2 * Copyright © 2019 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 DEALINGS
25 #include "nir_builder.h"
26 #include "nir_deref.h"
28 /** @file nir_lower_io_to_vector.c
30 * Merges compatible input/output variables residing in different components
31 * of the same location. It's expected that further passes such as
32 * nir_lower_io_to_temporaries will combine loads and stores of the merged
33 * variables, producing vector nir_load_input/nir_store_output instructions
34 * when all is said and done.
37 /* FRAG_RESULT_MAX+1 instead of just FRAG_RESULT_MAX because of how this pass
38 * handles dual source blending */
39 #define MAX_SLOTS MAX2(VARYING_SLOT_TESS_MAX, FRAG_RESULT_MAX+1)
42 get_slot(const nir_variable
*var
)
44 /* This handling of dual-source blending might not be correct when more than
45 * one render target is supported, but it seems no driver supports more than
47 return var
->data
.location
+ var
->data
.index
;
50 static const struct glsl_type
*
51 get_per_vertex_type(const nir_shader
*shader
, const nir_variable
*var
,
52 unsigned *num_vertices
)
54 if (nir_is_per_vertex_io(var
, shader
->info
.stage
)) {
55 assert(glsl_type_is_array(var
->type
));
57 *num_vertices
= glsl_get_length(var
->type
);
58 return glsl_get_array_element(var
->type
);
66 static const struct glsl_type
*
67 resize_array_vec_type(const struct glsl_type
*type
, unsigned num_components
)
69 if (glsl_type_is_array(type
)) {
70 const struct glsl_type
*arr_elem
=
71 resize_array_vec_type(glsl_get_array_element(type
), num_components
);
72 return glsl_array_type(arr_elem
, glsl_get_length(type
), 0);
74 assert(glsl_type_is_vector_or_scalar(type
));
75 return glsl_vector_type(glsl_get_base_type(type
), num_components
);
80 variables_can_merge(const nir_shader
*shader
,
81 const nir_variable
*a
, const nir_variable
*b
,
82 bool same_array_structure
)
84 if (a
->data
.compact
|| b
->data
.compact
)
87 const struct glsl_type
*a_type_tail
= a
->type
;
88 const struct glsl_type
*b_type_tail
= b
->type
;
90 if (nir_is_per_vertex_io(a
, shader
->info
.stage
) !=
91 nir_is_per_vertex_io(b
, shader
->info
.stage
))
94 /* They must have the same array structure */
95 if (same_array_structure
) {
96 while (glsl_type_is_array(a_type_tail
)) {
97 if (!glsl_type_is_array(b_type_tail
))
100 if (glsl_get_length(a_type_tail
) != glsl_get_length(b_type_tail
))
103 a_type_tail
= glsl_get_array_element(a_type_tail
);
104 b_type_tail
= glsl_get_array_element(b_type_tail
);
106 if (glsl_type_is_array(b_type_tail
))
109 a_type_tail
= glsl_without_array(a_type_tail
);
110 b_type_tail
= glsl_without_array(b_type_tail
);
113 if (!glsl_type_is_vector_or_scalar(a_type_tail
) ||
114 !glsl_type_is_vector_or_scalar(b_type_tail
))
117 if (glsl_get_base_type(a_type_tail
) != glsl_get_base_type(b_type_tail
))
120 /* TODO: add 64/16bit support ? */
121 if (glsl_get_bit_size(a_type_tail
) != 32)
124 assert(a
->data
.mode
== b
->data
.mode
);
125 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
126 a
->data
.mode
== nir_var_shader_in
&&
127 a
->data
.interpolation
!= b
->data
.interpolation
)
130 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
131 a
->data
.mode
== nir_var_shader_out
&&
132 a
->data
.index
!= b
->data
.index
)
138 static const struct glsl_type
*
139 get_flat_type(const nir_shader
*shader
, nir_variable
*old_vars
[MAX_SLOTS
][4],
140 unsigned *loc
, nir_variable
**first_var
, unsigned *num_vertices
)
144 unsigned num_vars
= 0;
145 enum glsl_base_type base
;
150 assert(*loc
< MAX_SLOTS
);
151 for (unsigned frac
= 0; frac
< 4; frac
++) {
152 nir_variable
*var
= old_vars
[*loc
][frac
];
156 !variables_can_merge(shader
, var
, *first_var
, false)) ||
163 if (!glsl_type_is_vector_or_scalar(glsl_without_array(var
->type
))) {
168 base
= glsl_get_base_type(
169 glsl_without_array(get_per_vertex_type(shader
, var
, NULL
)));
172 bool vs_in
= shader
->info
.stage
== MESA_SHADER_VERTEX
&&
173 var
->data
.mode
== nir_var_shader_in
;
174 unsigned var_slots
= glsl_count_attribute_slots(
175 get_per_vertex_type(shader
, var
, num_vertices
), vs_in
);
176 todo
= MAX2(todo
, var_slots
);
187 return glsl_array_type(glsl_vector_type(base
, 4), slots
, 0);
191 create_new_io_vars(nir_shader
*shader
, struct exec_list
*io_list
,
192 nir_variable
*new_vars
[MAX_SLOTS
][4],
193 bool flat_vars
[MAX_SLOTS
])
195 if (exec_list_is_empty(io_list
))
198 nir_variable
*old_vars
[MAX_SLOTS
][4] = {{0}};
200 nir_foreach_variable(var
, io_list
) {
201 unsigned frac
= var
->data
.location_frac
;
202 old_vars
[get_slot(var
)][frac
] = var
;
205 bool merged_any_vars
= false;
207 for (unsigned loc
= 0; loc
< MAX_SLOTS
; loc
++) {
210 nir_variable
*first_var
= old_vars
[loc
][frac
];
217 bool found_merge
= false;
220 nir_variable
*var
= old_vars
[loc
][frac
];
224 if (var
!= first_var
) {
225 if (!variables_can_merge(shader
, first_var
, var
, true))
231 const unsigned num_components
=
232 glsl_get_components(glsl_without_array(var
->type
));
233 if (!num_components
) {
236 break; /* The type was a struct. */
239 /* We had better not have any overlapping vars */
240 for (unsigned i
= 1; i
< num_components
; i
++)
241 assert(old_vars
[loc
][frac
+ i
] == NULL
);
243 frac
+= num_components
;
249 merged_any_vars
= true;
251 nir_variable
*var
= nir_variable_clone(old_vars
[loc
][first
], shader
);
252 var
->data
.location_frac
= first
;
253 var
->type
= resize_array_vec_type(var
->type
, frac
- first
);
255 nir_shader_add_variable(shader
, var
);
256 for (unsigned i
= first
; i
< frac
; i
++) {
257 new_vars
[loc
][i
] = var
;
258 old_vars
[loc
][i
] = NULL
;
261 old_vars
[loc
][first
] = var
;
265 /* "flat" mode: tries to ensure there is at most one variable per slot by
266 * merging variables into vec4s
268 for (unsigned loc
= 0; loc
< MAX_SLOTS
;) {
269 nir_variable
*first_var
;
270 unsigned num_vertices
;
271 unsigned new_loc
= loc
;
272 const struct glsl_type
*flat_type
=
273 get_flat_type(shader
, old_vars
, &new_loc
, &first_var
, &num_vertices
);
275 merged_any_vars
= true;
277 nir_variable
*var
= nir_variable_clone(first_var
, shader
);
278 var
->data
.location_frac
= 0;
280 var
->type
= glsl_array_type(flat_type
, num_vertices
, 0);
282 var
->type
= flat_type
;
284 nir_shader_add_variable(shader
, var
);
285 for (unsigned i
= 0; i
< glsl_get_length(flat_type
); i
++) {
286 for (unsigned j
= 0; j
< 4; j
++)
287 new_vars
[loc
+ i
][j
] = var
;
288 flat_vars
[loc
+ i
] = true;
294 return merged_any_vars
;
297 static nir_deref_instr
*
298 build_array_deref_of_new_var(nir_builder
*b
, nir_variable
*new_var
,
299 nir_deref_instr
*leader
)
301 if (leader
->deref_type
== nir_deref_type_var
)
302 return nir_build_deref_var(b
, new_var
);
304 nir_deref_instr
*parent
=
305 build_array_deref_of_new_var(b
, new_var
, nir_deref_instr_parent(leader
));
307 return nir_build_deref_follower(b
, parent
, leader
);
311 build_array_index(nir_builder
*b
, nir_deref_instr
*deref
, nir_ssa_def
*base
,
314 switch (deref
->deref_type
) {
315 case nir_deref_type_var
:
317 case nir_deref_type_array
: {
318 nir_ssa_def
*index
= nir_i2i(b
, deref
->arr
.index
.ssa
,
319 deref
->dest
.ssa
.bit_size
);
321 b
, build_array_index(b
, nir_deref_instr_parent(deref
), base
, vs_in
),
322 nir_imul_imm(b
, index
, glsl_count_attribute_slots(deref
->type
, vs_in
)));
325 unreachable("Invalid deref instruction type");
329 static nir_deref_instr
*
330 build_array_deref_of_new_var_flat(nir_shader
*shader
,
331 nir_builder
*b
, nir_variable
*new_var
,
332 nir_deref_instr
*leader
, unsigned base
)
334 nir_deref_instr
*deref
= nir_build_deref_var(b
, new_var
);
336 if (nir_is_per_vertex_io(new_var
, shader
->info
.stage
)) {
337 assert(leader
->deref_type
== nir_deref_type_array
);
338 nir_ssa_def
*index
= leader
->arr
.index
.ssa
;
339 leader
= nir_deref_instr_parent(leader
);
340 deref
= nir_build_deref_array(b
, deref
, index
);
343 bool vs_in
= shader
->info
.stage
== MESA_SHADER_VERTEX
&&
344 new_var
->data
.mode
== nir_var_shader_in
;
345 return nir_build_deref_array(
346 b
, deref
, build_array_index(b
, leader
, nir_imm_int(b
, base
), vs_in
));
350 nir_lower_io_to_vector_impl(nir_function_impl
*impl
, nir_variable_mode modes
)
352 assert(!(modes
& ~(nir_var_shader_in
| nir_var_shader_out
)));
355 nir_builder_init(&b
, impl
);
357 nir_metadata_require(impl
, nir_metadata_dominance
);
359 nir_shader
*shader
= impl
->function
->shader
;
360 nir_variable
*new_inputs
[MAX_SLOTS
][4] = {{0}};
361 nir_variable
*new_outputs
[MAX_SLOTS
][4] = {{0}};
362 bool flat_inputs
[MAX_SLOTS
] = {0};
363 bool flat_outputs
[MAX_SLOTS
] = {0};
365 if (modes
& nir_var_shader_in
) {
366 /* Vertex shaders support overlapping inputs. We don't do those */
367 assert(b
.shader
->info
.stage
!= MESA_SHADER_VERTEX
);
369 /* If we don't actually merge any variables, remove that bit from modes
370 * so we don't bother doing extra non-work.
372 if (!create_new_io_vars(shader
, &shader
->inputs
,
373 new_inputs
, flat_inputs
))
374 modes
&= ~nir_var_shader_in
;
377 if (modes
& nir_var_shader_out
) {
378 /* If we don't actually merge any variables, remove that bit from modes
379 * so we don't bother doing extra non-work.
381 if (!create_new_io_vars(shader
, &shader
->outputs
,
382 new_outputs
, flat_outputs
))
383 modes
&= ~nir_var_shader_out
;
389 bool progress
= false;
391 /* Actually lower all the IO load/store intrinsics. Load instructions are
392 * lowered to a vector load and an ALU instruction to grab the channels we
393 * want. Outputs are lowered to a write-masked store of the vector output.
394 * For non-TCS outputs, we then run nir_lower_io_to_temporaries at the end
395 * to clean up the partial writes.
397 nir_foreach_block(block
, impl
) {
398 nir_foreach_instr_safe(instr
, block
) {
399 if (instr
->type
!= nir_instr_type_intrinsic
)
402 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
404 switch (intrin
->intrinsic
) {
405 case nir_intrinsic_load_deref
:
406 case nir_intrinsic_interp_deref_at_centroid
:
407 case nir_intrinsic_interp_deref_at_sample
:
408 case nir_intrinsic_interp_deref_at_offset
: {
409 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
410 if (!(old_deref
->mode
& modes
))
413 if (old_deref
->mode
== nir_var_shader_out
)
414 assert(b
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
415 b
.shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
417 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
419 const unsigned loc
= get_slot(old_var
);
420 const unsigned old_frac
= old_var
->data
.location_frac
;
421 nir_variable
*new_var
= old_deref
->mode
== nir_var_shader_in
?
422 new_inputs
[loc
][old_frac
] :
423 new_outputs
[loc
][old_frac
];
424 bool flat
= old_deref
->mode
== nir_var_shader_in
?
425 flat_inputs
[loc
] : flat_outputs
[loc
];
429 const unsigned new_frac
= new_var
->data
.location_frac
;
431 nir_component_mask_t vec4_comp_mask
=
432 ((1 << intrin
->num_components
) - 1) << old_frac
;
434 b
.cursor
= nir_before_instr(&intrin
->instr
);
436 /* Rewrite the load to use the new variable and only select a
437 * portion of the result.
439 nir_deref_instr
*new_deref
;
441 new_deref
= build_array_deref_of_new_var_flat(
442 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
444 assert(get_slot(new_var
) == loc
);
445 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
446 assert(glsl_type_is_vector(new_deref
->type
));
448 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
449 nir_src_for_ssa(&new_deref
->dest
.ssa
));
451 intrin
->num_components
=
452 glsl_get_components(new_deref
->type
);
453 intrin
->dest
.ssa
.num_components
= intrin
->num_components
;
455 b
.cursor
= nir_after_instr(&intrin
->instr
);
457 nir_ssa_def
*new_vec
= nir_channels(&b
, &intrin
->dest
.ssa
,
458 vec4_comp_mask
>> new_frac
);
459 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
460 nir_src_for_ssa(new_vec
),
461 new_vec
->parent_instr
);
467 case nir_intrinsic_store_deref
: {
468 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
469 if (old_deref
->mode
!= nir_var_shader_out
)
472 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
474 const unsigned loc
= get_slot(old_var
);
475 const unsigned old_frac
= old_var
->data
.location_frac
;
476 nir_variable
*new_var
= new_outputs
[loc
][old_frac
];
477 bool flat
= flat_outputs
[loc
];
481 const unsigned new_frac
= new_var
->data
.location_frac
;
483 b
.cursor
= nir_before_instr(&intrin
->instr
);
485 /* Rewrite the store to be a masked store to the new variable */
486 nir_deref_instr
*new_deref
;
488 new_deref
= build_array_deref_of_new_var_flat(
489 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
491 assert(get_slot(new_var
) == loc
);
492 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
493 assert(glsl_type_is_vector(new_deref
->type
));
495 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
496 nir_src_for_ssa(&new_deref
->dest
.ssa
));
498 intrin
->num_components
=
499 glsl_get_components(new_deref
->type
);
501 nir_component_mask_t old_wrmask
= nir_intrinsic_write_mask(intrin
);
503 assert(intrin
->src
[1].is_ssa
);
504 nir_ssa_def
*old_value
= intrin
->src
[1].ssa
;
505 nir_ssa_def
*comps
[4];
506 for (unsigned c
= 0; c
< intrin
->num_components
; c
++) {
507 if (new_frac
+ c
>= old_frac
&&
508 (old_wrmask
& 1 << (new_frac
+ c
- old_frac
))) {
509 comps
[c
] = nir_channel(&b
, old_value
,
510 new_frac
+ c
- old_frac
);
512 comps
[c
] = nir_ssa_undef(&b
, old_value
->num_components
,
513 old_value
->bit_size
);
516 nir_ssa_def
*new_value
= nir_vec(&b
, comps
, intrin
->num_components
);
517 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
518 nir_src_for_ssa(new_value
));
520 nir_intrinsic_set_write_mask(intrin
,
521 old_wrmask
<< (old_frac
- new_frac
));
534 nir_metadata_preserve(impl
, nir_metadata_block_index
|
535 nir_metadata_dominance
);
542 nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode modes
)
544 bool progress
= false;
546 nir_foreach_function(function
, shader
) {
548 progress
|= nir_lower_io_to_vector_impl(function
->impl
, modes
);