2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
33 static struct vtn_ssa_value
*
34 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
36 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
39 if (glsl_type_is_vector_or_scalar(type
)) {
40 unsigned num_components
= glsl_get_vector_elements(val
->type
);
41 nir_ssa_undef_instr
*undef
=
42 nir_ssa_undef_instr_create(b
->shader
, num_components
);
44 nir_instr_insert_before_cf_list(&b
->impl
->body
, &undef
->instr
);
45 val
->def
= &undef
->def
;
47 unsigned elems
= glsl_get_length(val
->type
);
48 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
49 if (glsl_type_is_matrix(type
)) {
50 const struct glsl_type
*elem_type
=
51 glsl_vector_type(glsl_get_base_type(type
),
52 glsl_get_vector_elements(type
));
54 for (unsigned i
= 0; i
< elems
; i
++)
55 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
56 } else if (glsl_type_is_array(type
)) {
57 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
58 for (unsigned i
= 0; i
< elems
; i
++)
59 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
61 for (unsigned i
= 0; i
< elems
; i
++) {
62 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
63 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
71 static struct vtn_ssa_value
*
72 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
73 const struct glsl_type
*type
)
75 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
80 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
83 switch (glsl_get_base_type(type
)) {
88 case GLSL_TYPE_DOUBLE
:
89 if (glsl_type_is_vector_or_scalar(type
)) {
90 unsigned num_components
= glsl_get_vector_elements(val
->type
);
91 nir_load_const_instr
*load
=
92 nir_load_const_instr_create(b
->shader
, num_components
);
94 for (unsigned i
= 0; i
< num_components
; i
++)
95 load
->value
.u
[i
] = constant
->value
.u
[i
];
97 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
98 val
->def
= &load
->def
;
100 assert(glsl_type_is_matrix(type
));
101 unsigned rows
= glsl_get_vector_elements(val
->type
);
102 unsigned columns
= glsl_get_matrix_columns(val
->type
);
103 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
105 for (unsigned i
= 0; i
< columns
; i
++) {
106 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
107 col_val
->type
= glsl_get_column_type(val
->type
);
108 nir_load_const_instr
*load
=
109 nir_load_const_instr_create(b
->shader
, rows
);
111 for (unsigned j
= 0; j
< rows
; j
++)
112 load
->value
.u
[j
] = constant
->value
.u
[rows
* i
+ j
];
114 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
115 col_val
->def
= &load
->def
;
117 val
->elems
[i
] = col_val
;
122 case GLSL_TYPE_ARRAY
: {
123 unsigned elems
= glsl_get_length(val
->type
);
124 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
125 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
126 for (unsigned i
= 0; i
< elems
; i
++)
127 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
132 case GLSL_TYPE_STRUCT
: {
133 unsigned elems
= glsl_get_length(val
->type
);
134 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
135 for (unsigned i
= 0; i
< elems
; i
++) {
136 const struct glsl_type
*elem_type
=
137 glsl_get_struct_field(val
->type
, i
);
138 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
145 unreachable("bad constant type");
151 struct vtn_ssa_value
*
152 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
154 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
155 switch (val
->value_type
) {
156 case vtn_value_type_undef
:
157 return vtn_undef_ssa_value(b
, val
->type
->type
);
159 case vtn_value_type_constant
:
160 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
162 case vtn_value_type_ssa
:
165 case vtn_value_type_access_chain
:
166 /* This is needed for function parameters */
167 return vtn_variable_load(b
, val
->access_chain
);
170 unreachable("Invalid type for an SSA value");
175 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
176 unsigned word_count
, unsigned *words_used
)
178 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
180 /* Ammount of space taken by the string (including the null) */
181 unsigned len
= strlen(dup
) + 1;
182 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
188 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
189 const uint32_t *end
, vtn_instruction_handler handler
)
195 const uint32_t *w
= start
;
197 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
198 unsigned count
= w
[0] >> SpvWordCountShift
;
199 assert(count
>= 1 && w
+ count
<= end
);
203 break; /* Do nothing */
206 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
218 if (!handler(b
, opcode
, w
, count
))
230 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
231 const uint32_t *w
, unsigned count
)
234 case SpvOpExtInstImport
: {
235 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
236 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
237 val
->ext_handler
= vtn_handle_glsl450_instruction
;
239 assert(!"Unsupported extension");
245 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
246 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
253 unreachable("Unhandled opcode");
258 _foreach_decoration_helper(struct vtn_builder
*b
,
259 struct vtn_value
*base_value
,
261 struct vtn_value
*value
,
262 vtn_decoration_foreach_cb cb
, void *data
)
264 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
266 if (dec
->scope
== VTN_DEC_DECORATION
) {
267 member
= parent_member
;
268 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
269 assert(parent_member
== -1);
270 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
272 /* Not a decoration */
277 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
278 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
281 cb(b
, base_value
, member
, dec
, data
);
286 /** Iterates (recursively if needed) over all of the decorations on a value
288 * This function iterates over all of the decorations applied to a given
289 * value. If it encounters a decoration group, it recurses into the group
290 * and iterates over all of those decorations as well.
293 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
294 vtn_decoration_foreach_cb cb
, void *data
)
296 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
300 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
301 vtn_execution_mode_foreach_cb cb
, void *data
)
303 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
304 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
307 assert(dec
->group
== NULL
);
308 cb(b
, value
, dec
, data
);
313 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
314 const uint32_t *w
, unsigned count
)
316 const uint32_t *w_end
= w
+ count
;
317 const uint32_t target
= w
[1];
321 case SpvOpDecorationGroup
:
322 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
326 case SpvOpMemberDecorate
:
327 case SpvOpExecutionMode
: {
328 struct vtn_value
*val
= &b
->values
[target
];
330 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
333 dec
->scope
= VTN_DEC_DECORATION
;
335 case SpvOpMemberDecorate
:
336 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
338 case SpvOpExecutionMode
:
339 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
342 unreachable("Invalid decoration opcode");
344 dec
->decoration
= *(w
++);
347 /* Link into the list */
348 dec
->next
= val
->decoration
;
349 val
->decoration
= dec
;
353 case SpvOpGroupMemberDecorate
:
354 case SpvOpGroupDecorate
: {
355 struct vtn_value
*group
=
356 vtn_value(b
, target
, vtn_value_type_decoration_group
);
358 for (; w
< w_end
; w
++) {
359 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
360 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
363 if (opcode
== SpvOpGroupDecorate
) {
364 dec
->scope
= VTN_DEC_DECORATION
;
366 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
369 /* Link into the list */
370 dec
->next
= val
->decoration
;
371 val
->decoration
= dec
;
377 unreachable("Unhandled opcode");
381 struct member_decoration_ctx
{
383 struct glsl_struct_field
*fields
;
384 struct vtn_type
*type
;
387 /* does a shallow copy of a vtn_type */
389 static struct vtn_type
*
390 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
392 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
393 dest
->type
= src
->type
;
394 dest
->is_builtin
= src
->is_builtin
;
396 dest
->builtin
= src
->builtin
;
398 if (!glsl_type_is_scalar(src
->type
)) {
399 switch (glsl_get_base_type(src
->type
)) {
403 case GLSL_TYPE_FLOAT
:
404 case GLSL_TYPE_DOUBLE
:
405 case GLSL_TYPE_ARRAY
:
406 dest
->row_major
= src
->row_major
;
407 dest
->stride
= src
->stride
;
408 dest
->array_element
= src
->array_element
;
411 case GLSL_TYPE_STRUCT
: {
412 unsigned elems
= glsl_get_length(src
->type
);
414 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
415 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
417 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
418 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
423 unreachable("unhandled type");
430 static struct vtn_type
*
431 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
433 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
434 type
= type
->members
[member
];
436 /* We may have an array of matrices.... Oh, joy! */
437 while (glsl_type_is_array(type
->type
)) {
438 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
439 type
= type
->array_element
;
442 assert(glsl_type_is_matrix(type
->type
));
448 struct_member_decoration_cb(struct vtn_builder
*b
,
449 struct vtn_value
*val
, int member
,
450 const struct vtn_decoration
*dec
, void *void_ctx
)
452 struct member_decoration_ctx
*ctx
= void_ctx
;
457 assert(member
< ctx
->num_fields
);
459 switch (dec
->decoration
) {
460 case SpvDecorationRelaxedPrecision
:
461 break; /* FIXME: Do nothing with this for now. */
462 case SpvDecorationNoPerspective
:
463 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_NOPERSPECTIVE
;
465 case SpvDecorationFlat
:
466 ctx
->fields
[member
].interpolation
= INTERP_QUALIFIER_FLAT
;
468 case SpvDecorationCentroid
:
469 ctx
->fields
[member
].centroid
= true;
471 case SpvDecorationSample
:
472 ctx
->fields
[member
].sample
= true;
474 case SpvDecorationLocation
:
475 ctx
->fields
[member
].location
= dec
->literals
[0];
477 case SpvDecorationBuiltIn
:
478 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
479 ctx
->type
->members
[member
]->is_builtin
= true;
480 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
481 ctx
->type
->builtin_block
= true;
483 case SpvDecorationOffset
:
484 ctx
->type
->offsets
[member
] = dec
->literals
[0];
486 case SpvDecorationMatrixStride
:
487 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
489 case SpvDecorationColMajor
:
490 break; /* Nothing to do here. Column-major is the default. */
491 case SpvDecorationRowMajor
:
492 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
495 unreachable("Unhandled member decoration");
500 type_decoration_cb(struct vtn_builder
*b
,
501 struct vtn_value
*val
, int member
,
502 const struct vtn_decoration
*dec
, void *ctx
)
504 struct vtn_type
*type
= val
->type
;
509 switch (dec
->decoration
) {
510 case SpvDecorationArrayStride
:
511 type
->stride
= dec
->literals
[0];
513 case SpvDecorationBlock
:
516 case SpvDecorationBufferBlock
:
517 type
->buffer_block
= true;
519 case SpvDecorationGLSLShared
:
520 case SpvDecorationGLSLPacked
:
521 /* Ignore these, since we get explicit offsets anyways */
524 case SpvDecorationStream
:
525 assert(dec
->literals
[0] == 0);
529 unreachable("Unhandled type decoration");
534 translate_image_format(SpvImageFormat format
)
537 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
538 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
539 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
540 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
541 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
542 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
543 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
544 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
545 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
546 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
547 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
548 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
549 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
550 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
551 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
552 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
553 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
554 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
555 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
556 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
557 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
558 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
559 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
560 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
561 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
562 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
563 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
564 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
565 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
566 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
567 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
568 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
569 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
570 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
571 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
572 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
573 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
574 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
575 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
576 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
578 assert(!"Invalid image format");
584 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
585 const uint32_t *w
, unsigned count
)
587 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
589 val
->type
= rzalloc(b
, struct vtn_type
);
590 val
->type
->is_builtin
= false;
591 val
->type
->val
= val
;
595 val
->type
->type
= glsl_void_type();
598 val
->type
->type
= glsl_bool_type();
601 const bool signedness
= w
[3];
602 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
606 val
->type
->type
= glsl_float_type();
609 case SpvOpTypeVector
: {
610 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
611 unsigned elems
= w
[3];
613 assert(glsl_type_is_scalar(base
->type
));
614 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
616 /* Vectors implicitly have sizeof(base_type) stride. For now, this
617 * is always 4 bytes. This will have to change if we want to start
618 * supporting doubles or half-floats.
620 val
->type
->stride
= 4;
621 val
->type
->array_element
= base
;
625 case SpvOpTypeMatrix
: {
626 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
627 unsigned columns
= w
[3];
629 assert(glsl_type_is_vector(base
->type
));
630 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
631 glsl_get_vector_elements(base
->type
),
633 assert(!glsl_type_is_error(val
->type
->type
));
634 val
->type
->array_element
= base
;
635 val
->type
->row_major
= false;
636 val
->type
->stride
= 0;
640 case SpvOpTypeRuntimeArray
:
641 case SpvOpTypeArray
: {
642 struct vtn_type
*array_element
=
643 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
646 if (opcode
== SpvOpTypeRuntimeArray
) {
647 /* A length of 0 is used to denote unsized arrays */
651 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->value
.u
[0];
654 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
655 val
->type
->array_element
= array_element
;
656 val
->type
->stride
= 0;
660 case SpvOpTypeStruct
: {
661 unsigned num_fields
= count
- 2;
662 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
663 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
665 NIR_VLA(struct glsl_struct_field
, fields
, count
);
666 for (unsigned i
= 0; i
< num_fields
; i
++) {
667 val
->type
->members
[i
] =
668 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
669 fields
[i
] = (struct glsl_struct_field
) {
670 .type
= val
->type
->members
[i
]->type
,
671 .name
= ralloc_asprintf(b
, "field%d", i
),
676 struct member_decoration_ctx ctx
= {
677 .num_fields
= num_fields
,
682 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
684 const char *name
= val
->name
? val
->name
: "struct";
686 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
690 case SpvOpTypeFunction
: {
691 const struct glsl_type
*return_type
=
692 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
693 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
694 for (unsigned i
= 0; i
< count
- 3; i
++) {
695 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
699 params
[i
].out
= true;
701 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
705 case SpvOpTypePointer
:
706 /* FIXME: For now, we'll just do the really lame thing and return
707 * the same type. The validator should ensure that the proper number
708 * of dereferences happen
710 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
713 case SpvOpTypeImage
: {
714 const struct glsl_type
*sampled_type
=
715 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
717 assert(glsl_type_is_vector_or_scalar(sampled_type
));
719 enum glsl_sampler_dim dim
;
720 switch ((SpvDim
)w
[3]) {
721 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
722 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
723 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
724 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
725 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
726 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
728 unreachable("Invalid SPIR-V Sampler dimension");
731 bool is_shadow
= w
[4];
732 bool is_array
= w
[5];
733 bool multisampled
= w
[6];
734 unsigned sampled
= w
[7];
735 SpvImageFormat format
= w
[8];
738 val
->type
->access_qualifier
= w
[9];
740 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
742 assert(!multisampled
&& "FIXME: Handl multi-sampled textures");
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 * TODO: Eventually we should consider adding a "bare sampler" type
773 val
->type
->type
= glsl_sampler_type(GLSL_SAMPLER_DIM_2D
, false, false,
777 case SpvOpTypeOpaque
:
779 case SpvOpTypeDeviceEvent
:
780 case SpvOpTypeReserveId
:
784 unreachable("Unhandled opcode");
787 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
790 static nir_constant
*
791 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
793 nir_constant
*c
= rzalloc(b
, nir_constant
);
795 switch (glsl_get_base_type(type
)) {
799 case GLSL_TYPE_FLOAT
:
800 case GLSL_TYPE_DOUBLE
:
801 /* Nothing to do here. It's already initialized to zero */
804 case GLSL_TYPE_ARRAY
:
805 assert(glsl_get_length(type
) > 0);
806 c
->num_elements
= glsl_get_length(type
);
807 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
809 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
810 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
811 c
->elements
[i
] = c
->elements
[0];
814 case GLSL_TYPE_STRUCT
:
815 c
->num_elements
= glsl_get_length(type
);
816 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
818 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
819 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
824 unreachable("Invalid type for null constant");
831 spec_constant_deocoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
832 int member
, const struct vtn_decoration
*dec
,
835 assert(member
== -1);
836 if (dec
->decoration
!= SpvDecorationSpecId
)
839 uint32_t *const_value
= data
;
841 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
842 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
843 *const_value
= b
->specializations
[i
].data
;
850 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
851 uint32_t const_value
)
853 vtn_foreach_decoration(b
, val
, spec_constant_deocoration_cb
, &const_value
);
858 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
859 const uint32_t *w
, unsigned count
)
861 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
862 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
863 val
->constant
= rzalloc(b
, nir_constant
);
865 case SpvOpConstantTrue
:
866 assert(val
->const_type
== glsl_bool_type());
867 val
->constant
->value
.u
[0] = NIR_TRUE
;
869 case SpvOpConstantFalse
:
870 assert(val
->const_type
== glsl_bool_type());
871 val
->constant
->value
.u
[0] = NIR_FALSE
;
874 case SpvOpSpecConstantTrue
:
875 case SpvOpSpecConstantFalse
: {
876 assert(val
->const_type
== glsl_bool_type());
878 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
879 val
->constant
->value
.u
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
884 assert(glsl_type_is_scalar(val
->const_type
));
885 val
->constant
->value
.u
[0] = w
[3];
887 case SpvOpSpecConstant
:
888 assert(glsl_type_is_scalar(val
->const_type
));
889 val
->constant
->value
.u
[0] = get_specialization(b
, val
, w
[3]);
891 case SpvOpSpecConstantComposite
:
892 case SpvOpConstantComposite
: {
893 unsigned elem_count
= count
- 3;
894 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
895 for (unsigned i
= 0; i
< elem_count
; i
++)
896 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
898 switch (glsl_get_base_type(val
->const_type
)) {
901 case GLSL_TYPE_FLOAT
:
903 if (glsl_type_is_matrix(val
->const_type
)) {
904 unsigned rows
= glsl_get_vector_elements(val
->const_type
);
905 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
906 for (unsigned i
= 0; i
< elem_count
; i
++)
907 for (unsigned j
= 0; j
< rows
; j
++)
908 val
->constant
->value
.u
[rows
* i
+ j
] = elems
[i
]->value
.u
[j
];
910 assert(glsl_type_is_vector(val
->const_type
));
911 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
912 for (unsigned i
= 0; i
< elem_count
; i
++)
913 val
->constant
->value
.u
[i
] = elems
[i
]->value
.u
[0];
918 case GLSL_TYPE_STRUCT
:
919 case GLSL_TYPE_ARRAY
:
920 ralloc_steal(val
->constant
, elems
);
921 val
->constant
->num_elements
= elem_count
;
922 val
->constant
->elements
= elems
;
926 unreachable("Unsupported type for constants");
931 case SpvOpSpecConstantOp
: {
932 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
934 case SpvOpVectorShuffle
: {
935 struct vtn_value
*v0
= vtn_value(b
, w
[4], vtn_value_type_constant
);
936 struct vtn_value
*v1
= vtn_value(b
, w
[5], vtn_value_type_constant
);
937 unsigned len0
= glsl_get_vector_elements(v0
->const_type
);
938 unsigned len1
= glsl_get_vector_elements(v1
->const_type
);
941 for (unsigned i
= 0; i
< len0
; i
++)
942 u
[i
] = v0
->constant
->value
.u
[i
];
943 for (unsigned i
= 0; i
< len1
; i
++)
944 u
[len0
+ i
] = v1
->constant
->value
.u
[i
];
946 for (unsigned i
= 0; i
< count
- 6; i
++) {
947 uint32_t comp
= w
[i
+ 6];
948 if (comp
== (uint32_t)-1) {
949 val
->constant
->value
.u
[i
] = 0xdeadbeef;
951 val
->constant
->value
.u
[i
] = u
[comp
];
957 case SpvOpCompositeExtract
:
958 case SpvOpCompositeInsert
: {
959 struct vtn_value
*comp
;
960 unsigned deref_start
;
961 struct nir_constant
**c
;
962 if (opcode
== SpvOpCompositeExtract
) {
963 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
967 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
969 val
->constant
= nir_constant_clone(comp
->constant
,
975 const struct glsl_type
*type
= comp
->const_type
;
976 for (unsigned i
= deref_start
; i
< count
; i
++) {
977 switch (glsl_get_base_type(type
)) {
980 case GLSL_TYPE_FLOAT
:
982 /* If we hit this granularity, we're picking off an element */
986 if (glsl_type_is_matrix(type
)) {
987 elem
+= w
[i
] * glsl_get_vector_elements(type
);
988 type
= glsl_get_column_type(type
);
990 assert(glsl_type_is_vector(type
));
992 type
= glsl_scalar_type(glsl_get_base_type(type
));
996 case GLSL_TYPE_ARRAY
:
997 c
= &(*c
)->elements
[w
[i
]];
998 type
= glsl_get_array_element(type
);
1001 case GLSL_TYPE_STRUCT
:
1002 c
= &(*c
)->elements
[w
[i
]];
1003 type
= glsl_get_struct_field(type
, w
[i
]);
1007 unreachable("Invalid constant type");
1011 if (opcode
== SpvOpCompositeExtract
) {
1015 unsigned num_components
= glsl_get_vector_elements(type
);
1016 for (unsigned i
= 0; i
< num_components
; i
++)
1017 val
->constant
->value
.u
[i
] = (*c
)->value
.u
[elem
+ i
];
1020 struct vtn_value
*insert
=
1021 vtn_value(b
, w
[4], vtn_value_type_constant
);
1022 assert(insert
->const_type
== type
);
1024 *c
= insert
->constant
;
1026 unsigned num_components
= glsl_get_vector_elements(type
);
1027 for (unsigned i
= 0; i
< num_components
; i
++)
1028 (*c
)->value
.u
[elem
+ i
] = insert
->constant
->value
.u
[i
];
1036 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
);
1038 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1040 nir_const_value src
[3];
1042 for (unsigned i
= 0; i
< count
- 4; i
++) {
1044 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1046 unsigned j
= swap
? 1 - i
: i
;
1047 for (unsigned k
= 0; k
< num_components
; k
++)
1048 src
[j
].u
[k
] = c
->value
.u
[k
];
1051 nir_const_value res
= nir_eval_const_opcode(op
, num_components
, src
);
1053 for (unsigned k
= 0; k
< num_components
; k
++)
1054 val
->constant
->value
.u
[k
] = res
.u
[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 nir_tex_src srcs
[8]; /* 8 should be enough */
1203 nir_tex_src
*p
= srcs
;
1207 bool has_coord
= false;
1209 case SpvOpImageSampleImplicitLod
:
1210 case SpvOpImageSampleExplicitLod
:
1211 case SpvOpImageSampleDrefImplicitLod
:
1212 case SpvOpImageSampleDrefExplicitLod
:
1213 case SpvOpImageSampleProjImplicitLod
:
1214 case SpvOpImageSampleProjExplicitLod
:
1215 case SpvOpImageSampleProjDrefImplicitLod
:
1216 case SpvOpImageSampleProjDrefExplicitLod
:
1217 case SpvOpImageFetch
:
1218 case SpvOpImageGather
:
1219 case SpvOpImageDrefGather
:
1220 case SpvOpImageQueryLod
: {
1221 /* All these types have the coordinate as their first real argument */
1222 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, w
[idx
++]);
1224 p
->src
= nir_src_for_ssa(coord
->def
);
1225 p
->src_type
= nir_tex_src_coord
;
1234 /* These all have an explicit depth value as their next source */
1236 case SpvOpImageSampleDrefImplicitLod
:
1237 case SpvOpImageSampleDrefExplicitLod
:
1238 case SpvOpImageSampleProjDrefImplicitLod
:
1239 case SpvOpImageSampleProjDrefExplicitLod
:
1240 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparitor
);
1246 /* For OpImageQuerySizeLod, we always have an LOD */
1247 if (opcode
== SpvOpImageQuerySizeLod
)
1248 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1250 /* Figure out the base texture operation */
1253 case SpvOpImageSampleImplicitLod
:
1254 case SpvOpImageSampleDrefImplicitLod
:
1255 case SpvOpImageSampleProjImplicitLod
:
1256 case SpvOpImageSampleProjDrefImplicitLod
:
1257 texop
= nir_texop_tex
;
1260 case SpvOpImageSampleExplicitLod
:
1261 case SpvOpImageSampleDrefExplicitLod
:
1262 case SpvOpImageSampleProjExplicitLod
:
1263 case SpvOpImageSampleProjDrefExplicitLod
:
1264 texop
= nir_texop_txl
;
1267 case SpvOpImageFetch
:
1268 texop
= nir_texop_txf
;
1271 case SpvOpImageGather
:
1272 case SpvOpImageDrefGather
:
1273 texop
= nir_texop_tg4
;
1276 case SpvOpImageQuerySizeLod
:
1277 case SpvOpImageQuerySize
:
1278 texop
= nir_texop_txs
;
1281 case SpvOpImageQueryLod
:
1282 texop
= nir_texop_lod
;
1285 case SpvOpImageQueryLevels
:
1286 texop
= nir_texop_query_levels
;
1289 case SpvOpImageQuerySamples
:
1291 unreachable("Unhandled opcode");
1294 nir_constant
*const_offset
= NULL
;
1296 /* Now we need to handle some number of optional arguments */
1298 uint32_t operands
= w
[idx
++];
1300 if (operands
& SpvImageOperandsBiasMask
) {
1301 assert(texop
== nir_texop_tex
);
1302 texop
= nir_texop_txb
;
1303 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1306 if (operands
& SpvImageOperandsLodMask
) {
1307 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1308 texop
== nir_texop_txs
);
1309 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1312 if (operands
& SpvImageOperandsGradMask
) {
1313 assert(texop
== nir_texop_tex
);
1314 texop
= nir_texop_txd
;
1315 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1316 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1319 if (operands
& SpvImageOperandsConstOffsetMask
) {
1321 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
;
1324 if (operands
& SpvImageOperandsOffsetMask
)
1325 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1327 if (operands
& SpvImageOperandsConstOffsetsMask
)
1328 assert(!"Constant offsets to texture gather not yet implemented");
1330 if (operands
& SpvImageOperandsSampleMask
) {
1331 assert(texop
== nir_texop_txf
);
1332 texop
= nir_texop_txf_ms
;
1333 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1336 /* We should have now consumed exactly all of the arguments */
1337 assert(idx
== count
);
1339 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1342 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1344 const struct glsl_type
*image_type
;
1345 if (sampled
.image
) {
1346 image_type
= sampled
.image
->var
->var
->interface_type
;
1348 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1351 instr
->sampler_dim
= glsl_get_sampler_dim(image_type
);
1352 instr
->is_array
= glsl_sampler_type_is_array(image_type
);
1353 instr
->is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1354 instr
->is_new_style_shadow
= instr
->is_shadow
;
1357 for (unsigned i
= 0; i
< 4; i
++)
1358 instr
->const_offset
[i
] = const_offset
->value
.u
[i
];
1362 switch (instr
->sampler_dim
) {
1363 case GLSL_SAMPLER_DIM_1D
:
1364 case GLSL_SAMPLER_DIM_BUF
:
1365 instr
->coord_components
= 1;
1367 case GLSL_SAMPLER_DIM_2D
:
1368 case GLSL_SAMPLER_DIM_RECT
:
1369 instr
->coord_components
= 2;
1371 case GLSL_SAMPLER_DIM_3D
:
1372 case GLSL_SAMPLER_DIM_CUBE
:
1373 case GLSL_SAMPLER_DIM_MS
:
1374 instr
->coord_components
= 3;
1377 assert("Invalid sampler type");
1380 if (instr
->is_array
)
1381 instr
->coord_components
++;
1383 instr
->coord_components
= 0;
1386 switch (glsl_get_sampler_result_type(image_type
)) {
1387 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1388 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1389 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1390 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1392 unreachable("Invalid base type for sampler result");
1395 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1396 if (sampled
.image
) {
1397 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1398 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &image
->deref
));
1400 instr
->texture
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1403 switch (instr
->op
) {
1408 /* These operations require a sampler */
1409 instr
->sampler
= nir_deref_as_var(nir_copy_deref(instr
, &sampler
->deref
));
1412 case nir_texop_txf_ms
:
1416 case nir_texop_query_levels
:
1417 case nir_texop_texture_samples
:
1418 case nir_texop_samples_identical
:
1420 instr
->sampler
= NULL
;
1424 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1425 nir_tex_instr_dest_size(instr
), NULL
);
1427 assert(glsl_get_vector_elements(ret_type
->type
) ==
1428 nir_tex_instr_dest_size(instr
));
1430 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1431 val
->ssa
->def
= &instr
->dest
.ssa
;
1433 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1436 static nir_ssa_def
*
1437 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1439 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1441 /* The image_load_store intrinsics assume a 4-dim coordinate */
1442 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1443 unsigned swizzle
[4];
1444 for (unsigned i
= 0; i
< 4; i
++)
1445 swizzle
[i
] = MIN2(i
, dim
- 1);
1447 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1451 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1452 const uint32_t *w
, unsigned count
)
1454 /* Just get this one out of the way */
1455 if (opcode
== SpvOpImageTexelPointer
) {
1456 struct vtn_value
*val
=
1457 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1458 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1461 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1462 val
->image
->coord
= get_image_coord(b
, w
[4]);
1463 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1467 struct vtn_image_pointer image
;
1470 case SpvOpAtomicExchange
:
1471 case SpvOpAtomicCompareExchange
:
1472 case SpvOpAtomicCompareExchangeWeak
:
1473 case SpvOpAtomicIIncrement
:
1474 case SpvOpAtomicIDecrement
:
1475 case SpvOpAtomicIAdd
:
1476 case SpvOpAtomicISub
:
1477 case SpvOpAtomicSMin
:
1478 case SpvOpAtomicUMin
:
1479 case SpvOpAtomicSMax
:
1480 case SpvOpAtomicUMax
:
1481 case SpvOpAtomicAnd
:
1483 case SpvOpAtomicXor
:
1484 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1487 case SpvOpImageQuerySize
:
1489 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1491 image
.sample
= NULL
;
1494 case SpvOpImageRead
:
1496 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1497 image
.coord
= get_image_coord(b
, w
[4]);
1499 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1500 assert(w
[5] == SpvImageOperandsSampleMask
);
1501 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1503 image
.sample
= nir_ssa_undef(&b
->nb
, 1);
1507 case SpvOpImageWrite
:
1509 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1510 image
.coord
= get_image_coord(b
, w
[2]);
1514 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1515 assert(w
[4] == SpvImageOperandsSampleMask
);
1516 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1518 image
.sample
= nir_ssa_undef(&b
->nb
, 1);
1523 unreachable("Invalid image opcode");
1526 nir_intrinsic_op op
;
1528 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1529 OP(ImageQuerySize
, size
)
1531 OP(ImageWrite
, store
)
1532 OP(AtomicExchange
, atomic_exchange
)
1533 OP(AtomicCompareExchange
, atomic_comp_swap
)
1534 OP(AtomicIIncrement
, atomic_add
)
1535 OP(AtomicIDecrement
, atomic_add
)
1536 OP(AtomicIAdd
, atomic_add
)
1537 OP(AtomicISub
, atomic_add
)
1538 OP(AtomicSMin
, atomic_min
)
1539 OP(AtomicUMin
, atomic_min
)
1540 OP(AtomicSMax
, atomic_max
)
1541 OP(AtomicUMax
, atomic_max
)
1542 OP(AtomicAnd
, atomic_and
)
1543 OP(AtomicOr
, atomic_or
)
1544 OP(AtomicXor
, atomic_xor
)
1547 unreachable("Invalid image opcode");
1550 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1552 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1553 intrin
->variables
[0] =
1554 nir_deref_as_var(nir_copy_deref(&intrin
->instr
, &image_deref
->deref
));
1556 /* ImageQuerySize doesn't take any extra parameters */
1557 if (opcode
!= SpvOpImageQuerySize
) {
1558 /* The image coordinate is always 4 components but we may not have that
1559 * many. Swizzle to compensate.
1562 for (unsigned i
= 0; i
< 4; i
++)
1563 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1564 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1566 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1570 case SpvOpImageQuerySize
:
1571 case SpvOpImageRead
:
1573 case SpvOpImageWrite
:
1574 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1576 case SpvOpAtomicIIncrement
:
1577 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1579 case SpvOpAtomicIDecrement
:
1580 intrin
->src
[2] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1583 case SpvOpAtomicExchange
:
1584 case SpvOpAtomicIAdd
:
1585 case SpvOpAtomicSMin
:
1586 case SpvOpAtomicUMin
:
1587 case SpvOpAtomicSMax
:
1588 case SpvOpAtomicUMax
:
1589 case SpvOpAtomicAnd
:
1591 case SpvOpAtomicXor
:
1592 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1595 case SpvOpAtomicCompareExchange
:
1596 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1597 intrin
->src
[3] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1600 case SpvOpAtomicISub
:
1601 intrin
->src
[2] = nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1605 unreachable("Invalid image opcode");
1608 if (opcode
!= SpvOpImageWrite
) {
1609 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1610 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1611 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, NULL
);
1613 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1615 /* The image intrinsics always return 4 channels but we may not want
1616 * that many. Emit a mov to trim it down.
1618 unsigned swiz
[4] = {0, 1, 2, 3};
1619 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1620 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1621 glsl_get_vector_elements(type
->type
), false);
1623 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1627 static nir_intrinsic_op
1628 get_ssbo_nir_atomic_op(SpvOp opcode
)
1631 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
1632 OP(AtomicExchange
, atomic_exchange
)
1633 OP(AtomicCompareExchange
, atomic_comp_swap
)
1634 OP(AtomicIIncrement
, atomic_add
)
1635 OP(AtomicIDecrement
, atomic_add
)
1636 OP(AtomicIAdd
, atomic_add
)
1637 OP(AtomicISub
, atomic_add
)
1638 OP(AtomicSMin
, atomic_imin
)
1639 OP(AtomicUMin
, atomic_umin
)
1640 OP(AtomicSMax
, atomic_imax
)
1641 OP(AtomicUMax
, atomic_umax
)
1642 OP(AtomicAnd
, atomic_and
)
1643 OP(AtomicOr
, atomic_or
)
1644 OP(AtomicXor
, atomic_xor
)
1647 unreachable("Invalid SSBO atomic");
1651 static nir_intrinsic_op
1652 get_shared_nir_atomic_op(SpvOp opcode
)
1655 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
1656 OP(AtomicExchange
, atomic_exchange
)
1657 OP(AtomicCompareExchange
, atomic_comp_swap
)
1658 OP(AtomicIIncrement
, atomic_add
)
1659 OP(AtomicIDecrement
, atomic_add
)
1660 OP(AtomicIAdd
, atomic_add
)
1661 OP(AtomicISub
, atomic_add
)
1662 OP(AtomicSMin
, atomic_imin
)
1663 OP(AtomicUMin
, atomic_umin
)
1664 OP(AtomicSMax
, atomic_imax
)
1665 OP(AtomicUMax
, atomic_umax
)
1666 OP(AtomicAnd
, atomic_and
)
1667 OP(AtomicOr
, atomic_or
)
1668 OP(AtomicXor
, atomic_xor
)
1671 unreachable("Invalid shared atomic");
1676 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1677 const uint32_t *w
, nir_src
*src
)
1680 case SpvOpAtomicIIncrement
:
1681 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1684 case SpvOpAtomicIDecrement
:
1685 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1688 case SpvOpAtomicISub
:
1690 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1693 case SpvOpAtomicCompareExchange
:
1694 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1695 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1699 case SpvOpAtomicExchange
:
1700 case SpvOpAtomicIAdd
:
1701 case SpvOpAtomicSMin
:
1702 case SpvOpAtomicUMin
:
1703 case SpvOpAtomicSMax
:
1704 case SpvOpAtomicUMax
:
1705 case SpvOpAtomicAnd
:
1707 case SpvOpAtomicXor
:
1708 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1712 unreachable("Invalid SPIR-V atomic");
1717 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
1718 const uint32_t *w
, unsigned count
)
1720 struct vtn_access_chain
*chain
=
1721 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1722 nir_intrinsic_instr
*atomic
;
1725 SpvScope scope = w[4];
1726 SpvMemorySemanticsMask semantics = w[5];
1729 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
1730 nir_deref
*deref
= &vtn_access_chain_to_deref(b
, chain
)->deref
;
1731 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
1732 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1733 atomic
->variables
[0] = nir_deref_as_var(nir_copy_deref(atomic
, deref
));
1734 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
1736 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
1737 struct vtn_type
*type
;
1738 nir_ssa_def
*offset
, *index
;
1739 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
1741 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
1743 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
1744 atomic
->src
[0] = nir_src_for_ssa(index
);
1745 atomic
->src
[1] = nir_src_for_ssa(offset
);
1746 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
1749 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
, 1, NULL
);
1751 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1752 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1753 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
1754 val
->ssa
->def
= &atomic
->dest
.ssa
;
1755 val
->ssa
->type
= type
->type
;
1757 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
1760 static nir_alu_instr
*
1761 create_vec(nir_shader
*shader
, unsigned num_components
)
1764 switch (num_components
) {
1765 case 1: op
= nir_op_fmov
; break;
1766 case 2: op
= nir_op_vec2
; break;
1767 case 3: op
= nir_op_vec3
; break;
1768 case 4: op
= nir_op_vec4
; break;
1769 default: unreachable("bad vector size");
1772 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
1773 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
, NULL
);
1774 vec
->dest
.write_mask
= (1 << num_components
) - 1;
1779 struct vtn_ssa_value
*
1780 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
1782 if (src
->transposed
)
1783 return src
->transposed
;
1785 struct vtn_ssa_value
*dest
=
1786 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
1788 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
1789 nir_alu_instr
*vec
= create_vec(b
->shader
,
1790 glsl_get_matrix_columns(src
->type
));
1791 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1792 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
1793 vec
->src
[0].swizzle
[0] = i
;
1795 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
1796 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
1797 vec
->src
[j
].swizzle
[0] = i
;
1800 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1801 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
1804 dest
->transposed
= src
;
1810 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
1812 unsigned swiz
[4] = { index
};
1813 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
1817 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
1820 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
);
1822 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
1824 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
1826 vec
->src
[i
].src
= nir_src_for_ssa(src
);
1827 vec
->src
[i
].swizzle
[0] = i
;
1831 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1833 return &vec
->dest
.dest
.ssa
;
1837 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1840 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
1841 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1842 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1843 vtn_vector_extract(b
, src
, i
), dest
);
1849 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
1850 nir_ssa_def
*insert
, nir_ssa_def
*index
)
1852 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
1853 for (unsigned i
= 1; i
< src
->num_components
; i
++)
1854 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
1855 vtn_vector_insert(b
, src
, insert
, i
), dest
);
1860 static nir_ssa_def
*
1861 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
1862 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
1863 const uint32_t *indices
)
1865 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
);
1867 nir_ssa_undef_instr
*undef
= nir_ssa_undef_instr_create(b
->shader
, 1);
1868 nir_builder_instr_insert(&b
->nb
, &undef
->instr
);
1870 for (unsigned i
= 0; i
< num_components
; i
++) {
1871 uint32_t index
= indices
[i
];
1872 if (index
== 0xffffffff) {
1873 vec
->src
[i
].src
= nir_src_for_ssa(&undef
->def
);
1874 } else if (index
< src0
->num_components
) {
1875 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
1876 vec
->src
[i
].swizzle
[0] = index
;
1878 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
1879 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
1883 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1885 return &vec
->dest
.dest
.ssa
;
1889 * Concatentates a number of vectors/scalars together to produce a vector
1891 static nir_ssa_def
*
1892 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
1893 unsigned num_srcs
, nir_ssa_def
**srcs
)
1895 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
);
1897 unsigned dest_idx
= 0;
1898 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1899 nir_ssa_def
*src
= srcs
[i
];
1900 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
1901 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
1902 vec
->src
[dest_idx
].swizzle
[0] = j
;
1907 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
1909 return &vec
->dest
.dest
.ssa
;
1912 static struct vtn_ssa_value
*
1913 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
1915 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
1916 dest
->type
= src
->type
;
1918 if (glsl_type_is_vector_or_scalar(src
->type
)) {
1919 dest
->def
= src
->def
;
1921 unsigned elems
= glsl_get_length(src
->type
);
1923 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
1924 for (unsigned i
= 0; i
< elems
; i
++)
1925 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
1931 static struct vtn_ssa_value
*
1932 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1933 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
1934 unsigned num_indices
)
1936 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
1938 struct vtn_ssa_value
*cur
= dest
;
1940 for (i
= 0; i
< num_indices
- 1; i
++) {
1941 cur
= cur
->elems
[indices
[i
]];
1944 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1945 /* According to the SPIR-V spec, OpCompositeInsert may work down to
1946 * the component granularity. In that case, the last index will be
1947 * the index to insert the scalar into the vector.
1950 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
1952 cur
->elems
[indices
[i
]] = insert
;
1958 static struct vtn_ssa_value
*
1959 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1960 const uint32_t *indices
, unsigned num_indices
)
1962 struct vtn_ssa_value
*cur
= src
;
1963 for (unsigned i
= 0; i
< num_indices
; i
++) {
1964 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
1965 assert(i
== num_indices
- 1);
1966 /* According to the SPIR-V spec, OpCompositeExtract may work down to
1967 * the component granularity. The last index will be the index of the
1968 * vector to extract.
1971 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
1972 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
1973 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
1976 cur
= cur
->elems
[indices
[i
]];
1984 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
1985 const uint32_t *w
, unsigned count
)
1987 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1988 const struct glsl_type
*type
=
1989 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1990 val
->ssa
= vtn_create_ssa_value(b
, type
);
1993 case SpvOpVectorExtractDynamic
:
1994 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
1995 vtn_ssa_value(b
, w
[4])->def
);
1998 case SpvOpVectorInsertDynamic
:
1999 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2000 vtn_ssa_value(b
, w
[4])->def
,
2001 vtn_ssa_value(b
, w
[5])->def
);
2004 case SpvOpVectorShuffle
:
2005 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2006 vtn_ssa_value(b
, w
[3])->def
,
2007 vtn_ssa_value(b
, w
[4])->def
,
2011 case SpvOpCompositeConstruct
: {
2012 unsigned elems
= count
- 3;
2013 if (glsl_type_is_vector_or_scalar(type
)) {
2014 nir_ssa_def
*srcs
[4];
2015 for (unsigned i
= 0; i
< elems
; i
++)
2016 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2018 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2021 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2022 for (unsigned i
= 0; i
< elems
; i
++)
2023 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2027 case SpvOpCompositeExtract
:
2028 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2032 case SpvOpCompositeInsert
:
2033 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2034 vtn_ssa_value(b
, w
[3]),
2038 case SpvOpCopyObject
:
2039 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2043 unreachable("unknown composite operation");
2048 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2049 const uint32_t *w
, unsigned count
)
2051 nir_intrinsic_op intrinsic_op
;
2053 case SpvOpEmitVertex
:
2054 case SpvOpEmitStreamVertex
:
2055 intrinsic_op
= nir_intrinsic_emit_vertex
;
2057 case SpvOpEndPrimitive
:
2058 case SpvOpEndStreamPrimitive
:
2059 intrinsic_op
= nir_intrinsic_end_primitive
;
2061 case SpvOpMemoryBarrier
:
2062 intrinsic_op
= nir_intrinsic_memory_barrier
;
2064 case SpvOpControlBarrier
:
2065 intrinsic_op
= nir_intrinsic_barrier
;
2068 unreachable("unknown barrier instruction");
2071 nir_intrinsic_instr
*intrin
=
2072 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2074 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2075 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2077 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2081 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2084 case SpvExecutionModeInputPoints
:
2085 case SpvExecutionModeOutputPoints
:
2086 return 0; /* GL_POINTS */
2087 case SpvExecutionModeInputLines
:
2088 return 1; /* GL_LINES */
2089 case SpvExecutionModeInputLinesAdjacency
:
2090 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2091 case SpvExecutionModeTriangles
:
2092 return 4; /* GL_TRIANGLES */
2093 case SpvExecutionModeInputTrianglesAdjacency
:
2094 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2095 case SpvExecutionModeQuads
:
2096 return 7; /* GL_QUADS */
2097 case SpvExecutionModeIsolines
:
2098 return 0x8E7A; /* GL_ISOLINES */
2099 case SpvExecutionModeOutputLineStrip
:
2100 return 3; /* GL_LINE_STRIP */
2101 case SpvExecutionModeOutputTriangleStrip
:
2102 return 5; /* GL_TRIANGLE_STRIP */
2104 assert(!"Invalid primitive type");
2110 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2113 case SpvExecutionModeInputPoints
:
2115 case SpvExecutionModeInputLines
:
2117 case SpvExecutionModeInputLinesAdjacency
:
2119 case SpvExecutionModeTriangles
:
2121 case SpvExecutionModeInputTrianglesAdjacency
:
2124 assert(!"Invalid GS input mode");
2129 static gl_shader_stage
2130 stage_for_execution_model(SpvExecutionModel model
)
2133 case SpvExecutionModelVertex
:
2134 return MESA_SHADER_VERTEX
;
2135 case SpvExecutionModelTessellationControl
:
2136 return MESA_SHADER_TESS_CTRL
;
2137 case SpvExecutionModelTessellationEvaluation
:
2138 return MESA_SHADER_TESS_EVAL
;
2139 case SpvExecutionModelGeometry
:
2140 return MESA_SHADER_GEOMETRY
;
2141 case SpvExecutionModelFragment
:
2142 return MESA_SHADER_FRAGMENT
;
2143 case SpvExecutionModelGLCompute
:
2144 return MESA_SHADER_COMPUTE
;
2146 unreachable("Unsupported execution model");
2151 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2152 const uint32_t *w
, unsigned count
)
2156 case SpvOpSourceExtension
:
2157 case SpvOpSourceContinued
:
2158 case SpvOpExtension
:
2159 /* Unhandled, but these are for debug so that's ok. */
2162 case SpvOpCapability
:
2163 switch ((SpvCapability
)w
[1]) {
2164 case SpvCapabilityMatrix
:
2165 case SpvCapabilityShader
:
2166 case SpvCapabilityGeometry
:
2169 assert(!"Unsupported capability");
2173 case SpvOpExtInstImport
:
2174 vtn_handle_extension(b
, opcode
, w
, count
);
2177 case SpvOpMemoryModel
:
2178 assert(w
[1] == SpvAddressingModelLogical
);
2179 assert(w
[2] == SpvMemoryModelGLSL450
);
2182 case SpvOpEntryPoint
: {
2183 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2184 /* Let this be a name label regardless */
2185 unsigned name_words
;
2186 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2188 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2189 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2192 assert(b
->entry_point
== NULL
);
2193 b
->entry_point
= entry_point
;
2198 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2199 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2203 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2206 case SpvOpMemberName
:
2210 case SpvOpExecutionMode
:
2211 case SpvOpDecorationGroup
:
2213 case SpvOpMemberDecorate
:
2214 case SpvOpGroupDecorate
:
2215 case SpvOpGroupMemberDecorate
:
2216 vtn_handle_decoration(b
, opcode
, w
, count
);
2220 return false; /* End of preamble */
2227 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2228 const struct vtn_decoration
*mode
, void *data
)
2230 assert(b
->entry_point
== entry_point
);
2232 switch(mode
->exec_mode
) {
2233 case SpvExecutionModeOriginUpperLeft
:
2234 case SpvExecutionModeOriginLowerLeft
:
2235 b
->origin_upper_left
=
2236 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2239 case SpvExecutionModeEarlyFragmentTests
:
2240 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2241 b
->shader
->info
.fs
.early_fragment_tests
= true;
2244 case SpvExecutionModeInvocations
:
2245 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2246 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2249 case SpvExecutionModeDepthReplacing
:
2250 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2251 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2253 case SpvExecutionModeDepthGreater
:
2254 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2255 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2257 case SpvExecutionModeDepthLess
:
2258 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2259 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2261 case SpvExecutionModeDepthUnchanged
:
2262 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2263 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2266 case SpvExecutionModeLocalSize
:
2267 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2268 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2269 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2270 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2272 case SpvExecutionModeLocalSizeHint
:
2273 break; /* Nothing do do with this */
2275 case SpvExecutionModeOutputVertices
:
2276 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2277 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2280 case SpvExecutionModeInputPoints
:
2281 case SpvExecutionModeInputLines
:
2282 case SpvExecutionModeInputLinesAdjacency
:
2283 case SpvExecutionModeTriangles
:
2284 case SpvExecutionModeInputTrianglesAdjacency
:
2285 case SpvExecutionModeQuads
:
2286 case SpvExecutionModeIsolines
:
2287 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
) {
2288 b
->shader
->info
.gs
.vertices_in
=
2289 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2291 assert(!"Tesselation shaders not yet supported");
2295 case SpvExecutionModeOutputPoints
:
2296 case SpvExecutionModeOutputLineStrip
:
2297 case SpvExecutionModeOutputTriangleStrip
:
2298 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2299 b
->shader
->info
.gs
.output_primitive
=
2300 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2303 case SpvExecutionModeSpacingEqual
:
2304 case SpvExecutionModeSpacingFractionalEven
:
2305 case SpvExecutionModeSpacingFractionalOdd
:
2306 case SpvExecutionModeVertexOrderCw
:
2307 case SpvExecutionModeVertexOrderCcw
:
2308 case SpvExecutionModePointMode
:
2309 assert(!"TODO: Add tessellation metadata");
2312 case SpvExecutionModePixelCenterInteger
:
2313 case SpvExecutionModeXfb
:
2314 assert(!"Unhandled execution mode");
2317 case SpvExecutionModeVecTypeHint
:
2318 case SpvExecutionModeContractionOff
:
2324 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2325 const uint32_t *w
, unsigned count
)
2329 case SpvOpSourceContinued
:
2330 case SpvOpSourceExtension
:
2331 case SpvOpExtension
:
2332 case SpvOpCapability
:
2333 case SpvOpExtInstImport
:
2334 case SpvOpMemoryModel
:
2335 case SpvOpEntryPoint
:
2336 case SpvOpExecutionMode
:
2339 case SpvOpMemberName
:
2340 case SpvOpDecorationGroup
:
2342 case SpvOpMemberDecorate
:
2343 case SpvOpGroupDecorate
:
2344 case SpvOpGroupMemberDecorate
:
2345 assert(!"Invalid opcode types and variables section");
2351 case SpvOpTypeFloat
:
2352 case SpvOpTypeVector
:
2353 case SpvOpTypeMatrix
:
2354 case SpvOpTypeImage
:
2355 case SpvOpTypeSampler
:
2356 case SpvOpTypeSampledImage
:
2357 case SpvOpTypeArray
:
2358 case SpvOpTypeRuntimeArray
:
2359 case SpvOpTypeStruct
:
2360 case SpvOpTypeOpaque
:
2361 case SpvOpTypePointer
:
2362 case SpvOpTypeFunction
:
2363 case SpvOpTypeEvent
:
2364 case SpvOpTypeDeviceEvent
:
2365 case SpvOpTypeReserveId
:
2366 case SpvOpTypeQueue
:
2368 vtn_handle_type(b
, opcode
, w
, count
);
2371 case SpvOpConstantTrue
:
2372 case SpvOpConstantFalse
:
2374 case SpvOpConstantComposite
:
2375 case SpvOpConstantSampler
:
2376 case SpvOpConstantNull
:
2377 case SpvOpSpecConstantTrue
:
2378 case SpvOpSpecConstantFalse
:
2379 case SpvOpSpecConstant
:
2380 case SpvOpSpecConstantComposite
:
2381 case SpvOpSpecConstantOp
:
2382 vtn_handle_constant(b
, opcode
, w
, count
);
2386 vtn_handle_variables(b
, opcode
, w
, count
);
2390 return false; /* End of preamble */
2397 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2398 const uint32_t *w
, unsigned count
)
2404 case SpvOpLoopMerge
:
2405 case SpvOpSelectionMerge
:
2406 /* This is handled by cfg pre-pass and walk_blocks */
2410 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2411 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2416 vtn_handle_extension(b
, opcode
, w
, count
);
2422 case SpvOpCopyMemory
:
2423 case SpvOpCopyMemorySized
:
2424 case SpvOpAccessChain
:
2425 case SpvOpInBoundsAccessChain
:
2426 case SpvOpArrayLength
:
2427 vtn_handle_variables(b
, opcode
, w
, count
);
2430 case SpvOpFunctionCall
:
2431 vtn_handle_function_call(b
, opcode
, w
, count
);
2434 case SpvOpSampledImage
:
2436 case SpvOpImageSampleImplicitLod
:
2437 case SpvOpImageSampleExplicitLod
:
2438 case SpvOpImageSampleDrefImplicitLod
:
2439 case SpvOpImageSampleDrefExplicitLod
:
2440 case SpvOpImageSampleProjImplicitLod
:
2441 case SpvOpImageSampleProjExplicitLod
:
2442 case SpvOpImageSampleProjDrefImplicitLod
:
2443 case SpvOpImageSampleProjDrefExplicitLod
:
2444 case SpvOpImageFetch
:
2445 case SpvOpImageGather
:
2446 case SpvOpImageDrefGather
:
2447 case SpvOpImageQuerySizeLod
:
2448 case SpvOpImageQueryLod
:
2449 case SpvOpImageQueryLevels
:
2450 case SpvOpImageQuerySamples
:
2451 vtn_handle_texture(b
, opcode
, w
, count
);
2454 case SpvOpImageRead
:
2455 case SpvOpImageWrite
:
2456 case SpvOpImageTexelPointer
:
2457 vtn_handle_image(b
, opcode
, w
, count
);
2460 case SpvOpImageQuerySize
: {
2461 struct vtn_access_chain
*image
=
2462 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2463 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
2464 vtn_handle_image(b
, opcode
, w
, count
);
2466 vtn_handle_texture(b
, opcode
, w
, count
);
2471 case SpvOpAtomicExchange
:
2472 case SpvOpAtomicCompareExchange
:
2473 case SpvOpAtomicCompareExchangeWeak
:
2474 case SpvOpAtomicIIncrement
:
2475 case SpvOpAtomicIDecrement
:
2476 case SpvOpAtomicIAdd
:
2477 case SpvOpAtomicISub
:
2478 case SpvOpAtomicSMin
:
2479 case SpvOpAtomicUMin
:
2480 case SpvOpAtomicSMax
:
2481 case SpvOpAtomicUMax
:
2482 case SpvOpAtomicAnd
:
2484 case SpvOpAtomicXor
: {
2485 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
2486 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
2487 vtn_handle_image(b
, opcode
, w
, count
);
2489 assert(pointer
->value_type
== vtn_value_type_access_chain
);
2490 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
2500 case SpvOpConvertFToU
:
2501 case SpvOpConvertFToS
:
2502 case SpvOpConvertSToF
:
2503 case SpvOpConvertUToF
:
2507 case SpvOpQuantizeToF16
:
2508 case SpvOpConvertPtrToU
:
2509 case SpvOpConvertUToPtr
:
2510 case SpvOpPtrCastToGeneric
:
2511 case SpvOpGenericCastToPtr
:
2517 case SpvOpSignBitSet
:
2518 case SpvOpLessOrGreater
:
2520 case SpvOpUnordered
:
2535 case SpvOpVectorTimesScalar
:
2537 case SpvOpIAddCarry
:
2538 case SpvOpISubBorrow
:
2539 case SpvOpUMulExtended
:
2540 case SpvOpSMulExtended
:
2541 case SpvOpShiftRightLogical
:
2542 case SpvOpShiftRightArithmetic
:
2543 case SpvOpShiftLeftLogical
:
2544 case SpvOpLogicalEqual
:
2545 case SpvOpLogicalNotEqual
:
2546 case SpvOpLogicalOr
:
2547 case SpvOpLogicalAnd
:
2548 case SpvOpLogicalNot
:
2549 case SpvOpBitwiseOr
:
2550 case SpvOpBitwiseXor
:
2551 case SpvOpBitwiseAnd
:
2554 case SpvOpFOrdEqual
:
2555 case SpvOpFUnordEqual
:
2556 case SpvOpINotEqual
:
2557 case SpvOpFOrdNotEqual
:
2558 case SpvOpFUnordNotEqual
:
2559 case SpvOpULessThan
:
2560 case SpvOpSLessThan
:
2561 case SpvOpFOrdLessThan
:
2562 case SpvOpFUnordLessThan
:
2563 case SpvOpUGreaterThan
:
2564 case SpvOpSGreaterThan
:
2565 case SpvOpFOrdGreaterThan
:
2566 case SpvOpFUnordGreaterThan
:
2567 case SpvOpULessThanEqual
:
2568 case SpvOpSLessThanEqual
:
2569 case SpvOpFOrdLessThanEqual
:
2570 case SpvOpFUnordLessThanEqual
:
2571 case SpvOpUGreaterThanEqual
:
2572 case SpvOpSGreaterThanEqual
:
2573 case SpvOpFOrdGreaterThanEqual
:
2574 case SpvOpFUnordGreaterThanEqual
:
2580 case SpvOpFwidthFine
:
2581 case SpvOpDPdxCoarse
:
2582 case SpvOpDPdyCoarse
:
2583 case SpvOpFwidthCoarse
:
2584 case SpvOpBitFieldInsert
:
2585 case SpvOpBitFieldSExtract
:
2586 case SpvOpBitFieldUExtract
:
2587 case SpvOpBitReverse
:
2589 case SpvOpTranspose
:
2590 case SpvOpOuterProduct
:
2591 case SpvOpMatrixTimesScalar
:
2592 case SpvOpVectorTimesMatrix
:
2593 case SpvOpMatrixTimesVector
:
2594 case SpvOpMatrixTimesMatrix
:
2595 vtn_handle_alu(b
, opcode
, w
, count
);
2598 case SpvOpVectorExtractDynamic
:
2599 case SpvOpVectorInsertDynamic
:
2600 case SpvOpVectorShuffle
:
2601 case SpvOpCompositeConstruct
:
2602 case SpvOpCompositeExtract
:
2603 case SpvOpCompositeInsert
:
2604 case SpvOpCopyObject
:
2605 vtn_handle_composite(b
, opcode
, w
, count
);
2608 case SpvOpEmitVertex
:
2609 case SpvOpEndPrimitive
:
2610 case SpvOpEmitStreamVertex
:
2611 case SpvOpEndStreamPrimitive
:
2612 case SpvOpControlBarrier
:
2613 case SpvOpMemoryBarrier
:
2614 vtn_handle_barrier(b
, opcode
, w
, count
);
2618 unreachable("Unhandled opcode");
2625 spirv_to_nir(const uint32_t *words
, size_t word_count
,
2626 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
2627 gl_shader_stage stage
, const char *entry_point_name
,
2628 const nir_shader_compiler_options
*options
)
2630 const uint32_t *word_end
= words
+ word_count
;
2632 /* Handle the SPIR-V header (first 4 dwords) */
2633 assert(word_count
> 5);
2635 assert(words
[0] == SpvMagicNumber
);
2636 assert(words
[1] >= 0x10000);
2637 /* words[2] == generator magic */
2638 unsigned value_id_bound
= words
[3];
2639 assert(words
[4] == 0);
2643 /* Initialize the stn_builder object */
2644 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
2645 b
->value_id_bound
= value_id_bound
;
2646 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
2647 exec_list_make_empty(&b
->functions
);
2648 b
->entry_point_stage
= stage
;
2649 b
->entry_point_name
= entry_point_name
;
2651 /* Handle all the preamble instructions */
2652 words
= vtn_foreach_instruction(b
, words
, word_end
,
2653 vtn_handle_preamble_instruction
);
2655 if (b
->entry_point
== NULL
) {
2656 assert(!"Entry point not found");
2661 b
->shader
= nir_shader_create(NULL
, stage
, options
);
2663 /* Parse execution modes */
2664 vtn_foreach_execution_mode(b
, b
->entry_point
,
2665 vtn_handle_execution_mode
, NULL
);
2667 b
->specializations
= spec
;
2668 b
->num_specializations
= num_spec
;
2670 /* Handle all variable, type, and constant instructions */
2671 words
= vtn_foreach_instruction(b
, words
, word_end
,
2672 vtn_handle_variable_or_type_instruction
);
2674 vtn_build_cfg(b
, words
, word_end
);
2676 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
2677 b
->impl
= func
->impl
;
2678 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
2679 _mesa_key_pointer_equal
);
2681 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
2684 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
2685 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
2686 assert(entry_point
);