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
,
54 /* Check that the types match between stages.
56 const glsl_type
*type_to_match
= input
->type
;
57 if (consumer_type
== GL_GEOMETRY_SHADER
) {
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_glsl_shader_target_name(producer_type
),
88 _mesa_glsl_shader_target_name(consumer_type
),
94 /* Check that all of the qualifiers match between stages.
96 if (input
->centroid
!= output
->centroid
) {
98 "%s shader output `%s' %s centroid qualifier, "
99 "but %s shader input %s centroid qualifier\n",
100 _mesa_glsl_shader_target_name(producer_type
),
102 (output
->centroid
) ? "has" : "lacks",
103 _mesa_glsl_shader_target_name(consumer_type
),
104 (input
->centroid
) ? "has" : "lacks");
108 if (input
->invariant
!= output
->invariant
) {
110 "%s shader output `%s' %s invariant qualifier, "
111 "but %s shader input %s invariant qualifier\n",
112 _mesa_glsl_shader_target_name(producer_type
),
114 (output
->invariant
) ? "has" : "lacks",
115 _mesa_glsl_shader_target_name(consumer_type
),
116 (input
->invariant
) ? "has" : "lacks");
120 if (input
->interpolation
!= output
->interpolation
) {
122 "%s shader output `%s' specifies %s "
123 "interpolation qualifier, "
124 "but %s shader input specifies %s "
125 "interpolation qualifier\n",
126 _mesa_glsl_shader_target_name(producer_type
),
128 output
->interpolation_string(),
129 _mesa_glsl_shader_target_name(consumer_type
),
130 input
->interpolation_string());
136 * Validate front and back color outputs against single color input
139 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
140 const ir_variable
*input
,
141 const ir_variable
*front_color
,
142 const ir_variable
*back_color
,
143 GLenum consumer_type
,
144 GLenum producer_type
)
146 if (front_color
!= NULL
&& front_color
->assigned
)
147 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
148 consumer_type
, producer_type
);
150 if (back_color
!= NULL
&& back_color
->assigned
)
151 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
152 consumer_type
, producer_type
);
156 * Validate that outputs from one stage match inputs of another
159 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
160 gl_shader
*producer
, gl_shader
*consumer
)
162 glsl_symbol_table parameters
;
164 /* Find all shader outputs in the "producer" stage.
166 foreach_list(node
, producer
->ir
) {
167 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
169 if ((var
== NULL
) || (var
->mode
!= ir_var_shader_out
))
172 parameters
.add_variable(var
);
176 /* Find all shader inputs in the "consumer" stage. Any variables that have
177 * matching outputs already in the symbol table must have the same type and
180 * Exception: if the consumer is the geometry shader, then the inputs
181 * should be arrays and the type of the array element should match the type
182 * of the corresponding producer output.
184 foreach_list(node
, consumer
->ir
) {
185 ir_variable
*const input
= ((ir_instruction
*) node
)->as_variable();
187 if ((input
== NULL
) || (input
->mode
!= ir_var_shader_in
))
190 if (strcmp(input
->name
, "gl_Color") == 0 && input
->used
) {
191 const ir_variable
*const front_color
=
192 parameters
.get_variable("gl_FrontColor");
194 const ir_variable
*const back_color
=
195 parameters
.get_variable("gl_BackColor");
197 cross_validate_front_and_back_color(prog
, input
,
198 front_color
, back_color
,
199 consumer
->Type
, producer
->Type
);
200 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->used
) {
201 const ir_variable
*const front_color
=
202 parameters
.get_variable("gl_FrontSecondaryColor");
204 const ir_variable
*const back_color
=
205 parameters
.get_variable("gl_BackSecondaryColor");
207 cross_validate_front_and_back_color(prog
, input
,
208 front_color
, back_color
,
209 consumer
->Type
, producer
->Type
);
211 ir_variable
*const output
= parameters
.get_variable(input
->name
);
212 if (output
!= NULL
) {
213 cross_validate_types_and_qualifiers(prog
, input
, output
,
214 consumer
->Type
, producer
->Type
);
222 * Initialize this object based on a string that was passed to
223 * glTransformFeedbackVaryings.
225 * If the input is mal-formed, this call still succeeds, but it sets
226 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
227 * will fail to find any matching variable.
230 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
233 /* We don't have to be pedantic about what is a valid GLSL variable name,
234 * because any variable with an invalid name can't exist in the IR anyway.
238 this->orig_name
= input
;
239 this->is_clip_distance_mesa
= false;
240 this->skip_components
= 0;
241 this->next_buffer_separator
= false;
242 this->matched_candidate
= NULL
;
244 if (ctx
->Extensions
.ARB_transform_feedback3
) {
245 /* Parse gl_NextBuffer. */
246 if (strcmp(input
, "gl_NextBuffer") == 0) {
247 this->next_buffer_separator
= true;
251 /* Parse gl_SkipComponents. */
252 if (strcmp(input
, "gl_SkipComponents1") == 0)
253 this->skip_components
= 1;
254 else if (strcmp(input
, "gl_SkipComponents2") == 0)
255 this->skip_components
= 2;
256 else if (strcmp(input
, "gl_SkipComponents3") == 0)
257 this->skip_components
= 3;
258 else if (strcmp(input
, "gl_SkipComponents4") == 0)
259 this->skip_components
= 4;
261 if (this->skip_components
)
265 /* Parse a declaration. */
266 const char *base_name_end
;
267 long subscript
= parse_program_resource_name(input
, &base_name_end
);
268 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
269 if (subscript
>= 0) {
270 this->array_subscript
= subscript
;
271 this->is_subscripted
= true;
273 this->is_subscripted
= false;
276 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
277 * class must behave specially to account for the fact that gl_ClipDistance
278 * is converted from a float[8] to a vec4[2].
280 if (ctx
->ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerClipDistance
&&
281 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
282 this->is_clip_distance_mesa
= true;
288 * Determine whether two tfeedback_decl objects refer to the same variable and
289 * array index (if applicable).
292 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
294 assert(x
.is_varying() && y
.is_varying());
296 if (strcmp(x
.var_name
, y
.var_name
) != 0)
298 if (x
.is_subscripted
!= y
.is_subscripted
)
300 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
307 * Assign a location for this tfeedback_decl object based on the transform
308 * feedback candidate found by find_candidate.
310 * If an error occurs, the error is reported through linker_error() and false
314 tfeedback_decl::assign_location(struct gl_context
*ctx
,
315 struct gl_shader_program
*prog
)
317 assert(this->is_varying());
319 unsigned fine_location
320 = this->matched_candidate
->toplevel_var
->location
* 4
321 + this->matched_candidate
->toplevel_var
->location_frac
322 + this->matched_candidate
->offset
;
324 if (this->matched_candidate
->type
->is_array()) {
326 const unsigned matrix_cols
=
327 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
328 const unsigned vector_elements
=
329 this->matched_candidate
->type
->fields
.array
->vector_elements
;
330 unsigned actual_array_size
= this->is_clip_distance_mesa
?
331 prog
->LastClipDistanceArraySize
:
332 this->matched_candidate
->type
->array_size();
334 if (this->is_subscripted
) {
335 /* Check array bounds. */
336 if (this->array_subscript
>= actual_array_size
) {
337 linker_error(prog
, "Transform feedback varying %s has index "
338 "%i, but the array size is %u.",
339 this->orig_name
, this->array_subscript
,
343 unsigned array_elem_size
= this->is_clip_distance_mesa
?
344 1 : vector_elements
* matrix_cols
;
345 fine_location
+= array_elem_size
* this->array_subscript
;
348 this->size
= actual_array_size
;
350 this->vector_elements
= vector_elements
;
351 this->matrix_columns
= matrix_cols
;
352 if (this->is_clip_distance_mesa
)
353 this->type
= GL_FLOAT
;
355 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
357 /* Regular variable (scalar, vector, or matrix) */
358 if (this->is_subscripted
) {
359 linker_error(prog
, "Transform feedback varying %s requested, "
360 "but %s is not an array.",
361 this->orig_name
, this->var_name
);
365 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
366 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
367 this->type
= this->matched_candidate
->type
->gl_type
;
369 this->location
= fine_location
/ 4;
370 this->location_frac
= fine_location
% 4;
372 /* From GL_EXT_transform_feedback:
373 * A program will fail to link if:
375 * * the total number of components to capture in any varying
376 * variable in <varyings> is greater than the constant
377 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
378 * buffer mode is SEPARATE_ATTRIBS_EXT;
380 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
381 this->num_components() >
382 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
383 linker_error(prog
, "Transform feedback varying %s exceeds "
384 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
394 tfeedback_decl::get_num_outputs() const
396 if (!this->is_varying()) {
400 return (this->num_components() + this->location_frac
+ 3)/4;
405 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
407 * If an error occurs, the error is reported through linker_error() and false
411 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
412 struct gl_transform_feedback_info
*info
,
413 unsigned buffer
, const unsigned max_outputs
) const
415 assert(!this->next_buffer_separator
);
417 /* Handle gl_SkipComponents. */
418 if (this->skip_components
) {
419 info
->BufferStride
[buffer
] += this->skip_components
;
423 /* From GL_EXT_transform_feedback:
424 * A program will fail to link if:
426 * * the total number of components to capture is greater than
427 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
428 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
430 if (prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
&&
431 info
->BufferStride
[buffer
] + this->num_components() >
432 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
433 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
434 "limit has been exceeded.");
438 unsigned location
= this->location
;
439 unsigned location_frac
= this->location_frac
;
440 unsigned num_components
= this->num_components();
441 while (num_components
> 0) {
442 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
443 assert(info
->NumOutputs
< max_outputs
);
444 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
445 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
446 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
447 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
448 info
->Outputs
[info
->NumOutputs
].DstOffset
= info
->BufferStride
[buffer
];
450 info
->BufferStride
[buffer
] += output_size
;
451 num_components
-= output_size
;
456 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
, this->orig_name
);
457 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
458 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
465 const tfeedback_candidate
*
466 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
467 hash_table
*tfeedback_candidates
)
469 const char *name
= this->is_clip_distance_mesa
470 ? "gl_ClipDistanceMESA" : this->var_name
;
471 this->matched_candidate
= (const tfeedback_candidate
*)
472 hash_table_find(tfeedback_candidates
, name
);
473 if (!this->matched_candidate
) {
474 /* From GL_EXT_transform_feedback:
475 * A program will fail to link if:
477 * * any variable name specified in the <varyings> array is not
478 * declared as an output in the geometry shader (if present) or
479 * the vertex shader (if no geometry shader is present);
481 linker_error(prog
, "Transform feedback varying %s undeclared.",
484 return this->matched_candidate
;
489 * Parse all the transform feedback declarations that were passed to
490 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
492 * If an error occurs, the error is reported through linker_error() and false
496 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
497 const void *mem_ctx
, unsigned num_names
,
498 char **varying_names
, tfeedback_decl
*decls
)
500 for (unsigned i
= 0; i
< num_names
; ++i
) {
501 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
503 if (!decls
[i
].is_varying())
506 /* From GL_EXT_transform_feedback:
507 * A program will fail to link if:
509 * * any two entries in the <varyings> array specify the same varying
512 * We interpret this to mean "any two entries in the <varyings> array
513 * specify the same varying variable and array index", since transform
514 * feedback of arrays would be useless otherwise.
516 for (unsigned j
= 0; j
< i
; ++j
) {
517 if (!decls
[j
].is_varying())
520 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
521 linker_error(prog
, "Transform feedback varying %s specified "
522 "more than once.", varying_names
[i
]);
532 * Store transform feedback location assignments into
533 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
535 * If an error occurs, the error is reported through linker_error() and false
539 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
540 unsigned num_tfeedback_decls
,
541 tfeedback_decl
*tfeedback_decls
)
543 bool separate_attribs_mode
=
544 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
546 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
547 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
549 memset(&prog
->LinkedTransformFeedback
, 0,
550 sizeof(prog
->LinkedTransformFeedback
));
552 prog
->LinkedTransformFeedback
.Varyings
=
554 struct gl_transform_feedback_varying_info
,
555 num_tfeedback_decls
);
557 unsigned num_outputs
= 0;
558 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
)
559 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
561 prog
->LinkedTransformFeedback
.Outputs
=
563 struct gl_transform_feedback_output
,
566 unsigned num_buffers
= 0;
568 if (separate_attribs_mode
) {
569 /* GL_SEPARATE_ATTRIBS */
570 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
571 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
572 num_buffers
, num_outputs
))
579 /* GL_INVERLEAVED_ATTRIBS */
580 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
581 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
586 if (!tfeedback_decls
[i
].store(ctx
, prog
,
587 &prog
->LinkedTransformFeedback
,
588 num_buffers
, num_outputs
))
594 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
596 prog
->LinkedTransformFeedback
.NumBuffers
= num_buffers
;
603 * Data structure recording the relationship between outputs of one shader
604 * stage (the "producer") and inputs of another (the "consumer").
606 class varying_matches
609 varying_matches(bool disable_varying_packing
, bool consumer_is_fs
);
611 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
612 unsigned assign_locations();
613 void store_locations(unsigned producer_base
, unsigned consumer_base
) const;
617 * If true, this driver disables varying packing, so all varyings need to
618 * be aligned on slot boundaries, and take up a number of slots equal to
619 * their number of matrix columns times their array size.
621 const bool disable_varying_packing
;
624 * Enum representing the order in which varyings are packed within a
627 * Currently we pack vec4's first, then vec2's, then scalar values, then
628 * vec3's. This order ensures that the only vectors that are at risk of
629 * having to be "double parked" (split between two adjacent varying slots)
632 enum packing_order_enum
{
635 PACKING_ORDER_SCALAR
,
639 static unsigned compute_packing_class(ir_variable
*var
);
640 static packing_order_enum
compute_packing_order(ir_variable
*var
);
641 static int match_comparator(const void *x_generic
, const void *y_generic
);
644 * Structure recording the relationship between a single producer output
645 * and a single consumer input.
649 * Packing class for this varying, computed by compute_packing_class().
651 unsigned packing_class
;
654 * Packing order for this varying, computed by compute_packing_order().
656 packing_order_enum packing_order
;
657 unsigned num_components
;
660 * The output variable in the producer stage.
662 ir_variable
*producer_var
;
665 * The input variable in the consumer stage.
667 ir_variable
*consumer_var
;
670 * The location which has been assigned for this varying. This is
671 * expressed in multiples of a float, with the first generic varying
672 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
675 unsigned generic_location
;
679 * The number of elements in the \c matches array that are currently in
682 unsigned num_matches
;
685 * The number of elements that were set aside for the \c matches array when
688 unsigned matches_capacity
;
690 const bool consumer_is_fs
;
693 } /* anonymous namespace */
695 varying_matches::varying_matches(bool disable_varying_packing
,
697 : disable_varying_packing(disable_varying_packing
),
698 consumer_is_fs(consumer_is_fs
)
700 /* Note: this initial capacity is rather arbitrarily chosen to be large
701 * enough for many cases without wasting an unreasonable amount of space.
702 * varying_matches::record() will resize the array if there are more than
703 * this number of varyings.
705 this->matches_capacity
= 8;
706 this->matches
= (match
*)
707 malloc(sizeof(*this->matches
) * this->matches_capacity
);
708 this->num_matches
= 0;
712 varying_matches::~varying_matches()
719 * Record the given producer/consumer variable pair in the list of variables
720 * that should later be assigned locations.
722 * It is permissible for \c consumer_var to be NULL (this happens if a
723 * variable is output by the producer and consumed by transform feedback, but
724 * not consumed by the consumer).
726 * If \c producer_var has already been paired up with a consumer_var, or
727 * producer_var is part of fixed pipeline functionality (and hence already has
728 * a location assigned), this function has no effect.
730 * Note: as a side effect this function may change the interpolation type of
731 * \c producer_var, but only when the change couldn't possibly affect
735 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
737 if (!producer_var
->is_unmatched_generic_inout
) {
738 /* Either a location already exists for this variable (since it is part
739 * of fixed functionality), or it has already been recorded as part of a
745 if ((consumer_var
== NULL
&& producer_var
->type
->contains_integer()) ||
747 /* Since this varying is not being consumed by the fragment shader, its
748 * interpolation type varying cannot possibly affect rendering. Also,
749 * this variable is non-flat and is (or contains) an integer.
751 * lower_packed_varyings requires all integer varyings to flat,
752 * regardless of where they appear. We can trivially satisfy that
753 * requirement by changing the interpolation type to flat here.
755 producer_var
->centroid
= false;
756 producer_var
->interpolation
= INTERP_QUALIFIER_FLAT
;
759 consumer_var
->centroid
= false;
760 consumer_var
->interpolation
= INTERP_QUALIFIER_FLAT
;
764 if (this->num_matches
== this->matches_capacity
) {
765 this->matches_capacity
*= 2;
766 this->matches
= (match
*)
767 realloc(this->matches
,
768 sizeof(*this->matches
) * this->matches_capacity
);
770 this->matches
[this->num_matches
].packing_class
771 = this->compute_packing_class(producer_var
);
772 this->matches
[this->num_matches
].packing_order
773 = this->compute_packing_order(producer_var
);
774 if (this->disable_varying_packing
) {
775 unsigned slots
= producer_var
->type
->is_array()
776 ? (producer_var
->type
->length
777 * producer_var
->type
->fields
.array
->matrix_columns
)
778 : producer_var
->type
->matrix_columns
;
779 this->matches
[this->num_matches
].num_components
= 4 * slots
;
781 this->matches
[this->num_matches
].num_components
782 = producer_var
->type
->component_slots();
784 this->matches
[this->num_matches
].producer_var
= producer_var
;
785 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
787 producer_var
->is_unmatched_generic_inout
= 0;
789 consumer_var
->is_unmatched_generic_inout
= 0;
794 * Choose locations for all of the variable matches that were previously
795 * passed to varying_matches::record().
798 varying_matches::assign_locations()
800 /* Sort varying matches into an order that makes them easy to pack. */
801 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
802 &varying_matches::match_comparator
);
804 unsigned generic_location
= 0;
806 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
807 /* Advance to the next slot if this varying has a different packing
808 * class than the previous one, and we're not already on a slot
812 this->matches
[i
- 1].packing_class
813 != this->matches
[i
].packing_class
) {
814 generic_location
= ALIGN(generic_location
, 4);
817 this->matches
[i
].generic_location
= generic_location
;
819 generic_location
+= this->matches
[i
].num_components
;
822 return (generic_location
+ 3) / 4;
827 * Update the producer and consumer shaders to reflect the locations
828 * assignments that were made by varying_matches::assign_locations().
831 varying_matches::store_locations(unsigned producer_base
,
832 unsigned consumer_base
) const
834 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
835 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
836 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
837 unsigned generic_location
= this->matches
[i
].generic_location
;
838 unsigned slot
= generic_location
/ 4;
839 unsigned offset
= generic_location
% 4;
841 producer_var
->location
= producer_base
+ slot
;
842 producer_var
->location_frac
= offset
;
844 assert(consumer_var
->location
== -1);
845 consumer_var
->location
= consumer_base
+ slot
;
846 consumer_var
->location_frac
= offset
;
853 * Compute the "packing class" of the given varying. This is an unsigned
854 * integer with the property that two variables in the same packing class can
855 * be safely backed into the same vec4.
858 varying_matches::compute_packing_class(ir_variable
*var
)
860 /* Without help from the back-end, there is no way to pack together
861 * variables with different interpolation types, because
862 * lower_packed_varyings must choose exactly one interpolation type for
863 * each packed varying it creates.
865 * However, we can safely pack together floats, ints, and uints, because:
867 * - varyings of base type "int" and "uint" must use the "flat"
868 * interpolation type, which can only occur in GLSL 1.30 and above.
870 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
871 * can store flat floats as ints without losing any information (using
872 * the ir_unop_bitcast_* opcodes).
874 * Therefore, the packing class depends only on the interpolation type.
876 unsigned packing_class
= var
->centroid
? 1 : 0;
878 packing_class
+= var
->interpolation
;
879 return packing_class
;
884 * Compute the "packing order" of the given varying. This is a sort key we
885 * use to determine when to attempt to pack the given varying relative to
886 * other varyings in the same packing class.
888 varying_matches::packing_order_enum
889 varying_matches::compute_packing_order(ir_variable
*var
)
891 const glsl_type
*element_type
= var
->type
;
893 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
894 element_type
= element_type
->fields
.array
;
897 switch (element_type
->component_slots() % 4) {
898 case 1: return PACKING_ORDER_SCALAR
;
899 case 2: return PACKING_ORDER_VEC2
;
900 case 3: return PACKING_ORDER_VEC3
;
901 case 0: return PACKING_ORDER_VEC4
;
903 assert(!"Unexpected value of vector_elements");
904 return PACKING_ORDER_VEC4
;
910 * Comparison function passed to qsort() to sort varyings by packing_class and
911 * then by packing_order.
914 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
916 const match
*x
= (const match
*) x_generic
;
917 const match
*y
= (const match
*) y_generic
;
919 if (x
->packing_class
!= y
->packing_class
)
920 return x
->packing_class
- y
->packing_class
;
921 return x
->packing_order
- y
->packing_order
;
926 * Is the given variable a varying variable to be counted against the
927 * limit in ctx->Const.MaxVarying?
928 * This includes variables such as texcoords, colors and generic
929 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
932 is_varying_var(GLenum shaderType
, const ir_variable
*var
)
934 /* Only fragment shaders will take a varying variable as an input */
935 if (shaderType
== GL_FRAGMENT_SHADER
&&
936 var
->mode
== ir_var_shader_in
) {
937 switch (var
->location
) {
938 case VARYING_SLOT_POS
:
939 case VARYING_SLOT_FACE
:
940 case VARYING_SLOT_PNTC
:
951 * Visitor class that generates tfeedback_candidate structs describing all
952 * possible targets of transform feedback.
954 * tfeedback_candidate structs are stored in the hash table
955 * tfeedback_candidates, which is passed to the constructor. This hash table
956 * maps varying names to instances of the tfeedback_candidate struct.
958 class tfeedback_candidate_generator
: public program_resource_visitor
961 tfeedback_candidate_generator(void *mem_ctx
,
962 hash_table
*tfeedback_candidates
)
964 tfeedback_candidates(tfeedback_candidates
),
970 void process(ir_variable
*var
)
972 this->toplevel_var
= var
;
973 this->varying_floats
= 0;
974 if (var
->is_interface_instance())
975 program_resource_visitor::process(var
->get_interface_type(),
976 var
->get_interface_type()->name
);
978 program_resource_visitor::process(var
);
982 virtual void visit_field(const glsl_type
*type
, const char *name
,
985 assert(!type
->is_record());
986 assert(!(type
->is_array() && type
->fields
.array
->is_record()));
987 assert(!type
->is_interface());
988 assert(!(type
->is_array() && type
->fields
.array
->is_interface()));
992 tfeedback_candidate
*candidate
993 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
994 candidate
->toplevel_var
= this->toplevel_var
;
995 candidate
->type
= type
;
996 candidate
->offset
= this->varying_floats
;
997 hash_table_insert(this->tfeedback_candidates
, candidate
,
998 ralloc_strdup(this->mem_ctx
, name
));
999 this->varying_floats
+= type
->component_slots();
1003 * Memory context used to allocate hash table keys and values.
1005 void * const mem_ctx
;
1008 * Hash table in which tfeedback_candidate objects should be stored.
1010 hash_table
* const tfeedback_candidates
;
1013 * Pointer to the toplevel variable that is being traversed.
1015 ir_variable
*toplevel_var
;
1018 * Total number of varying floats that have been visited so far. This is
1019 * used to determine the offset to each varying within the toplevel
1022 unsigned varying_floats
;
1027 * Assign locations for all variables that are produced in one pipeline stage
1028 * (the "producer") and consumed in the next stage (the "consumer").
1030 * Variables produced by the producer may also be consumed by transform
1033 * \param num_tfeedback_decls is the number of declarations indicating
1034 * variables that may be consumed by transform feedback.
1036 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1037 * representing the result of parsing the strings passed to
1038 * glTransformFeedbackVaryings(). assign_location() will be called for
1039 * each of these objects that matches one of the outputs of the
1042 * \param gs_input_vertices: if \c consumer is a geometry shader, this is the
1043 * number of input vertices it accepts. Otherwise zero.
1045 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1046 * be NULL. In this case, varying locations are assigned solely based on the
1047 * requirements of transform feedback.
1050 assign_varying_locations(struct gl_context
*ctx
,
1052 struct gl_shader_program
*prog
,
1053 gl_shader
*producer
, gl_shader
*consumer
,
1054 unsigned num_tfeedback_decls
,
1055 tfeedback_decl
*tfeedback_decls
,
1056 unsigned gs_input_vertices
)
1058 const unsigned producer_base
= VARYING_SLOT_VAR0
;
1059 const unsigned consumer_base
= VARYING_SLOT_VAR0
;
1060 varying_matches
matches(ctx
->Const
.DisableVaryingPacking
,
1061 consumer
&& consumer
->Type
== GL_FRAGMENT_SHADER
);
1062 hash_table
*tfeedback_candidates
1063 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1064 hash_table
*consumer_inputs
1065 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1066 hash_table
*consumer_interface_inputs
1067 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1069 /* Operate in a total of three passes.
1071 * 1. Assign locations for any matching inputs and outputs.
1073 * 2. Mark output variables in the producer that do not have locations as
1074 * not being outputs. This lets the optimizer eliminate them.
1076 * 3. Mark input variables in the consumer that do not have locations as
1077 * not being inputs. This lets the optimizer eliminate them.
1081 foreach_list(node
, consumer
->ir
) {
1082 ir_variable
*const input_var
=
1083 ((ir_instruction
*) node
)->as_variable();
1085 if ((input_var
!= NULL
) && (input_var
->mode
== ir_var_shader_in
)) {
1086 if (input_var
->get_interface_type() != NULL
) {
1087 char *const iface_field_name
=
1088 ralloc_asprintf(mem_ctx
, "%s.%s",
1089 input_var
->get_interface_type()->name
,
1091 hash_table_insert(consumer_interface_inputs
, input_var
,
1094 hash_table_insert(consumer_inputs
, input_var
,
1095 ralloc_strdup(mem_ctx
, input_var
->name
));
1101 foreach_list(node
, producer
->ir
) {
1102 ir_variable
*const output_var
= ((ir_instruction
*) node
)->as_variable();
1104 if ((output_var
== NULL
) || (output_var
->mode
!= ir_var_shader_out
))
1107 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
1108 g
.process(output_var
);
1110 ir_variable
*input_var
;
1111 if (output_var
->get_interface_type() != NULL
) {
1112 char *const iface_field_name
=
1113 ralloc_asprintf(mem_ctx
, "%s.%s",
1114 output_var
->get_interface_type()->name
,
1117 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1121 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1124 if (input_var
&& input_var
->mode
!= ir_var_shader_in
)
1128 matches
.record(output_var
, input_var
);
1132 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1133 if (!tfeedback_decls
[i
].is_varying())
1136 const tfeedback_candidate
*matched_candidate
1137 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
1139 if (matched_candidate
== NULL
) {
1140 hash_table_dtor(tfeedback_candidates
);
1141 hash_table_dtor(consumer_inputs
);
1142 hash_table_dtor(consumer_interface_inputs
);
1146 if (matched_candidate
->toplevel_var
->is_unmatched_generic_inout
)
1147 matches
.record(matched_candidate
->toplevel_var
, NULL
);
1150 const unsigned slots_used
= matches
.assign_locations();
1151 matches
.store_locations(producer_base
, consumer_base
);
1153 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1154 if (!tfeedback_decls
[i
].is_varying())
1157 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
1158 hash_table_dtor(tfeedback_candidates
);
1159 hash_table_dtor(consumer_inputs
);
1160 hash_table_dtor(consumer_interface_inputs
);
1165 hash_table_dtor(tfeedback_candidates
);
1166 hash_table_dtor(consumer_inputs
);
1167 hash_table_dtor(consumer_interface_inputs
);
1169 if (ctx
->Const
.DisableVaryingPacking
) {
1170 /* Transform feedback code assumes varyings are packed, so if the driver
1171 * has disabled varying packing, make sure it does not support transform
1174 assert(!ctx
->Extensions
.EXT_transform_feedback
);
1176 lower_packed_varyings(mem_ctx
, producer_base
, slots_used
,
1177 ir_var_shader_out
, 0, producer
);
1179 lower_packed_varyings(mem_ctx
, consumer_base
, slots_used
,
1180 ir_var_shader_in
, gs_input_vertices
, consumer
);
1185 foreach_list(node
, consumer
->ir
) {
1186 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1188 if (var
&& var
->mode
== ir_var_shader_in
&&
1189 var
->is_unmatched_generic_inout
) {
1190 if (prog
->Version
<= 120) {
1191 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1193 * Only those varying variables used (i.e. read) in
1194 * the fragment shader executable must be written to
1195 * by the vertex shader executable; declaring
1196 * superfluous varying variables in a vertex shader is
1199 * We interpret this text as meaning that the VS must
1200 * write the variable for the FS to read it. See
1201 * "glsl1-varying read but not written" in piglit.
1204 linker_error(prog
, "%s shader varying %s not written "
1206 _mesa_glsl_shader_target_name(consumer
->Type
),
1208 _mesa_glsl_shader_target_name(producer
->Type
));
1211 /* An 'in' variable is only really a shader input if its
1212 * value is written by the previous stage.
1214 var
->mode
= ir_var_auto
;
1223 check_against_output_limit(struct gl_context
*ctx
,
1224 struct gl_shader_program
*prog
,
1225 gl_shader
*producer
)
1227 unsigned output_vectors
= 0;
1229 foreach_list(node
, producer
->ir
) {
1230 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1232 if (var
&& var
->mode
== ir_var_shader_out
&&
1233 is_varying_var(producer
->Type
, var
)) {
1234 output_vectors
+= var
->type
->count_attribute_slots();
1238 unsigned max_output_components
;
1239 switch (producer
->Type
) {
1240 case GL_VERTEX_SHADER
:
1241 max_output_components
= ctx
->Const
.VertexProgram
.MaxOutputComponents
;
1243 case GL_GEOMETRY_SHADER
:
1244 max_output_components
= ctx
->Const
.GeometryProgram
.MaxOutputComponents
;
1246 case GL_FRAGMENT_SHADER
:
1248 assert(!"Should not get here.");
1252 const unsigned output_components
= output_vectors
* 4;
1253 if (output_components
> max_output_components
) {
1254 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1255 linker_error(prog
, "shader uses too many output vectors "
1258 max_output_components
/ 4);
1260 linker_error(prog
, "shader uses too many output components "
1263 max_output_components
);
1272 check_against_input_limit(struct gl_context
*ctx
,
1273 struct gl_shader_program
*prog
,
1274 gl_shader
*consumer
)
1276 unsigned input_vectors
= 0;
1278 foreach_list(node
, consumer
->ir
) {
1279 ir_variable
*const var
= ((ir_instruction
*) node
)->as_variable();
1281 if (var
&& var
->mode
== ir_var_shader_in
&&
1282 is_varying_var(consumer
->Type
, var
)) {
1283 input_vectors
+= var
->type
->count_attribute_slots();
1287 unsigned max_input_components
;
1288 switch (consumer
->Type
) {
1289 case GL_GEOMETRY_SHADER
:
1290 max_input_components
= ctx
->Const
.GeometryProgram
.MaxInputComponents
;
1292 case GL_FRAGMENT_SHADER
:
1293 max_input_components
= ctx
->Const
.FragmentProgram
.MaxInputComponents
;
1295 case GL_VERTEX_SHADER
:
1297 assert(!"Should not get here.");
1301 const unsigned input_components
= input_vectors
* 4;
1302 if (input_components
> max_input_components
) {
1303 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
1304 linker_error(prog
, "shader uses too many input vectors "
1307 max_input_components
/ 4);
1309 linker_error(prog
, "shader uses too many input components "
1312 max_input_components
);