2 * Copyright © 2012 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
21 * DEALINGS IN THE SOFTWARE.
25 * \file link_varyings.cpp
27 * Linker functions related specifically to linking varyings between shader
32 #include "main/mtypes.h"
33 #include "glsl_symbol_table.h"
34 #include "glsl_parser_extras.h"
35 #include "ir_optimization.h"
37 #include "link_varyings.h"
38 #include "main/macros.h"
39 #include "program/hash_table.h"
44 * Get the varying type stripped of the outermost array if we're processing
45 * a stage whose varyings are arrays indexed by a vertex number (such as
46 * geometry shader inputs).
48 static const glsl_type
*
49 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
51 const glsl_type
*type
= var
->type
;
53 if (!var
->data
.patch
&&
54 ((var
->data
.mode
== ir_var_shader_out
&&
55 stage
== MESA_SHADER_TESS_CTRL
) ||
56 (var
->data
.mode
== ir_var_shader_in
&&
57 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
58 stage
== MESA_SHADER_GEOMETRY
)))) {
59 assert(type
->is_array());
60 type
= type
->fields
.array
;
67 * Validate the types and qualifiers of an output from one stage against the
68 * matching input to another stage.
71 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
72 const ir_variable
*input
,
73 const ir_variable
*output
,
74 gl_shader_stage consumer_stage
,
75 gl_shader_stage producer_stage
)
77 /* Check that the types match between stages.
79 const glsl_type
*type_to_match
= input
->type
;
81 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
82 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
83 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
84 consumer_stage
== MESA_SHADER_GEOMETRY
;
85 if (extra_array_level
) {
86 assert(type_to_match
->is_array());
87 type_to_match
= type_to_match
->fields
.array
;
90 if (type_to_match
!= output
->type
) {
91 /* There is a bit of a special case for gl_TexCoord. This
92 * built-in is unsized by default. Applications that variable
93 * access it must redeclare it with a size. There is some
94 * language in the GLSL spec that implies the fragment shader
95 * and vertex shader do not have to agree on this size. Other
96 * driver behave this way, and one or two applications seem to
99 * Neither declaration needs to be modified here because the array
100 * sizes are fixed later when update_array_sizes is called.
102 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
104 * "Unlike user-defined varying variables, the built-in
105 * varying variables don't have a strict one-to-one
106 * correspondence between the vertex language and the
107 * fragment language."
109 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
111 "%s shader output `%s' declared as type `%s', "
112 "but %s shader input declared as type `%s'\n",
113 _mesa_shader_stage_to_string(producer_stage
),
116 _mesa_shader_stage_to_string(consumer_stage
),
122 /* Check that all of the qualifiers match between stages.
124 if (input
->data
.centroid
!= output
->data
.centroid
) {
126 "%s shader output `%s' %s centroid qualifier, "
127 "but %s shader input %s centroid qualifier\n",
128 _mesa_shader_stage_to_string(producer_stage
),
130 (output
->data
.centroid
) ? "has" : "lacks",
131 _mesa_shader_stage_to_string(consumer_stage
),
132 (input
->data
.centroid
) ? "has" : "lacks");
136 if (input
->data
.sample
!= output
->data
.sample
) {
138 "%s shader output `%s' %s sample qualifier, "
139 "but %s shader input %s sample qualifier\n",
140 _mesa_shader_stage_to_string(producer_stage
),
142 (output
->data
.sample
) ? "has" : "lacks",
143 _mesa_shader_stage_to_string(consumer_stage
),
144 (input
->data
.sample
) ? "has" : "lacks");
148 if (input
->data
.patch
!= output
->data
.patch
) {
150 "%s shader output `%s' %s patch qualifier, "
151 "but %s shader input %s patch qualifier\n",
152 _mesa_shader_stage_to_string(producer_stage
),
154 (output
->data
.patch
) ? "has" : "lacks",
155 _mesa_shader_stage_to_string(consumer_stage
),
156 (input
->data
.patch
) ? "has" : "lacks");
160 if (!prog
->IsES
&& input
->data
.invariant
!= output
->data
.invariant
) {
162 "%s shader output `%s' %s invariant qualifier, "
163 "but %s shader input %s invariant qualifier\n",
164 _mesa_shader_stage_to_string(producer_stage
),
166 (output
->data
.invariant
) ? "has" : "lacks",
167 _mesa_shader_stage_to_string(consumer_stage
),
168 (input
->data
.invariant
) ? "has" : "lacks");
172 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
173 * to match cross stage, they must only match within the same stage.
175 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
177 * "It is a link-time error if, within the same stage, the interpolation
178 * qualifiers of variables of the same name do not match.
181 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
182 prog
->Version
< 440) {
184 "%s shader output `%s' specifies %s "
185 "interpolation qualifier, "
186 "but %s shader input specifies %s "
187 "interpolation qualifier\n",
188 _mesa_shader_stage_to_string(producer_stage
),
190 interpolation_string(output
->data
.interpolation
),
191 _mesa_shader_stage_to_string(consumer_stage
),
192 interpolation_string(input
->data
.interpolation
));
198 * Validate front and back color outputs against single color input
201 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
202 const ir_variable
*input
,
203 const ir_variable
*front_color
,
204 const ir_variable
*back_color
,
205 gl_shader_stage consumer_stage
,
206 gl_shader_stage producer_stage
)
208 if (front_color
!= NULL
&& front_color
->data
.assigned
)
209 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
210 consumer_stage
, producer_stage
);
212 if (back_color
!= NULL
&& back_color
->data
.assigned
)
213 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
214 consumer_stage
, producer_stage
);
218 * Validate that outputs from one stage match inputs of another
221 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
222 gl_shader
*producer
, gl_shader
*consumer
)
224 glsl_symbol_table parameters
;
225 ir_variable
*explicit_locations
[MAX_VARYING
] = { NULL
, };
227 /* Find all shader outputs in the "producer" stage.
229 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
230 ir_variable
*const var
= node
->as_variable();
232 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_shader_out
))
235 if (!var
->data
.explicit_location
236 || var
->data
.location
< VARYING_SLOT_VAR0
)
237 parameters
.add_variable(var
);
239 /* User-defined varyings with explicit locations are handled
240 * differently because they do not need to have matching names.
242 const unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
244 if (explicit_locations
[idx
] != NULL
) {
246 "%s shader has multiple outputs explicitly "
247 "assigned to location %d\n",
248 _mesa_shader_stage_to_string(producer
->Stage
),
253 explicit_locations
[idx
] = var
;
258 /* Find all shader inputs in the "consumer" stage. Any variables that have
259 * matching outputs already in the symbol table must have the same type and
262 * Exception: if the consumer is the geometry shader, then the inputs
263 * should be arrays and the type of the array element should match the type
264 * of the corresponding producer output.
266 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
267 ir_variable
*const input
= node
->as_variable();
269 if ((input
== NULL
) || (input
->data
.mode
!= ir_var_shader_in
))
272 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
273 const ir_variable
*const front_color
=
274 parameters
.get_variable("gl_FrontColor");
276 const ir_variable
*const back_color
=
277 parameters
.get_variable("gl_BackColor");
279 cross_validate_front_and_back_color(prog
, input
,
280 front_color
, back_color
,
281 consumer
->Stage
, producer
->Stage
);
282 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
283 const ir_variable
*const front_color
=
284 parameters
.get_variable("gl_FrontSecondaryColor");
286 const ir_variable
*const back_color
=
287 parameters
.get_variable("gl_BackSecondaryColor");
289 cross_validate_front_and_back_color(prog
, input
,
290 front_color
, back_color
,
291 consumer
->Stage
, producer
->Stage
);
293 /* The rules for connecting inputs and outputs change in the presence
294 * of explicit locations. In this case, we no longer care about the
295 * names of the variables. Instead, we care only about the
296 * explicitly assigned location.
298 ir_variable
*output
= NULL
;
299 if (input
->data
.explicit_location
300 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
301 output
= explicit_locations
[input
->data
.location
- VARYING_SLOT_VAR0
];
303 if (output
== NULL
) {
305 "%s shader input `%s' with explicit location "
306 "has no matching output\n",
307 _mesa_shader_stage_to_string(consumer
->Stage
),
311 output
= parameters
.get_variable(input
->name
);
314 if (output
!= NULL
) {
315 cross_validate_types_and_qualifiers(prog
, input
, output
,
316 consumer
->Stage
, producer
->Stage
);
318 /* Check for input vars with unmatched output vars in prev stage
319 * taking into account that interface blocks could have a matching
320 * output but with different name, so we ignore them.
322 assert(!input
->data
.assigned
);
323 if (input
->data
.used
&& !input
->get_interface_type() &&
324 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
326 "%s shader input `%s' "
327 "has no matching output in the previous stage\n",
328 _mesa_shader_stage_to_string(consumer
->Stage
),
336 * Demote shader inputs and outputs that are not used in other stages, and
337 * remove them via dead code elimination.
340 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
342 enum ir_variable_mode mode
)
344 if (is_separate_shader_object
)
347 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
348 ir_variable
*const var
= node
->as_variable();
350 if ((var
== NULL
) || (var
->data
.mode
!= int(mode
)))
353 /* A shader 'in' or 'out' variable is only really an input or output if
354 * its value is used by other shader stages. This will cause the
355 * variable to have a location assigned.
357 if (var
->data
.is_unmatched_generic_inout
) {
358 assert(var
->data
.mode
!= ir_var_temporary
);
359 var
->data
.mode
= ir_var_auto
;
363 /* Eliminate code that is now dead due to unused inputs/outputs being
366 while (do_dead_code(sh
->ir
, false))
372 * Initialize this object based on a string that was passed to
373 * glTransformFeedbackVaryings.
375 * If the input is mal-formed, this call still succeeds, but it sets
376 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
377 * will fail to find any matching variable.
380 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
383 /* We don't have to be pedantic about what is a valid GLSL variable name,
384 * because any variable with an invalid name can't exist in the IR anyway.
388 this->orig_name
= input
;
389 this->lowered_builtin_array_variable
= none
;
390 this->skip_components
= 0;
391 this->next_buffer_separator
= false;
392 this->matched_candidate
= NULL
;
395 if (ctx
->Extensions
.ARB_transform_feedback3
) {
396 /* Parse gl_NextBuffer. */
397 if (strcmp(input
, "gl_NextBuffer") == 0) {
398 this->next_buffer_separator
= true;
402 /* Parse gl_SkipComponents. */
403 if (strcmp(input
, "gl_SkipComponents1") == 0)
404 this->skip_components
= 1;
405 else if (strcmp(input
, "gl_SkipComponents2") == 0)
406 this->skip_components
= 2;
407 else if (strcmp(input
, "gl_SkipComponents3") == 0)
408 this->skip_components
= 3;
409 else if (strcmp(input
, "gl_SkipComponents4") == 0)
410 this->skip_components
= 4;
412 if (this->skip_components
)
416 /* Parse a declaration. */
417 const char *base_name_end
;
418 long subscript
= parse_program_resource_name(input
, &base_name_end
);
419 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
420 if (this->var_name
== NULL
) {
421 _mesa_error_no_memory(__func__
);
425 if (subscript
>= 0) {
426 this->array_subscript
= subscript
;
427 this->is_subscripted
= true;
429 this->is_subscripted
= false;
432 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
433 * class must behave specially to account for the fact that gl_ClipDistance
434 * is converted from a float[8] to a vec4[2].
436 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerClipDistance
&&
437 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
438 this->lowered_builtin_array_variable
= clip_distance
;
441 if (ctx
->Const
.LowerTessLevel
&&
442 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
443 this->lowered_builtin_array_variable
= tess_level_outer
;
444 if (ctx
->Const
.LowerTessLevel
&&
445 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
446 this->lowered_builtin_array_variable
= tess_level_inner
;
451 * Determine whether two tfeedback_decl objects refer to the same variable and
452 * array index (if applicable).
455 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
457 assert(x
.is_varying() && y
.is_varying());
459 if (strcmp(x
.var_name
, y
.var_name
) != 0)
461 if (x
.is_subscripted
!= y
.is_subscripted
)
463 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
470 * Assign a location and stream ID for this tfeedback_decl object based on the
471 * transform feedback candidate found by find_candidate.
473 * If an error occurs, the error is reported through linker_error() and false
477 tfeedback_decl::assign_location(struct gl_context
*ctx
,
478 struct gl_shader_program
*prog
)
480 assert(this->is_varying());
482 unsigned fine_location
483 = this->matched_candidate
->toplevel_var
->data
.location
* 4
484 + this->matched_candidate
->toplevel_var
->data
.location_frac
485 + this->matched_candidate
->offset
;
487 if (this->matched_candidate
->type
->is_array()) {
489 const unsigned matrix_cols
=
490 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
491 const unsigned vector_elements
=
492 this->matched_candidate
->type
->fields
.array
->vector_elements
;
493 const unsigned dmul
=
494 this->matched_candidate
->type
->fields
.array
->is_double() ? 2 : 1;
495 unsigned actual_array_size
;
496 switch (this->lowered_builtin_array_variable
) {
498 actual_array_size
= prog
->LastClipDistanceArraySize
;
500 case tess_level_outer
:
501 actual_array_size
= 4;
503 case tess_level_inner
:
504 actual_array_size
= 2;
508 actual_array_size
= this->matched_candidate
->type
->array_size();
512 if (this->is_subscripted
) {
513 /* Check array bounds. */
514 if (this->array_subscript
>= actual_array_size
) {
515 linker_error(prog
, "Transform feedback varying %s has index "
516 "%i, but the array size is %u.",
517 this->orig_name
, this->array_subscript
,
521 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
522 1 : vector_elements
* matrix_cols
* dmul
;
523 fine_location
+= array_elem_size
* this->array_subscript
;
526 this->size
= actual_array_size
;
528 this->vector_elements
= vector_elements
;
529 this->matrix_columns
= matrix_cols
;
530 if (this->lowered_builtin_array_variable
)
531 this->type
= GL_FLOAT
;
533 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
535 /* Regular variable (scalar, vector, or matrix) */
536 if (this->is_subscripted
) {
537 linker_error(prog
, "Transform feedback varying %s requested, "
538 "but %s is not an array.",
539 this->orig_name
, this->var_name
);
543 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
544 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
545 this->type
= this->matched_candidate
->type
->gl_type
;
547 this->location
= fine_location
/ 4;
548 this->location_frac
= fine_location
% 4;
550 /* From GL_EXT_transform_feedback:
551 * A program will fail to link if:
553 * * the total number of components to capture in any varying
554 * variable in <varyings> is greater than the constant
555 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
556 * buffer mode is SEPARATE_ATTRIBS_EXT;
558 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
559 this->num_components() >
560 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
561 linker_error(prog
, "Transform feedback varying %s exceeds "
562 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
567 /* Only transform feedback varyings can be assigned to non-zero streams,
568 * so assign the stream id here.
570 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
577 tfeedback_decl::get_num_outputs() const
579 if (!this->is_varying()) {
582 return (this->num_components() + this->location_frac
+ 3)/4;
587 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
589 * If an error occurs, the error is reported through linker_error() and false
593 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
594 struct gl_transform_feedback_info
*info
,
595 unsigned buffer
, const unsigned max_outputs
) const
597 assert(!this->next_buffer_separator
);
599 /* Handle gl_SkipComponents. */
600 if (this->skip_components
) {
601 info
->BufferStride
[buffer
] += this->skip_components
;
605 /* From GL_EXT_transform_feedback:
606 * A program will fail to link if:
608 * * the total number of components to capture is greater than
609 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
610 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
612 if (prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
&&
613 info
->BufferStride
[buffer
] + this->num_components() >
614 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
615 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
616 "limit has been exceeded.");
620 unsigned location
= this->location
;
621 unsigned location_frac
= this->location_frac
;
622 unsigned num_components
= this->num_components();
623 while (num_components
> 0) {
624 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
625 assert(info
->NumOutputs
< max_outputs
);
626 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
627 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
628 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
629 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
630 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
631 info
->Outputs
[info
->NumOutputs
].DstOffset
= info
->BufferStride
[buffer
];
633 info
->BufferStride
[buffer
] += output_size
;
634 info
->BufferStream
[buffer
] = this->stream_id
;
635 num_components
-= output_size
;
640 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
, this->orig_name
);
641 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
642 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
649 const tfeedback_candidate
*
650 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
651 hash_table
*tfeedback_candidates
)
653 const char *name
= this->var_name
;
654 switch (this->lowered_builtin_array_variable
) {
656 name
= this->var_name
;
659 name
= "gl_ClipDistanceMESA";
661 case tess_level_outer
:
662 name
= "gl_TessLevelOuterMESA";
664 case tess_level_inner
:
665 name
= "gl_TessLevelInnerMESA";
668 this->matched_candidate
= (const tfeedback_candidate
*)
669 hash_table_find(tfeedback_candidates
, name
);
670 if (!this->matched_candidate
) {
671 /* From GL_EXT_transform_feedback:
672 * A program will fail to link if:
674 * * any variable name specified in the <varyings> array is not
675 * declared as an output in the geometry shader (if present) or
676 * the vertex shader (if no geometry shader is present);
678 linker_error(prog
, "Transform feedback varying %s undeclared.",
681 return this->matched_candidate
;
686 * Parse all the transform feedback declarations that were passed to
687 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
689 * If an error occurs, the error is reported through linker_error() and false
693 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
694 const void *mem_ctx
, unsigned num_names
,
695 char **varying_names
, tfeedback_decl
*decls
)
697 for (unsigned i
= 0; i
< num_names
; ++i
) {
698 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
700 if (!decls
[i
].is_varying())
703 /* From GL_EXT_transform_feedback:
704 * A program will fail to link if:
706 * * any two entries in the <varyings> array specify the same varying
709 * We interpret this to mean "any two entries in the <varyings> array
710 * specify the same varying variable and array index", since transform
711 * feedback of arrays would be useless otherwise.
713 for (unsigned j
= 0; j
< i
; ++j
) {
714 if (!decls
[j
].is_varying())
717 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
718 linker_error(prog
, "Transform feedback varying %s specified "
719 "more than once.", varying_names
[i
]);
729 * Store transform feedback location assignments into
730 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
732 * If an error occurs, the error is reported through linker_error() and false
736 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
737 unsigned num_tfeedback_decls
,
738 tfeedback_decl
*tfeedback_decls
)
740 bool separate_attribs_mode
=
741 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
743 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
744 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
746 memset(&prog
->LinkedTransformFeedback
, 0,
747 sizeof(prog
->LinkedTransformFeedback
));
749 prog
->LinkedTransformFeedback
.Varyings
=
751 struct gl_transform_feedback_varying_info
,
752 num_tfeedback_decls
);
754 unsigned num_outputs
= 0;
755 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
)
756 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
758 prog
->LinkedTransformFeedback
.Outputs
=
760 struct gl_transform_feedback_output
,
763 unsigned num_buffers
= 0;
765 if (separate_attribs_mode
) {
766 /* GL_SEPARATE_ATTRIBS */
767 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
768 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
769 num_buffers
, num_outputs
))
776 /* GL_INVERLEAVED_ATTRIBS */
777 int buffer_stream_id
= -1;
778 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
779 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
781 buffer_stream_id
= -1;
783 } else if (buffer_stream_id
== -1) {
784 /* First varying writing to this buffer: remember its stream */
785 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
786 } else if (buffer_stream_id
!=
787 (int) tfeedback_decls
[i
].get_stream_id()) {
788 /* Varying writes to the same buffer from a different stream */
790 "Transform feedback can't capture varyings belonging "
791 "to different vertex streams in a single buffer. "
792 "Varying %s writes to buffer from stream %u, other "
793 "varyings in the same buffer write from stream %u.",
794 tfeedback_decls
[i
].name(),
795 tfeedback_decls
[i
].get_stream_id(),
800 if (!tfeedback_decls
[i
].store(ctx
, prog
,
801 &prog
->LinkedTransformFeedback
,
802 num_buffers
, num_outputs
))
808 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
810 prog
->LinkedTransformFeedback
.NumBuffers
= num_buffers
;
817 * Data structure recording the relationship between outputs of one shader
818 * stage (the "producer") and inputs of another (the "consumer").
820 class varying_matches
823 varying_matches(bool disable_varying_packing
,
824 gl_shader_stage producer_stage
,
825 gl_shader_stage consumer_stage
);
827 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
828 unsigned assign_locations(struct gl_shader_program
*prog
,
829 uint64_t reserved_slots
, bool separate_shader
);
830 void store_locations() const;
834 * If true, this driver disables varying packing, so all varyings need to
835 * be aligned on slot boundaries, and take up a number of slots equal to
836 * their number of matrix columns times their array size.
838 const bool disable_varying_packing
;
841 * Enum representing the order in which varyings are packed within a
844 * Currently we pack vec4's first, then vec2's, then scalar values, then
845 * vec3's. This order ensures that the only vectors that are at risk of
846 * having to be "double parked" (split between two adjacent varying slots)
849 enum packing_order_enum
{
852 PACKING_ORDER_SCALAR
,
856 static unsigned compute_packing_class(const ir_variable
*var
);
857 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
858 static int match_comparator(const void *x_generic
, const void *y_generic
);
861 * Structure recording the relationship between a single producer output
862 * and a single consumer input.
866 * Packing class for this varying, computed by compute_packing_class().
868 unsigned packing_class
;
871 * Packing order for this varying, computed by compute_packing_order().
873 packing_order_enum packing_order
;
874 unsigned num_components
;
877 * The output variable in the producer stage.
879 ir_variable
*producer_var
;
882 * The input variable in the consumer stage.
884 ir_variable
*consumer_var
;
887 * The location which has been assigned for this varying. This is
888 * expressed in multiples of a float, with the first generic varying
889 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
892 unsigned generic_location
;
896 * The number of elements in the \c matches array that are currently in
899 unsigned num_matches
;
902 * The number of elements that were set aside for the \c matches array when
905 unsigned matches_capacity
;
907 gl_shader_stage producer_stage
;
908 gl_shader_stage consumer_stage
;
911 } /* anonymous namespace */
913 varying_matches::varying_matches(bool disable_varying_packing
,
914 gl_shader_stage producer_stage
,
915 gl_shader_stage consumer_stage
)
916 : disable_varying_packing(disable_varying_packing
),
917 producer_stage(producer_stage
),
918 consumer_stage(consumer_stage
)
920 /* Note: this initial capacity is rather arbitrarily chosen to be large
921 * enough for many cases without wasting an unreasonable amount of space.
922 * varying_matches::record() will resize the array if there are more than
923 * this number of varyings.
925 this->matches_capacity
= 8;
926 this->matches
= (match
*)
927 malloc(sizeof(*this->matches
) * this->matches_capacity
);
928 this->num_matches
= 0;
932 varying_matches::~varying_matches()
939 * Record the given producer/consumer variable pair in the list of variables
940 * that should later be assigned locations.
942 * It is permissible for \c consumer_var to be NULL (this happens if a
943 * variable is output by the producer and consumed by transform feedback, but
944 * not consumed by the consumer).
946 * If \c producer_var has already been paired up with a consumer_var, or
947 * producer_var is part of fixed pipeline functionality (and hence already has
948 * a location assigned), this function has no effect.
950 * Note: as a side effect this function may change the interpolation type of
951 * \c producer_var, but only when the change couldn't possibly affect
955 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
957 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
959 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
960 producer_var
->data
.explicit_location
)) ||
961 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
962 consumer_var
->data
.explicit_location
))) {
963 /* Either a location already exists for this variable (since it is part
964 * of fixed functionality), or it has already been recorded as part of a
970 bool needs_flat_qualifier
= consumer_var
== NULL
&&
971 (producer_var
->type
->contains_integer() ||
972 producer_var
->type
->contains_double());
974 if (needs_flat_qualifier
||
975 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
)) {
976 /* Since this varying is not being consumed by the fragment shader, its
977 * interpolation type varying cannot possibly affect rendering.
978 * Also, this variable is non-flat and is (or contains) an integer
980 * If the consumer stage is unknown, don't modify the interpolation
981 * type as it could affect rendering later with separate shaders.
983 * lower_packed_varyings requires all integer varyings to flat,
984 * regardless of where they appear. We can trivially satisfy that
985 * requirement by changing the interpolation type to flat here.
988 producer_var
->data
.centroid
= false;
989 producer_var
->data
.sample
= false;
990 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
994 consumer_var
->data
.centroid
= false;
995 consumer_var
->data
.sample
= false;
996 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1000 if (this->num_matches
== this->matches_capacity
) {
1001 this->matches_capacity
*= 2;
1002 this->matches
= (match
*)
1003 realloc(this->matches
,
1004 sizeof(*this->matches
) * this->matches_capacity
);
1007 const ir_variable
*const var
= (producer_var
!= NULL
)
1008 ? producer_var
: consumer_var
;
1009 const gl_shader_stage stage
= (producer_var
!= NULL
)
1010 ? producer_stage
: consumer_stage
;
1011 const glsl_type
*type
= get_varying_type(var
, stage
);
1013 this->matches
[this->num_matches
].packing_class
1014 = this->compute_packing_class(var
);
1015 this->matches
[this->num_matches
].packing_order
1016 = this->compute_packing_order(var
);
1017 if (this->disable_varying_packing
) {
1018 unsigned slots
= type
->count_attribute_slots(false);
1019 this->matches
[this->num_matches
].num_components
= slots
* 4;
1021 this->matches
[this->num_matches
].num_components
1022 = type
->component_slots();
1024 this->matches
[this->num_matches
].producer_var
= producer_var
;
1025 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1026 this->num_matches
++;
1028 producer_var
->data
.is_unmatched_generic_inout
= 0;
1030 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1035 * Choose locations for all of the variable matches that were previously
1036 * passed to varying_matches::record().
1039 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1040 uint64_t reserved_slots
,
1041 bool separate_shader
)
1043 /* We disable varying sorting for separate shader programs for the
1044 * following reasons:
1046 * 1/ All programs must sort the code in the same order to guarantee the
1047 * interface matching. However varying_matches::record() will change the
1048 * interpolation qualifier of some stages.
1050 * 2/ GLSL version 4.50 removes the matching constrain on the interpolation
1053 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.40 spec:
1055 * "The type and presence of interpolation qualifiers of variables with
1056 * the same name declared in all linked shaders for the same cross-stage
1057 * interface must match, otherwise the link command will fail.
1059 * When comparing an output from one stage to an input of a subsequent
1060 * stage, the input and output don't match if their interpolation
1061 * qualifiers (or lack thereof) are not the same."
1063 * "It is a link-time error if, within the same stage, the interpolation
1064 * qualifiers of variables of the same name do not match."
1066 if (!separate_shader
) {
1067 /* Sort varying matches into an order that makes them easy to pack. */
1068 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1069 &varying_matches::match_comparator
);
1072 unsigned generic_location
= 0;
1073 unsigned generic_patch_location
= MAX_VARYING
*4;
1075 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1076 unsigned *location
= &generic_location
;
1078 const ir_variable
*var
;
1079 const glsl_type
*type
;
1080 bool is_vertex_input
= false;
1081 if (matches
[i
].consumer_var
) {
1082 var
= matches
[i
].consumer_var
;
1083 type
= get_varying_type(var
, consumer_stage
);
1084 if (consumer_stage
== MESA_SHADER_VERTEX
)
1085 is_vertex_input
= true;
1087 var
= matches
[i
].producer_var
;
1088 type
= get_varying_type(var
, producer_stage
);
1091 if (var
->data
.patch
)
1092 location
= &generic_patch_location
;
1094 /* Advance to the next slot if this varying has a different packing
1095 * class than the previous one, and we're not already on a slot
1099 this->matches
[i
- 1].packing_class
1100 != this->matches
[i
].packing_class
) {
1101 *location
= ALIGN(*location
, 4);
1104 unsigned num_elements
= type
->count_attribute_slots(is_vertex_input
);
1105 unsigned slot_end
= this->disable_varying_packing
? 4 :
1106 type
->without_array()->vector_elements
;
1107 slot_end
+= *location
- 1;
1109 /* FIXME: We could be smarter in the below code and loop back over
1110 * trying to fill any locations that we skipped because we couldn't pack
1111 * the varying between an explicit location. For now just let the user
1112 * hit the linking error if we run out of room and suggest they use
1113 * explicit locations.
1115 for (unsigned j
= 0; j
< num_elements
; j
++) {
1116 while ((slot_end
< MAX_VARYING
* 4u) &&
1117 ((reserved_slots
& (UINT64_C(1) << *location
/ 4u) ||
1118 (reserved_slots
& (UINT64_C(1) << slot_end
/ 4u))))) {
1120 *location
= ALIGN(*location
+ 1, 4);
1121 slot_end
= *location
;
1123 /* reset the counter and try again */
1127 /* Increase the slot to make sure there is enough room for next
1130 if (this->disable_varying_packing
)
1133 slot_end
+= type
->without_array()->vector_elements
;
1136 if (!var
->data
.patch
&& *location
>= MAX_VARYING
* 4u) {
1137 linker_error(prog
, "insufficient contiguous locations available for "
1138 "%s it is possible an array or struct could not be "
1139 "packed between varyings with explicit locations. Try "
1140 "using an explicit location for arrays and structs.",
1144 this->matches
[i
].generic_location
= *location
;
1146 *location
+= this->matches
[i
].num_components
;
1149 return (generic_location
+ 3) / 4;
1154 * Update the producer and consumer shaders to reflect the locations
1155 * assignments that were made by varying_matches::assign_locations().
1158 varying_matches::store_locations() const
1160 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1161 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1162 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1163 unsigned generic_location
= this->matches
[i
].generic_location
;
1164 unsigned slot
= generic_location
/ 4;
1165 unsigned offset
= generic_location
% 4;
1168 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1169 producer_var
->data
.location_frac
= offset
;
1173 assert(consumer_var
->data
.location
== -1);
1174 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1175 consumer_var
->data
.location_frac
= offset
;
1182 * Compute the "packing class" of the given varying. This is an unsigned
1183 * integer with the property that two variables in the same packing class can
1184 * be safely backed into the same vec4.
1187 varying_matches::compute_packing_class(const ir_variable
*var
)
1189 /* Without help from the back-end, there is no way to pack together
1190 * variables with different interpolation types, because
1191 * lower_packed_varyings must choose exactly one interpolation type for
1192 * each packed varying it creates.
1194 * However, we can safely pack together floats, ints, and uints, because:
1196 * - varyings of base type "int" and "uint" must use the "flat"
1197 * interpolation type, which can only occur in GLSL 1.30 and above.
1199 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1200 * can store flat floats as ints without losing any information (using
1201 * the ir_unop_bitcast_* opcodes).
1203 * Therefore, the packing class depends only on the interpolation type.
1205 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1206 (var
->data
.patch
<< 2);
1208 packing_class
+= var
->data
.interpolation
;
1209 return packing_class
;
1214 * Compute the "packing order" of the given varying. This is a sort key we
1215 * use to determine when to attempt to pack the given varying relative to
1216 * other varyings in the same packing class.
1218 varying_matches::packing_order_enum
1219 varying_matches::compute_packing_order(const ir_variable
*var
)
1221 const glsl_type
*element_type
= var
->type
;
1223 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
1224 element_type
= element_type
->fields
.array
;
1227 switch (element_type
->component_slots() % 4) {
1228 case 1: return PACKING_ORDER_SCALAR
;
1229 case 2: return PACKING_ORDER_VEC2
;
1230 case 3: return PACKING_ORDER_VEC3
;
1231 case 0: return PACKING_ORDER_VEC4
;
1233 assert(!"Unexpected value of vector_elements");
1234 return PACKING_ORDER_VEC4
;
1240 * Comparison function passed to qsort() to sort varyings by packing_class and
1241 * then by packing_order.
1244 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1246 const match
*x
= (const match
*) x_generic
;
1247 const match
*y
= (const match
*) y_generic
;
1249 if (x
->packing_class
!= y
->packing_class
)
1250 return x
->packing_class
- y
->packing_class
;
1251 return x
->packing_order
- y
->packing_order
;
1256 * Is the given variable a varying variable to be counted against the
1257 * limit in ctx->Const.MaxVarying?
1258 * This includes variables such as texcoords, colors and generic
1259 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1262 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1264 /* Only fragment shaders will take a varying variable as an input */
1265 if (stage
== MESA_SHADER_FRAGMENT
&&
1266 var
->data
.mode
== ir_var_shader_in
) {
1267 switch (var
->data
.location
) {
1268 case VARYING_SLOT_POS
:
1269 case VARYING_SLOT_FACE
:
1270 case VARYING_SLOT_PNTC
:
1281 * Visitor class that generates tfeedback_candidate structs describing all
1282 * possible targets of transform feedback.
1284 * tfeedback_candidate structs are stored in the hash table
1285 * tfeedback_candidates, which is passed to the constructor. This hash table
1286 * maps varying names to instances of the tfeedback_candidate struct.
1288 class tfeedback_candidate_generator
: public program_resource_visitor
1291 tfeedback_candidate_generator(void *mem_ctx
,
1292 hash_table
*tfeedback_candidates
)
1294 tfeedback_candidates(tfeedback_candidates
),
1300 void process(ir_variable
*var
)
1302 /* All named varying interface blocks should be flattened by now */
1303 assert(!var
->is_interface_instance());
1305 this->toplevel_var
= var
;
1306 this->varying_floats
= 0;
1307 program_resource_visitor::process(var
);
1311 virtual void visit_field(const glsl_type
*type
, const char *name
,
1314 assert(!type
->without_array()->is_record());
1315 assert(!type
->without_array()->is_interface());
1319 tfeedback_candidate
*candidate
1320 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1321 candidate
->toplevel_var
= this->toplevel_var
;
1322 candidate
->type
= type
;
1323 candidate
->offset
= this->varying_floats
;
1324 hash_table_insert(this->tfeedback_candidates
, candidate
,
1325 ralloc_strdup(this->mem_ctx
, name
));
1326 this->varying_floats
+= type
->component_slots();
1330 * Memory context used to allocate hash table keys and values.
1332 void * const mem_ctx
;
1335 * Hash table in which tfeedback_candidate objects should be stored.
1337 hash_table
* const tfeedback_candidates
;
1340 * Pointer to the toplevel variable that is being traversed.
1342 ir_variable
*toplevel_var
;
1345 * Total number of varying floats that have been visited so far. This is
1346 * used to determine the offset to each varying within the toplevel
1349 unsigned varying_floats
;
1356 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1357 hash_table
*consumer_inputs
,
1358 hash_table
*consumer_interface_inputs
,
1359 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1361 memset(consumer_inputs_with_locations
,
1363 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1365 foreach_in_list(ir_instruction
, node
, ir
) {
1366 ir_variable
*const input_var
= node
->as_variable();
1368 if ((input_var
!= NULL
) && (input_var
->data
.mode
== ir_var_shader_in
)) {
1369 if (input_var
->type
->is_interface())
1372 if (input_var
->data
.explicit_location
) {
1373 /* assign_varying_locations only cares about finding the
1374 * ir_variable at the start of a contiguous location block.
1376 * - For !producer, consumer_inputs_with_locations isn't used.
1378 * - For !consumer, consumer_inputs_with_locations is empty.
1380 * For consumer && producer, if you were trying to set some
1381 * ir_variable to the middle of a location block on the other side
1382 * of producer/consumer, cross_validate_outputs_to_inputs() should
1383 * be link-erroring due to either type mismatch or location
1384 * overlaps. If the variables do match up, then they've got a
1385 * matching data.location and you only looked at
1386 * consumer_inputs_with_locations[var->data.location], not any
1387 * following entries for the array/structure.
1389 consumer_inputs_with_locations
[input_var
->data
.location
] =
1391 } else if (input_var
->get_interface_type() != NULL
) {
1392 char *const iface_field_name
=
1393 ralloc_asprintf(mem_ctx
, "%s.%s",
1394 input_var
->get_interface_type()->name
,
1396 hash_table_insert(consumer_interface_inputs
, input_var
,
1399 hash_table_insert(consumer_inputs
, input_var
,
1400 ralloc_strdup(mem_ctx
, input_var
->name
));
1409 * Find a variable from the consumer that "matches" the specified variable
1411 * This function only finds inputs with names that match. There is no
1412 * validation (here) that the types, etc. are compatible.
1415 get_matching_input(void *mem_ctx
,
1416 const ir_variable
*output_var
,
1417 hash_table
*consumer_inputs
,
1418 hash_table
*consumer_interface_inputs
,
1419 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1421 ir_variable
*input_var
;
1423 if (output_var
->data
.explicit_location
) {
1424 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1425 } else if (output_var
->get_interface_type() != NULL
) {
1426 char *const iface_field_name
=
1427 ralloc_asprintf(mem_ctx
, "%s.%s",
1428 output_var
->get_interface_type()->name
,
1431 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1435 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1438 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1445 io_variable_cmp(const void *_a
, const void *_b
)
1447 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1448 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1450 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1451 return b
->data
.location
- a
->data
.location
;
1453 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1456 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1459 return -strcmp(a
->name
, b
->name
);
1463 * Sort the shader IO variables into canonical order
1466 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1468 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1469 unsigned num_variables
= 0;
1471 foreach_in_list(ir_instruction
, node
, ir
) {
1472 ir_variable
*const var
= node
->as_variable();
1474 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1477 /* If we have already encountered more I/O variables that could
1478 * successfully link, bail.
1480 if (num_variables
== ARRAY_SIZE(var_table
))
1483 var_table
[num_variables
++] = var
;
1486 if (num_variables
== 0)
1489 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1490 * we're going to push the variables on to the IR list as a stack, so we
1491 * want the last variable (in canonical order) to be first in the list.
1493 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1495 /* Remove the variable from it's current location in the IR, and put it at
1498 for (unsigned i
= 0; i
< num_variables
; i
++) {
1499 var_table
[i
]->remove();
1500 ir
->push_head(var_table
[i
]);
1505 * Generate a bitfield map of the explicit locations for shader varyings.
1507 * In theory a 32 bits value will be enough but a 64 bits value is future proof.
1510 reserved_varying_slot(struct gl_shader
*stage
, ir_variable_mode io_mode
)
1512 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
1513 assert(MAX_VARYING
<= 64); /* avoid an overflow of the returned value */
1521 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
1522 ir_variable
*const var
= node
->as_variable();
1524 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
1525 !var
->data
.explicit_location
||
1526 var
->data
.location
< VARYING_SLOT_VAR0
)
1529 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
1531 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
1532 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
1533 for (unsigned i
= 0; i
< num_elements
; i
++) {
1534 if (var_slot
>= 0 && var_slot
< MAX_VARYING
)
1535 slots
|= UINT64_C(1) << var_slot
;
1545 * Assign locations for all variables that are produced in one pipeline stage
1546 * (the "producer") and consumed in the next stage (the "consumer").
1548 * Variables produced by the producer may also be consumed by transform
1551 * \param num_tfeedback_decls is the number of declarations indicating
1552 * variables that may be consumed by transform feedback.
1554 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1555 * representing the result of parsing the strings passed to
1556 * glTransformFeedbackVaryings(). assign_location() will be called for
1557 * each of these objects that matches one of the outputs of the
1560 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1561 * be NULL. In this case, varying locations are assigned solely based on the
1562 * requirements of transform feedback.
1565 assign_varying_locations(struct gl_context
*ctx
,
1567 struct gl_shader_program
*prog
,
1568 gl_shader
*producer
, gl_shader
*consumer
,
1569 unsigned num_tfeedback_decls
,
1570 tfeedback_decl
*tfeedback_decls
)
1572 if (ctx
->Const
.DisableVaryingPacking
) {
1573 /* Transform feedback code assumes varyings are packed, so if the driver
1574 * has disabled varying packing, make sure it does not support transform
1577 assert(!ctx
->Extensions
.EXT_transform_feedback
);
1580 /* Tessellation shaders treat inputs and outputs as shared memory and can
1581 * access inputs and outputs of other invocations.
1582 * Therefore, they can't be lowered to temps easily (and definitely not
1585 bool disable_varying_packing
=
1586 ctx
->Const
.DisableVaryingPacking
||
1587 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
1588 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
1589 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
1591 varying_matches
matches(disable_varying_packing
,
1592 producer
? producer
->Stage
: (gl_shader_stage
)-1,
1593 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
1594 hash_table
*tfeedback_candidates
1595 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1596 hash_table
*consumer_inputs
1597 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1598 hash_table
*consumer_interface_inputs
1599 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1600 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
1604 unsigned consumer_vertices
= 0;
1605 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
1606 consumer_vertices
= prog
->Geom
.VerticesIn
;
1608 /* Operate in a total of four passes.
1610 * 1. Sort inputs / outputs into a canonical order. This is necessary so
1611 * that inputs / outputs of separable shaders will be assigned
1612 * predictable locations regardless of the order in which declarations
1613 * appeared in the shader source.
1615 * 2. Assign locations for any matching inputs and outputs.
1617 * 3. Mark output variables in the producer that do not have locations as
1618 * not being outputs. This lets the optimizer eliminate them.
1620 * 4. Mark input variables in the consumer that do not have locations as
1621 * not being inputs. This lets the optimizer eliminate them.
1624 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
1627 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
1630 && !linker::populate_consumer_input_sets(mem_ctx
,
1633 consumer_interface_inputs
,
1634 consumer_inputs_with_locations
)) {
1635 assert(!"populate_consumer_input_sets failed");
1636 hash_table_dtor(tfeedback_candidates
);
1637 hash_table_dtor(consumer_inputs
);
1638 hash_table_dtor(consumer_interface_inputs
);
1643 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
1644 ir_variable
*const output_var
= node
->as_variable();
1646 if ((output_var
== NULL
) ||
1647 (output_var
->data
.mode
!= ir_var_shader_out
))
1650 /* Only geometry shaders can use non-zero streams */
1651 assert(output_var
->data
.stream
== 0 ||
1652 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
1653 producer
->Stage
== MESA_SHADER_GEOMETRY
));
1655 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
1656 g
.process(output_var
);
1658 ir_variable
*const input_var
=
1659 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
1660 consumer_interface_inputs
,
1661 consumer_inputs_with_locations
);
1663 /* If a matching input variable was found, add this ouptut (and the
1664 * input) to the set. If this is a separable program and there is no
1665 * consumer stage, add the output.
1667 * Always add TCS outputs. They are shared by all invocations
1668 * within a patch and can be used as shared memory.
1670 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
1671 producer
->Type
== GL_TESS_CONTROL_SHADER
) {
1672 matches
.record(output_var
, input_var
);
1675 /* Only stream 0 outputs can be consumed in the next stage */
1676 if (input_var
&& output_var
->data
.stream
!= 0) {
1677 linker_error(prog
, "output %s is assigned to stream=%d but "
1678 "is linked to an input, which requires stream=0",
1679 output_var
->name
, output_var
->data
.stream
);
1684 /* If there's no producer stage, then this must be a separable program.
1685 * For example, we may have a program that has just a fragment shader.
1686 * Later this program will be used with some arbitrary vertex (or
1687 * geometry) shader program. This means that locations must be assigned
1688 * for all the inputs.
1690 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
1691 ir_variable
*const input_var
= node
->as_variable();
1693 if ((input_var
== NULL
) ||
1694 (input_var
->data
.mode
!= ir_var_shader_in
))
1697 matches
.record(NULL
, input_var
);
1701 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1702 if (!tfeedback_decls
[i
].is_varying())
1705 const tfeedback_candidate
*matched_candidate
1706 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
1708 if (matched_candidate
== NULL
) {
1709 hash_table_dtor(tfeedback_candidates
);
1710 hash_table_dtor(consumer_inputs
);
1711 hash_table_dtor(consumer_interface_inputs
);
1715 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
)
1716 matches
.record(matched_candidate
->toplevel_var
, NULL
);
1719 const uint64_t reserved_slots
=
1720 reserved_varying_slot(producer
, ir_var_shader_out
) |
1721 reserved_varying_slot(consumer
, ir_var_shader_in
);
1723 const unsigned slots_used
= matches
.assign_locations(prog
, reserved_slots
,
1724 prog
->SeparateShader
);
1725 matches
.store_locations();
1727 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1728 if (!tfeedback_decls
[i
].is_varying())
1731 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
1732 hash_table_dtor(tfeedback_candidates
);
1733 hash_table_dtor(consumer_inputs
);
1734 hash_table_dtor(consumer_interface_inputs
);
1739 hash_table_dtor(tfeedback_candidates
);
1740 hash_table_dtor(consumer_inputs
);
1741 hash_table_dtor(consumer_interface_inputs
);
1743 if (consumer
&& producer
) {
1744 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
1745 ir_variable
*const var
= node
->as_variable();
1747 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1748 var
->data
.is_unmatched_generic_inout
) {
1751 * On Page 91 (Page 97 of the PDF) of the GLSL ES 1.0 spec:
1753 * If the vertex shader declares but doesn't write to a
1754 * varying and the fragment shader declares and reads it,
1759 linker_warning(prog
, "%s shader varying %s not written "
1761 _mesa_shader_stage_to_string(consumer
->Stage
),
1763 _mesa_shader_stage_to_string(producer
->Stage
));
1764 } else if (prog
->Version
<= 120) {
1765 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1767 * Only those varying variables used (i.e. read) in
1768 * the fragment shader executable must be written to
1769 * by the vertex shader executable; declaring
1770 * superfluous varying variables in a vertex shader is
1773 * We interpret this text as meaning that the VS must
1774 * write the variable for the FS to read it. See
1775 * "glsl1-varying read but not written" in piglit.
1777 linker_error(prog
, "%s shader varying %s not written "
1779 _mesa_shader_stage_to_string(consumer
->Stage
),
1781 _mesa_shader_stage_to_string(producer
->Stage
));
1786 /* Now that validation is done its safe to remove unused varyings. As
1787 * we have both a producer and consumer its safe to remove unused
1788 * varyings even if the program is a SSO because the stages are being
1789 * linked together i.e. we have a multi-stage SSO.
1791 remove_unused_shader_inputs_and_outputs(false, producer
,
1793 remove_unused_shader_inputs_and_outputs(false, consumer
,
1797 if (!disable_varying_packing
) {
1799 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
1803 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
1804 consumer_vertices
, consumer
);
1812 check_against_output_limit(struct gl_context
*ctx
,
1813 struct gl_shader_program
*prog
,
1814 gl_shader
*producer
)
1816 unsigned output_vectors
= 0;
1818 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
1819 ir_variable
*const var
= node
->as_variable();
1821 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
1822 var_counts_against_varying_limit(producer
->Stage
, var
)) {
1823 /* outputs for fragment shader can't be doubles */
1824 output_vectors
+= var
->type
->count_attribute_slots(false);
1828 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
1829 unsigned max_output_components
=
1830 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
1832 const unsigned output_components
= output_vectors
* 4;
1833 if (output_components
> max_output_components
) {
1834 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1835 linker_error(prog
, "%s shader uses too many output vectors "
1837 _mesa_shader_stage_to_string(producer
->Stage
),
1839 max_output_components
/ 4);
1841 linker_error(prog
, "%s shader uses too many output components "
1843 _mesa_shader_stage_to_string(producer
->Stage
),
1845 max_output_components
);
1854 check_against_input_limit(struct gl_context
*ctx
,
1855 struct gl_shader_program
*prog
,
1856 gl_shader
*consumer
)
1858 unsigned input_vectors
= 0;
1860 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
1861 ir_variable
*const var
= node
->as_variable();
1863 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1864 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
1865 /* vertex inputs aren't varying counted */
1866 input_vectors
+= var
->type
->count_attribute_slots(false);
1870 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
1871 unsigned max_input_components
=
1872 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
1874 const unsigned input_components
= input_vectors
* 4;
1875 if (input_components
> max_input_components
) {
1876 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1877 linker_error(prog
, "%s shader uses too many input vectors "
1879 _mesa_shader_stage_to_string(consumer
->Stage
),
1881 max_input_components
/ 4);
1883 linker_error(prog
, "%s shader uses too many input components "
1885 _mesa_shader_stage_to_string(consumer
->Stage
),
1887 max_input_components
);