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()
81 || (strncmp("gl_", output
->name
, 3) != 0)) {
83 "%s shader output `%s' declared as type `%s', "
84 "but %s shader input declared as type `%s'\n",
85 _mesa_shader_stage_to_string(producer_stage
),
88 _mesa_shader_stage_to_string(consumer_stage
),
94 /* Check that all of the qualifiers match between stages.
96 if (input
->data
.centroid
!= output
->data
.centroid
) {
98 "%s shader output `%s' %s centroid qualifier, "
99 "but %s shader input %s centroid qualifier\n",
100 _mesa_shader_stage_to_string(producer_stage
),
102 (output
->data
.centroid
) ? "has" : "lacks",
103 _mesa_shader_stage_to_string(consumer_stage
),
104 (input
->data
.centroid
) ? "has" : "lacks");
108 if (input
->data
.sample
!= output
->data
.sample
) {
110 "%s shader output `%s' %s sample qualifier, "
111 "but %s shader input %s sample qualifier\n",
112 _mesa_shader_stage_to_string(producer_stage
),
114 (output
->data
.sample
) ? "has" : "lacks",
115 _mesa_shader_stage_to_string(consumer_stage
),
116 (input
->data
.sample
) ? "has" : "lacks");
120 if (input
->data
.invariant
!= output
->data
.invariant
) {
122 "%s shader output `%s' %s invariant qualifier, "
123 "but %s shader input %s invariant qualifier\n",
124 _mesa_shader_stage_to_string(producer_stage
),
126 (output
->data
.invariant
) ? "has" : "lacks",
127 _mesa_shader_stage_to_string(consumer_stage
),
128 (input
->data
.invariant
) ? "has" : "lacks");
132 if (input
->data
.interpolation
!= output
->data
.interpolation
) {
134 "%s shader output `%s' specifies %s "
135 "interpolation qualifier, "
136 "but %s shader input specifies %s "
137 "interpolation qualifier\n",
138 _mesa_shader_stage_to_string(producer_stage
),
140 interpolation_string(output
->data
.interpolation
),
141 _mesa_shader_stage_to_string(consumer_stage
),
142 interpolation_string(input
->data
.interpolation
));
148 * Validate front and back color outputs against single color input
151 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
152 const ir_variable
*input
,
153 const ir_variable
*front_color
,
154 const ir_variable
*back_color
,
155 gl_shader_stage consumer_stage
,
156 gl_shader_stage producer_stage
)
158 if (front_color
!= NULL
&& front_color
->data
.assigned
)
159 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
160 consumer_stage
, producer_stage
);
162 if (back_color
!= NULL
&& back_color
->data
.assigned
)
163 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
164 consumer_stage
, producer_stage
);
168 * Validate that outputs from one stage match inputs of another
171 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
172 gl_shader
*producer
, gl_shader
*consumer
)
174 glsl_symbol_table parameters
;
175 ir_variable
*explicit_locations
[MAX_VARYING
] = { NULL
, };
177 /* Find all shader outputs in the "producer" stage.
179 foreach_list(node
, producer
->ir
) {
180 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
182 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_shader_out
))
185 if (!var
->data
.explicit_location
186 || var
->data
.location
< VARYING_SLOT_VAR0
)
187 parameters
.add_variable(var
);
189 /* User-defined varyings with explicit locations are handled
190 * differently because they do not need to have matching names.
192 const unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
194 if (explicit_locations
[idx
] != NULL
) {
196 "%s shader has multiple outputs explicitly "
197 "assigned to location %d\n",
198 _mesa_shader_stage_to_string(producer
->Stage
),
203 explicit_locations
[idx
] = var
;
208 /* Find all shader inputs in the "consumer" stage. Any variables that have
209 * matching outputs already in the symbol table must have the same type and
212 * Exception: if the consumer is the geometry shader, then the inputs
213 * should be arrays and the type of the array element should match the type
214 * of the corresponding producer output.
216 foreach_list(node
, consumer
->ir
) {
217 ir_variable
*const input
= ((ir_instruction
*) node
)->as_variable();
219 if ((input
== NULL
) || (input
->data
.mode
!= ir_var_shader_in
))
222 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
223 const ir_variable
*const front_color
=
224 parameters
.get_variable("gl_FrontColor");
226 const ir_variable
*const back_color
=
227 parameters
.get_variable("gl_BackColor");
229 cross_validate_front_and_back_color(prog
, input
,
230 front_color
, back_color
,
231 consumer
->Stage
, producer
->Stage
);
232 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
233 const ir_variable
*const front_color
=
234 parameters
.get_variable("gl_FrontSecondaryColor");
236 const ir_variable
*const back_color
=
237 parameters
.get_variable("gl_BackSecondaryColor");
239 cross_validate_front_and_back_color(prog
, input
,
240 front_color
, back_color
,
241 consumer
->Stage
, producer
->Stage
);
243 /* The rules for connecting inputs and outputs change in the presence
244 * of explicit locations. In this case, we no longer care about the
245 * names of the variables. Instead, we care only about the
246 * explicitly assigned location.
248 ir_variable
*output
= NULL
;
249 if (input
->data
.explicit_location
250 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
251 output
= explicit_locations
[input
->data
.location
- VARYING_SLOT_VAR0
];
253 if (output
== NULL
) {
255 "%s shader input `%s' with explicit location "
256 "has no matching output\n",
257 _mesa_shader_stage_to_string(consumer
->Stage
),
261 output
= parameters
.get_variable(input
->name
);
264 if (output
!= NULL
) {
265 cross_validate_types_and_qualifiers(prog
, input
, output
,
266 consumer
->Stage
, producer
->Stage
);
274 * Initialize this object based on a string that was passed to
275 * glTransformFeedbackVaryings.
277 * If the input is mal-formed, this call still succeeds, but it sets
278 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
279 * will fail to find any matching variable.
282 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
285 /* We don't have to be pedantic about what is a valid GLSL variable name,
286 * because any variable with an invalid name can't exist in the IR anyway.
290 this->orig_name
= input
;
291 this->is_clip_distance_mesa
= false;
292 this->skip_components
= 0;
293 this->next_buffer_separator
= false;
294 this->matched_candidate
= NULL
;
296 if (ctx
->Extensions
.ARB_transform_feedback3
) {
297 /* Parse gl_NextBuffer. */
298 if (strcmp(input
, "gl_NextBuffer") == 0) {
299 this->next_buffer_separator
= true;
303 /* Parse gl_SkipComponents. */
304 if (strcmp(input
, "gl_SkipComponents1") == 0)
305 this->skip_components
= 1;
306 else if (strcmp(input
, "gl_SkipComponents2") == 0)
307 this->skip_components
= 2;
308 else if (strcmp(input
, "gl_SkipComponents3") == 0)
309 this->skip_components
= 3;
310 else if (strcmp(input
, "gl_SkipComponents4") == 0)
311 this->skip_components
= 4;
313 if (this->skip_components
)
317 /* Parse a declaration. */
318 const char *base_name_end
;
319 long subscript
= parse_program_resource_name(input
, &base_name_end
);
320 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
321 if (subscript
>= 0) {
322 this->array_subscript
= subscript
;
323 this->is_subscripted
= true;
325 this->is_subscripted
= false;
328 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
329 * class must behave specially to account for the fact that gl_ClipDistance
330 * is converted from a float[8] to a vec4[2].
332 if (ctx
->ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerClipDistance
&&
333 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
334 this->is_clip_distance_mesa
= true;
340 * Determine whether two tfeedback_decl objects refer to the same variable and
341 * array index (if applicable).
344 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
346 assert(x
.is_varying() && y
.is_varying());
348 if (strcmp(x
.var_name
, y
.var_name
) != 0)
350 if (x
.is_subscripted
!= y
.is_subscripted
)
352 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
359 * Assign a location for this tfeedback_decl object based on the transform
360 * feedback candidate found by find_candidate.
362 * If an error occurs, the error is reported through linker_error() and false
366 tfeedback_decl::assign_location(struct gl_context
*ctx
,
367 struct gl_shader_program
*prog
)
369 assert(this->is_varying());
371 unsigned fine_location
372 = this->matched_candidate
->toplevel_var
->data
.location
* 4
373 + this->matched_candidate
->toplevel_var
->data
.location_frac
374 + this->matched_candidate
->offset
;
376 if (this->matched_candidate
->type
->is_array()) {
378 const unsigned matrix_cols
=
379 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
380 const unsigned vector_elements
=
381 this->matched_candidate
->type
->fields
.array
->vector_elements
;
382 unsigned actual_array_size
= this->is_clip_distance_mesa
?
383 prog
->LastClipDistanceArraySize
:
384 this->matched_candidate
->type
->array_size();
386 if (this->is_subscripted
) {
387 /* Check array bounds. */
388 if (this->array_subscript
>= actual_array_size
) {
389 linker_error(prog
, "Transform feedback varying %s has index "
390 "%i, but the array size is %u.",
391 this->orig_name
, this->array_subscript
,
395 unsigned array_elem_size
= this->is_clip_distance_mesa
?
396 1 : vector_elements
* matrix_cols
;
397 fine_location
+= array_elem_size
* this->array_subscript
;
400 this->size
= actual_array_size
;
402 this->vector_elements
= vector_elements
;
403 this->matrix_columns
= matrix_cols
;
404 if (this->is_clip_distance_mesa
)
405 this->type
= GL_FLOAT
;
407 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
409 /* Regular variable (scalar, vector, or matrix) */
410 if (this->is_subscripted
) {
411 linker_error(prog
, "Transform feedback varying %s requested, "
412 "but %s is not an array.",
413 this->orig_name
, this->var_name
);
417 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
418 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
419 this->type
= this->matched_candidate
->type
->gl_type
;
421 this->location
= fine_location
/ 4;
422 this->location_frac
= fine_location
% 4;
424 /* From GL_EXT_transform_feedback:
425 * A program will fail to link if:
427 * * the total number of components to capture in any varying
428 * variable in <varyings> is greater than the constant
429 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
430 * buffer mode is SEPARATE_ATTRIBS_EXT;
432 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
433 this->num_components() >
434 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
435 linker_error(prog
, "Transform feedback varying %s exceeds "
436 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
446 tfeedback_decl::get_num_outputs() const
448 if (!this->is_varying()) {
452 return (this->num_components() + this->location_frac
+ 3)/4;
457 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
459 * If an error occurs, the error is reported through linker_error() and false
463 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
464 struct gl_transform_feedback_info
*info
,
465 unsigned buffer
, const unsigned max_outputs
) const
467 assert(!this->next_buffer_separator
);
469 /* Handle gl_SkipComponents. */
470 if (this->skip_components
) {
471 info
->BufferStride
[buffer
] += this->skip_components
;
475 /* From GL_EXT_transform_feedback:
476 * A program will fail to link if:
478 * * the total number of components to capture is greater than
479 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
480 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
482 if (prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
&&
483 info
->BufferStride
[buffer
] + this->num_components() >
484 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
485 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
486 "limit has been exceeded.");
490 unsigned location
= this->location
;
491 unsigned location_frac
= this->location_frac
;
492 unsigned num_components
= this->num_components();
493 while (num_components
> 0) {
494 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
495 assert(info
->NumOutputs
< max_outputs
);
496 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
497 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
498 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
499 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
500 info
->Outputs
[info
->NumOutputs
].DstOffset
= info
->BufferStride
[buffer
];
502 info
->BufferStride
[buffer
] += output_size
;
503 num_components
-= output_size
;
508 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
, this->orig_name
);
509 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
510 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
517 const tfeedback_candidate
*
518 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
519 hash_table
*tfeedback_candidates
)
521 const char *name
= this->is_clip_distance_mesa
522 ? "gl_ClipDistanceMESA" : this->var_name
;
523 this->matched_candidate
= (const tfeedback_candidate
*)
524 hash_table_find(tfeedback_candidates
, name
);
525 if (!this->matched_candidate
) {
526 /* From GL_EXT_transform_feedback:
527 * A program will fail to link if:
529 * * any variable name specified in the <varyings> array is not
530 * declared as an output in the geometry shader (if present) or
531 * the vertex shader (if no geometry shader is present);
533 linker_error(prog
, "Transform feedback varying %s undeclared.",
536 return this->matched_candidate
;
541 * Parse all the transform feedback declarations that were passed to
542 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
544 * If an error occurs, the error is reported through linker_error() and false
548 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
549 const void *mem_ctx
, unsigned num_names
,
550 char **varying_names
, tfeedback_decl
*decls
)
552 for (unsigned i
= 0; i
< num_names
; ++i
) {
553 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
555 if (!decls
[i
].is_varying())
558 /* From GL_EXT_transform_feedback:
559 * A program will fail to link if:
561 * * any two entries in the <varyings> array specify the same varying
564 * We interpret this to mean "any two entries in the <varyings> array
565 * specify the same varying variable and array index", since transform
566 * feedback of arrays would be useless otherwise.
568 for (unsigned j
= 0; j
< i
; ++j
) {
569 if (!decls
[j
].is_varying())
572 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
573 linker_error(prog
, "Transform feedback varying %s specified "
574 "more than once.", varying_names
[i
]);
584 * Store transform feedback location assignments into
585 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
587 * If an error occurs, the error is reported through linker_error() and false
591 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
592 unsigned num_tfeedback_decls
,
593 tfeedback_decl
*tfeedback_decls
)
595 bool separate_attribs_mode
=
596 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
598 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
599 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
601 memset(&prog
->LinkedTransformFeedback
, 0,
602 sizeof(prog
->LinkedTransformFeedback
));
604 prog
->LinkedTransformFeedback
.Varyings
=
606 struct gl_transform_feedback_varying_info
,
607 num_tfeedback_decls
);
609 unsigned num_outputs
= 0;
610 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
)
611 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
613 prog
->LinkedTransformFeedback
.Outputs
=
615 struct gl_transform_feedback_output
,
618 unsigned num_buffers
= 0;
620 if (separate_attribs_mode
) {
621 /* GL_SEPARATE_ATTRIBS */
622 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
623 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
624 num_buffers
, num_outputs
))
631 /* GL_INVERLEAVED_ATTRIBS */
632 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
633 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
638 if (!tfeedback_decls
[i
].store(ctx
, prog
,
639 &prog
->LinkedTransformFeedback
,
640 num_buffers
, num_outputs
))
646 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
648 prog
->LinkedTransformFeedback
.NumBuffers
= num_buffers
;
655 * Data structure recording the relationship between outputs of one shader
656 * stage (the "producer") and inputs of another (the "consumer").
658 class varying_matches
661 varying_matches(bool disable_varying_packing
, bool consumer_is_fs
);
663 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
664 unsigned assign_locations();
665 void store_locations() const;
669 * If true, this driver disables varying packing, so all varyings need to
670 * be aligned on slot boundaries, and take up a number of slots equal to
671 * their number of matrix columns times their array size.
673 const bool disable_varying_packing
;
676 * Enum representing the order in which varyings are packed within a
679 * Currently we pack vec4's first, then vec2's, then scalar values, then
680 * vec3's. This order ensures that the only vectors that are at risk of
681 * having to be "double parked" (split between two adjacent varying slots)
684 enum packing_order_enum
{
687 PACKING_ORDER_SCALAR
,
691 static unsigned compute_packing_class(const ir_variable
*var
);
692 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
693 static int match_comparator(const void *x_generic
, const void *y_generic
);
696 * Structure recording the relationship between a single producer output
697 * and a single consumer input.
701 * Packing class for this varying, computed by compute_packing_class().
703 unsigned packing_class
;
706 * Packing order for this varying, computed by compute_packing_order().
708 packing_order_enum packing_order
;
709 unsigned num_components
;
712 * The output variable in the producer stage.
714 ir_variable
*producer_var
;
717 * The input variable in the consumer stage.
719 ir_variable
*consumer_var
;
722 * The location which has been assigned for this varying. This is
723 * expressed in multiples of a float, with the first generic varying
724 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
727 unsigned generic_location
;
731 * The number of elements in the \c matches array that are currently in
734 unsigned num_matches
;
737 * The number of elements that were set aside for the \c matches array when
740 unsigned matches_capacity
;
742 const bool consumer_is_fs
;
745 } /* anonymous namespace */
747 varying_matches::varying_matches(bool disable_varying_packing
,
749 : disable_varying_packing(disable_varying_packing
),
750 consumer_is_fs(consumer_is_fs
)
752 /* Note: this initial capacity is rather arbitrarily chosen to be large
753 * enough for many cases without wasting an unreasonable amount of space.
754 * varying_matches::record() will resize the array if there are more than
755 * this number of varyings.
757 this->matches_capacity
= 8;
758 this->matches
= (match
*)
759 malloc(sizeof(*this->matches
) * this->matches_capacity
);
760 this->num_matches
= 0;
764 varying_matches::~varying_matches()
771 * Record the given producer/consumer variable pair in the list of variables
772 * that should later be assigned locations.
774 * It is permissible for \c consumer_var to be NULL (this happens if a
775 * variable is output by the producer and consumed by transform feedback, but
776 * not consumed by the consumer).
778 * If \c producer_var has already been paired up with a consumer_var, or
779 * producer_var is part of fixed pipeline functionality (and hence already has
780 * a location assigned), this function has no effect.
782 * Note: as a side effect this function may change the interpolation type of
783 * \c producer_var, but only when the change couldn't possibly affect
787 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
789 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
791 if ((producer_var
&& !producer_var
->data
.is_unmatched_generic_inout
)
792 || (consumer_var
&& !consumer_var
->data
.is_unmatched_generic_inout
)) {
793 /* Either a location already exists for this variable (since it is part
794 * of fixed functionality), or it has already been recorded as part of a
800 if ((consumer_var
== NULL
&& producer_var
->type
->contains_integer()) ||
802 /* Since this varying is not being consumed by the fragment shader, its
803 * interpolation type varying cannot possibly affect rendering. Also,
804 * this variable is non-flat and is (or contains) an integer.
806 * lower_packed_varyings requires all integer varyings to flat,
807 * regardless of where they appear. We can trivially satisfy that
808 * requirement by changing the interpolation type to flat here.
810 producer_var
->data
.centroid
= false;
811 producer_var
->data
.sample
= false;
812 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
815 consumer_var
->data
.centroid
= false;
816 consumer_var
->data
.sample
= false;
817 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
821 if (this->num_matches
== this->matches_capacity
) {
822 this->matches_capacity
*= 2;
823 this->matches
= (match
*)
824 realloc(this->matches
,
825 sizeof(*this->matches
) * this->matches_capacity
);
828 const ir_variable
*const var
= (producer_var
!= NULL
)
829 ? producer_var
: consumer_var
;
831 this->matches
[this->num_matches
].packing_class
832 = this->compute_packing_class(var
);
833 this->matches
[this->num_matches
].packing_order
834 = this->compute_packing_order(var
);
835 if (this->disable_varying_packing
) {
836 unsigned slots
= var
->type
->is_array()
837 ? (var
->type
->length
* var
->type
->fields
.array
->matrix_columns
)
838 : var
->type
->matrix_columns
;
839 this->matches
[this->num_matches
].num_components
= 4 * slots
;
841 this->matches
[this->num_matches
].num_components
842 = var
->type
->component_slots();
844 this->matches
[this->num_matches
].producer_var
= producer_var
;
845 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
848 producer_var
->data
.is_unmatched_generic_inout
= 0;
850 consumer_var
->data
.is_unmatched_generic_inout
= 0;
855 * Choose locations for all of the variable matches that were previously
856 * passed to varying_matches::record().
859 varying_matches::assign_locations()
861 /* Sort varying matches into an order that makes them easy to pack. */
862 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
863 &varying_matches::match_comparator
);
865 unsigned generic_location
= 0;
867 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
868 /* Advance to the next slot if this varying has a different packing
869 * class than the previous one, and we're not already on a slot
873 this->matches
[i
- 1].packing_class
874 != this->matches
[i
].packing_class
) {
875 generic_location
= ALIGN(generic_location
, 4);
878 this->matches
[i
].generic_location
= generic_location
;
880 generic_location
+= this->matches
[i
].num_components
;
883 return (generic_location
+ 3) / 4;
888 * Update the producer and consumer shaders to reflect the locations
889 * assignments that were made by varying_matches::assign_locations().
892 varying_matches::store_locations() const
894 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
895 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
896 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
897 unsigned generic_location
= this->matches
[i
].generic_location
;
898 unsigned slot
= generic_location
/ 4;
899 unsigned offset
= generic_location
% 4;
902 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
903 producer_var
->data
.location_frac
= offset
;
907 assert(consumer_var
->data
.location
== -1);
908 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
909 consumer_var
->data
.location_frac
= offset
;
916 * Compute the "packing class" of the given varying. This is an unsigned
917 * integer with the property that two variables in the same packing class can
918 * be safely backed into the same vec4.
921 varying_matches::compute_packing_class(const ir_variable
*var
)
923 /* Without help from the back-end, there is no way to pack together
924 * variables with different interpolation types, because
925 * lower_packed_varyings must choose exactly one interpolation type for
926 * each packed varying it creates.
928 * However, we can safely pack together floats, ints, and uints, because:
930 * - varyings of base type "int" and "uint" must use the "flat"
931 * interpolation type, which can only occur in GLSL 1.30 and above.
933 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
934 * can store flat floats as ints without losing any information (using
935 * the ir_unop_bitcast_* opcodes).
937 * Therefore, the packing class depends only on the interpolation type.
939 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1);
941 packing_class
+= var
->data
.interpolation
;
942 return packing_class
;
947 * Compute the "packing order" of the given varying. This is a sort key we
948 * use to determine when to attempt to pack the given varying relative to
949 * other varyings in the same packing class.
951 varying_matches::packing_order_enum
952 varying_matches::compute_packing_order(const ir_variable
*var
)
954 const glsl_type
*element_type
= var
->type
;
956 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
957 element_type
= element_type
->fields
.array
;
960 switch (element_type
->component_slots() % 4) {
961 case 1: return PACKING_ORDER_SCALAR
;
962 case 2: return PACKING_ORDER_VEC2
;
963 case 3: return PACKING_ORDER_VEC3
;
964 case 0: return PACKING_ORDER_VEC4
;
966 assert(!"Unexpected value of vector_elements");
967 return PACKING_ORDER_VEC4
;
973 * Comparison function passed to qsort() to sort varyings by packing_class and
974 * then by packing_order.
977 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
979 const match
*x
= (const match
*) x_generic
;
980 const match
*y
= (const match
*) y_generic
;
982 if (x
->packing_class
!= y
->packing_class
)
983 return x
->packing_class
- y
->packing_class
;
984 return x
->packing_order
- y
->packing_order
;
989 * Is the given variable a varying variable to be counted against the
990 * limit in ctx->Const.MaxVarying?
991 * This includes variables such as texcoords, colors and generic
992 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
995 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
997 /* Only fragment shaders will take a varying variable as an input */
998 if (stage
== MESA_SHADER_FRAGMENT
&&
999 var
->data
.mode
== ir_var_shader_in
) {
1000 switch (var
->data
.location
) {
1001 case VARYING_SLOT_POS
:
1002 case VARYING_SLOT_FACE
:
1003 case VARYING_SLOT_PNTC
:
1014 * Visitor class that generates tfeedback_candidate structs describing all
1015 * possible targets of transform feedback.
1017 * tfeedback_candidate structs are stored in the hash table
1018 * tfeedback_candidates, which is passed to the constructor. This hash table
1019 * maps varying names to instances of the tfeedback_candidate struct.
1021 class tfeedback_candidate_generator
: public program_resource_visitor
1024 tfeedback_candidate_generator(void *mem_ctx
,
1025 hash_table
*tfeedback_candidates
)
1027 tfeedback_candidates(tfeedback_candidates
),
1033 void process(ir_variable
*var
)
1035 this->toplevel_var
= var
;
1036 this->varying_floats
= 0;
1037 if (var
->is_interface_instance())
1038 program_resource_visitor::process(var
->get_interface_type(),
1039 var
->get_interface_type()->name
);
1041 program_resource_visitor::process(var
);
1045 virtual void visit_field(const glsl_type
*type
, const char *name
,
1048 assert(!type
->is_record());
1049 assert(!(type
->is_array() && type
->fields
.array
->is_record()));
1050 assert(!type
->is_interface());
1051 assert(!(type
->is_array() && type
->fields
.array
->is_interface()));
1055 tfeedback_candidate
*candidate
1056 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1057 candidate
->toplevel_var
= this->toplevel_var
;
1058 candidate
->type
= type
;
1059 candidate
->offset
= this->varying_floats
;
1060 hash_table_insert(this->tfeedback_candidates
, candidate
,
1061 ralloc_strdup(this->mem_ctx
, name
));
1062 this->varying_floats
+= type
->component_slots();
1066 * Memory context used to allocate hash table keys and values.
1068 void * const mem_ctx
;
1071 * Hash table in which tfeedback_candidate objects should be stored.
1073 hash_table
* const tfeedback_candidates
;
1076 * Pointer to the toplevel variable that is being traversed.
1078 ir_variable
*toplevel_var
;
1081 * Total number of varying floats that have been visited so far. This is
1082 * used to determine the offset to each varying within the toplevel
1085 unsigned varying_floats
;
1092 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1093 hash_table
*consumer_inputs
,
1094 hash_table
*consumer_interface_inputs
,
1095 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
])
1097 memset(consumer_inputs_with_locations
,
1099 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_MAX
);
1101 foreach_list(node
, ir
) {
1102 ir_variable
*const input_var
= ((ir_instruction
*) node
)->as_variable();
1104 if ((input_var
!= NULL
) && (input_var
->data
.mode
== ir_var_shader_in
)) {
1105 if (input_var
->type
->is_interface())
1108 if (input_var
->data
.explicit_location
) {
1109 /* assign_varying_locations only cares about finding the
1110 * ir_variable at the start of a contiguous location block.
1112 * - For !producer, consumer_inputs_with_locations isn't used.
1114 * - For !consumer, consumer_inputs_with_locations is empty.
1116 * For consumer && producer, if you were trying to set some
1117 * ir_variable to the middle of a location block on the other side
1118 * of producer/consumer, cross_validate_outputs_to_inputs() should
1119 * be link-erroring due to either type mismatch or location
1120 * overlaps. If the variables do match up, then they've got a
1121 * matching data.location and you only looked at
1122 * consumer_inputs_with_locations[var->data.location], not any
1123 * following entries for the array/structure.
1125 consumer_inputs_with_locations
[input_var
->data
.location
] =
1127 } else if (input_var
->get_interface_type() != NULL
) {
1128 char *const iface_field_name
=
1129 ralloc_asprintf(mem_ctx
, "%s.%s",
1130 input_var
->get_interface_type()->name
,
1132 hash_table_insert(consumer_interface_inputs
, input_var
,
1135 hash_table_insert(consumer_inputs
, input_var
,
1136 ralloc_strdup(mem_ctx
, input_var
->name
));
1145 * Find a variable from the consumer that "matches" the specified variable
1147 * This function only finds inputs with names that match. There is no
1148 * validation (here) that the types, etc. are compatible.
1151 get_matching_input(void *mem_ctx
,
1152 const ir_variable
*output_var
,
1153 hash_table
*consumer_inputs
,
1154 hash_table
*consumer_interface_inputs
,
1155 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
])
1157 ir_variable
*input_var
;
1159 if (output_var
->data
.explicit_location
) {
1160 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1161 } else if (output_var
->get_interface_type() != NULL
) {
1162 char *const iface_field_name
=
1163 ralloc_asprintf(mem_ctx
, "%s.%s",
1164 output_var
->get_interface_type()->name
,
1167 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1171 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1174 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1181 io_variable_cmp(const void *_a
, const void *_b
)
1183 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1184 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1186 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1187 return b
->data
.location
- a
->data
.location
;
1189 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1192 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1195 return -strcmp(a
->name
, b
->name
);
1199 * Sort the shader IO variables into canonical order
1202 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1204 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1205 unsigned num_variables
= 0;
1207 foreach_list(node
, ir
) {
1208 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1210 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1213 /* If we have already encountered more I/O variables that could
1214 * successfully link, bail.
1216 if (num_variables
== ARRAY_SIZE(var_table
))
1219 var_table
[num_variables
++] = var
;
1222 if (num_variables
== 0)
1225 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1226 * we're going to push the variables on to the IR list as a stack, so we
1227 * want the last variable (in canonical order) to be first in the list.
1229 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1231 /* Remove the variable from it's current location in the IR, and put it at
1234 for (unsigned i
= 0; i
< num_variables
; i
++) {
1235 var_table
[i
]->remove();
1236 ir
->push_head(var_table
[i
]);
1241 * Assign locations for all variables that are produced in one pipeline stage
1242 * (the "producer") and consumed in the next stage (the "consumer").
1244 * Variables produced by the producer may also be consumed by transform
1247 * \param num_tfeedback_decls is the number of declarations indicating
1248 * variables that may be consumed by transform feedback.
1250 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1251 * representing the result of parsing the strings passed to
1252 * glTransformFeedbackVaryings(). assign_location() will be called for
1253 * each of these objects that matches one of the outputs of the
1256 * \param gs_input_vertices: if \c consumer is a geometry shader, this is the
1257 * number of input vertices it accepts. Otherwise zero.
1259 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1260 * be NULL. In this case, varying locations are assigned solely based on the
1261 * requirements of transform feedback.
1264 assign_varying_locations(struct gl_context
*ctx
,
1266 struct gl_shader_program
*prog
,
1267 gl_shader
*producer
, gl_shader
*consumer
,
1268 unsigned num_tfeedback_decls
,
1269 tfeedback_decl
*tfeedback_decls
,
1270 unsigned gs_input_vertices
)
1272 varying_matches
matches(ctx
->Const
.DisableVaryingPacking
,
1273 consumer
&& consumer
->Stage
== MESA_SHADER_FRAGMENT
);
1274 hash_table
*tfeedback_candidates
1275 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1276 hash_table
*consumer_inputs
1277 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1278 hash_table
*consumer_interface_inputs
1279 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1280 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_MAX
] = {
1284 /* Operate in a total of four passes.
1286 * 1. Sort inputs / outputs into a canonical order. This is necessary so
1287 * that inputs / outputs of separable shaders will be assigned
1288 * predictable locations regardless of the order in which declarations
1289 * appeared in the shader source.
1291 * 2. Assign locations for any matching inputs and outputs.
1293 * 3. Mark output variables in the producer that do not have locations as
1294 * not being outputs. This lets the optimizer eliminate them.
1296 * 4. Mark input variables in the consumer that do not have locations as
1297 * not being inputs. This lets the optimizer eliminate them.
1300 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
1303 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
1306 && !linker::populate_consumer_input_sets(mem_ctx
,
1309 consumer_interface_inputs
,
1310 consumer_inputs_with_locations
)) {
1311 assert(!"populate_consumer_input_sets failed");
1312 hash_table_dtor(tfeedback_candidates
);
1313 hash_table_dtor(consumer_inputs
);
1314 hash_table_dtor(consumer_interface_inputs
);
1319 foreach_list(node
, producer
->ir
) {
1320 ir_variable
*const output_var
=
1321 ((ir_instruction
*) node
)->as_variable();
1323 if ((output_var
== NULL
) ||
1324 (output_var
->data
.mode
!= ir_var_shader_out
))
1327 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
1328 g
.process(output_var
);
1330 ir_variable
*const input_var
=
1331 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
1332 consumer_interface_inputs
,
1333 consumer_inputs_with_locations
);
1335 /* If a matching input variable was found, add this ouptut (and the
1336 * input) to the set. If this is a separable program and there is no
1337 * consumer stage, add the output.
1339 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
)) {
1340 matches
.record(output_var
, input_var
);
1344 /* If there's no producer stage, then this must be a separable program.
1345 * For example, we may have a program that has just a fragment shader.
1346 * Later this program will be used with some arbitrary vertex (or
1347 * geometry) shader program. This means that locations must be assigned
1348 * for all the inputs.
1350 foreach_list(node
, consumer
->ir
) {
1351 ir_variable
*const input_var
=
1352 ((ir_instruction
*) node
)->as_variable();
1354 if ((input_var
== NULL
) ||
1355 (input_var
->data
.mode
!= ir_var_shader_in
))
1358 matches
.record(NULL
, input_var
);
1362 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1363 if (!tfeedback_decls
[i
].is_varying())
1366 const tfeedback_candidate
*matched_candidate
1367 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
1369 if (matched_candidate
== NULL
) {
1370 hash_table_dtor(tfeedback_candidates
);
1371 hash_table_dtor(consumer_inputs
);
1372 hash_table_dtor(consumer_interface_inputs
);
1376 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
)
1377 matches
.record(matched_candidate
->toplevel_var
, NULL
);
1380 const unsigned slots_used
= matches
.assign_locations();
1381 matches
.store_locations();
1383 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1384 if (!tfeedback_decls
[i
].is_varying())
1387 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
1388 hash_table_dtor(tfeedback_candidates
);
1389 hash_table_dtor(consumer_inputs
);
1390 hash_table_dtor(consumer_interface_inputs
);
1395 hash_table_dtor(tfeedback_candidates
);
1396 hash_table_dtor(consumer_inputs
);
1397 hash_table_dtor(consumer_interface_inputs
);
1399 if (ctx
->Const
.DisableVaryingPacking
) {
1400 /* Transform feedback code assumes varyings are packed, so if the driver
1401 * has disabled varying packing, make sure it does not support transform
1404 assert(!ctx
->Extensions
.EXT_transform_feedback
);
1407 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
1411 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
1412 gs_input_vertices
, consumer
);
1416 if (consumer
&& producer
) {
1417 foreach_list(node
, consumer
->ir
) {
1418 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1420 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1421 var
->data
.is_unmatched_generic_inout
) {
1422 if (prog
->Version
<= 120) {
1423 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1425 * Only those varying variables used (i.e. read) in
1426 * the fragment shader executable must be written to
1427 * by the vertex shader executable; declaring
1428 * superfluous varying variables in a vertex shader is
1431 * We interpret this text as meaning that the VS must
1432 * write the variable for the FS to read it. See
1433 * "glsl1-varying read but not written" in piglit.
1436 linker_error(prog
, "%s shader varying %s not written "
1438 _mesa_shader_stage_to_string(consumer
->Stage
),
1440 _mesa_shader_stage_to_string(producer
->Stage
));
1443 /* An 'in' variable is only really a shader input if its
1444 * value is written by the previous stage.
1446 var
->data
.mode
= ir_var_auto
;
1455 check_against_output_limit(struct gl_context
*ctx
,
1456 struct gl_shader_program
*prog
,
1457 gl_shader
*producer
)
1459 unsigned output_vectors
= 0;
1461 foreach_list(node
, producer
->ir
) {
1462 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1464 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
1465 var_counts_against_varying_limit(producer
->Stage
, var
)) {
1466 output_vectors
+= var
->type
->count_attribute_slots();
1470 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
1471 unsigned max_output_components
=
1472 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
1474 const unsigned output_components
= output_vectors
* 4;
1475 if (output_components
> max_output_components
) {
1476 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1477 linker_error(prog
, "shader uses too many output vectors "
1480 max_output_components
/ 4);
1482 linker_error(prog
, "shader uses too many output components "
1485 max_output_components
);
1494 check_against_input_limit(struct gl_context
*ctx
,
1495 struct gl_shader_program
*prog
,
1496 gl_shader
*consumer
)
1498 unsigned input_vectors
= 0;
1500 foreach_list(node
, consumer
->ir
) {
1501 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1503 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1504 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
1505 input_vectors
+= var
->type
->count_attribute_slots();
1509 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
1510 unsigned max_input_components
=
1511 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
1513 const unsigned input_components
= input_vectors
* 4;
1514 if (input_components
> max_input_components
) {
1515 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1516 linker_error(prog
, "shader uses too many input vectors "
1519 max_input_components
/ 4);
1521 linker_error(prog
, "shader uses too many input components "
1524 max_input_components
);