2 * Copyright © 2011 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 lower_varyings_to_packed.cpp
27 * This lowering pass generates GLSL code that manually packs varyings into
28 * vec4 slots, for the benefit of back-ends that don't support packed varyings
31 * For example, the following shader:
33 * out mat3x2 foo; // location=4, location_frac=0
34 * out vec3 bar[2]; // location=5, location_frac=2
45 * out vec4 packed4; // location=4, location_frac=0
46 * out vec4 packed5; // location=5, location_frac=0
47 * out vec4 packed6; // location=6, location_frac=0
52 * packed4.xy = foo[0];
53 * packed4.zw = foo[1];
54 * packed5.xy = foo[2];
55 * packed5.zw = bar[0].xy;
56 * packed6.x = bar[0].z;
57 * packed6.yzw = bar[1];
60 * This lowering pass properly handles "double parking" of a varying vector
61 * across two varying slots. For example, in the code above, two of the
62 * components of bar[0] are stored in packed5, and the remaining component is
65 * Note that in theory, the extra instructions may cause some loss of
66 * performance. However, hopefully in most cases the performance loss will
67 * either be absorbed by a later optimization pass, or it will be offset by
68 * memory bandwidth savings (because fewer varyings are used).
70 * This lowering pass also packs flat floats, ints, and uints together, by
71 * using ivec4 as the base type of flat "varyings", and using appropriate
72 * casts to convert floats and uints into ints.
74 * This lowering pass also handles varyings whose type is a struct or an array
75 * of struct. Structs are packed in order and with no gaps, so there may be a
76 * performance penalty due to structure elements being double-parked.
78 * Lowering of geometry shader inputs is slightly more complex, since geometry
79 * inputs are always arrays, so we need to lower arrays to arrays. For
80 * example, the following input:
86 * } arr[3]; // location=4, location_frac=0
88 * Would get lowered like this if it occurred in a fragment shader:
95 * in vec4 packed4; // location=4, location_frac=0
96 * in vec4 packed5; // location=5, location_frac=0
97 * in vec4 packed6; // location=6, location_frac=0
98 * in vec4 packed7; // location=7, location_frac=0
99 * in vec4 packed8; // location=8, location_frac=0
100 * in vec4 packed9; // location=9, location_frac=0
104 * arr[0].f = packed4.x;
105 * arr[0].v = packed4.yzw;
106 * arr[0].a[0] = packed5.xy;
107 * arr[0].a[1] = packed5.zw;
108 * arr[1].f = packed6.x;
109 * arr[1].v = packed6.yzw;
110 * arr[1].a[0] = packed7.xy;
111 * arr[1].a[1] = packed7.zw;
112 * arr[2].f = packed8.x;
113 * arr[2].v = packed8.yzw;
114 * arr[2].a[0] = packed9.xy;
115 * arr[2].a[1] = packed9.zw;
119 * But it would get lowered like this if it occurred in a geometry shader:
126 * in vec4 packed4[3]; // location=4, location_frac=0
127 * in vec4 packed5[3]; // location=5, location_frac=0
131 * arr[0].f = packed4[0].x;
132 * arr[0].v = packed4[0].yzw;
133 * arr[0].a[0] = packed5[0].xy;
134 * arr[0].a[1] = packed5[0].zw;
135 * arr[1].f = packed4[1].x;
136 * arr[1].v = packed4[1].yzw;
137 * arr[1].a[0] = packed5[1].xy;
138 * arr[1].a[1] = packed5[1].zw;
139 * arr[2].f = packed4[2].x;
140 * arr[2].v = packed4[2].yzw;
141 * arr[2].a[0] = packed5[2].xy;
142 * arr[2].a[1] = packed5[2].zw;
147 #include "glsl_symbol_table.h"
149 #include "ir_builder.h"
150 #include "ir_optimization.h"
151 #include "program/prog_instruction.h"
153 using namespace ir_builder
;
158 * Visitor that performs varying packing. For each varying declared in the
159 * shader, this visitor determines whether it needs to be packed. If so, it
160 * demotes it to an ordinary global, creates new packed varyings, and
161 * generates assignments to convert between the original varying and the
164 class lower_packed_varyings_visitor
167 lower_packed_varyings_visitor(void *mem_ctx
, unsigned locations_used
,
168 ir_variable_mode mode
,
169 unsigned gs_input_vertices
,
170 exec_list
*out_instructions
,
171 exec_list
*out_variables
,
172 bool disable_varying_packing
,
175 void run(struct gl_shader
*shader
);
178 void bitwise_assign_pack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
179 void bitwise_assign_unpack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
180 unsigned lower_rvalue(ir_rvalue
*rvalue
, unsigned fine_location
,
181 ir_variable
*unpacked_var
, const char *name
,
182 bool gs_input_toplevel
, unsigned vertex_index
);
183 unsigned lower_arraylike(ir_rvalue
*rvalue
, unsigned array_size
,
184 unsigned fine_location
,
185 ir_variable
*unpacked_var
, const char *name
,
186 bool gs_input_toplevel
, unsigned vertex_index
);
187 ir_dereference
*get_packed_varying_deref(unsigned location
,
188 ir_variable
*unpacked_var
,
190 unsigned vertex_index
);
191 bool needs_lowering(ir_variable
*var
);
194 * Memory context used to allocate new instructions for the shader.
196 void * const mem_ctx
;
199 * Number of generic varying slots which are used by this shader. This is
200 * used to allocate temporary intermediate data structures. If any varying
201 * used by this shader has a location greater than or equal to
202 * VARYING_SLOT_VAR0 + locations_used, an assertion will fire.
204 const unsigned locations_used
;
207 * Array of pointers to the packed varyings that have been created for each
208 * generic varying slot. NULL entries in this array indicate varying slots
209 * for which a packed varying has not been created yet.
211 ir_variable
**packed_varyings
;
214 * Type of varying which is being lowered in this pass (either
215 * ir_var_shader_in or ir_var_shader_out).
217 const ir_variable_mode mode
;
220 * If we are currently lowering geometry shader inputs, the number of input
221 * vertices the geometry shader accepts. Otherwise zero.
223 const unsigned gs_input_vertices
;
226 * Exec list into which the visitor should insert the packing instructions.
227 * Caller provides this list; it should insert the instructions into the
228 * appropriate place in the shader once the visitor has finished running.
230 exec_list
*out_instructions
;
233 * Exec list into which the visitor should insert any new variables.
235 exec_list
*out_variables
;
237 bool disable_varying_packing
;
241 } /* anonymous namespace */
243 lower_packed_varyings_visitor::lower_packed_varyings_visitor(
244 void *mem_ctx
, unsigned locations_used
, ir_variable_mode mode
,
245 unsigned gs_input_vertices
, exec_list
*out_instructions
,
246 exec_list
*out_variables
, bool disable_varying_packing
,
249 locations_used(locations_used
),
250 packed_varyings((ir_variable
**)
251 rzalloc_array_size(mem_ctx
, sizeof(*packed_varyings
),
254 gs_input_vertices(gs_input_vertices
),
255 out_instructions(out_instructions
),
256 out_variables(out_variables
),
257 disable_varying_packing(disable_varying_packing
),
258 xfb_enabled(xfb_enabled
)
263 lower_packed_varyings_visitor::run(struct gl_shader
*shader
)
265 foreach_in_list(ir_instruction
, node
, shader
->ir
) {
266 ir_variable
*var
= node
->as_variable();
270 if (var
->data
.mode
!= this->mode
||
271 var
->data
.location
< VARYING_SLOT_VAR0
||
272 !this->needs_lowering(var
))
275 /* This lowering pass is only capable of packing floats and ints
276 * together when their interpolation mode is "flat". Therefore, to be
277 * safe, caller should ensure that integral varyings always use flat
278 * interpolation, even when this is not required by GLSL.
280 assert(var
->data
.interpolation
== INTERP_QUALIFIER_FLAT
||
281 !var
->type
->contains_integer());
283 /* Clone the variable for program resource list before
284 * it gets modified and lost.
286 if (!shader
->packed_varyings
)
287 shader
->packed_varyings
= new (shader
) exec_list
;
289 shader
->packed_varyings
->push_tail(var
->clone(shader
, NULL
));
291 /* Change the old varying into an ordinary global. */
292 assert(var
->data
.mode
!= ir_var_temporary
);
293 var
->data
.mode
= ir_var_auto
;
295 /* Create a reference to the old varying. */
296 ir_dereference_variable
*deref
297 = new(this->mem_ctx
) ir_dereference_variable(var
);
299 /* Recursively pack or unpack it. */
300 this->lower_rvalue(deref
, var
->data
.location
* 4 + var
->data
.location_frac
, var
,
301 var
->name
, this->gs_input_vertices
!= 0, 0);
305 #define SWIZZLE_ZWZW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W)
308 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
309 * bitcasts if necessary to match up types.
311 * This function is called when packing varyings.
314 lower_packed_varyings_visitor::bitwise_assign_pack(ir_rvalue
*lhs
,
317 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
318 /* Since we only mix types in flat varyings, and we always store flat
319 * varyings as type ivec4, we need only produce conversions from (uint
322 assert(lhs
->type
->base_type
== GLSL_TYPE_INT
);
323 switch (rhs
->type
->base_type
) {
325 rhs
= new(this->mem_ctx
)
326 ir_expression(ir_unop_u2i
, lhs
->type
, rhs
);
328 case GLSL_TYPE_FLOAT
:
329 rhs
= new(this->mem_ctx
)
330 ir_expression(ir_unop_bitcast_f2i
, lhs
->type
, rhs
);
332 case GLSL_TYPE_DOUBLE
:
333 assert(rhs
->type
->vector_elements
<= 2);
334 if (rhs
->type
->vector_elements
== 2) {
335 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "pack", ir_var_temporary
);
337 assert(lhs
->type
->vector_elements
== 4);
338 this->out_variables
->push_tail(t
);
339 this->out_instructions
->push_tail(
340 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_x(rhs
->clone(mem_ctx
, NULL
)))), 0x3));
341 this->out_instructions
->push_tail(
342 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_y(rhs
))), 0xc));
345 rhs
= u2i(expr(ir_unop_unpack_double_2x32
, rhs
));
349 assert(!"Unexpected type conversion while lowering varyings");
353 this->out_instructions
->push_tail(new (this->mem_ctx
) ir_assignment(lhs
, rhs
));
358 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
359 * bitcasts if necessary to match up types.
361 * This function is called when unpacking varyings.
364 lower_packed_varyings_visitor::bitwise_assign_unpack(ir_rvalue
*lhs
,
367 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
368 /* Since we only mix types in flat varyings, and we always store flat
369 * varyings as type ivec4, we need only produce conversions from int to
372 assert(rhs
->type
->base_type
== GLSL_TYPE_INT
);
373 switch (lhs
->type
->base_type
) {
375 rhs
= new(this->mem_ctx
)
376 ir_expression(ir_unop_i2u
, lhs
->type
, rhs
);
378 case GLSL_TYPE_FLOAT
:
379 rhs
= new(this->mem_ctx
)
380 ir_expression(ir_unop_bitcast_i2f
, lhs
->type
, rhs
);
382 case GLSL_TYPE_DOUBLE
:
383 assert(lhs
->type
->vector_elements
<= 2);
384 if (lhs
->type
->vector_elements
== 2) {
385 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "unpack", ir_var_temporary
);
386 assert(rhs
->type
->vector_elements
== 4);
387 this->out_variables
->push_tail(t
);
388 this->out_instructions
->push_tail(
389 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle_xy(rhs
->clone(mem_ctx
, NULL
)))), 0x1));
390 this->out_instructions
->push_tail(
391 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle(rhs
->clone(mem_ctx
, NULL
), SWIZZLE_ZWZW
, 2))), 0x2));
394 rhs
= expr(ir_unop_pack_double_2x32
, i2u(rhs
));
398 assert(!"Unexpected type conversion while lowering varyings");
402 this->out_instructions
->push_tail(new(this->mem_ctx
) ir_assignment(lhs
, rhs
));
407 * Recursively pack or unpack the given varying (or portion of a varying) by
408 * traversing all of its constituent vectors.
410 * \param fine_location is the location where the first constituent vector
411 * should be packed--the word "fine" indicates that this location is expressed
412 * in multiples of a float, rather than multiples of a vec4 as is used
415 * \param gs_input_toplevel should be set to true if we are lowering geometry
416 * shader inputs, and we are currently lowering the whole input variable
417 * (i.e. we are lowering the array whose index selects the vertex).
419 * \param vertex_index: if we are lowering geometry shader inputs, and the
420 * level of the array that we are currently lowering is *not* the top level,
421 * then this indicates which vertex we are currently lowering. Otherwise it
424 * \return the location where the next constituent vector (after this one)
428 lower_packed_varyings_visitor::lower_rvalue(ir_rvalue
*rvalue
,
429 unsigned fine_location
,
430 ir_variable
*unpacked_var
,
432 bool gs_input_toplevel
,
433 unsigned vertex_index
)
435 unsigned dmul
= rvalue
->type
->is_double() ? 2 : 1;
436 /* When gs_input_toplevel is set, we should be looking at a geometry shader
439 assert(!gs_input_toplevel
|| rvalue
->type
->is_array());
441 if (rvalue
->type
->is_record()) {
442 for (unsigned i
= 0; i
< rvalue
->type
->length
; i
++) {
444 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
445 const char *field_name
= rvalue
->type
->fields
.structure
[i
].name
;
446 ir_dereference_record
*dereference_record
= new(this->mem_ctx
)
447 ir_dereference_record(rvalue
, field_name
);
449 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, field_name
);
450 fine_location
= this->lower_rvalue(dereference_record
, fine_location
,
451 unpacked_var
, deref_name
, false,
454 return fine_location
;
455 } else if (rvalue
->type
->is_array()) {
456 /* Arrays are packed/unpacked by considering each array element in
459 return this->lower_arraylike(rvalue
, rvalue
->type
->array_size(),
460 fine_location
, unpacked_var
, name
,
461 gs_input_toplevel
, vertex_index
);
462 } else if (rvalue
->type
->is_matrix()) {
463 /* Matrices are packed/unpacked by considering each column vector in
466 return this->lower_arraylike(rvalue
, rvalue
->type
->matrix_columns
,
467 fine_location
, unpacked_var
, name
,
468 false, vertex_index
);
469 } else if (rvalue
->type
->vector_elements
* dmul
+
470 fine_location
% 4 > 4) {
471 /* This vector is going to be "double parked" across two varying slots,
472 * so handle it as two separate assignments. For doubles, a dvec3/dvec4
473 * can end up being spread over 3 slots. However the second splitting
474 * will happen later, here we just always want to split into 2.
476 unsigned left_components
, right_components
;
477 unsigned left_swizzle_values
[4] = { 0, 0, 0, 0 };
478 unsigned right_swizzle_values
[4] = { 0, 0, 0, 0 };
479 char left_swizzle_name
[4] = { 0, 0, 0, 0 };
480 char right_swizzle_name
[4] = { 0, 0, 0, 0 };
482 left_components
= 4 - fine_location
% 4;
483 if (rvalue
->type
->is_double()) {
484 /* We might actually end up with 0 left components! */
485 left_components
/= 2;
487 right_components
= rvalue
->type
->vector_elements
- left_components
;
489 for (unsigned i
= 0; i
< left_components
; i
++) {
490 left_swizzle_values
[i
] = i
;
491 left_swizzle_name
[i
] = "xyzw"[i
];
493 for (unsigned i
= 0; i
< right_components
; i
++) {
494 right_swizzle_values
[i
] = i
+ left_components
;
495 right_swizzle_name
[i
] = "xyzw"[i
+ left_components
];
497 ir_swizzle
*left_swizzle
= new(this->mem_ctx
)
498 ir_swizzle(rvalue
, left_swizzle_values
, left_components
);
499 ir_swizzle
*right_swizzle
= new(this->mem_ctx
)
500 ir_swizzle(rvalue
->clone(this->mem_ctx
, NULL
), right_swizzle_values
,
503 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, left_swizzle_name
);
505 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, right_swizzle_name
);
507 fine_location
= this->lower_rvalue(left_swizzle
, fine_location
,
508 unpacked_var
, left_name
, false,
511 /* Top up the fine location to the next slot */
513 return this->lower_rvalue(right_swizzle
, fine_location
, unpacked_var
,
514 right_name
, false, vertex_index
);
516 /* No special handling is necessary; pack the rvalue into the
519 unsigned swizzle_values
[4] = { 0, 0, 0, 0 };
520 unsigned components
= rvalue
->type
->vector_elements
* dmul
;
521 unsigned location
= fine_location
/ 4;
522 unsigned location_frac
= fine_location
% 4;
523 for (unsigned i
= 0; i
< components
; ++i
)
524 swizzle_values
[i
] = i
+ location_frac
;
525 ir_dereference
*packed_deref
=
526 this->get_packed_varying_deref(location
, unpacked_var
, name
,
528 ir_swizzle
*swizzle
= new(this->mem_ctx
)
529 ir_swizzle(packed_deref
, swizzle_values
, components
);
530 if (this->mode
== ir_var_shader_out
) {
531 this->bitwise_assign_pack(swizzle
, rvalue
);
533 this->bitwise_assign_unpack(rvalue
, swizzle
);
535 return fine_location
+ components
;
540 * Recursively pack or unpack a varying for which we need to iterate over its
541 * constituent elements, accessing each one using an ir_dereference_array.
542 * This takes care of both arrays and matrices, since ir_dereference_array
543 * treats a matrix like an array of its column vectors.
545 * \param gs_input_toplevel should be set to true if we are lowering geometry
546 * shader inputs, and we are currently lowering the whole input variable
547 * (i.e. we are lowering the array whose index selects the vertex).
549 * \param vertex_index: if we are lowering geometry shader inputs, and the
550 * level of the array that we are currently lowering is *not* the top level,
551 * then this indicates which vertex we are currently lowering. Otherwise it
555 lower_packed_varyings_visitor::lower_arraylike(ir_rvalue
*rvalue
,
557 unsigned fine_location
,
558 ir_variable
*unpacked_var
,
560 bool gs_input_toplevel
,
561 unsigned vertex_index
)
563 for (unsigned i
= 0; i
< array_size
; i
++) {
565 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
566 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(i
);
567 ir_dereference_array
*dereference_array
= new(this->mem_ctx
)
568 ir_dereference_array(rvalue
, constant
);
569 if (gs_input_toplevel
) {
570 /* Geometry shader inputs are a special case. Instead of storing
571 * each element of the array at a different location, all elements
572 * are at the same location, but with a different vertex index.
574 (void) this->lower_rvalue(dereference_array
, fine_location
,
575 unpacked_var
, name
, false, i
);
577 char *subscripted_name
578 = ralloc_asprintf(this->mem_ctx
, "%s[%d]", name
, i
);
580 this->lower_rvalue(dereference_array
, fine_location
,
581 unpacked_var
, subscripted_name
,
582 false, vertex_index
);
585 return fine_location
;
589 * Retrieve the packed varying corresponding to the given varying location.
590 * If no packed varying has been created for the given varying location yet,
591 * create it and add it to the shader before returning it.
593 * The newly created varying inherits its interpolation parameters from \c
594 * unpacked_var. Its base type is ivec4 if we are lowering a flat varying,
597 * \param vertex_index: if we are lowering geometry shader inputs, then this
598 * indicates which vertex we are currently lowering. Otherwise it is ignored.
601 lower_packed_varyings_visitor::get_packed_varying_deref(
602 unsigned location
, ir_variable
*unpacked_var
, const char *name
,
603 unsigned vertex_index
)
605 unsigned slot
= location
- VARYING_SLOT_VAR0
;
606 assert(slot
< locations_used
);
607 if (this->packed_varyings
[slot
] == NULL
) {
608 char *packed_name
= ralloc_asprintf(this->mem_ctx
, "packed:%s", name
);
609 const glsl_type
*packed_type
;
610 if (unpacked_var
->data
.interpolation
== INTERP_QUALIFIER_FLAT
)
611 packed_type
= glsl_type::ivec4_type
;
613 packed_type
= glsl_type::vec4_type
;
614 if (this->gs_input_vertices
!= 0) {
616 glsl_type::get_array_instance(packed_type
,
617 this->gs_input_vertices
);
619 ir_variable
*packed_var
= new(this->mem_ctx
)
620 ir_variable(packed_type
, packed_name
, this->mode
);
621 if (this->gs_input_vertices
!= 0) {
622 /* Prevent update_array_sizes() from messing with the size of the
625 packed_var
->data
.max_array_access
= this->gs_input_vertices
- 1;
627 packed_var
->data
.centroid
= unpacked_var
->data
.centroid
;
628 packed_var
->data
.sample
= unpacked_var
->data
.sample
;
629 packed_var
->data
.patch
= unpacked_var
->data
.patch
;
630 packed_var
->data
.interpolation
= unpacked_var
->data
.interpolation
;
631 packed_var
->data
.location
= location
;
632 packed_var
->data
.precision
= unpacked_var
->data
.precision
;
633 packed_var
->data
.always_active_io
= unpacked_var
->data
.always_active_io
;
634 unpacked_var
->insert_before(packed_var
);
635 this->packed_varyings
[slot
] = packed_var
;
637 /* For geometry shader inputs, only update the packed variable name the
638 * first time we visit each component.
640 if (this->gs_input_vertices
== 0 || vertex_index
== 0) {
641 ralloc_asprintf_append((char **) &this->packed_varyings
[slot
]->name
,
646 ir_dereference
*deref
= new(this->mem_ctx
)
647 ir_dereference_variable(this->packed_varyings
[slot
]);
648 if (this->gs_input_vertices
!= 0) {
649 /* When lowering GS inputs, the packed variable is an array, so we need
650 * to dereference it using vertex_index.
652 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(vertex_index
);
653 deref
= new(this->mem_ctx
) ir_dereference_array(deref
, constant
);
659 lower_packed_varyings_visitor::needs_lowering(ir_variable
*var
)
661 /* Things composed of vec4's and varyings with explicitly assigned
662 * locations don't need lowering. Everything else does.
664 if (var
->data
.explicit_location
)
667 /* Override disable_varying_packing if the var is only used by transform
668 * feedback. Also override it if transform feedback is enabled and the
669 * variable is an array, struct or matrix as the elements of these types
670 * will always has the same interpolation and therefore asre safe to pack.
672 const glsl_type
*type
= var
->type
;
673 if (disable_varying_packing
&& !var
->data
.is_xfb_only
&&
674 !((type
->is_array() || type
->is_record() || type
->is_matrix()) &&
678 type
= type
->without_array();
679 if (type
->vector_elements
== 4 && !type
->is_double())
686 * Visitor that splices varying packing code before every use of EmitVertex()
687 * in a geometry shader.
689 class lower_packed_varyings_gs_splicer
: public ir_hierarchical_visitor
692 explicit lower_packed_varyings_gs_splicer(void *mem_ctx
,
693 const exec_list
*instructions
);
695 virtual ir_visitor_status
visit_leave(ir_emit_vertex
*ev
);
699 * Memory context used to allocate new instructions for the shader.
701 void * const mem_ctx
;
704 * Instructions that should be spliced into place before each EmitVertex()
707 const exec_list
*instructions
;
711 lower_packed_varyings_gs_splicer::lower_packed_varyings_gs_splicer(
712 void *mem_ctx
, const exec_list
*instructions
)
713 : mem_ctx(mem_ctx
), instructions(instructions
)
719 lower_packed_varyings_gs_splicer::visit_leave(ir_emit_vertex
*ev
)
721 foreach_in_list(ir_instruction
, ir
, this->instructions
) {
722 ev
->insert_before(ir
->clone(this->mem_ctx
, NULL
));
724 return visit_continue
;
729 lower_packed_varyings(void *mem_ctx
, unsigned locations_used
,
730 ir_variable_mode mode
, unsigned gs_input_vertices
,
731 gl_shader
*shader
, bool disable_varying_packing
,
734 exec_list
*instructions
= shader
->ir
;
735 ir_function
*main_func
= shader
->symbols
->get_function("main");
736 exec_list void_parameters
;
737 ir_function_signature
*main_func_sig
738 = main_func
->matching_signature(NULL
, &void_parameters
, false);
739 exec_list new_instructions
, new_variables
;
740 lower_packed_varyings_visitor
visitor(mem_ctx
, locations_used
, mode
,
744 disable_varying_packing
,
747 if (mode
== ir_var_shader_out
) {
748 if (shader
->Stage
== MESA_SHADER_GEOMETRY
) {
749 /* For geometry shaders, outputs need to be lowered before each call
752 lower_packed_varyings_gs_splicer
splicer(mem_ctx
, &new_instructions
);
754 /* Add all the variables in first. */
755 main_func_sig
->body
.head
->insert_before(&new_variables
);
757 /* Now update all the EmitVertex instances */
758 splicer
.run(instructions
);
760 /* For other shader types, outputs need to be lowered at the end of
763 main_func_sig
->body
.append_list(&new_variables
);
764 main_func_sig
->body
.append_list(&new_instructions
);
767 /* Shader inputs need to be lowered at the beginning of main() */
768 main_func_sig
->body
.head
->insert_before(&new_instructions
);
769 main_func_sig
->body
.head
->insert_before(&new_variables
);