field
) {
667 .type
= val
->type
->members
[i
]->type
,
668 .name
= ralloc_asprintf(b
, "field%d", i
),
673 struct member_decoration_ctx ctx
= {
674 .num_fields
= num_fields
,
679 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
681 const char *name
= val
->name
? val
->name
: "struct";
683 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
687 case SpvOpTypeFunction
: {
688 const struct glsl_type
*return_type
=
689 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
690 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
691 for (unsigned i
= 0; i
< count
- 3; i
++) {
692 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
696 params
[i
].out
= true;
698 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
702 case SpvOpTypePointer
:
703 /* FIXME: For now, we'll just do the really lame thing and return
704 * the same type. The validator should ensure that the proper number
705 * of dereferences happen
707 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
710 case SpvOpTypeImage
: {
711 const struct glsl_type
*sampled_type
=
712 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
714 assert(glsl_type_is_vector_or_scalar(sampled_type
));
716 enum glsl_sampler_dim dim
;
717 switch ((SpvDim
)w
[3]) {
718 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
719 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
720 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
721 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
722 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
723 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
725 unreachable("Invalid SPIR-V Sampler dimension");
728 bool is_shadow
= w
[4];
729 bool is_array
= w
[5];
730 bool multisampled
= w
[6];
731 unsigned sampled
= w
[7];
732 SpvImageFormat format
= w
[8];
735 val
->type
->access_qualifier
= w
[9];
737 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
740 assert(dim
== GLSL_SAMPLER_DIM_2D
);
741 dim
= GLSL_SAMPLER_DIM_MS
;
744 val
->type
->image_format
= translate_image_format(format
);
747 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
748 glsl_get_base_type(sampled_type
));
749 } else if (sampled
== 2) {
752 val
->type
->type
= glsl_image_type(dim
, is_array
,
753 glsl_get_base_type(sampled_type
));
755 assert(!"We need to know if the image will be sampled");
760 case SpvOpTypeSampledImage
:
761 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
764 case SpvOpTypeSampler
:
765 /* The actual sampler type here doesn't really matter. It gets
766 * thrown away the moment you combine it with an image. What really
767 * matters is that it's a sampler type as opposed to an integer type
768 * so the backend knows what to do.
770 val
->type
->type
= glsl_bare_sampler_type();
773 case SpvOpTypeOpaque
:
775 case SpvOpTypeDeviceEvent
:
776 case SpvOpTypeReserveId
:
780 unreachable("Unhandled opcode");
783 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
786 static nir_constant
*
787 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
789 nir_constant
*c
= rzalloc(b
, nir_constant
);
791 switch (glsl_get_base_type(type
)) {
795 case GLSL_TYPE_FLOAT
:
796 case GLSL_TYPE_DOUBLE
:
797 /* Nothing to do here. It's already initialized to zero */
800 case GLSL_TYPE_ARRAY
:
801 assert(glsl_get_length(type
) > 0);
802 c
->num_elements
= glsl_get_length(type
);
803 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
805 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
806 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
807 c
->elements
[i
] = c
->elements
[0];
810 case GLSL_TYPE_STRUCT
:
811 c
->num_elements
= glsl_get_length(type
);
812 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
814 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
815 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
820 unreachable("Invalid type for null constant");
827 spec_constant_deocoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
828 int member
, const struct vtn_decoration
*dec
,
831 assert(member
== -1);
832 if (dec
->decoration
!= SpvDecorationSpecId
)
835 uint32_t *const_value
= data
;
837 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
838 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
839 *const_value
= b
->specializations
[i
].data
;
846 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
847 uint32_t const_value
)
849 vtn_foreach_decoration(b
, val
, spec_constant_deocoration_cb
, &const_value
);
854 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
855 const uint32_t *w
, unsigned count
)
857 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
858 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
859 val
->constant
= rzalloc(b
, nir_constant
);
861 case SpvOpConstantTrue
:
862 assert(val
->const_type
== glsl_bool_type());
863 val
->constant
->value
.u
[0] = NIR_TRUE
;
865 case SpvOpConstantFalse
:
866 assert(val
->const_type
== glsl_bool_type());
867 val
->constant
->value
.u
[0] = NIR_FALSE
;
870 case SpvOpSpecConstantTrue
:
871 case SpvOpSpecConstantFalse
: {
872 assert(val
->const_type
== glsl_bool_type());
874 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
875 val
->constant
->value
.u
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
880 assert(glsl_type_is_scalar(val
->const_type
));
881 val
->constant
->value
.u
[0] = w
[3];
883 case SpvOpSpecConstant
:
884 assert(glsl_type_is_scalar(val
->const_type
));
885 val
->constant
->value
.u
[0] = get_specialization(b
, val
, w
[3]);
887 case SpvOpSpecConstantComposite
:
888 case SpvOpConstantComposite
: {
889 unsigned elem_count
= count
- 3;
890 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
891 for (unsigned i
= 0; i
< elem_count
; i
++)
892 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
894 switch (glsl_get_base_type(val
->const_type
)) {
897 case GLSL_TYPE_FLOAT
:
899 if (glsl_type_is_matrix(val
->const_type
)) {
900 unsigned rows
= glsl_get_vector_elements(val
->const_type
);
901 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
902 for (unsigned i
= 0; i
< elem_count
; i
++)
903 for (unsigned j
= 0; j
< rows
; j
++)
904 val
->constant
->value
.u
[rows
* i
+ j
] = elems
[i
]->value
.u
[j
];
906 assert(glsl_type_is_vector(val
->const_type
));
907 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
908 for (unsigned i
= 0; i
< elem_count
; i
++)
909 val
->constant
->value
.u
[i
] = elems
[i
]->value
.u
[0];
914 case GLSL_TYPE_STRUCT
:
915 case GLSL_TYPE_ARRAY
:
916 ralloc_steal(val
->constant
, elems
);
917 val
->constant
->num_elements
= elem_count
;
918 val
->constant
->elements
= elems
;
922 unreachable("Unsupported type for constants");
927 case SpvOpSpecConstantOp
: {
928 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
930 case SpvOpVectorShuffle
: {
931 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
932 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
933 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
934 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
937 for (unsigned i
= 0; i
< len0
; i
++)
938 u
[i
] = v0
->constant
->value
.u
[i
];
939 for (unsigned i
= 0; i
< len1
; i
++)
940 u
[len0
+ i
] = v1
->constant
->value
.u
[i
];
942 for (unsigned i
= 0; i
< count
- 6; i
++) {
943 uint32_t comp
= w
[i
+ 6];
944 if (comp
== (uint32_t)-1) {
945 val
->constant
->value
.u
[i
] = 0xdeadbeef;
947 val
->constant
->value
.u
[i
] = u
[comp
];
953 case SpvOpCompositeExtract
:
954 case SpvOpCompositeInsert
: {
955 struct vtn_value
*comp
;
956 unsigned deref_start
;
957 struct nir_constant
**c
;
958 if (opcode
== SpvOpCompositeExtract
) {
959 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
963 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
965 val
->constant
= nir_constant_clone(comp
->constant
,
971 const struct glsl_type
*type
= comp
->const_type
;
972 for (unsigned i
= deref_start
; i
< count
; i
++) {
973 switch (glsl_get_base_type(type
)) {
976 case GLSL_TYPE_FLOAT
:
978 /* If we hit this granularity, we're picking off an element */
982 if (glsl_type_is_matrix(type
)) {
983 elem
+= w
[i
] * glsl_get_vector_elements(type
);
984 type
= glsl_get_column_type(type
);
986 assert(glsl_type_is_vector(type
));
988 type
= glsl_scalar_type(glsl_get_base_type(type
));
992 case GLSL_TYPE_ARRAY
:
993 c
= &(*c
)->elements
[w
[i
]];
994 type
= glsl_get_array_element(type
);
997 case GLSL_TYPE_STRUCT
:
998 c
= &(*c
)->elements
[w
[i
]];
999 type
= glsl_get_struct_field(type
, w
[i
]);
1003 unreachable("Invalid constant type");
1007 if (opcode
== SpvOpCompositeExtract
) {
1011 unsigned num_components
= glsl_get_vector_elements(type
);
1012 for (unsigned i
= 0; i
< num_components
; i
++)
1013 val
->constant
->value
.u
[i
] = (*c
)->value
.u
[elem
+ i
];
1016 struct vtn_value
*insert
=
1017 vtn_value(b
, w
[4], vtn_value_type_constant
);
1018 assert(insert
->const_type
== type
);
1020 *c
= insert
->constant
;
1022 unsigned num_components
= glsl_get_vector_elements(type
);
1023 for (unsigned i
= 0; i
< num_components
; i
++)
1024 (*c
)->value
.u
[elem
+ i
] = insert
->constant
->value
.u
[i
];
1032 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
);
1034 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1036 glsl_get_bit_size(glsl_get_base_type(val
->const_type
));
1038 nir_const_value src
[3];
1040 for (unsigned i
= 0; i
< count
- 4; i
++) {
1042 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1044 unsigned j
= swap
? 1 - i
: i
;
1045 assert(bit_size
== 32);
1046 for (unsigned k
= 0; k
< num_components
; k
++)
1047 src
[j
].u32
[k
] = c
->value
.u
[k
];
1050 nir_const_value res
= nir_eval_const_opcode(op
, num_components
,
1053 for (unsigned k
= 0; k
< num_components
; k
++)
1054 val
->constant
->value
.u
[k
] = res
.u32
[k
];
1061 case SpvOpConstantNull
:
1062 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1065 case SpvOpConstantSampler
:
1066 assert(!"OpConstantSampler requires Kernel Capability");
1070 unreachable("Unhandled opcode");
1075 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1076 const uint32_t *w
, unsigned count
)
1078 struct nir_function
*callee
=
1079 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1081 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1082 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1083 unsigned arg_id
= w
[4 + i
];
1084 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1085 if (arg
->value_type
== vtn_value_type_access_chain
) {
1086 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1087 call
->params
[i
] = nir_deref_as_var(nir_copy_deref(call
, &d
->deref
));
1089 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1091 /* Make a temporary to store the argument in */
1093 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1094 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1096 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1100 nir_variable
*out_tmp
= NULL
;
1101 if (!glsl_type_is_void(callee
->return_type
)) {
1102 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1104 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1107 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1109 if (glsl_type_is_void(callee
->return_type
)) {
1110 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1112 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1113 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1117 struct vtn_ssa_value
*
1118 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1120 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1123 if (!glsl_type_is_vector_or_scalar(type
)) {
1124 unsigned elems
= glsl_get_length(type
);
1125 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1126 for (unsigned i
= 0; i
< elems
; i
++) {
1127 const struct glsl_type
*child_type
;
1129 switch (glsl_get_base_type(type
)) {
1131 case GLSL_TYPE_UINT
:
1132 case GLSL_TYPE_BOOL
:
1133 case GLSL_TYPE_FLOAT
:
1134 case GLSL_TYPE_DOUBLE
:
1135 child_type
= glsl_get_column_type(type
);
1137 case GLSL_TYPE_ARRAY
:
1138 child_type
= glsl_get_array_element(type
);
1140 case GLSL_TYPE_STRUCT
:
1141 child_type
= glsl_get_struct_field(type
, i
);
1144 unreachable("unkown base type");
1147 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1155 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1158 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1159 src
.src_type
= type
;
1164 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1165 const uint32_t *w
, unsigned count
)
1167 if (opcode
== SpvOpSampledImage
) {
1168 struct vtn_value
*val
=
1169 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1170 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1171 val
->sampled_image
->image
=
1172 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1173 val
->sampled_image
->sampler
=
1174 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1176 } else if (opcode
== SpvOpImage
) {
1177 struct vtn_value
*val
=
1178 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1179 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1180 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1181 val
->access_chain
= src_val
->sampled_image
->image
;
1183 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1184 val
->access_chain
= src_val
->access_chain
;
1189 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1190 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1192 struct vtn_sampled_image sampled
;
1193 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1194 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1195 sampled
= *sampled_val
->sampled_image
;
1197 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1198 sampled
.image
= NULL
;
1199 sampled
.sampler
= sampled_val
->access_chain
;
1202 const struct glsl_type
*image_type
;
1203 if (sampled
.image
) {
1204 image_type
= sampled
.image
->var
->var
->interface_type
;
1206 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1209 nir_tex_src srcs
[8]; /* 8 should be enough */
1210 nir_tex_src
*p
= srcs
;
1214 bool has_coord
= false;
1216 case SpvOpImageSampleImplicitLod
:
1217 case SpvOpImageSampleExplicitLod
:
1218 case SpvOpImageSampleDrefImplicitLod
:
1219 case SpvOpImageSampleDrefExplicitLod
:
1220 case SpvOpImageSampleProjImplicitLod
:
1221 case SpvOpImageSampleProjExplicitLod
:
1222 case SpvOpImageSampleProjDrefImplicitLod
:
1223 case SpvOpImageSampleProjDrefExplicitLod
:
1224 case SpvOpImageFetch
:
1225 case SpvOpImageGather
:
1226 case SpvOpImageDrefGather
:
1227 case SpvOpImageQueryLod
: {
1228 /* All these types have the coordinate as their first real argument */
1229 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, w
[idx
++]);
1231 p
->src
= nir_src_for_ssa(coord
->def
);
1232 p
->src_type
= nir_tex_src_coord
;
1241 /* These all have an explicit depth value as their next source */
1243 case SpvOpImageSampleDrefImplicitLod
:
1244 case SpvOpImageSampleDrefExplicitLod
:
1245 case SpvOpImageSampleProjDrefImplicitLod
:
1246 case SpvOpImageSampleProjDrefExplicitLod
:
1247 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparitor
);
1253 /* For OpImageQuerySizeLod, we always have an LOD */
1254 if (opcode
== SpvOpImageQuerySizeLod
)
1255 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1257 /* Figure out the base texture operation */
1260 case SpvOpImageSampleImplicitLod
:
1261 case SpvOpImageSampleDrefImplicitLod
:
1262 case SpvOpImageSampleProjImplicitLod
:
1263 case SpvOpImageSampleProjDrefImplicitLod
:
1264 texop
= nir_texop_tex
;
1267 case SpvOpImageSampleExplicitLod
:
1268 case SpvOpImageSampleDrefExplicitLod
:
1269 case SpvOpImageSampleProjExplicitLod
:
1270 case SpvOpImageSampleProjDrefExplicitLod
:
1271 texop
= nir_texop_txl
;
1274 case SpvOpImageFetch
:
1275 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1276 texop
= nir_texop_txf_ms
;
1278 texop
= nir_texop_txf
;
1282 case SpvOpImageGather
:
1283 case SpvOpImageDrefGather
:
1284 texop
= nir_texop_tg4
;
1287 case SpvOpImageQuerySizeLod
:
1288 case SpvOpImageQuerySize
:
1289 texop
= nir_texop_txs
;
1292 case SpvOpImageQueryLod
:
1293 texop
= nir_texop_lod
;
1296 case SpvOpImageQueryLevels
:
1297 texop
= nir_texop_query_levels
;
1300 case SpvOpImageQuerySamples
:
1302 unreachable("Unhandled opcode");
1305 /* Now we need to handle some number of optional arguments */
1307 uint32_t operands
= w
[idx
++];
1309 if (operands
& SpvImageOperandsBiasMask
) {
1310 assert(texop
== nir_texop_tex
);
1311 texop
= nir_texop_txb
;
1312 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1315 if (operands
& SpvImageOperandsLodMask
) {
1316 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1317 texop
== nir_texop_txf_ms
|| texop
== nir_texop_txs
);
1318 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1321 if (operands
& SpvImageOperandsGradMask
) {
1322 assert(texop
== nir_texop_tex
);
1323 texop
= nir_texop_txd
;
1324 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1325 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1328 if (operands
& SpvImageOperandsOffsetMask
||
1329 operands
& SpvImageOperandsConstOffsetMask
)
1330 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1332 if (operands
& SpvImageOperandsConstOffsetsMask
)
1333 assert(!"Constant offsets to texture gather not yet implemented");
1335 if (operands
& SpvImageOperandsSampleMask
) {
1336 assert(texop
== nir_texop_txf_ms
);
1337 texop
= nir_texop_txf_ms
;
1338 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1341 /* We should have now consumed exactly all of the arguments */
1342 assert(idx
== count
);
1344 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1347 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1349 instr
->sampler_dim
= glsl_get_sampler_dim(image_type
);
1350 instr
->is_array
= glsl_sampler_type_is_array(image_type
);
1351 instr
->is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1352 instr
->is_new_style_shadow
= instr
->is_shadow
;
1355 switch (instr
->sampler_dim
) {
1356 case GLSL_SAMPLER_DIM_1D
:
1357 case GLSL_SAMPLER_DIM_BUF
:
1358 instr
->coord_components
= 1;
1360 case GLSL_SAMPLER_DIM_2D
:
1361 case GLSL_SAMPLER_DIM_RECT
:
1362 case GLSL_SAMPLER_DIM_MS
:
1363 instr
->coord_components
= 2;
1365 case GLSL_SAMPLER_DIM_3D
:
1366 case GLSL_SAMPLER_DIM_CUBE
:
1367 instr
->coord_components
= 3;
1370 assert("Invalid sampler type");
1373 if (instr
->is_array
)
1374 instr
->coord_components
++;
1376 instr
->coord_components
= 0;
1379 switch (glsl_get_sampler_result_type(image_type
)) {
1380 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1381 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1382 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1383 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1385 unreachable("Invalid base type for sampler result");
1388 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1389 if (sampled
.image
) {
1390 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1391 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &image
->deref
));
1393 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1396 switch (instr
->op
) {
1401 /* These operations require a sampler */
1402 instr
->sampler
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1405 case nir_texop_txf_ms
:
1409 case nir_texop_query_levels
:
1410 case nir_texop_texture_samples
:
1411 case nir_texop_samples_identical
:
1413 instr
->sampler
= NULL
;
1417 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1418 nir_tex_instr_dest_size(instr
), 32, NULL
);
1420 assert(glsl_get_vector_elements(ret_type
->type
) ==
1421 nir_tex_instr_dest_size(instr
));
1423 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1424 val
->ssa
->def
= &instr
->dest
.ssa
;
1426 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1429 static nir_ssa_def
*
1430 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1432 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1434 /* The image_load_store intrinsics assume a 4-dim coordinate */
1435 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1436 unsigned swizzle
[4];
1437 for (unsigned i
= 0; i
< 4; i
++)
1438 swizzle
[i
] = MIN2(i
, dim
- 1);
1440 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1444 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1445 const uint32_t *w
, unsigned count
)
1447 /* Just get this one out of the way */
1448 if (opcode
== SpvOpImageTexelPointer
) {
1449 struct vtn_value
*val
=
1450 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1451 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1454 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1455 val
->image
->coord
= get_image_coord(b
, w
[4]);
1456 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1460 struct vtn_image_pointer image
;
1463 case SpvOpAtomicExchange
:
1464 case SpvOpAtomicCompareExchange
:
1465 case SpvOpAtomicCompareExchangeWeak
:
1466 case SpvOpAtomicIIncrement
:
1467 case SpvOpAtomicIDecrement
:
1468 case SpvOpAtomicIAdd
:
1469 case SpvOpAtomicISub
:
1470 case SpvOpAtomicSMin
:
1471 case SpvOpAtomicUMin
:
1472 case SpvOpAtomicSMax
:
1473 case SpvOpAtomicUMax
:
1474 case SpvOpAtomicAnd
:
1476 case SpvOpAtomicXor
:
1477 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1480 case SpvOpImageQuerySize
:
1482 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1484 image
.sample
= NULL
;
1487 case SpvOpImageRead
:
1489 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1490 image
.coord
= get_image_coord(b
, w
[4]);
1492 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1493 assert(w
[5] == SpvImageOperandsSampleMask
);
1494 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1496 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1500 case SpvOpImageWrite
:
1502 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1503 image
.coord
= get_image_coord(b
, w
[2]);
1507 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1508 assert(w
[4] == SpvImageOperandsSampleMask
);
1509 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1511 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1516 unreachable("Invalid image opcode");
1519 nir_intrinsic_op op
;
1521 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1522 OP(ImageQuerySize
, size
)
1524 OP(ImageWrite
, store
)
1525 OP(AtomicExchange
, atomic_exchange
)
1526 OP(AtomicCompareExchange
, atomic_comp_swap
)
1527 OP(AtomicIIncrement
, atomic_add
)
1528 OP(AtomicIDecrement
, atomic_add
)
1529 OP(AtomicIAdd
, atomic_add
)
1530 OP(AtomicISub
, atomic_add
)
1531 OP(AtomicSMin
, atomic_min
)
1532 OP(AtomicUMin
, atomic_min
)
1533 OP(AtomicSMax
, atomic_max
)
1534 OP(AtomicUMax
, atomic_max
)
1535 OP(AtomicAnd
, atomic_and
)
1536 OP(AtomicOr
, atomic_or
)
1537 OP(AtomicXor
, atomic_xor
)
1540 unreachable("Invalid image opcode");
1543 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1545 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1546 intrin
->variables
[0] =
1547 nir_deref_as_var(nir_copy_deref(&intrin
->instr
, &image_deref
->deref
));
1549 /* ImageQuerySize doesn't take any extra parameters */
1550 if (opcode
!= SpvOpImageQuerySize
) {
1551 /* The image coordinate is always 4 components but we may not have that
1552 * many. Swizzle to compensate.
1555 for (unsigned i
= 0; i
< 4; i
++)
1556 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1557 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1559 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1563 case SpvOpImageQuerySize
:
1564 case SpvOpImageRead
:
1566 case SpvOpImageWrite
:
1567 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1569 case SpvOpAtomicIIncrement
:
1570 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1572 case SpvOpAtomicIDecrement
:
1573 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1576 case SpvOpAtomicExchange
:
1577 case SpvOpAtomicIAdd
:
1578 case SpvOpAtomicSMin
:
1579 case SpvOpAtomicUMin
:
1580 case SpvOpAtomicSMax
:
1581 case SpvOpAtomicUMax
:
1582 case SpvOpAtomicAnd
:
1584 case SpvOpAtomicXor
:
1585 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1588 case SpvOpAtomicCompareExchange
:
1589 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1590 intrin
->src
[3] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1593 case SpvOpAtomicISub
:
1594 intrin
->src
[2] = nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1598 unreachable("Invalid image opcode");
1601 if (opcode
!= SpvOpImageWrite
) {
1602 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1603 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1604 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, 32, NULL
);
1606 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1608 /* The image intrinsics always return 4 channels but we may not want
1609 * that many. Emit a mov to trim it down.
1611 unsigned swiz
[4] = {0, 1, 2, 3};
1612 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1613 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1614 glsl_get_vector_elements(type
->type
), false);
1616 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1620 static nir_intrinsic_op
1621 get_ssbo_nir_atomic_op(SpvOp opcode
)
1624 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1625 OP(AtomicExchange
, atomic_exchange
)
1626 OP(AtomicCompareExchange
, atomic_comp_swap
)
1627 OP(AtomicIIncrement
, atomic_add
)
1628 OP(AtomicIDecrement
, atomic_add
)
1629 OP(AtomicIAdd
, atomic_add
)
1630 OP(AtomicISub
, atomic_add
)
1631 OP(AtomicSMin
, atomic_imin
)
1632 OP(AtomicUMin
, atomic_umin
)
1633 OP(AtomicSMax
, atomic_imax
)
1634 OP(AtomicUMax
, atomic_umax
)
1635 OP(AtomicAnd
, atomic_and
)
1636 OP(AtomicOr
, atomic_or
)
1637 OP(AtomicXor
, atomic_xor
)
1640 unreachable("Invalid SSBO atomic");
1644 static nir_intrinsic_op
1645 get_shared_nir_atomic_op(SpvOp opcode
)
1648 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
1649 OP(AtomicExchange
, atomic_exchange
)
1650 OP(AtomicCompareExchange
, atomic_comp_swap
)
1651 OP(AtomicIIncrement
, atomic_add
)
1652 OP(AtomicIDecrement
, atomic_add
)
1653 OP(AtomicIAdd
, atomic_add
)
1654 OP(AtomicISub
, atomic_add
)
1655 OP(AtomicSMin
, atomic_imin
)
1656 OP(AtomicUMin
, atomic_umin
)
1657 OP(AtomicSMax
, atomic_imax
)
1658 OP(AtomicUMax
, atomic_umax
)
1659 OP(AtomicAnd
, atomic_and
)
1660 OP(AtomicOr
, atomic_or
)
1661 OP(AtomicXor
, atomic_xor
)
1664 unreachable("Invalid shared atomic");
1669 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1670 const uint32_t *w
, nir_src
*src
)
1673 case SpvOpAtomicIIncrement
:
1674 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1677 case SpvOpAtomicIDecrement
:
1678 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1681 case SpvOpAtomicISub
:
1683 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1686 case SpvOpAtomicCompareExchange
:
1687 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1688 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1692 case SpvOpAtomicExchange
:
1693 case SpvOpAtomicIAdd
:
1694 case SpvOpAtomicSMin
:
1695 case SpvOpAtomicUMin
:
1696 case SpvOpAtomicSMax
:
1697 case SpvOpAtomicUMax
:
1698 case SpvOpAtomicAnd
:
1700 case SpvOpAtomicXor
:
1701 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1705 unreachable("Invalid SPIR-V atomic");
1710 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
1711 const uint32_t *w
, unsigned count
)
1713 struct vtn_access_chain
*chain
=
1714 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1715 nir_intrinsic_instr
*atomic
;
1718 SpvScope scope = w[4];
1719 SpvMemorySemanticsMask semantics = w[5];
1722 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
1723 nir_deref
*deref
= &vtn_access_chain_to_deref(b
, chain
)->deref
;
1724 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
1725 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1726 atomic
->variables
[0] = nir_deref_as_var(nir_copy_deref(atomic
, deref
));
1727 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
1729 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
1730 struct vtn_type
*type
;
1731 nir_ssa_def
*offset
, *index
;
1732 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
1734 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
1736 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1737 atomic
->src
[0] = nir_src_for_ssa(index
);
1738 atomic
->src
[1] = nir_src_for_ssa(offset
);
1739 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
1742 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
, 1, 32, NULL
);
1744 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1745 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1746 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
1747 val
->ssa
->def
= &atomic
->dest
.ssa
;
1748 val
->ssa
->type
= type
->type
;
1750 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
1753 static nir_alu_instr
*
1754 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
1757 switch (num_components
) {
1758 case 1: op
= nir_op_fmov
; break;
1759 case 2: op
= nir_op_vec2
; break;
1760 case 3: op
= nir_op_vec3
; break;
1761 case 4: op
= nir_op_vec4
; break;
1762 default: unreachable("bad vector size");
1765 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
1766 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
1768 vec
->dest
.write_mask
= (1 << num_components
) - 1;
1773 struct vtn_ssa_value
*
1774 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
1776 if (src
->transposed
)
1777 return src
->transposed
;
1779 struct vtn_ssa_value
*dest
=
1780 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
1782 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
1783 nir_alu_instr
*vec
= create_vec(b
->shader
,
1784 glsl_get_matrix_columns(src
->type
),
1785 glsl_get_bit_size(glsl_get_base_type(src
->type
)));
1786 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1787 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
1788 vec
->src
[0].swizzle
[0] = i
;
1790 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
1791 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
1792 vec
->src
[j
].swizzle
[0] = i
;
1795 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1796 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
1799 dest
->transposed
= src
;
1805 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
1807 unsigned swiz
[4] = { index
};
1808 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
1812 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
1815 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
1818 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
1820 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
1822 vec
->src
[i
].src
= nir_src_for_ssa(src
);
1823 vec
->src
[i
].swizzle
[0] = i
;
1827 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1829 return &vec
->dest
.dest
.ssa
;
1833 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1836 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
1837 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1838 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1839 vtn_vector_extract(b
, src
, i
), dest
);
1845 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1846 nir_ssa_def
*insert
, nir_ssa_def
*index
)
1848 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
1849 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1850 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1851 vtn_vector_insert(b
, src
, insert
, i
), dest
);
1856 static nir_ssa_def
*
1857 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
1858 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
1859 const uint32_t *indices
)
1861 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
1863 for (unsigned i
= 0; i
< num_components
; i
++) {
1864 uint32_t index
= indices
[i
];
1865 if (index
== 0xffffffff) {
1867 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
1868 } else if (index
< src0
->num_components
) {
1869 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
1870 vec
->src
[i
].swizzle
[0] = index
;
1872 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
1873 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
1877 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1879 return &vec
->dest
.dest
.ssa
;
1883 * Concatentates a number of vectors/scalars together to produce a vector
1885 static nir_ssa_def
*
1886 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
1887 unsigned num_srcs
, nir_ssa_def
**srcs
)
1889 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
1892 unsigned dest_idx
= 0;
1893 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1894 nir_ssa_def
*src
= srcs
[i
];
1895 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
1896 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
1897 vec
->src
[dest_idx
].swizzle
[0] = j
;
1902 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1904 return &vec
->dest
.dest
.ssa
;
1907 static struct vtn_ssa_value
*
1908 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
1910 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
1911 dest
->type
= src
->type
;
1913 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1914 dest
->def
= src
->def
;
1916 unsigned elems
= glsl_get_length(src
->type
);
1918 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
1919 for (unsigned i
= 0; i
< elems
; i
++)
1920 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
1926 static struct vtn_ssa_value
*
1927 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1928 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
1929 unsigned num_indices
)
1931 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
1933 struct vtn_ssa_value
*cur
= dest
;
1935 for (i
= 0; i
< num_indices
- 1; i
++) {
1936 cur
= cur
->elems
[indices
[i
]];
1939 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1940 /* According to the SPIR-V spec, OpCompositeInsert may work down to
1941 * the component granularity. In that case, the last index will be
1942 * the index to insert the scalar into the vector.
1945 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
1947 cur
->elems
[indices
[i
]] = insert
;
1953 static struct vtn_ssa_value
*
1954 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1955 const uint32_t *indices
, unsigned num_indices
)
1957 struct vtn_ssa_value
*cur
= src
;
1958 for (unsigned i
= 0; i
< num_indices
; i
++) {
1959 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1960 assert(i
== num_indices
- 1);
1961 /* According to the SPIR-V spec, OpCompositeExtract may work down to
1962 * the component granularity. The last index will be the index of the
1963 * vector to extract.
1966 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
1967 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
1968 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
1971 cur
= cur
->elems
[indices
[i
]];
1979 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
1980 const uint32_t *w
, unsigned count
)
1982 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1983 const struct glsl_type
*type
=
1984 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1985 val
->ssa
= vtn_create_ssa_value(b
, type
);
1988 case SpvOpVectorExtractDynamic
:
1989 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1990 vtn_ssa_value(b
, w
[4])->def
);
1993 case SpvOpVectorInsertDynamic
:
1994 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1995 vtn_ssa_value(b
, w
[4])->def
,
1996 vtn_ssa_value(b
, w
[5])->def
);
1999 case SpvOpVectorShuffle
:
2000 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2001 vtn_ssa_value(b
, w
[3])->def
,
2002 vtn_ssa_value(b
, w
[4])->def
,
2006 case SpvOpCompositeConstruct
: {
2007 unsigned elems
= count
- 3;
2008 if (glsl_type_is_vector_or_scalar(type
)) {
2009 nir_ssa_def
*srcs
[4];
2010 for (unsigned i
= 0; i
< elems
; i
++)
2011 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2013 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2016 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2017 for (unsigned i
= 0; i
< elems
; i
++)
2018 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2022 case SpvOpCompositeExtract
:
2023 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2027 case SpvOpCompositeInsert
:
2028 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2029 vtn_ssa_value(b
, w
[3]),
2033 case SpvOpCopyObject
:
2034 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2038 unreachable("unknown composite operation");
2043 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2044 const uint32_t *w
, unsigned count
)
2046 nir_intrinsic_op intrinsic_op
;
2048 case SpvOpEmitVertex
:
2049 case SpvOpEmitStreamVertex
:
2050 intrinsic_op
= nir_intrinsic_emit_vertex
;
2052 case SpvOpEndPrimitive
:
2053 case SpvOpEndStreamPrimitive
:
2054 intrinsic_op
= nir_intrinsic_end_primitive
;
2056 case SpvOpMemoryBarrier
:
2057 intrinsic_op
= nir_intrinsic_memory_barrier
;
2059 case SpvOpControlBarrier
:
2060 intrinsic_op
= nir_intrinsic_barrier
;
2063 unreachable("unknown barrier instruction");
2066 nir_intrinsic_instr
*intrin
=
2067 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2069 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2070 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2072 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2076 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2079 case SpvExecutionModeInputPoints
:
2080 case SpvExecutionModeOutputPoints
:
2081 return 0; /* GL_POINTS */
2082 case SpvExecutionModeInputLines
:
2083 return 1; /* GL_LINES */
2084 case SpvExecutionModeInputLinesAdjacency
:
2085 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2086 case SpvExecutionModeTriangles
:
2087 return 4; /* GL_TRIANGLES */
2088 case SpvExecutionModeInputTrianglesAdjacency
:
2089 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2090 case SpvExecutionModeQuads
:
2091 return 7; /* GL_QUADS */
2092 case SpvExecutionModeIsolines
:
2093 return 0x8E7A; /* GL_ISOLINES */
2094 case SpvExecutionModeOutputLineStrip
:
2095 return 3; /* GL_LINE_STRIP */
2096 case SpvExecutionModeOutputTriangleStrip
:
2097 return 5; /* GL_TRIANGLE_STRIP */
2099 assert(!"Invalid primitive type");
2105 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2108 case SpvExecutionModeInputPoints
:
2110 case SpvExecutionModeInputLines
:
2112 case SpvExecutionModeInputLinesAdjacency
:
2114 case SpvExecutionModeTriangles
:
2116 case SpvExecutionModeInputTrianglesAdjacency
:
2119 assert(!"Invalid GS input mode");
2124 static gl_shader_stage
2125 stage_for_execution_model(SpvExecutionModel model
)
2128 case SpvExecutionModelVertex
:
2129 return MESA_SHADER_VERTEX
;
2130 case SpvExecutionModelTessellationControl
:
2131 return MESA_SHADER_TESS_CTRL
;
2132 case SpvExecutionModelTessellationEvaluation
:
2133 return MESA_SHADER_TESS_EVAL
;
2134 case SpvExecutionModelGeometry
:
2135 return MESA_SHADER_GEOMETRY
;
2136 case SpvExecutionModelFragment
:
2137 return MESA_SHADER_FRAGMENT
;
2138 case SpvExecutionModelGLCompute
:
2139 return MESA_SHADER_COMPUTE
;
2141 unreachable("Unsupported execution model");
2146 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2147 const uint32_t *w
, unsigned count
)
2151 case SpvOpSourceExtension
:
2152 case SpvOpSourceContinued
:
2153 case SpvOpExtension
:
2154 /* Unhandled, but these are for debug so that's ok. */
2157 case SpvOpCapability
: {
2158 SpvCapability cap
= w
[1];
2160 case SpvCapabilityMatrix
:
2161 case SpvCapabilityShader
:
2162 case SpvCapabilityGeometry
:
2163 case SpvCapabilityTessellationPointSize
:
2164 case SpvCapabilityGeometryPointSize
:
2165 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2166 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2167 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2168 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2169 case SpvCapabilityImageRect
:
2170 case SpvCapabilitySampledRect
:
2171 case SpvCapabilitySampled1D
:
2172 case SpvCapabilityImage1D
:
2173 case SpvCapabilitySampledCubeArray
:
2174 case SpvCapabilitySampledBuffer
:
2175 case SpvCapabilityImageBuffer
:
2176 case SpvCapabilityImageQuery
:
2178 case SpvCapabilityClipDistance
:
2179 case SpvCapabilityCullDistance
:
2180 case SpvCapabilityGeometryStreams
:
2181 fprintf(stderr
, "WARNING: Unsupported SPIR-V Capability\n");
2184 assert(!"Unsupported capability");
2189 case SpvOpExtInstImport
:
2190 vtn_handle_extension(b
, opcode
, w
, count
);
2193 case SpvOpMemoryModel
:
2194 assert(w
[1] == SpvAddressingModelLogical
);
2195 assert(w
[2] == SpvMemoryModelGLSL450
);
2198 case SpvOpEntryPoint
: {
2199 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2200 /* Let this be a name label regardless */
2201 unsigned name_words
;
2202 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2204 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2205 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2208 assert(b
->entry_point
== NULL
);
2209 b
->entry_point
= entry_point
;
2214 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2215 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2219 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2222 case SpvOpMemberName
:
2226 case SpvOpExecutionMode
:
2227 case SpvOpDecorationGroup
:
2229 case SpvOpMemberDecorate
:
2230 case SpvOpGroupDecorate
:
2231 case SpvOpGroupMemberDecorate
:
2232 vtn_handle_decoration(b
, opcode
, w
, count
);
2236 return false; /* End of preamble */
2243 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2244 const struct vtn_decoration
*mode
, void *data
)
2246 assert(b
->entry_point
== entry_point
);
2248 switch(mode
->exec_mode
) {
2249 case SpvExecutionModeOriginUpperLeft
:
2250 case SpvExecutionModeOriginLowerLeft
:
2251 b
->origin_upper_left
=
2252 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2255 case SpvExecutionModeEarlyFragmentTests
:
2256 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2257 b
->shader
->info
.fs
.early_fragment_tests
= true;
2260 case SpvExecutionModeInvocations
:
2261 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2262 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2265 case SpvExecutionModeDepthReplacing
:
2266 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2267 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2269 case SpvExecutionModeDepthGreater
:
2270 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2271 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2273 case SpvExecutionModeDepthLess
:
2274 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2275 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2277 case SpvExecutionModeDepthUnchanged
:
2278 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2279 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2282 case SpvExecutionModeLocalSize
:
2283 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2284 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2285 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2286 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2288 case SpvExecutionModeLocalSizeHint
:
2289 break; /* Nothing do do with this */
2291 case SpvExecutionModeOutputVertices
:
2292 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2293 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2296 case SpvExecutionModeInputPoints
:
2297 case SpvExecutionModeInputLines
:
2298 case SpvExecutionModeInputLinesAdjacency
:
2299 case SpvExecutionModeTriangles
:
2300 case SpvExecutionModeInputTrianglesAdjacency
:
2301 case SpvExecutionModeQuads
:
2302 case SpvExecutionModeIsolines
:
2303 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2304 b
->shader
->info
.gs
.vertices_in
=
2305 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2307 assert(!"Tesselation shaders not yet supported");
2311 case SpvExecutionModeOutputPoints
:
2312 case SpvExecutionModeOutputLineStrip
:
2313 case SpvExecutionModeOutputTriangleStrip
:
2314 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2315 b
->shader
->info
.gs
.output_primitive
=
2316 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2319 case SpvExecutionModeSpacingEqual
:
2320 case SpvExecutionModeSpacingFractionalEven
:
2321 case SpvExecutionModeSpacingFractionalOdd
:
2322 case SpvExecutionModeVertexOrderCw
:
2323 case SpvExecutionModeVertexOrderCcw
:
2324 case SpvExecutionModePointMode
:
2325 assert(!"TODO: Add tessellation metadata");
2328 case SpvExecutionModePixelCenterInteger
:
2329 case SpvExecutionModeXfb
:
2330 assert(!"Unhandled execution mode");
2333 case SpvExecutionModeVecTypeHint
:
2334 case SpvExecutionModeContractionOff
:
2340 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2341 const uint32_t *w
, unsigned count
)
2345 case SpvOpSourceContinued
:
2346 case SpvOpSourceExtension
:
2347 case SpvOpExtension
:
2348 case SpvOpCapability
:
2349 case SpvOpExtInstImport
:
2350 case SpvOpMemoryModel
:
2351 case SpvOpEntryPoint
:
2352 case SpvOpExecutionMode
:
2355 case SpvOpMemberName
:
2356 case SpvOpDecorationGroup
:
2358 case SpvOpMemberDecorate
:
2359 case SpvOpGroupDecorate
:
2360 case SpvOpGroupMemberDecorate
:
2361 assert(!"Invalid opcode types and variables section");
2367 case SpvOpTypeFloat
:
2368 case SpvOpTypeVector
:
2369 case SpvOpTypeMatrix
:
2370 case SpvOpTypeImage
:
2371 case SpvOpTypeSampler
:
2372 case SpvOpTypeSampledImage
:
2373 case SpvOpTypeArray
:
2374 case SpvOpTypeRuntimeArray
:
2375 case SpvOpTypeStruct
:
2376 case SpvOpTypeOpaque
:
2377 case SpvOpTypePointer
:
2378 case SpvOpTypeFunction
:
2379 case SpvOpTypeEvent
:
2380 case SpvOpTypeDeviceEvent
:
2381 case SpvOpTypeReserveId
:
2382 case SpvOpTypeQueue
:
2384 vtn_handle_type(b
, opcode
, w
, count
);
2387 case SpvOpConstantTrue
:
2388 case SpvOpConstantFalse
:
2390 case SpvOpConstantComposite
:
2391 case SpvOpConstantSampler
:
2392 case SpvOpConstantNull
:
2393 case SpvOpSpecConstantTrue
:
2394 case SpvOpSpecConstantFalse
:
2395 case SpvOpSpecConstant
:
2396 case SpvOpSpecConstantComposite
:
2397 case SpvOpSpecConstantOp
:
2398 vtn_handle_constant(b
, opcode
, w
, count
);
2402 vtn_handle_variables(b
, opcode
, w
, count
);
2406 return false; /* End of preamble */
2413 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2414 const uint32_t *w
, unsigned count
)
2420 case SpvOpLoopMerge
:
2421 case SpvOpSelectionMerge
:
2422 /* This is handled by cfg pre-pass and walk_blocks */
2426 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2427 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2432 vtn_handle_extension(b
, opcode
, w
, count
);
2438 case SpvOpCopyMemory
:
2439 case SpvOpCopyMemorySized
:
2440 case SpvOpAccessChain
:
2441 case SpvOpInBoundsAccessChain
:
2442 case SpvOpArrayLength
:
2443 vtn_handle_variables(b
, opcode
, w
, count
);
2446 case SpvOpFunctionCall
:
2447 vtn_handle_function_call(b
, opcode
, w
, count
);
2450 case SpvOpSampledImage
:
2452 case SpvOpImageSampleImplicitLod
:
2453 case SpvOpImageSampleExplicitLod
:
2454 case SpvOpImageSampleDrefImplicitLod
:
2455 case SpvOpImageSampleDrefExplicitLod
:
2456 case SpvOpImageSampleProjImplicitLod
:
2457 case SpvOpImageSampleProjExplicitLod
:
2458 case SpvOpImageSampleProjDrefImplicitLod
:
2459 case SpvOpImageSampleProjDrefExplicitLod
:
2460 case SpvOpImageFetch
:
2461 case SpvOpImageGather
:
2462 case SpvOpImageDrefGather
:
2463 case SpvOpImageQuerySizeLod
:
2464 case SpvOpImageQueryLod
:
2465 case SpvOpImageQueryLevels
:
2466 case SpvOpImageQuerySamples
:
2467 vtn_handle_texture(b
, opcode
, w
, count
);
2470 case SpvOpImageRead
:
2471 case SpvOpImageWrite
:
2472 case SpvOpImageTexelPointer
:
2473 vtn_handle_image(b
, opcode
, w
, count
);
2476 case SpvOpImageQuerySize
: {
2477 struct vtn_access_chain
*image
=
2478 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2479 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2480 vtn_handle_image(b
, opcode
, w
, count
);
2482 vtn_handle_texture(b
, opcode
, w
, count
);
2487 case SpvOpAtomicExchange
:
2488 case SpvOpAtomicCompareExchange
:
2489 case SpvOpAtomicCompareExchangeWeak
:
2490 case SpvOpAtomicIIncrement
:
2491 case SpvOpAtomicIDecrement
:
2492 case SpvOpAtomicIAdd
:
2493 case SpvOpAtomicISub
:
2494 case SpvOpAtomicSMin
:
2495 case SpvOpAtomicUMin
:
2496 case SpvOpAtomicSMax
:
2497 case SpvOpAtomicUMax
:
2498 case SpvOpAtomicAnd
:
2500 case SpvOpAtomicXor
: {
2501 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2502 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2503 vtn_handle_image(b
, opcode
, w
, count
);
2505 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2506 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2516 case SpvOpConvertFToU
:
2517 case SpvOpConvertFToS
:
2518 case SpvOpConvertSToF
:
2519 case SpvOpConvertUToF
:
2523 case SpvOpQuantizeToF16
:
2524 case SpvOpConvertPtrToU
:
2525 case SpvOpConvertUToPtr
:
2526 case SpvOpPtrCastToGeneric
:
2527 case SpvOpGenericCastToPtr
:
2533 case SpvOpSignBitSet
:
2534 case SpvOpLessOrGreater
:
2536 case SpvOpUnordered
:
2551 case SpvOpVectorTimesScalar
:
2553 case SpvOpIAddCarry
:
2554 case SpvOpISubBorrow
:
2555 case SpvOpUMulExtended
:
2556 case SpvOpSMulExtended
:
2557 case SpvOpShiftRightLogical
:
2558 case SpvOpShiftRightArithmetic
:
2559 case SpvOpShiftLeftLogical
:
2560 case SpvOpLogicalEqual
:
2561 case SpvOpLogicalNotEqual
:
2562 case SpvOpLogicalOr
:
2563 case SpvOpLogicalAnd
:
2564 case SpvOpLogicalNot
:
2565 case SpvOpBitwiseOr
:
2566 case SpvOpBitwiseXor
:
2567 case SpvOpBitwiseAnd
:
2570 case SpvOpFOrdEqual
:
2571 case SpvOpFUnordEqual
:
2572 case SpvOpINotEqual
:
2573 case SpvOpFOrdNotEqual
:
2574 case SpvOpFUnordNotEqual
:
2575 case SpvOpULessThan
:
2576 case SpvOpSLessThan
:
2577 case SpvOpFOrdLessThan
:
2578 case SpvOpFUnordLessThan
:
2579 case SpvOpUGreaterThan
:
2580 case SpvOpSGreaterThan
:
2581 case SpvOpFOrdGreaterThan
:
2582 case SpvOpFUnordGreaterThan
:
2583 case SpvOpULessThanEqual
:
2584 case SpvOpSLessThanEqual
:
2585 case SpvOpFOrdLessThanEqual
:
2586 case SpvOpFUnordLessThanEqual
:
2587 case SpvOpUGreaterThanEqual
:
2588 case SpvOpSGreaterThanEqual
:
2589 case SpvOpFOrdGreaterThanEqual
:
2590 case SpvOpFUnordGreaterThanEqual
:
2596 case SpvOpFwidthFine
:
2597 case SpvOpDPdxCoarse
:
2598 case SpvOpDPdyCoarse
:
2599 case SpvOpFwidthCoarse
:
2600 case SpvOpBitFieldInsert
:
2601 case SpvOpBitFieldSExtract
:
2602 case SpvOpBitFieldUExtract
:
2603 case SpvOpBitReverse
:
2605 case SpvOpTranspose
:
2606 case SpvOpOuterProduct
:
2607 case SpvOpMatrixTimesScalar
:
2608 case SpvOpVectorTimesMatrix
:
2609 case SpvOpMatrixTimesVector
:
2610 case SpvOpMatrixTimesMatrix
:
2611 vtn_handle_alu(b
, opcode
, w
, count
);
2614 case SpvOpVectorExtractDynamic
:
2615 case SpvOpVectorInsertDynamic
:
2616 case SpvOpVectorShuffle
:
2617 case SpvOpCompositeConstruct
:
2618 case SpvOpCompositeExtract
:
2619 case SpvOpCompositeInsert
:
2620 case SpvOpCopyObject
:
2621 vtn_handle_composite(b
, opcode
, w
, count
);
2624 case SpvOpEmitVertex
:
2625 case SpvOpEndPrimitive
:
2626 case SpvOpEmitStreamVertex
:
2627 case SpvOpEndStreamPrimitive
:
2628 case SpvOpControlBarrier
:
2629 case SpvOpMemoryBarrier
:
2630 vtn_handle_barrier(b
, opcode
, w
, count
);
2634 unreachable("Unhandled opcode");
2641 spirv_to_nir(const uint32_t *words
, size_t word_count
,
2642 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
2643 gl_shader_stage stage
, const char *entry_point_name
,
2644 const nir_shader_compiler_options
*options
)
2646 const uint32_t *word_end
= words
+ word_count
;
2648 /* Handle the SPIR-V header (first 4 dwords) */
2649 assert(word_count
> 5);
2651 assert(words
[0] == SpvMagicNumber
);
2652 assert(words
[1] >= 0x10000);
2653 /* words[2] == generator magic */
2654 unsigned value_id_bound
= words
[3];
2655 assert(words
[4] == 0);
2659 /* Initialize the stn_builder object */
2660 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
2661 b
->value_id_bound
= value_id_bound
;
2662 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
2663 exec_list_make_empty(&b
->functions
);
2664 b
->entry_point_stage
= stage
;
2665 b
->entry_point_name
= entry_point_name
;
2667 /* Handle all the preamble instructions */
2668 words
= vtn_foreach_instruction(b
, words
, word_end
,
2669 vtn_handle_preamble_instruction
);
2671 if (b
->entry_point
== NULL
) {
2672 assert(!"Entry point not found");
2677 b
->shader
= nir_shader_create(NULL
, stage
, options
);
2679 /* Set shader info defaults */
2680 b
->shader
->info
.gs
.invocations
= 1;
2682 /* Parse execution modes */
2683 vtn_foreach_execution_mode(b
, b
->entry_point
,
2684 vtn_handle_execution_mode
, NULL
);
2686 b
->specializations
= spec
;
2687 b
->num_specializations
= num_spec
;
2689 /* Handle all variable, type, and constant instructions */
2690 words
= vtn_foreach_instruction(b
, words
, word_end
,
2691 vtn_handle_variable_or_type_instruction
);
2693 vtn_build_cfg(b
, words
, word_end
);
2695 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
2696 b
->impl
= func
->impl
;
2697 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
2698 _mesa_key_pointer_equal
);
2700 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
2703 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
2704 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
2705 assert(entry_point
);