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 * Validate the types and qualifiers of an output from one stage against the
45 * matching input to another stage.
48 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
49 const ir_variable
*input
,
50 const ir_variable
*output
,
51 gl_shader_stage consumer_stage
,
52 gl_shader_stage producer_stage
)
54 /* Check that the types match between stages.
56 const glsl_type
*type_to_match
= input
->type
;
57 if (consumer_stage
== MESA_SHADER_GEOMETRY
) {
58 assert(type_to_match
->is_array()); /* Enforced by ast_to_hir */
59 type_to_match
= type_to_match
->element_type();
61 if (type_to_match
!= output
->type
) {
62 /* There is a bit of a special case for gl_TexCoord. This
63 * built-in is unsized by default. Applications that variable
64 * access it must redeclare it with a size. There is some
65 * language in the GLSL spec that implies the fragment shader
66 * and vertex shader do not have to agree on this size. Other
67 * driver behave this way, and one or two applications seem to
70 * Neither declaration needs to be modified here because the array
71 * sizes are fixed later when update_array_sizes is called.
73 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
75 * "Unlike user-defined varying variables, the built-in
76 * varying variables don't have a strict one-to-one
77 * correspondence between the vertex language and the
80 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
82 "%s shader output `%s' declared as type `%s', "
83 "but %s shader input declared as type `%s'\n",
84 _mesa_shader_stage_to_string(producer_stage
),
87 _mesa_shader_stage_to_string(consumer_stage
),
93 /* Check that all of the qualifiers match between stages.
95 if (input
->data
.centroid
!= output
->data
.centroid
) {
97 "%s shader output `%s' %s centroid qualifier, "
98 "but %s shader input %s centroid qualifier\n",
99 _mesa_shader_stage_to_string(producer_stage
),
101 (output
->data
.centroid
) ? "has" : "lacks",
102 _mesa_shader_stage_to_string(consumer_stage
),
103 (input
->data
.centroid
) ? "has" : "lacks");
107 if (input
->data
.sample
!= output
->data
.sample
) {
109 "%s shader output `%s' %s sample qualifier, "
110 "but %s shader input %s sample qualifier\n",
111 _mesa_shader_stage_to_string(producer_stage
),
113 (output
->data
.sample
) ? "has" : "lacks",
114 _mesa_shader_stage_to_string(consumer_stage
),
115 (input
->data
.sample
) ? "has" : "lacks");
119 if (input
->data
.invariant
!= output
->data
.invariant
) {
121 "%s shader output `%s' %s invariant qualifier, "
122 "but %s shader input %s invariant qualifier\n",
123 _mesa_shader_stage_to_string(producer_stage
),
125 (output
->data
.invariant
) ? "has" : "lacks",
126 _mesa_shader_stage_to_string(consumer_stage
),
127 (input
->data
.invariant
) ? "has" : "lacks");
131 if (input
->data
.interpolation
!= output
->data
.interpolation
) {
133 "%s shader output `%s' specifies %s "
134 "interpolation qualifier, "
135 "but %s shader input specifies %s "
136 "interpolation qualifier\n",
137 _mesa_shader_stage_to_string(producer_stage
),
139 interpolation_string(output
->data
.interpolation
),
140 _mesa_shader_stage_to_string(consumer_stage
),
141 interpolation_string(input
->data
.interpolation
));
147 * Validate front and back color outputs against single color input
150 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
151 const ir_variable
*input
,
152 const ir_variable
*front_color
,
153 const ir_variable
*back_color
,
154 gl_shader_stage consumer_stage
,
155 gl_shader_stage producer_stage
)
157 if (front_color
!= NULL
&& front_color
->data
.assigned
)
158 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
159 consumer_stage
, producer_stage
);
161 if (back_color
!= NULL
&& back_color
->data
.assigned
)
162 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
163 consumer_stage
, producer_stage
);
167 * Validate that outputs from one stage match inputs of another
170 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
171 gl_shader
*producer
, gl_shader
*consumer
)
173 glsl_symbol_table parameters
;
174 ir_variable
*explicit_locations
[MAX_VARYING
] = { NULL
, };
176 /* Find all shader outputs in the "producer" stage.
178 foreach_list(node
, producer
->ir
) {
179 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
181 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_shader_out
))
184 if (!var
->data
.explicit_location
185 || var
->data
.location
< VARYING_SLOT_VAR0
)
186 parameters
.add_variable(var
);
188 /* User-defined varyings with explicit locations are handled
189 * differently because they do not need to have matching names.
191 const unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
193 if (explicit_locations
[idx
] != NULL
) {
195 "%s shader has multiple outputs explicitly "
196 "assigned to location %d\n",
197 _mesa_shader_stage_to_string(producer
->Stage
),
202 explicit_locations
[idx
] = var
;
207 /* Find all shader inputs in the "consumer" stage. Any variables that have
208 * matching outputs already in the symbol table must have the same type and
211 * Exception: if the consumer is the geometry shader, then the inputs
212 * should be arrays and the type of the array element should match the type
213 * of the corresponding producer output.
215 foreach_list(node
, consumer
->ir
) {
216 ir_variable
*const input
= ((ir_instruction
*) node
)->as_variable();
218 if ((input
== NULL
) || (input
->data
.mode
!= ir_var_shader_in
))
221 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
222 const ir_variable
*const front_color
=
223 parameters
.get_variable("gl_FrontColor");
225 const ir_variable
*const back_color
=
226 parameters
.get_variable("gl_BackColor");
228 cross_validate_front_and_back_color(prog
, input
,
229 front_color
, back_color
,
230 consumer
->Stage
, producer
->Stage
);
231 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
232 const ir_variable
*const front_color
=
233 parameters
.get_variable("gl_FrontSecondaryColor");
235 const ir_variable
*const back_color
=
236 parameters
.get_variable("gl_BackSecondaryColor");
238 cross_validate_front_and_back_color(prog
, input
,
239 front_color
, back_color
,
240 consumer
->Stage
, producer
->Stage
);
242 /* The rules for connecting inputs and outputs change in the presence
243 * of explicit locations. In this case, we no longer care about the
244 * names of the variables. Instead, we care only about the
245 * explicitly assigned location.
247 ir_variable
*output
= NULL
;
248 if (input
->data
.explicit_location
249 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
250 output
= explicit_locations
[input
->data
.location
- VARYING_SLOT_VAR0
];
252 if (output
== NULL
) {
254 "%s shader input `%s' with explicit location "
255 "has no matching output\n",
256 _mesa_shader_stage_to_string(consumer
->Stage
),
260 output
= parameters
.get_variable(input
->name
);
263 if (output
!= NULL
) {
264 cross_validate_types_and_qualifiers(prog
, input
, output
,
265 consumer
->Stage
, producer
->Stage
);
273 * Initialize this object based on a string that was passed to
274 * glTransformFeedbackVaryings.
276 * If the input is mal-formed, this call still succeeds, but it sets
277 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
278 * will fail to find any matching variable.
281 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
284 /* We don't have to be pedantic about what is a valid GLSL variable name,
285 * because any variable with an invalid name can't exist in the IR anyway.
289 this->orig_name
= input
;
290 this->is_clip_distance_mesa
= false;
291 this->skip_components
= 0;
292 this->next_buffer_separator
= false;
293 this->matched_candidate
= NULL
;
295 if (ctx
->Extensions
.ARB_transform_feedback3
) {
296 /* Parse gl_NextBuffer. */
297 if (strcmp(input
, "gl_NextBuffer") == 0) {
298 this->next_buffer_separator
= true;
302 /* Parse gl_SkipComponents. */
303 if (strcmp(input
, "gl_SkipComponents1") == 0)
304 this->skip_components
= 1;
305 else if (strcmp(input
, "gl_SkipComponents2") == 0)
306 this->skip_components
= 2;
307 else if (strcmp(input
, "gl_SkipComponents3") == 0)
308 this->skip_components
= 3;
309 else if (strcmp(input
, "gl_SkipComponents4") == 0)
310 this->skip_components
= 4;
312 if (this->skip_components
)
316 /* Parse a declaration. */
317 const char *base_name_end
;
318 long subscript
= parse_program_resource_name(input
, &base_name_end
);
319 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
320 if (subscript
>= 0) {
321 this->array_subscript
= subscript
;
322 this->is_subscripted
= true;
324 this->is_subscripted
= false;
327 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
328 * class must behave specially to account for the fact that gl_ClipDistance
329 * is converted from a float[8] to a vec4[2].
331 if (ctx
->ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerClipDistance
&&
332 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
333 this->is_clip_distance_mesa
= true;
339 * Determine whether two tfeedback_decl objects refer to the same variable and
340 * array index (if applicable).
343 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
345 assert(x
.is_varying() && y
.is_varying());
347 if (strcmp(x
.var_name
, y
.var_name
) != 0)
349 if (x
.is_subscripted
!= y
.is_subscripted
)
351 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
358 * Assign a location for this tfeedback_decl object based on the transform
359 * feedback candidate found by find_candidate.
361 * If an error occurs, the error is reported through linker_error() and false
365 tfeedback_decl::assign_location(struct gl_context
*ctx
,
366 struct gl_shader_program
*prog
)
368 assert(this->is_varying());
370 unsigned fine_location
371 = this->matched_candidate
->toplevel_var
->data
.location
* 4
372 + this->matched_candidate
->toplevel_var
->data
.location_frac
373 + this->matched_candidate
->offset
;
375 if (this->matched_candidate
->type
->is_array()) {
377 const unsigned matrix_cols
=
378 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
379 const unsigned vector_elements
=
380 this->matched_candidate
->type
->fields
.array
->vector_elements
;
381 unsigned actual_array_size
= this->is_clip_distance_mesa
?
382 prog
->LastClipDistanceArraySize
:
383 this->matched_candidate
->type
->array_size();
385 if (this->is_subscripted
) {
386 /* Check array bounds. */
387 if (this->array_subscript
>= actual_array_size
) {
388 linker_error(prog
, "Transform feedback varying %s has index "
389 "%i, but the array size is %u.",
390 this->orig_name
, this->array_subscript
,
394 unsigned array_elem_size
= this->is_clip_distance_mesa
?
395 1 : vector_elements
* matrix_cols
;
396 fine_location
+= array_elem_size
* this->array_subscript
;
399 this->size
= actual_array_size
;
401 this->vector_elements
= vector_elements
;
402 this->matrix_columns
= matrix_cols
;
403 if (this->is_clip_distance_mesa
)
404 this->type
= GL_FLOAT
;
406 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
408 /* Regular variable (scalar, vector, or matrix) */
409 if (this->is_subscripted
) {
410 linker_error(prog
, "Transform feedback varying %s requested, "
411 "but %s is not an array.",
412 this->orig_name
, this->var_name
);
416 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
417 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
418 this->type
= this->matched_candidate
->type
->gl_type
;
420 this->location
= fine_location
/ 4;
421 this->location_frac
= fine_location
% 4;
423 /* From GL_EXT_transform_feedback:
424 * A program will fail to link if:
426 * * the total number of components to capture in any varying
427 * variable in <varyings> is greater than the constant
428 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
429 * buffer mode is SEPARATE_ATTRIBS_EXT;
431 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
432 this->num_components() >
433 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
434 linker_error(prog
, "Transform feedback varying %s exceeds "
435 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
445 tfeedback_decl::get_num_outputs() const
447 if (!this->is_varying()) {
451 return (this->num_components() + this->location_frac
+ 3)/4;
456 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
458 * If an error occurs, the error is reported through linker_error() and false
462 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
463 struct gl_transform_feedback_info
*info
,
464 unsigned buffer
, const unsigned max_outputs
) const
466 assert(!this->next_buffer_separator
);
468 /* Handle gl_SkipComponents. */
469 if (this->skip_components
) {
470 info
->BufferStride
[buffer
] += this->skip_components
;
474 /* From GL_EXT_transform_feedback:
475 * A program will fail to link if:
477 * * the total number of components to capture is greater than
478 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
479 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
481 if (prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
&&
482 info
->BufferStride
[buffer
] + this->num_components() >
483 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
484 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
485 "limit has been exceeded.");
489 unsigned location
= this->location
;
490 unsigned location_frac
= this->location_frac
;
491 unsigned num_components
= this->num_components();
492 while (num_components
> 0) {
493 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
494 assert(info
->NumOutputs
< max_outputs
);
495 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
496 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
497 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
498 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
499 info
->Outputs
[info
->NumOutputs
].DstOffset
= info
->BufferStride
[buffer
];
501 info
->BufferStride
[buffer
] += output_size
;
502 num_components
-= output_size
;
507 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
, this->orig_name
);
508 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
509 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
516 const tfeedback_candidate
*
517 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
518 hash_table
*tfeedback_candidates
)
520 const char *name
= this->is_clip_distance_mesa
521 ? "gl_ClipDistanceMESA" : this->var_name
;
522 this->matched_candidate
= (const tfeedback_candidate
*)
523 hash_table_find(tfeedback_candidates
, name
);
524 if (!this->matched_candidate
) {
525 /* From GL_EXT_transform_feedback:
526 * A program will fail to link if:
528 * * any variable name specified in the <varyings> array is not
529 * declared as an output in the geometry shader (if present) or
530 * the vertex shader (if no geometry shader is present);
532 linker_error(prog
, "Transform feedback varying %s undeclared.",
535 return this->matched_candidate
;
540 * Parse all the transform feedback declarations that were passed to
541 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
543 * If an error occurs, the error is reported through linker_error() and false
547 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
548 const void *mem_ctx
, unsigned num_names
,
549 char **varying_names
, tfeedback_decl
*decls
)
551 for (unsigned i
= 0; i
< num_names
; ++i
) {
552 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
554 if (!decls
[i
].is_varying())
557 /* From GL_EXT_transform_feedback:
558 * A program will fail to link if:
560 * * any two entries in the <varyings> array specify the same varying
563 * We interpret this to mean "any two entries in the <varyings> array
564 * specify the same varying variable and array index", since transform
565 * feedback of arrays would be useless otherwise.
567 for (unsigned j
= 0; j
< i
; ++j
) {
568 if (!decls
[j
].is_varying())
571 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
572 linker_error(prog
, "Transform feedback varying %s specified "
573 "more than once.", varying_names
[i
]);
583 * Store transform feedback location assignments into
584 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
586 * If an error occurs, the error is reported through linker_error() and false
590 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
591 unsigned num_tfeedback_decls
,
592 tfeedback_decl
*tfeedback_decls
)
594 bool separate_attribs_mode
=
595 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
597 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
598 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
600 memset(&prog
->LinkedTransformFeedback
, 0,
601 sizeof(prog
->LinkedTransformFeedback
));
603 prog
->LinkedTransformFeedback
.Varyings
=
605 struct gl_transform_feedback_varying_info
,
606 num_tfeedback_decls
);
608 unsigned num_outputs
= 0;
609 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
)
610 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
612 prog
->LinkedTransformFeedback
.Outputs
=
614 struct gl_transform_feedback_output
,
617 unsigned num_buffers
= 0;
619 if (separate_attribs_mode
) {
620 /* GL_SEPARATE_ATTRIBS */
621 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
622 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
623 num_buffers
, num_outputs
))
630 /* GL_INVERLEAVED_ATTRIBS */
631 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
632 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
637 if (!tfeedback_decls
[i
].store(ctx
, prog
,
638 &prog
->LinkedTransformFeedback
,
639 num_buffers
, num_outputs
))
645 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
647 prog
->LinkedTransformFeedback
.NumBuffers
= num_buffers
;
654 * Data structure recording the relationship between outputs of one shader
655 * stage (the "producer") and inputs of another (the "consumer").
657 class varying_matches
660 varying_matches(bool disable_varying_packing
, bool consumer_is_fs
);
662 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
663 unsigned assign_locations();
664 void store_locations() const;
668 * If true, this driver disables varying packing, so all varyings need to
669 * be aligned on slot boundaries, and take up a number of slots equal to
670 * their number of matrix columns times their array size.
672 const bool disable_varying_packing
;
675 * Enum representing the order in which varyings are packed within a
678 * Currently we pack vec4's first, then vec2's, then scalar values, then
679 * vec3's. This order ensures that the only vectors that are at risk of
680 * having to be "double parked" (split between two adjacent varying slots)
683 enum packing_order_enum
{
686 PACKING_ORDER_SCALAR
,
690 static unsigned compute_packing_class(const ir_variable
*var
);
691 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
692 static int match_comparator(const void *x_generic
, const void *y_generic
);
695 * Structure recording the relationship between a single producer output
696 * and a single consumer input.
700 * Packing class for this varying, computed by compute_packing_class().
702 unsigned packing_class
;
705 * Packing order for this varying, computed by compute_packing_order().
707 packing_order_enum packing_order
;
708 unsigned num_components
;
711 * The output variable in the producer stage.
713 ir_variable
*producer_var
;
716 * The input variable in the consumer stage.
718 ir_variable
*consumer_var
;
721 * The location which has been assigned for this varying. This is
722 * expressed in multiples of a float, with the first generic varying
723 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
726 unsigned generic_location
;
730 * The number of elements in the \c matches array that are currently in
733 unsigned num_matches
;
736 * The number of elements that were set aside for the \c matches array when
739 unsigned matches_capacity
;
741 const bool consumer_is_fs
;
744 } /* anonymous namespace */
746 varying_matches::varying_matches(bool disable_varying_packing
,
748 : disable_varying_packing(disable_varying_packing
),
749 consumer_is_fs(consumer_is_fs
)
751 /* Note: this initial capacity is rather arbitrarily chosen to be large
752 * enough for many cases without wasting an unreasonable amount of space.
753 * varying_matches::record() will resize the array if there are more than
754 * this number of varyings.
756 this->matches_capacity
= 8;
757 this->matches
= (match
*)
758 malloc(sizeof(*this->matches
) * this->matches_capacity
);
759 this->num_matches
= 0;
763 varying_matches::~varying_matches()
770 * Record the given producer/consumer variable pair in the list of variables
771 * that should later be assigned locations.
773 * It is permissible for \c consumer_var to be NULL (this happens if a
774 * variable is output by the producer and consumed by transform feedback, but
775 * not consumed by the consumer).
777 * If \c producer_var has already been paired up with a consumer_var, or
778 * producer_var is part of fixed pipeline functionality (and hence already has
779 * a location assigned), this function has no effect.
781 * Note: as a side effect this function may change the interpolation type of
782 * \c producer_var, but only when the change couldn't possibly affect
786 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
788 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
790 if ((producer_var
&& !producer_var
->data
.is_unmatched_generic_inout
)
791 || (consumer_var
&& !consumer_var
->data
.is_unmatched_generic_inout
)) {
792 /* Either a location already exists for this variable (since it is part
793 * of fixed functionality), or it has already been recorded as part of a
799 if ((consumer_var
== NULL
&& producer_var
->type
->contains_integer()) ||
801 /* Since this varying is not being consumed by the fragment shader, its
802 * interpolation type varying cannot possibly affect rendering. Also,
803 * this variable is non-flat and is (or contains) an integer.
805 * lower_packed_varyings requires all integer varyings to flat,
806 * regardless of where they appear. We can trivially satisfy that
807 * requirement by changing the interpolation type to flat here.
809 producer_var
->data
.centroid
= false;
810 producer_var
->data
.sample
= false;
811 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
814 consumer_var
->data
.centroid
= false;
815 consumer_var
->data
.sample
= false;
816 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
820 if (this->num_matches
== this->matches_capacity
) {
821 this->matches_capacity
*= 2;
822 this->matches
= (match
*)
823 realloc(this->matches
,
824 sizeof(*this->matches
) * this->matches_capacity
);
827 const ir_variable
*const var
= (producer_var
!= NULL
)
828 ? producer_var
: consumer_var
;
830 this->matches
[this->num_matches
].packing_class
831 = this->compute_packing_class(var
);
832 this->matches
[this->num_matches
].packing_order
833 = this->compute_packing_order(var
);
834 if (this->disable_varying_packing
) {
835 unsigned slots
= var
->type
->is_array()
836 ? (var
->type
->length
* var
->type
->fields
.array
->matrix_columns
)
837 : var
->type
->matrix_columns
;
838 this->matches
[this->num_matches
].num_components
= 4 * slots
;
840 this->matches
[this->num_matches
].num_components
841 = var
->type
->component_slots();
843 this->matches
[this->num_matches
].producer_var
= producer_var
;
844 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
847 producer_var
->data
.is_unmatched_generic_inout
= 0;
849 consumer_var
->data
.is_unmatched_generic_inout
= 0;
854 * Choose locations for all of the variable matches that were previously
855 * passed to varying_matches::record().
858 varying_matches::assign_locations()
860 /* Sort varying matches into an order that makes them easy to pack. */
861 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
862 &varying_matches::match_comparator
);
864 unsigned generic_location
= 0;
866 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
867 /* Advance to the next slot if this varying has a different packing
868 * class than the previous one, and we're not already on a slot
872 this->matches
[i
- 1].packing_class
873 != this->matches
[i
].packing_class
) {
874 generic_location
= ALIGN(generic_location
, 4);
877 this->matches
[i
].generic_location
= generic_location
;
879 generic_location
+= this->matches
[i
].num_components
;
882 return (generic_location
+ 3) / 4;
887 * Update the producer and consumer shaders to reflect the locations
888 * assignments that were made by varying_matches::assign_locations().
891 varying_matches::store_locations() const
893 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
894 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
895 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
896 unsigned generic_location
= this->matches
[i
].generic_location
;
897 unsigned slot
= generic_location
/ 4;
898 unsigned offset
= generic_location
% 4;
901 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
902 producer_var
->data
.location_frac
= offset
;
906 assert(consumer_var
->data
.location
== -1);
907 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
908 consumer_var
->data
.location_frac
= offset
;
915 * Compute the "packing class" of the given varying. This is an unsigned
916 * integer with the property that two variables in the same packing class can
917 * be safely backed into the same vec4.
920 varying_matches::compute_packing_class(const ir_variable
*var
)
922 /* Without help from the back-end, there is no way to pack together
923 * variables with different interpolation types, because
924 * lower_packed_varyings must choose exactly one interpolation type for
925 * each packed varying it creates.
927 * However, we can safely pack together floats, ints, and uints, because:
929 * - varyings of base type "int" and "uint" must use the "flat"
930 * interpolation type, which can only occur in GLSL 1.30 and above.
932 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
933 * can store flat floats as ints without losing any information (using
934 * the ir_unop_bitcast_* opcodes).
936 * Therefore, the packing class depends only on the interpolation type.
938 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1);
940 packing_class
+= var
->data
.interpolation
;
941 return packing_class
;
946 * Compute the "packing order" of the given varying. This is a sort key we
947 * use to determine when to attempt to pack the given varying relative to
948 * other varyings in the same packing class.
950 varying_matches::packing_order_enum
951 varying_matches::compute_packing_order(const ir_variable
*var
)
953 const glsl_type
*element_type
= var
->type
;
955 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
956 element_type
= element_type
->fields
.array
;
959 switch (element_type
->component_slots() % 4) {
960 case 1: return PACKING_ORDER_SCALAR
;
961 case 2: return PACKING_ORDER_VEC2
;
962 case 3: return PACKING_ORDER_VEC3
;
963 case 0: return PACKING_ORDER_VEC4
;
965 assert(!"Unexpected value of vector_elements");
966 return PACKING_ORDER_VEC4
;
972 * Comparison function passed to qsort() to sort varyings by packing_class and
973 * then by packing_order.
976 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
978 const match
*x
= (const match
*) x_generic
;
979 const match
*y
= (const match
*) y_generic
;
981 if (x
->packing_class
!= y
->packing_class
)
982 return x
->packing_class
- y
->packing_class
;
983 return x
->packing_order
- y
->packing_order
;
988 * Is the given variable a varying variable to be counted against the
989 * limit in ctx->Const.MaxVarying?
990 * This includes variables such as texcoords, colors and generic
991 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
994 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
996 /* Only fragment shaders will take a varying variable as an input */
997 if (stage
== MESA_SHADER_FRAGMENT
&&
998 var
->data
.mode
== ir_var_shader_in
) {
999 switch (var
->data
.location
) {
1000 case VARYING_SLOT_POS
:
1001 case VARYING_SLOT_FACE
:
1002 case VARYING_SLOT_PNTC
:
1013 * Visitor class that generates tfeedback_candidate structs describing all
1014 * possible targets of transform feedback.
1016 * tfeedback_candidate structs are stored in the hash table
1017 * tfeedback_candidates, which is passed to the constructor. This hash table
1018 * maps varying names to instances of the tfeedback_candidate struct.
1020 class tfeedback_candidate_generator
: public program_resource_visitor
1023 tfeedback_candidate_generator(void *mem_ctx
,
1024 hash_table
*tfeedback_candidates
)
1026 tfeedback_candidates(tfeedback_candidates
),
1032 void process(ir_variable
*var
)
1034 this->toplevel_var
= var
;
1035 this->varying_floats
= 0;
1036 if (var
->is_interface_instance())
1037 program_resource_visitor::process(var
->get_interface_type(),
1038 var
->get_interface_type()->name
);
1040 program_resource_visitor::process(var
);
1044 virtual void visit_field(const glsl_type
*type
, const char *name
,
1047 assert(!type
->is_record());
1048 assert(!(type
->is_array() && type
->fields
.array
->is_record()));
1049 assert(!type
->is_interface());
1050 assert(!(type
->is_array() && type
->fields
.array
->is_interface()));
1054 tfeedback_candidate
*candidate
1055 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1056 candidate
->toplevel_var
= this->toplevel_var
;
1057 candidate
->type
= type
;
1058 candidate
->offset
= this->varying_floats
;
1059 hash_table_insert(this->tfeedback_candidates
, candidate
,
1060 ralloc_strdup(this->mem_ctx
, name
));
1061 this->varying_floats
+= type
->component_slots();
1065 * Memory context used to allocate hash table keys and values.
1067 void * const mem_ctx
;
1070 * Hash table in which tfeedback_candidate objects should be stored.
1072 hash_table
* const tfeedback_candidates
;
1075 * Pointer to the toplevel variable that is being traversed.
1077 ir_variable
*toplevel_var
;
1080 * Total number of varying floats that have been visited so far. This is
1081 * used to determine the offset to each varying within the toplevel
1084 unsigned varying_floats
;
1091 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1092 hash_table
*consumer_inputs
,
1093 hash_table
*consumer_interface_inputs
,
1094 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
])
1096 memset(consumer_inputs_with_locations
,
1098 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_MAX
);
1100 foreach_list(node
, ir
) {
1101 ir_variable
*const input_var
= ((ir_instruction
*) node
)->as_variable();
1103 if ((input_var
!= NULL
) && (input_var
->data
.mode
== ir_var_shader_in
)) {
1104 if (input_var
->type
->is_interface())
1107 if (input_var
->data
.explicit_location
) {
1108 /* assign_varying_locations only cares about finding the
1109 * ir_variable at the start of a contiguous location block.
1111 * - For !producer, consumer_inputs_with_locations isn't used.
1113 * - For !consumer, consumer_inputs_with_locations is empty.
1115 * For consumer && producer, if you were trying to set some
1116 * ir_variable to the middle of a location block on the other side
1117 * of producer/consumer, cross_validate_outputs_to_inputs() should
1118 * be link-erroring due to either type mismatch or location
1119 * overlaps. If the variables do match up, then they've got a
1120 * matching data.location and you only looked at
1121 * consumer_inputs_with_locations[var->data.location], not any
1122 * following entries for the array/structure.
1124 consumer_inputs_with_locations
[input_var
->data
.location
] =
1126 } else if (input_var
->get_interface_type() != NULL
) {
1127 char *const iface_field_name
=
1128 ralloc_asprintf(mem_ctx
, "%s.%s",
1129 input_var
->get_interface_type()->name
,
1131 hash_table_insert(consumer_interface_inputs
, input_var
,
1134 hash_table_insert(consumer_inputs
, input_var
,
1135 ralloc_strdup(mem_ctx
, input_var
->name
));
1144 * Find a variable from the consumer that "matches" the specified variable
1146 * This function only finds inputs with names that match. There is no
1147 * validation (here) that the types, etc. are compatible.
1150 get_matching_input(void *mem_ctx
,
1151 const ir_variable
*output_var
,
1152 hash_table
*consumer_inputs
,
1153 hash_table
*consumer_interface_inputs
,
1154 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
])
1156 ir_variable
*input_var
;
1158 if (output_var
->data
.explicit_location
) {
1159 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1160 } else if (output_var
->get_interface_type() != NULL
) {
1161 char *const iface_field_name
=
1162 ralloc_asprintf(mem_ctx
, "%s.%s",
1163 output_var
->get_interface_type()->name
,
1166 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1170 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1173 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1180 io_variable_cmp(const void *_a
, const void *_b
)
1182 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1183 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1185 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1186 return b
->data
.location
- a
->data
.location
;
1188 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1191 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1194 return -strcmp(a
->name
, b
->name
);
1198 * Sort the shader IO variables into canonical order
1201 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1203 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1204 unsigned num_variables
= 0;
1206 foreach_list(node
, ir
) {
1207 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1209 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1212 /* If we have already encountered more I/O variables that could
1213 * successfully link, bail.
1215 if (num_variables
== ARRAY_SIZE(var_table
))
1218 var_table
[num_variables
++] = var
;
1221 if (num_variables
== 0)
1224 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1225 * we're going to push the variables on to the IR list as a stack, so we
1226 * want the last variable (in canonical order) to be first in the list.
1228 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1230 /* Remove the variable from it's current location in the IR, and put it at
1233 for (unsigned i
= 0; i
< num_variables
; i
++) {
1234 var_table
[i
]->remove();
1235 ir
->push_head(var_table
[i
]);
1240 * Assign locations for all variables that are produced in one pipeline stage
1241 * (the "producer") and consumed in the next stage (the "consumer").
1243 * Variables produced by the producer may also be consumed by transform
1246 * \param num_tfeedback_decls is the number of declarations indicating
1247 * variables that may be consumed by transform feedback.
1249 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1250 * representing the result of parsing the strings passed to
1251 * glTransformFeedbackVaryings(). assign_location() will be called for
1252 * each of these objects that matches one of the outputs of the
1255 * \param gs_input_vertices: if \c consumer is a geometry shader, this is the
1256 * number of input vertices it accepts. Otherwise zero.
1258 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1259 * be NULL. In this case, varying locations are assigned solely based on the
1260 * requirements of transform feedback.
1263 assign_varying_locations(struct gl_context
*ctx
,
1265 struct gl_shader_program
*prog
,
1266 gl_shader
*producer
, gl_shader
*consumer
,
1267 unsigned num_tfeedback_decls
,
1268 tfeedback_decl
*tfeedback_decls
,
1269 unsigned gs_input_vertices
)
1271 varying_matches
matches(ctx
->Const
.DisableVaryingPacking
,
1272 consumer
&& consumer
->Stage
== MESA_SHADER_FRAGMENT
);
1273 hash_table
*tfeedback_candidates
1274 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1275 hash_table
*consumer_inputs
1276 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1277 hash_table
*consumer_interface_inputs
1278 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1279 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
] = {
1283 /* Operate in a total of four passes.
1285 * 1. Sort inputs / outputs into a canonical order. This is necessary so
1286 * that inputs / outputs of separable shaders will be assigned
1287 * predictable locations regardless of the order in which declarations
1288 * appeared in the shader source.
1290 * 2. Assign locations for any matching inputs and outputs.
1292 * 3. Mark output variables in the producer that do not have locations as
1293 * not being outputs. This lets the optimizer eliminate them.
1295 * 4. Mark input variables in the consumer that do not have locations as
1296 * not being inputs. This lets the optimizer eliminate them.
1299 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
1302 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
1305 && !linker::populate_consumer_input_sets(mem_ctx
,
1308 consumer_interface_inputs
,
1309 consumer_inputs_with_locations
)) {
1310 assert(!"populate_consumer_input_sets failed");
1311 hash_table_dtor(tfeedback_candidates
);
1312 hash_table_dtor(consumer_inputs
);
1313 hash_table_dtor(consumer_interface_inputs
);
1318 foreach_list(node
, producer
->ir
) {
1319 ir_variable
*const output_var
=
1320 ((ir_instruction
*) node
)->as_variable();
1322 if ((output_var
== NULL
) ||
1323 (output_var
->data
.mode
!= ir_var_shader_out
))
1326 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
1327 g
.process(output_var
);
1329 ir_variable
*const input_var
=
1330 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
1331 consumer_interface_inputs
,
1332 consumer_inputs_with_locations
);
1334 /* If a matching input variable was found, add this ouptut (and the
1335 * input) to the set. If this is a separable program and there is no
1336 * consumer stage, add the output.
1338 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
)) {
1339 matches
.record(output_var
, input_var
);
1343 /* If there's no producer stage, then this must be a separable program.
1344 * For example, we may have a program that has just a fragment shader.
1345 * Later this program will be used with some arbitrary vertex (or
1346 * geometry) shader program. This means that locations must be assigned
1347 * for all the inputs.
1349 foreach_list(node
, consumer
->ir
) {
1350 ir_variable
*const input_var
=
1351 ((ir_instruction
*) node
)->as_variable();
1353 if ((input_var
== NULL
) ||
1354 (input_var
->data
.mode
!= ir_var_shader_in
))
1357 matches
.record(NULL
, input_var
);
1361 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1362 if (!tfeedback_decls
[i
].is_varying())
1365 const tfeedback_candidate
*matched_candidate
1366 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
1368 if (matched_candidate
== NULL
) {
1369 hash_table_dtor(tfeedback_candidates
);
1370 hash_table_dtor(consumer_inputs
);
1371 hash_table_dtor(consumer_interface_inputs
);
1375 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
)
1376 matches
.record(matched_candidate
->toplevel_var
, NULL
);
1379 const unsigned slots_used
= matches
.assign_locations();
1380 matches
.store_locations();
1382 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1383 if (!tfeedback_decls
[i
].is_varying())
1386 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
1387 hash_table_dtor(tfeedback_candidates
);
1388 hash_table_dtor(consumer_inputs
);
1389 hash_table_dtor(consumer_interface_inputs
);
1394 hash_table_dtor(tfeedback_candidates
);
1395 hash_table_dtor(consumer_inputs
);
1396 hash_table_dtor(consumer_interface_inputs
);
1398 if (ctx
->Const
.DisableVaryingPacking
) {
1399 /* Transform feedback code assumes varyings are packed, so if the driver
1400 * has disabled varying packing, make sure it does not support transform
1403 assert(!ctx
->Extensions
.EXT_transform_feedback
);
1406 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
1410 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
1411 gs_input_vertices
, consumer
);
1415 if (consumer
&& producer
) {
1416 foreach_list(node
, consumer
->ir
) {
1417 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1419 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1420 var
->data
.is_unmatched_generic_inout
) {
1421 if (prog
->Version
<= 120) {
1422 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1424 * Only those varying variables used (i.e. read) in
1425 * the fragment shader executable must be written to
1426 * by the vertex shader executable; declaring
1427 * superfluous varying variables in a vertex shader is
1430 * We interpret this text as meaning that the VS must
1431 * write the variable for the FS to read it. See
1432 * "glsl1-varying read but not written" in piglit.
1435 linker_error(prog
, "%s shader varying %s not written "
1437 _mesa_shader_stage_to_string(consumer
->Stage
),
1439 _mesa_shader_stage_to_string(producer
->Stage
));
1442 /* An 'in' variable is only really a shader input if its
1443 * value is written by the previous stage.
1445 var
->data
.mode
= ir_var_auto
;
1454 check_against_output_limit(struct gl_context
*ctx
,
1455 struct gl_shader_program
*prog
,
1456 gl_shader
*producer
)
1458 unsigned output_vectors
= 0;
1460 foreach_list(node
, producer
->ir
) {
1461 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1463 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
1464 var_counts_against_varying_limit(producer
->Stage
, var
)) {
1465 output_vectors
+= var
->type
->count_attribute_slots();
1469 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
1470 unsigned max_output_components
=
1471 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
1473 const unsigned output_components
= output_vectors
* 4;
1474 if (output_components
> max_output_components
) {
1475 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1476 linker_error(prog
, "shader uses too many output vectors "
1479 max_output_components
/ 4);
1481 linker_error(prog
, "shader uses too many output components "
1484 max_output_components
);
1493 check_against_input_limit(struct gl_context
*ctx
,
1494 struct gl_shader_program
*prog
,
1495 gl_shader
*consumer
)
1497 unsigned input_vectors
= 0;
1499 foreach_list(node
, consumer
->ir
) {
1500 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1502 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1503 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
1504 input_vectors
+= var
->type
->count_attribute_slots();
1508 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
1509 unsigned max_input_components
=
1510 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
1512 const unsigned input_components
= input_vectors
* 4;
1513 if (input_components
> max_input_components
) {
1514 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1515 linker_error(prog
, "shader uses too many input vectors "
1518 max_input_components
/ 4);
1520 linker_error(prog
, "shader uses too many input components "
1523 max_input_components
);