2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file link_varyings.cpp
27 * Linker functions related specifically to linking varyings between shader
32 #include "main/mtypes.h"
33 #include "glsl_symbol_table.h"
34 #include "glsl_parser_extras.h"
35 #include "ir_optimization.h"
37 #include "link_varyings.h"
38 #include "main/macros.h"
39 #include "program/hash_table.h"
44 * Get the varying type stripped of the outermost array if we're processing
45 * a stage whose varyings are arrays indexed by a vertex number (such as
46 * geometry shader inputs).
48 static const glsl_type
*
49 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
51 const glsl_type
*type
= var
->type
;
53 if (!var
->data
.patch
&&
54 ((var
->data
.mode
== ir_var_shader_out
&&
55 stage
== MESA_SHADER_TESS_CTRL
) ||
56 (var
->data
.mode
== ir_var_shader_in
&&
57 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
58 stage
== MESA_SHADER_GEOMETRY
)))) {
59 assert(type
->is_array());
60 type
= type
->fields
.array
;
67 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
68 size_t name_length
, unsigned *count
,
69 const char *ifc_member_name
,
70 const glsl_type
*ifc_member_t
, char ***varying_names
)
72 if (t
->is_interface()) {
73 size_t new_length
= name_length
;
75 assert(ifc_member_name
&& ifc_member_t
);
76 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
78 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
79 NULL
, NULL
, varying_names
);
80 } else if (t
->is_record()) {
81 for (unsigned i
= 0; i
< t
->length
; i
++) {
82 const char *field
= t
->fields
.structure
[i
].name
;
83 size_t new_length
= name_length
;
85 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
87 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
88 new_length
, count
, NULL
, NULL
,
91 } else if (t
->without_array()->is_record() ||
92 t
->without_array()->is_interface() ||
93 (t
->is_array() && t
->fields
.array
->is_array())) {
94 for (unsigned i
= 0; i
< t
->length
; i
++) {
95 size_t new_length
= name_length
;
97 /* Append the subscript to the current variable name */
98 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
100 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
101 count
, ifc_member_name
, ifc_member_t
,
105 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
110 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_shader
*sh
,
111 unsigned *num_tfeedback_decls
,
112 char ***varying_names
)
114 bool has_xfb_qualifiers
= false;
116 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
117 ir_variable
*var
= node
->as_variable();
118 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
121 /* From the ARB_enhanced_layouts spec:
123 * "Any shader making any static use (after preprocessing) of any of
124 * these *xfb_* qualifiers will cause the shader to be in a
125 * transform feedback capturing mode and hence responsible for
126 * describing the transform feedback setup. This mode will capture
127 * any output selected by *xfb_offset*, directly or indirectly, to
128 * a transform feedback buffer."
130 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
131 has_xfb_qualifiers
= true;
134 if (var
->data
.explicit_xfb_offset
) {
135 *num_tfeedback_decls
+= var
->type
->varying_count();
136 has_xfb_qualifiers
= true;
140 if (*num_tfeedback_decls
== 0)
141 return has_xfb_qualifiers
;
144 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
145 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
146 ir_variable
*var
= node
->as_variable();
147 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
150 if (var
->data
.explicit_xfb_offset
) {
152 const glsl_type
*type
, *member_type
;
154 if (var
->data
.from_named_ifc_block
) {
155 type
= var
->get_interface_type();
156 /* Find the member type before it was altered by lowering */
158 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
159 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
163 name
= ralloc_strdup(NULL
, var
->name
);
165 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
166 var
->name
, member_type
, varying_names
);
171 assert(i
== *num_tfeedback_decls
);
172 return has_xfb_qualifiers
;
176 * Validate the types and qualifiers of an output from one stage against the
177 * matching input to another stage.
180 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
181 const ir_variable
*input
,
182 const ir_variable
*output
,
183 gl_shader_stage consumer_stage
,
184 gl_shader_stage producer_stage
)
186 /* Check that the types match between stages.
188 const glsl_type
*type_to_match
= input
->type
;
190 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
191 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
192 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
193 consumer_stage
== MESA_SHADER_GEOMETRY
;
194 if (extra_array_level
) {
195 assert(type_to_match
->is_array());
196 type_to_match
= type_to_match
->fields
.array
;
199 if (type_to_match
!= output
->type
) {
200 /* There is a bit of a special case for gl_TexCoord. This
201 * built-in is unsized by default. Applications that variable
202 * access it must redeclare it with a size. There is some
203 * language in the GLSL spec that implies the fragment shader
204 * and vertex shader do not have to agree on this size. Other
205 * driver behave this way, and one or two applications seem to
208 * Neither declaration needs to be modified here because the array
209 * sizes are fixed later when update_array_sizes is called.
211 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
213 * "Unlike user-defined varying variables, the built-in
214 * varying variables don't have a strict one-to-one
215 * correspondence between the vertex language and the
216 * fragment language."
218 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
220 "%s shader output `%s' declared as type `%s', "
221 "but %s shader input declared as type `%s'\n",
222 _mesa_shader_stage_to_string(producer_stage
),
225 _mesa_shader_stage_to_string(consumer_stage
),
231 /* Check that all of the qualifiers match between stages.
233 if (input
->data
.centroid
!= output
->data
.centroid
) {
235 "%s shader output `%s' %s centroid qualifier, "
236 "but %s shader input %s centroid qualifier\n",
237 _mesa_shader_stage_to_string(producer_stage
),
239 (output
->data
.centroid
) ? "has" : "lacks",
240 _mesa_shader_stage_to_string(consumer_stage
),
241 (input
->data
.centroid
) ? "has" : "lacks");
245 if (input
->data
.sample
!= output
->data
.sample
) {
247 "%s shader output `%s' %s sample qualifier, "
248 "but %s shader input %s sample qualifier\n",
249 _mesa_shader_stage_to_string(producer_stage
),
251 (output
->data
.sample
) ? "has" : "lacks",
252 _mesa_shader_stage_to_string(consumer_stage
),
253 (input
->data
.sample
) ? "has" : "lacks");
257 if (input
->data
.patch
!= output
->data
.patch
) {
259 "%s shader output `%s' %s patch qualifier, "
260 "but %s shader input %s patch qualifier\n",
261 _mesa_shader_stage_to_string(producer_stage
),
263 (output
->data
.patch
) ? "has" : "lacks",
264 _mesa_shader_stage_to_string(consumer_stage
),
265 (input
->data
.patch
) ? "has" : "lacks");
269 if (!prog
->IsES
&& input
->data
.invariant
!= output
->data
.invariant
) {
271 "%s shader output `%s' %s invariant qualifier, "
272 "but %s shader input %s invariant qualifier\n",
273 _mesa_shader_stage_to_string(producer_stage
),
275 (output
->data
.invariant
) ? "has" : "lacks",
276 _mesa_shader_stage_to_string(consumer_stage
),
277 (input
->data
.invariant
) ? "has" : "lacks");
281 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
282 * to match cross stage, they must only match within the same stage.
284 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
286 * "It is a link-time error if, within the same stage, the interpolation
287 * qualifiers of variables of the same name do not match.
290 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
291 prog
->Version
< 440) {
293 "%s shader output `%s' specifies %s "
294 "interpolation qualifier, "
295 "but %s shader input specifies %s "
296 "interpolation qualifier\n",
297 _mesa_shader_stage_to_string(producer_stage
),
299 interpolation_string(output
->data
.interpolation
),
300 _mesa_shader_stage_to_string(consumer_stage
),
301 interpolation_string(input
->data
.interpolation
));
307 * Validate front and back color outputs against single color input
310 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
311 const ir_variable
*input
,
312 const ir_variable
*front_color
,
313 const ir_variable
*back_color
,
314 gl_shader_stage consumer_stage
,
315 gl_shader_stage producer_stage
)
317 if (front_color
!= NULL
&& front_color
->data
.assigned
)
318 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
319 consumer_stage
, producer_stage
);
321 if (back_color
!= NULL
&& back_color
->data
.assigned
)
322 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
323 consumer_stage
, producer_stage
);
327 * Validate that outputs from one stage match inputs of another
330 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
331 gl_shader
*producer
, gl_shader
*consumer
)
333 glsl_symbol_table parameters
;
334 ir_variable
*explicit_locations
[MAX_VARYING
] = { NULL
, };
336 /* Find all shader outputs in the "producer" stage.
338 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
339 ir_variable
*const var
= node
->as_variable();
341 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_shader_out
))
344 if (!var
->data
.explicit_location
345 || var
->data
.location
< VARYING_SLOT_VAR0
)
346 parameters
.add_variable(var
);
348 /* User-defined varyings with explicit locations are handled
349 * differently because they do not need to have matching names.
351 const unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
353 if (explicit_locations
[idx
] != NULL
) {
355 "%s shader has multiple outputs explicitly "
356 "assigned to location %d\n",
357 _mesa_shader_stage_to_string(producer
->Stage
),
362 explicit_locations
[idx
] = var
;
367 /* Find all shader inputs in the "consumer" stage. Any variables that have
368 * matching outputs already in the symbol table must have the same type and
371 * Exception: if the consumer is the geometry shader, then the inputs
372 * should be arrays and the type of the array element should match the type
373 * of the corresponding producer output.
375 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
376 ir_variable
*const input
= node
->as_variable();
378 if ((input
== NULL
) || (input
->data
.mode
!= ir_var_shader_in
))
381 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
382 const ir_variable
*const front_color
=
383 parameters
.get_variable("gl_FrontColor");
385 const ir_variable
*const back_color
=
386 parameters
.get_variable("gl_BackColor");
388 cross_validate_front_and_back_color(prog
, input
,
389 front_color
, back_color
,
390 consumer
->Stage
, producer
->Stage
);
391 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
392 const ir_variable
*const front_color
=
393 parameters
.get_variable("gl_FrontSecondaryColor");
395 const ir_variable
*const back_color
=
396 parameters
.get_variable("gl_BackSecondaryColor");
398 cross_validate_front_and_back_color(prog
, input
,
399 front_color
, back_color
,
400 consumer
->Stage
, producer
->Stage
);
402 /* The rules for connecting inputs and outputs change in the presence
403 * of explicit locations. In this case, we no longer care about the
404 * names of the variables. Instead, we care only about the
405 * explicitly assigned location.
407 ir_variable
*output
= NULL
;
408 if (input
->data
.explicit_location
409 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
410 output
= explicit_locations
[input
->data
.location
- VARYING_SLOT_VAR0
];
412 if (output
== NULL
) {
414 "%s shader input `%s' with explicit location "
415 "has no matching output\n",
416 _mesa_shader_stage_to_string(consumer
->Stage
),
420 output
= parameters
.get_variable(input
->name
);
423 if (output
!= NULL
) {
424 /* Interface blocks have their own validation elsewhere so don't
425 * try validating them here.
427 if (!(input
->get_interface_type() &&
428 output
->get_interface_type()))
429 cross_validate_types_and_qualifiers(prog
, input
, output
,
433 /* Check for input vars with unmatched output vars in prev stage
434 * taking into account that interface blocks could have a matching
435 * output but with different name, so we ignore them.
437 assert(!input
->data
.assigned
);
438 if (input
->data
.used
&& !input
->get_interface_type() &&
439 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
441 "%s shader input `%s' "
442 "has no matching output in the previous stage\n",
443 _mesa_shader_stage_to_string(consumer
->Stage
),
451 * Demote shader inputs and outputs that are not used in other stages, and
452 * remove them via dead code elimination.
455 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
457 enum ir_variable_mode mode
)
459 if (is_separate_shader_object
)
462 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
463 ir_variable
*const var
= node
->as_variable();
465 if ((var
== NULL
) || (var
->data
.mode
!= int(mode
)))
468 /* A shader 'in' or 'out' variable is only really an input or output if
469 * its value is used by other shader stages. This will cause the
470 * variable to have a location assigned.
472 if (var
->data
.is_unmatched_generic_inout
) {
473 assert(var
->data
.mode
!= ir_var_temporary
);
474 var
->data
.mode
= ir_var_auto
;
478 /* Eliminate code that is now dead due to unused inputs/outputs being
481 while (do_dead_code(sh
->ir
, false))
487 * Initialize this object based on a string that was passed to
488 * glTransformFeedbackVaryings.
490 * If the input is mal-formed, this call still succeeds, but it sets
491 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
492 * will fail to find any matching variable.
495 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
498 /* We don't have to be pedantic about what is a valid GLSL variable name,
499 * because any variable with an invalid name can't exist in the IR anyway.
503 this->orig_name
= input
;
504 this->lowered_builtin_array_variable
= none
;
505 this->skip_components
= 0;
506 this->next_buffer_separator
= false;
507 this->matched_candidate
= NULL
;
512 if (ctx
->Extensions
.ARB_transform_feedback3
) {
513 /* Parse gl_NextBuffer. */
514 if (strcmp(input
, "gl_NextBuffer") == 0) {
515 this->next_buffer_separator
= true;
519 /* Parse gl_SkipComponents. */
520 if (strcmp(input
, "gl_SkipComponents1") == 0)
521 this->skip_components
= 1;
522 else if (strcmp(input
, "gl_SkipComponents2") == 0)
523 this->skip_components
= 2;
524 else if (strcmp(input
, "gl_SkipComponents3") == 0)
525 this->skip_components
= 3;
526 else if (strcmp(input
, "gl_SkipComponents4") == 0)
527 this->skip_components
= 4;
529 if (this->skip_components
)
533 /* Parse a declaration. */
534 const char *base_name_end
;
535 long subscript
= parse_program_resource_name(input
, &base_name_end
);
536 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
537 if (this->var_name
== NULL
) {
538 _mesa_error_no_memory(__func__
);
542 if (subscript
>= 0) {
543 this->array_subscript
= subscript
;
544 this->is_subscripted
= true;
546 this->is_subscripted
= false;
549 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
550 * class must behave specially to account for the fact that gl_ClipDistance
551 * is converted from a float[8] to a vec4[2].
553 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerClipDistance
&&
554 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
555 this->lowered_builtin_array_variable
= clip_distance
;
558 if (ctx
->Const
.LowerTessLevel
&&
559 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
560 this->lowered_builtin_array_variable
= tess_level_outer
;
561 if (ctx
->Const
.LowerTessLevel
&&
562 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
563 this->lowered_builtin_array_variable
= tess_level_inner
;
568 * Determine whether two tfeedback_decl objects refer to the same variable and
569 * array index (if applicable).
572 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
574 assert(x
.is_varying() && y
.is_varying());
576 if (strcmp(x
.var_name
, y
.var_name
) != 0)
578 if (x
.is_subscripted
!= y
.is_subscripted
)
580 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
587 * Assign a location and stream ID for this tfeedback_decl object based on the
588 * transform feedback candidate found by find_candidate.
590 * If an error occurs, the error is reported through linker_error() and false
594 tfeedback_decl::assign_location(struct gl_context
*ctx
,
595 struct gl_shader_program
*prog
)
597 assert(this->is_varying());
599 unsigned fine_location
600 = this->matched_candidate
->toplevel_var
->data
.location
* 4
601 + this->matched_candidate
->toplevel_var
->data
.location_frac
602 + this->matched_candidate
->offset
;
603 const unsigned dmul
=
604 this->matched_candidate
->type
->without_array()->is_double() ? 2 : 1;
606 if (this->matched_candidate
->type
->is_array()) {
608 const unsigned matrix_cols
=
609 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
610 const unsigned vector_elements
=
611 this->matched_candidate
->type
->fields
.array
->vector_elements
;
612 unsigned actual_array_size
;
613 switch (this->lowered_builtin_array_variable
) {
615 actual_array_size
= prog
->LastClipDistanceArraySize
;
617 case tess_level_outer
:
618 actual_array_size
= 4;
620 case tess_level_inner
:
621 actual_array_size
= 2;
625 actual_array_size
= this->matched_candidate
->type
->array_size();
629 if (this->is_subscripted
) {
630 /* Check array bounds. */
631 if (this->array_subscript
>= actual_array_size
) {
632 linker_error(prog
, "Transform feedback varying %s has index "
633 "%i, but the array size is %u.",
634 this->orig_name
, this->array_subscript
,
638 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
639 1 : vector_elements
* matrix_cols
* dmul
;
640 fine_location
+= array_elem_size
* this->array_subscript
;
643 this->size
= actual_array_size
;
645 this->vector_elements
= vector_elements
;
646 this->matrix_columns
= matrix_cols
;
647 if (this->lowered_builtin_array_variable
)
648 this->type
= GL_FLOAT
;
650 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
652 /* Regular variable (scalar, vector, or matrix) */
653 if (this->is_subscripted
) {
654 linker_error(prog
, "Transform feedback varying %s requested, "
655 "but %s is not an array.",
656 this->orig_name
, this->var_name
);
660 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
661 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
662 this->type
= this->matched_candidate
->type
->gl_type
;
664 this->location
= fine_location
/ 4;
665 this->location_frac
= fine_location
% 4;
667 /* From GL_EXT_transform_feedback:
668 * A program will fail to link if:
670 * * the total number of components to capture in any varying
671 * variable in <varyings> is greater than the constant
672 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
673 * buffer mode is SEPARATE_ATTRIBS_EXT;
675 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
676 this->num_components() >
677 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
678 linker_error(prog
, "Transform feedback varying %s exceeds "
679 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
684 /* Only transform feedback varyings can be assigned to non-zero streams,
685 * so assign the stream id here.
687 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
689 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
690 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
691 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
692 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
693 array_offset
+ struct_offset
;
700 tfeedback_decl::get_num_outputs() const
702 if (!this->is_varying()) {
705 return (this->num_components() + this->location_frac
+ 3)/4;
710 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
712 * If an error occurs, the error is reported through linker_error() and false
716 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
717 struct gl_transform_feedback_info
*info
,
718 unsigned buffer
, const unsigned max_outputs
,
719 bool has_xfb_qualifiers
) const
721 assert(!this->next_buffer_separator
);
723 /* Handle gl_SkipComponents. */
724 if (this->skip_components
) {
725 info
->BufferStride
[buffer
] += this->skip_components
;
729 /* From GL_EXT_transform_feedback:
730 * A program will fail to link if:
732 * * the total number of components to capture is greater than
733 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
734 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
736 if (prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
&&
737 info
->BufferStride
[buffer
] + this->num_components() >
738 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
739 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
740 "limit has been exceeded.");
744 unsigned location
= this->location
;
745 unsigned location_frac
= this->location_frac
;
746 unsigned num_components
= this->num_components();
747 while (num_components
> 0) {
748 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
749 assert(info
->NumOutputs
< max_outputs
);
750 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
751 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
752 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
753 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
754 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
755 info
->Outputs
[info
->NumOutputs
].DstOffset
= info
->BufferStride
[buffer
];
757 info
->BufferStride
[buffer
] += output_size
;
758 info
->BufferStream
[buffer
] = this->stream_id
;
759 num_components
-= output_size
;
764 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
, this->orig_name
);
765 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
766 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
773 const tfeedback_candidate
*
774 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
775 hash_table
*tfeedback_candidates
)
777 const char *name
= this->var_name
;
778 switch (this->lowered_builtin_array_variable
) {
780 name
= this->var_name
;
783 name
= "gl_ClipDistanceMESA";
785 case tess_level_outer
:
786 name
= "gl_TessLevelOuterMESA";
788 case tess_level_inner
:
789 name
= "gl_TessLevelInnerMESA";
792 this->matched_candidate
= (const tfeedback_candidate
*)
793 hash_table_find(tfeedback_candidates
, name
);
794 if (!this->matched_candidate
) {
795 /* From GL_EXT_transform_feedback:
796 * A program will fail to link if:
798 * * any variable name specified in the <varyings> array is not
799 * declared as an output in the geometry shader (if present) or
800 * the vertex shader (if no geometry shader is present);
802 linker_error(prog
, "Transform feedback varying %s undeclared.",
805 return this->matched_candidate
;
810 * Parse all the transform feedback declarations that were passed to
811 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
813 * If an error occurs, the error is reported through linker_error() and false
817 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
818 const void *mem_ctx
, unsigned num_names
,
819 char **varying_names
, tfeedback_decl
*decls
)
821 for (unsigned i
= 0; i
< num_names
; ++i
) {
822 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
824 if (!decls
[i
].is_varying())
827 /* From GL_EXT_transform_feedback:
828 * A program will fail to link if:
830 * * any two entries in the <varyings> array specify the same varying
833 * We interpret this to mean "any two entries in the <varyings> array
834 * specify the same varying variable and array index", since transform
835 * feedback of arrays would be useless otherwise.
837 for (unsigned j
= 0; j
< i
; ++j
) {
838 if (!decls
[j
].is_varying())
841 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
842 linker_error(prog
, "Transform feedback varying %s specified "
843 "more than once.", varying_names
[i
]);
853 * Store transform feedback location assignments into
854 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
856 * If an error occurs, the error is reported through linker_error() and false
860 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
861 unsigned num_tfeedback_decls
,
862 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
864 bool separate_attribs_mode
=
865 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
867 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
868 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
870 memset(&prog
->LinkedTransformFeedback
, 0,
871 sizeof(prog
->LinkedTransformFeedback
));
873 prog
->LinkedTransformFeedback
.Varyings
=
875 struct gl_transform_feedback_varying_info
,
876 num_tfeedback_decls
);
878 unsigned num_outputs
= 0;
879 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
)
880 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
882 prog
->LinkedTransformFeedback
.Outputs
=
884 struct gl_transform_feedback_output
,
887 unsigned num_buffers
= 0;
889 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
890 /* GL_SEPARATE_ATTRIBS */
891 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
892 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
893 num_buffers
, num_outputs
,
901 /* GL_INVERLEAVED_ATTRIBS */
902 int buffer_stream_id
= -1;
904 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
906 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
907 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
909 buffer_stream_id
= -1;
911 } else if (buffer_stream_id
== -1) {
912 /* First varying writing to this buffer: remember its stream */
913 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
914 } else if (buffer_stream_id
!=
915 (int) tfeedback_decls
[i
].get_stream_id()) {
916 /* Varying writes to the same buffer from a different stream */
918 "Transform feedback can't capture varyings belonging "
919 "to different vertex streams in a single buffer. "
920 "Varying %s writes to buffer from stream %u, other "
921 "varyings in the same buffer write from stream %u.",
922 tfeedback_decls
[i
].name(),
923 tfeedback_decls
[i
].get_stream_id(),
928 if (has_xfb_qualifiers
) {
929 buffer
= tfeedback_decls
[i
].get_buffer();
931 buffer
= num_buffers
;
934 if (!tfeedback_decls
[i
].store(ctx
, prog
,
935 &prog
->LinkedTransformFeedback
,
936 num_buffers
, num_outputs
,
943 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
945 prog
->LinkedTransformFeedback
.NumBuffers
= num_buffers
;
952 * Data structure recording the relationship between outputs of one shader
953 * stage (the "producer") and inputs of another (the "consumer").
955 class varying_matches
958 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
959 gl_shader_stage producer_stage
,
960 gl_shader_stage consumer_stage
);
962 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
963 unsigned assign_locations(struct gl_shader_program
*prog
,
964 uint64_t reserved_slots
, bool separate_shader
);
965 void store_locations() const;
968 bool is_varying_packing_safe(const glsl_type
*type
,
969 const ir_variable
*var
);
972 * If true, this driver disables varying packing, so all varyings need to
973 * be aligned on slot boundaries, and take up a number of slots equal to
974 * their number of matrix columns times their array size.
976 * Packing may also be disabled because our current packing method is not
977 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
978 * guaranteed to match across stages.
980 const bool disable_varying_packing
;
983 * If true, this driver has transform feedback enabled. The transform
984 * feedback code requires at least some packing be done even when varying
985 * packing is disabled, fortunately where transform feedback requires
986 * packing it's safe to override the disabled setting. See
987 * is_varying_packing_safe().
989 const bool xfb_enabled
;
992 * Enum representing the order in which varyings are packed within a
995 * Currently we pack vec4's first, then vec2's, then scalar values, then
996 * vec3's. This order ensures that the only vectors that are at risk of
997 * having to be "double parked" (split between two adjacent varying slots)
1000 enum packing_order_enum
{
1003 PACKING_ORDER_SCALAR
,
1007 static unsigned compute_packing_class(const ir_variable
*var
);
1008 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1009 static int match_comparator(const void *x_generic
, const void *y_generic
);
1010 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1013 * Structure recording the relationship between a single producer output
1014 * and a single consumer input.
1018 * Packing class for this varying, computed by compute_packing_class().
1020 unsigned packing_class
;
1023 * Packing order for this varying, computed by compute_packing_order().
1025 packing_order_enum packing_order
;
1026 unsigned num_components
;
1029 * The output variable in the producer stage.
1031 ir_variable
*producer_var
;
1034 * The input variable in the consumer stage.
1036 ir_variable
*consumer_var
;
1039 * The location which has been assigned for this varying. This is
1040 * expressed in multiples of a float, with the first generic varying
1041 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1044 unsigned generic_location
;
1048 * The number of elements in the \c matches array that are currently in
1051 unsigned num_matches
;
1054 * The number of elements that were set aside for the \c matches array when
1057 unsigned matches_capacity
;
1059 gl_shader_stage producer_stage
;
1060 gl_shader_stage consumer_stage
;
1063 } /* anonymous namespace */
1065 varying_matches::varying_matches(bool disable_varying_packing
,
1067 gl_shader_stage producer_stage
,
1068 gl_shader_stage consumer_stage
)
1069 : disable_varying_packing(disable_varying_packing
),
1070 xfb_enabled(xfb_enabled
),
1071 producer_stage(producer_stage
),
1072 consumer_stage(consumer_stage
)
1074 /* Note: this initial capacity is rather arbitrarily chosen to be large
1075 * enough for many cases without wasting an unreasonable amount of space.
1076 * varying_matches::record() will resize the array if there are more than
1077 * this number of varyings.
1079 this->matches_capacity
= 8;
1080 this->matches
= (match
*)
1081 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1082 this->num_matches
= 0;
1086 varying_matches::~varying_matches()
1088 free(this->matches
);
1093 * Packing is always safe on individual arrays, structure and matices. It is
1094 * also safe if the varying is only used for transform feedback.
1097 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1098 const ir_variable
*var
)
1100 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1101 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1102 producer_stage
== MESA_SHADER_TESS_CTRL
)
1105 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1106 type
->is_matrix() || var
->data
.is_xfb_only
);
1111 * Record the given producer/consumer variable pair in the list of variables
1112 * that should later be assigned locations.
1114 * It is permissible for \c consumer_var to be NULL (this happens if a
1115 * variable is output by the producer and consumed by transform feedback, but
1116 * not consumed by the consumer).
1118 * If \c producer_var has already been paired up with a consumer_var, or
1119 * producer_var is part of fixed pipeline functionality (and hence already has
1120 * a location assigned), this function has no effect.
1122 * Note: as a side effect this function may change the interpolation type of
1123 * \c producer_var, but only when the change couldn't possibly affect
1127 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1129 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1131 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1132 producer_var
->data
.explicit_location
)) ||
1133 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1134 consumer_var
->data
.explicit_location
))) {
1135 /* Either a location already exists for this variable (since it is part
1136 * of fixed functionality), or it has already been recorded as part of a
1142 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1143 (producer_var
->type
->contains_integer() ||
1144 producer_var
->type
->contains_double());
1146 if (needs_flat_qualifier
||
1147 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
)) {
1148 /* Since this varying is not being consumed by the fragment shader, its
1149 * interpolation type varying cannot possibly affect rendering.
1150 * Also, this variable is non-flat and is (or contains) an integer
1152 * If the consumer stage is unknown, don't modify the interpolation
1153 * type as it could affect rendering later with separate shaders.
1155 * lower_packed_varyings requires all integer varyings to flat,
1156 * regardless of where they appear. We can trivially satisfy that
1157 * requirement by changing the interpolation type to flat here.
1160 producer_var
->data
.centroid
= false;
1161 producer_var
->data
.sample
= false;
1162 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1166 consumer_var
->data
.centroid
= false;
1167 consumer_var
->data
.sample
= false;
1168 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1172 if (this->num_matches
== this->matches_capacity
) {
1173 this->matches_capacity
*= 2;
1174 this->matches
= (match
*)
1175 realloc(this->matches
,
1176 sizeof(*this->matches
) * this->matches_capacity
);
1179 const ir_variable
*const var
= (producer_var
!= NULL
)
1180 ? producer_var
: consumer_var
;
1181 const gl_shader_stage stage
= (producer_var
!= NULL
)
1182 ? producer_stage
: consumer_stage
;
1183 const glsl_type
*type
= get_varying_type(var
, stage
);
1185 this->matches
[this->num_matches
].packing_class
1186 = this->compute_packing_class(var
);
1187 this->matches
[this->num_matches
].packing_order
1188 = this->compute_packing_order(var
);
1189 if (this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) {
1190 unsigned slots
= type
->count_attribute_slots(false);
1191 this->matches
[this->num_matches
].num_components
= slots
* 4;
1193 this->matches
[this->num_matches
].num_components
1194 = type
->component_slots();
1196 this->matches
[this->num_matches
].producer_var
= producer_var
;
1197 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1198 this->num_matches
++;
1200 producer_var
->data
.is_unmatched_generic_inout
= 0;
1202 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1207 * Choose locations for all of the variable matches that were previously
1208 * passed to varying_matches::record().
1211 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1212 uint64_t reserved_slots
,
1213 bool separate_shader
)
1215 /* If packing has been disabled then we cannot safely sort the varyings by
1216 * class as it may mean we are using a version of OpenGL where
1217 * interpolation qualifiers are not guaranteed to be matching across
1218 * shaders, sorting in this case could result in mismatching shader
1220 * When packing is disabled the sort orders varyings used by transform
1221 * feedback first, but also depends on *undefined behaviour* of qsort to
1222 * reverse the order of the varyings. See: xfb_comparator().
1224 if (!this->disable_varying_packing
) {
1225 /* Sort varying matches into an order that makes them easy to pack. */
1226 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1227 &varying_matches::match_comparator
);
1229 /* Only sort varyings that are only used by transform feedback. */
1230 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1231 &varying_matches::xfb_comparator
);
1234 unsigned generic_location
= 0;
1235 unsigned generic_patch_location
= MAX_VARYING
*4;
1236 bool previous_var_xfb_only
= false;
1238 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1239 unsigned *location
= &generic_location
;
1241 const ir_variable
*var
;
1242 const glsl_type
*type
;
1243 bool is_vertex_input
= false;
1244 if (matches
[i
].consumer_var
) {
1245 var
= matches
[i
].consumer_var
;
1246 type
= get_varying_type(var
, consumer_stage
);
1247 if (consumer_stage
== MESA_SHADER_VERTEX
)
1248 is_vertex_input
= true;
1250 var
= matches
[i
].producer_var
;
1251 type
= get_varying_type(var
, producer_stage
);
1254 if (var
->data
.patch
)
1255 location
= &generic_patch_location
;
1257 /* Advance to the next slot if this varying has a different packing
1258 * class than the previous one, and we're not already on a slot
1261 * Also advance to the next slot if packing is disabled. This makes sure
1262 * we don't assign varyings the same locations which is possible
1263 * because we still pack individual arrays, records and matrices even
1264 * when packing is disabled. Note we don't advance to the next slot if
1265 * we can pack varyings together that are only used for transform
1268 if ((this->disable_varying_packing
&&
1269 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1270 (i
> 0 && this->matches
[i
- 1].packing_class
1271 != this->matches
[i
].packing_class
)) {
1272 *location
= ALIGN(*location
, 4);
1275 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1277 unsigned num_elements
= type
->count_attribute_slots(is_vertex_input
);
1279 if (this->disable_varying_packing
&&
1280 !is_varying_packing_safe(type
, var
))
1283 slot_end
= type
->without_array()->vector_elements
;
1284 slot_end
+= *location
- 1;
1286 /* FIXME: We could be smarter in the below code and loop back over
1287 * trying to fill any locations that we skipped because we couldn't pack
1288 * the varying between an explicit location. For now just let the user
1289 * hit the linking error if we run out of room and suggest they use
1290 * explicit locations.
1292 for (unsigned j
= 0; j
< num_elements
; j
++) {
1293 while ((slot_end
< MAX_VARYING
* 4u) &&
1294 ((reserved_slots
& (UINT64_C(1) << *location
/ 4u) ||
1295 (reserved_slots
& (UINT64_C(1) << slot_end
/ 4u))))) {
1297 *location
= ALIGN(*location
+ 1, 4);
1298 slot_end
= *location
;
1300 /* reset the counter and try again */
1304 /* Increase the slot to make sure there is enough room for next
1307 if (this->disable_varying_packing
&&
1308 !is_varying_packing_safe(type
, var
))
1311 slot_end
+= type
->without_array()->vector_elements
;
1314 if (!var
->data
.patch
&& *location
>= MAX_VARYING
* 4u) {
1315 linker_error(prog
, "insufficient contiguous locations available for "
1316 "%s it is possible an array or struct could not be "
1317 "packed between varyings with explicit locations. Try "
1318 "using an explicit location for arrays and structs.",
1322 this->matches
[i
].generic_location
= *location
;
1324 *location
+= this->matches
[i
].num_components
;
1327 return (generic_location
+ 3) / 4;
1332 * Update the producer and consumer shaders to reflect the locations
1333 * assignments that were made by varying_matches::assign_locations().
1336 varying_matches::store_locations() const
1338 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1339 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1340 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1341 unsigned generic_location
= this->matches
[i
].generic_location
;
1342 unsigned slot
= generic_location
/ 4;
1343 unsigned offset
= generic_location
% 4;
1346 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1347 producer_var
->data
.location_frac
= offset
;
1351 assert(consumer_var
->data
.location
== -1);
1352 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1353 consumer_var
->data
.location_frac
= offset
;
1360 * Compute the "packing class" of the given varying. This is an unsigned
1361 * integer with the property that two variables in the same packing class can
1362 * be safely backed into the same vec4.
1365 varying_matches::compute_packing_class(const ir_variable
*var
)
1367 /* Without help from the back-end, there is no way to pack together
1368 * variables with different interpolation types, because
1369 * lower_packed_varyings must choose exactly one interpolation type for
1370 * each packed varying it creates.
1372 * However, we can safely pack together floats, ints, and uints, because:
1374 * - varyings of base type "int" and "uint" must use the "flat"
1375 * interpolation type, which can only occur in GLSL 1.30 and above.
1377 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1378 * can store flat floats as ints without losing any information (using
1379 * the ir_unop_bitcast_* opcodes).
1381 * Therefore, the packing class depends only on the interpolation type.
1383 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1384 (var
->data
.patch
<< 2);
1386 packing_class
+= var
->data
.interpolation
;
1387 return packing_class
;
1392 * Compute the "packing order" of the given varying. This is a sort key we
1393 * use to determine when to attempt to pack the given varying relative to
1394 * other varyings in the same packing class.
1396 varying_matches::packing_order_enum
1397 varying_matches::compute_packing_order(const ir_variable
*var
)
1399 const glsl_type
*element_type
= var
->type
;
1401 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
1402 element_type
= element_type
->fields
.array
;
1405 switch (element_type
->component_slots() % 4) {
1406 case 1: return PACKING_ORDER_SCALAR
;
1407 case 2: return PACKING_ORDER_VEC2
;
1408 case 3: return PACKING_ORDER_VEC3
;
1409 case 0: return PACKING_ORDER_VEC4
;
1411 assert(!"Unexpected value of vector_elements");
1412 return PACKING_ORDER_VEC4
;
1418 * Comparison function passed to qsort() to sort varyings by packing_class and
1419 * then by packing_order.
1422 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1424 const match
*x
= (const match
*) x_generic
;
1425 const match
*y
= (const match
*) y_generic
;
1427 if (x
->packing_class
!= y
->packing_class
)
1428 return x
->packing_class
- y
->packing_class
;
1429 return x
->packing_order
- y
->packing_order
;
1434 * Comparison function passed to qsort() to sort varyings used only by
1435 * transform feedback when packing of other varyings is disabled.
1438 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1440 const match
*x
= (const match
*) x_generic
;
1442 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1443 return match_comparator(x_generic
, y_generic
);
1445 /* FIXME: When the comparator returns 0 it means the elements being
1446 * compared are equivalent. However the qsort documentation says:
1448 * "The order of equivalent elements is undefined."
1450 * In practice the sort ends up reversing the order of the varyings which
1451 * means locations are also assigned in this reversed order and happens to
1452 * be what we want. This is also whats happening in
1453 * varying_matches::match_comparator().
1460 * Is the given variable a varying variable to be counted against the
1461 * limit in ctx->Const.MaxVarying?
1462 * This includes variables such as texcoords, colors and generic
1463 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1466 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1468 /* Only fragment shaders will take a varying variable as an input */
1469 if (stage
== MESA_SHADER_FRAGMENT
&&
1470 var
->data
.mode
== ir_var_shader_in
) {
1471 switch (var
->data
.location
) {
1472 case VARYING_SLOT_POS
:
1473 case VARYING_SLOT_FACE
:
1474 case VARYING_SLOT_PNTC
:
1485 * Visitor class that generates tfeedback_candidate structs describing all
1486 * possible targets of transform feedback.
1488 * tfeedback_candidate structs are stored in the hash table
1489 * tfeedback_candidates, which is passed to the constructor. This hash table
1490 * maps varying names to instances of the tfeedback_candidate struct.
1492 class tfeedback_candidate_generator
: public program_resource_visitor
1495 tfeedback_candidate_generator(void *mem_ctx
,
1496 hash_table
*tfeedback_candidates
)
1498 tfeedback_candidates(tfeedback_candidates
),
1504 void process(ir_variable
*var
)
1506 /* All named varying interface blocks should be flattened by now */
1507 assert(!var
->is_interface_instance());
1509 this->toplevel_var
= var
;
1510 this->varying_floats
= 0;
1511 program_resource_visitor::process(var
);
1515 virtual void visit_field(const glsl_type
*type
, const char *name
,
1518 assert(!type
->without_array()->is_record());
1519 assert(!type
->without_array()->is_interface());
1523 tfeedback_candidate
*candidate
1524 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1525 candidate
->toplevel_var
= this->toplevel_var
;
1526 candidate
->type
= type
;
1527 candidate
->offset
= this->varying_floats
;
1528 hash_table_insert(this->tfeedback_candidates
, candidate
,
1529 ralloc_strdup(this->mem_ctx
, name
));
1530 this->varying_floats
+= type
->component_slots();
1534 * Memory context used to allocate hash table keys and values.
1536 void * const mem_ctx
;
1539 * Hash table in which tfeedback_candidate objects should be stored.
1541 hash_table
* const tfeedback_candidates
;
1544 * Pointer to the toplevel variable that is being traversed.
1546 ir_variable
*toplevel_var
;
1549 * Total number of varying floats that have been visited so far. This is
1550 * used to determine the offset to each varying within the toplevel
1553 unsigned varying_floats
;
1560 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1561 hash_table
*consumer_inputs
,
1562 hash_table
*consumer_interface_inputs
,
1563 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1565 memset(consumer_inputs_with_locations
,
1567 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1569 foreach_in_list(ir_instruction
, node
, ir
) {
1570 ir_variable
*const input_var
= node
->as_variable();
1572 if ((input_var
!= NULL
) && (input_var
->data
.mode
== ir_var_shader_in
)) {
1573 /* All interface blocks should have been lowered by this point */
1574 assert(!input_var
->type
->is_interface());
1576 if (input_var
->data
.explicit_location
) {
1577 /* assign_varying_locations only cares about finding the
1578 * ir_variable at the start of a contiguous location block.
1580 * - For !producer, consumer_inputs_with_locations isn't used.
1582 * - For !consumer, consumer_inputs_with_locations is empty.
1584 * For consumer && producer, if you were trying to set some
1585 * ir_variable to the middle of a location block on the other side
1586 * of producer/consumer, cross_validate_outputs_to_inputs() should
1587 * be link-erroring due to either type mismatch or location
1588 * overlaps. If the variables do match up, then they've got a
1589 * matching data.location and you only looked at
1590 * consumer_inputs_with_locations[var->data.location], not any
1591 * following entries for the array/structure.
1593 consumer_inputs_with_locations
[input_var
->data
.location
] =
1595 } else if (input_var
->get_interface_type() != NULL
) {
1596 char *const iface_field_name
=
1597 ralloc_asprintf(mem_ctx
, "%s.%s",
1598 input_var
->get_interface_type()->without_array()->name
,
1600 hash_table_insert(consumer_interface_inputs
, input_var
,
1603 hash_table_insert(consumer_inputs
, input_var
,
1604 ralloc_strdup(mem_ctx
, input_var
->name
));
1611 * Find a variable from the consumer that "matches" the specified variable
1613 * This function only finds inputs with names that match. There is no
1614 * validation (here) that the types, etc. are compatible.
1617 get_matching_input(void *mem_ctx
,
1618 const ir_variable
*output_var
,
1619 hash_table
*consumer_inputs
,
1620 hash_table
*consumer_interface_inputs
,
1621 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1623 ir_variable
*input_var
;
1625 if (output_var
->data
.explicit_location
) {
1626 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1627 } else if (output_var
->get_interface_type() != NULL
) {
1628 char *const iface_field_name
=
1629 ralloc_asprintf(mem_ctx
, "%s.%s",
1630 output_var
->get_interface_type()->without_array()->name
,
1633 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1637 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1640 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1647 io_variable_cmp(const void *_a
, const void *_b
)
1649 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1650 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1652 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1653 return b
->data
.location
- a
->data
.location
;
1655 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1658 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1661 return -strcmp(a
->name
, b
->name
);
1665 * Sort the shader IO variables into canonical order
1668 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1670 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1671 unsigned num_variables
= 0;
1673 foreach_in_list(ir_instruction
, node
, ir
) {
1674 ir_variable
*const var
= node
->as_variable();
1676 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1679 /* If we have already encountered more I/O variables that could
1680 * successfully link, bail.
1682 if (num_variables
== ARRAY_SIZE(var_table
))
1685 var_table
[num_variables
++] = var
;
1688 if (num_variables
== 0)
1691 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1692 * we're going to push the variables on to the IR list as a stack, so we
1693 * want the last variable (in canonical order) to be first in the list.
1695 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1697 /* Remove the variable from it's current location in the IR, and put it at
1700 for (unsigned i
= 0; i
< num_variables
; i
++) {
1701 var_table
[i
]->remove();
1702 ir
->push_head(var_table
[i
]);
1707 * Generate a bitfield map of the explicit locations for shader varyings.
1709 * In theory a 32 bits value will be enough but a 64 bits value is future proof.
1712 reserved_varying_slot(struct gl_shader
*stage
, ir_variable_mode io_mode
)
1714 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
1715 assert(MAX_VARYING
<= 64); /* avoid an overflow of the returned value */
1723 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
1724 ir_variable
*const var
= node
->as_variable();
1726 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
1727 !var
->data
.explicit_location
||
1728 var
->data
.location
< VARYING_SLOT_VAR0
)
1731 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
1733 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
1734 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
1735 for (unsigned i
= 0; i
< num_elements
; i
++) {
1736 if (var_slot
>= 0 && var_slot
< MAX_VARYING
)
1737 slots
|= UINT64_C(1) << var_slot
;
1747 * Assign locations for all variables that are produced in one pipeline stage
1748 * (the "producer") and consumed in the next stage (the "consumer").
1750 * Variables produced by the producer may also be consumed by transform
1753 * \param num_tfeedback_decls is the number of declarations indicating
1754 * variables that may be consumed by transform feedback.
1756 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1757 * representing the result of parsing the strings passed to
1758 * glTransformFeedbackVaryings(). assign_location() will be called for
1759 * each of these objects that matches one of the outputs of the
1762 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1763 * be NULL. In this case, varying locations are assigned solely based on the
1764 * requirements of transform feedback.
1767 assign_varying_locations(struct gl_context
*ctx
,
1769 struct gl_shader_program
*prog
,
1770 gl_shader
*producer
, gl_shader
*consumer
,
1771 unsigned num_tfeedback_decls
,
1772 tfeedback_decl
*tfeedback_decls
)
1774 /* Tessellation shaders treat inputs and outputs as shared memory and can
1775 * access inputs and outputs of other invocations.
1776 * Therefore, they can't be lowered to temps easily (and definitely not
1779 bool unpackable_tess
=
1780 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
1781 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
1782 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
1784 /* Transform feedback code assumes varying arrays are packed, so if the
1785 * driver has disabled varying packing, make sure to at least enable
1786 * packing required by transform feedback.
1789 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
1791 /* Disable varying packing for GL 4.4+ as there is no guarantee
1792 * that interpolation qualifiers will match between shaders in these
1793 * versions. We also disable packing on outerward facing interfaces for
1794 * SSO because in ES we need to retain the unpacked varying information
1795 * for draw time validation. For desktop GL we could allow packing for
1796 * versions < 4.4 but its just safer not to do packing.
1798 * Packing is still enabled on individual arrays, structs, and matrices as
1799 * these are required by the transform feedback code and it is still safe
1800 * to do so. We also enable packing when a varying is only used for
1801 * transform feedback and its not a SSO.
1803 * Varying packing currently only packs together varyings with matching
1804 * interpolation qualifiers as the backends assume all packed components
1805 * are to be processed in the same way. Therefore we cannot do packing in
1806 * these versions of GL without the risk of mismatching interfaces.
1808 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1810 * "The type and presence of interpolation qualifiers of variables with
1811 * the same name declared in all linked shaders for the same cross-stage
1812 * interface must match, otherwise the link command will fail.
1814 * When comparing an output from one stage to an input of a subsequent
1815 * stage, the input and output don't match if their interpolation
1816 * qualifiers (or lack thereof) are not the same."
1818 * This text was also in at least revison 7 of the 4.40 spec but is no
1819 * longer in revision 9 and not in the 4.50 spec.
1821 bool disable_varying_packing
=
1822 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
1823 if ((ctx
->API
== API_OPENGL_CORE
&& ctx
->Version
>= 44) ||
1824 (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
)))
1825 disable_varying_packing
= true;
1827 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
1828 producer
? producer
->Stage
: (gl_shader_stage
)-1,
1829 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
1830 hash_table
*tfeedback_candidates
1831 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1832 hash_table
*consumer_inputs
1833 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1834 hash_table
*consumer_interface_inputs
1835 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
1836 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
1840 unsigned consumer_vertices
= 0;
1841 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
1842 consumer_vertices
= prog
->Geom
.VerticesIn
;
1844 /* Operate in a total of four passes.
1846 * 1. Sort inputs / outputs into a canonical order. This is necessary so
1847 * that inputs / outputs of separable shaders will be assigned
1848 * predictable locations regardless of the order in which declarations
1849 * appeared in the shader source.
1851 * 2. Assign locations for any matching inputs and outputs.
1853 * 3. Mark output variables in the producer that do not have locations as
1854 * not being outputs. This lets the optimizer eliminate them.
1856 * 4. Mark input variables in the consumer that do not have locations as
1857 * not being inputs. This lets the optimizer eliminate them.
1860 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
1863 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
1866 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
1868 consumer_interface_inputs
,
1869 consumer_inputs_with_locations
);
1872 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
1873 ir_variable
*const output_var
= node
->as_variable();
1875 if ((output_var
== NULL
) ||
1876 (output_var
->data
.mode
!= ir_var_shader_out
))
1879 /* Only geometry shaders can use non-zero streams */
1880 assert(output_var
->data
.stream
== 0 ||
1881 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
1882 producer
->Stage
== MESA_SHADER_GEOMETRY
));
1884 if (num_tfeedback_decls
> 0) {
1885 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
1886 g
.process(output_var
);
1889 ir_variable
*const input_var
=
1890 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
1891 consumer_interface_inputs
,
1892 consumer_inputs_with_locations
);
1894 /* If a matching input variable was found, add this ouptut (and the
1895 * input) to the set. If this is a separable program and there is no
1896 * consumer stage, add the output.
1898 * Always add TCS outputs. They are shared by all invocations
1899 * within a patch and can be used as shared memory.
1901 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
1902 producer
->Type
== GL_TESS_CONTROL_SHADER
) {
1903 matches
.record(output_var
, input_var
);
1906 /* Only stream 0 outputs can be consumed in the next stage */
1907 if (input_var
&& output_var
->data
.stream
!= 0) {
1908 linker_error(prog
, "output %s is assigned to stream=%d but "
1909 "is linked to an input, which requires stream=0",
1910 output_var
->name
, output_var
->data
.stream
);
1915 /* If there's no producer stage, then this must be a separable program.
1916 * For example, we may have a program that has just a fragment shader.
1917 * Later this program will be used with some arbitrary vertex (or
1918 * geometry) shader program. This means that locations must be assigned
1919 * for all the inputs.
1921 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
1922 ir_variable
*const input_var
= node
->as_variable();
1924 if ((input_var
== NULL
) ||
1925 (input_var
->data
.mode
!= ir_var_shader_in
))
1928 matches
.record(NULL
, input_var
);
1932 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1933 if (!tfeedback_decls
[i
].is_varying())
1936 const tfeedback_candidate
*matched_candidate
1937 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
1939 if (matched_candidate
== NULL
) {
1940 hash_table_dtor(tfeedback_candidates
);
1941 hash_table_dtor(consumer_inputs
);
1942 hash_table_dtor(consumer_interface_inputs
);
1946 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
1947 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
1948 matches
.record(matched_candidate
->toplevel_var
, NULL
);
1952 const uint64_t reserved_slots
=
1953 reserved_varying_slot(producer
, ir_var_shader_out
) |
1954 reserved_varying_slot(consumer
, ir_var_shader_in
);
1956 const unsigned slots_used
= matches
.assign_locations(prog
, reserved_slots
,
1957 prog
->SeparateShader
);
1958 matches
.store_locations();
1960 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1961 if (!tfeedback_decls
[i
].is_varying())
1964 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
1965 hash_table_dtor(tfeedback_candidates
);
1966 hash_table_dtor(consumer_inputs
);
1967 hash_table_dtor(consumer_interface_inputs
);
1972 hash_table_dtor(tfeedback_candidates
);
1973 hash_table_dtor(consumer_inputs
);
1974 hash_table_dtor(consumer_interface_inputs
);
1976 if (consumer
&& producer
) {
1977 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
1978 ir_variable
*const var
= node
->as_variable();
1980 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
1981 var
->data
.is_unmatched_generic_inout
) {
1982 if (!prog
->IsES
&& prog
->Version
<= 120) {
1983 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1985 * Only those varying variables used (i.e. read) in
1986 * the fragment shader executable must be written to
1987 * by the vertex shader executable; declaring
1988 * superfluous varying variables in a vertex shader is
1991 * We interpret this text as meaning that the VS must
1992 * write the variable for the FS to read it. See
1993 * "glsl1-varying read but not written" in piglit.
1995 linker_error(prog
, "%s shader varying %s not written "
1997 _mesa_shader_stage_to_string(consumer
->Stage
),
1999 _mesa_shader_stage_to_string(producer
->Stage
));
2001 linker_warning(prog
, "%s shader varying %s not written "
2003 _mesa_shader_stage_to_string(consumer
->Stage
),
2005 _mesa_shader_stage_to_string(producer
->Stage
));
2010 /* Now that validation is done its safe to remove unused varyings. As
2011 * we have both a producer and consumer its safe to remove unused
2012 * varyings even if the program is a SSO because the stages are being
2013 * linked together i.e. we have a multi-stage SSO.
2015 remove_unused_shader_inputs_and_outputs(false, producer
,
2017 remove_unused_shader_inputs_and_outputs(false, consumer
,
2022 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
2023 0, producer
, disable_varying_packing
,
2028 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
2029 consumer_vertices
, consumer
,
2030 disable_varying_packing
, xfb_enabled
);
2037 check_against_output_limit(struct gl_context
*ctx
,
2038 struct gl_shader_program
*prog
,
2039 gl_shader
*producer
)
2041 unsigned output_vectors
= 0;
2043 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2044 ir_variable
*const var
= node
->as_variable();
2046 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
2047 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2048 /* outputs for fragment shader can't be doubles */
2049 output_vectors
+= var
->type
->count_attribute_slots(false);
2053 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2054 unsigned max_output_components
=
2055 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2057 const unsigned output_components
= output_vectors
* 4;
2058 if (output_components
> max_output_components
) {
2059 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2060 linker_error(prog
, "%s shader uses too many output vectors "
2062 _mesa_shader_stage_to_string(producer
->Stage
),
2064 max_output_components
/ 4);
2066 linker_error(prog
, "%s shader uses too many output components "
2068 _mesa_shader_stage_to_string(producer
->Stage
),
2070 max_output_components
);
2079 check_against_input_limit(struct gl_context
*ctx
,
2080 struct gl_shader_program
*prog
,
2081 gl_shader
*consumer
)
2083 unsigned input_vectors
= 0;
2085 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2086 ir_variable
*const var
= node
->as_variable();
2088 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2089 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2090 /* vertex inputs aren't varying counted */
2091 input_vectors
+= var
->type
->count_attribute_slots(false);
2095 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2096 unsigned max_input_components
=
2097 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2099 const unsigned input_components
= input_vectors
* 4;
2100 if (input_components
> max_input_components
) {
2101 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2102 linker_error(prog
, "%s shader uses too many input vectors "
2104 _mesa_shader_stage_to_string(consumer
->Stage
),
2106 max_input_components
/ 4);
2108 linker_error(prog
, "%s shader uses too many input components "
2110 _mesa_shader_stage_to_string(consumer
->Stage
),
2112 max_input_components
);