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
);
188 return glsl_vector_type(base
, 4);
190 return glsl_array_type(glsl_vector_type(base
, 4), slots
, 0);
194 create_new_io_vars(nir_shader
*shader
, struct exec_list
*io_list
,
195 nir_variable
*new_vars
[MAX_SLOTS
][4],
196 bool flat_vars
[MAX_SLOTS
])
198 if (exec_list_is_empty(io_list
))
201 nir_variable
*old_vars
[MAX_SLOTS
][4] = {{0}};
203 nir_foreach_variable(var
, io_list
) {
204 unsigned frac
= var
->data
.location_frac
;
205 old_vars
[get_slot(var
)][frac
] = var
;
208 bool merged_any_vars
= false;
210 for (unsigned loc
= 0; loc
< MAX_SLOTS
; loc
++) {
213 nir_variable
*first_var
= old_vars
[loc
][frac
];
220 bool found_merge
= false;
223 nir_variable
*var
= old_vars
[loc
][frac
];
227 if (var
!= first_var
) {
228 if (!variables_can_merge(shader
, first_var
, var
, true))
234 const unsigned num_components
=
235 glsl_get_components(glsl_without_array(var
->type
));
236 if (!num_components
) {
239 break; /* The type was a struct. */
242 /* We had better not have any overlapping vars */
243 for (unsigned i
= 1; i
< num_components
; i
++)
244 assert(old_vars
[loc
][frac
+ i
] == NULL
);
246 frac
+= num_components
;
252 merged_any_vars
= true;
254 nir_variable
*var
= nir_variable_clone(old_vars
[loc
][first
], shader
);
255 var
->data
.location_frac
= first
;
256 var
->type
= resize_array_vec_type(var
->type
, frac
- first
);
258 nir_shader_add_variable(shader
, var
);
259 for (unsigned i
= first
; i
< frac
; i
++) {
260 new_vars
[loc
][i
] = var
;
261 old_vars
[loc
][i
] = NULL
;
264 old_vars
[loc
][first
] = var
;
268 /* "flat" mode: tries to ensure there is at most one variable per slot by
269 * merging variables into vec4s
271 for (unsigned loc
= 0; loc
< MAX_SLOTS
;) {
272 nir_variable
*first_var
;
273 unsigned num_vertices
;
274 unsigned new_loc
= loc
;
275 const struct glsl_type
*flat_type
=
276 get_flat_type(shader
, old_vars
, &new_loc
, &first_var
, &num_vertices
);
278 merged_any_vars
= true;
280 nir_variable
*var
= nir_variable_clone(first_var
, shader
);
281 var
->data
.location_frac
= 0;
283 var
->type
= glsl_array_type(flat_type
, num_vertices
, 0);
285 var
->type
= flat_type
;
287 nir_shader_add_variable(shader
, var
);
288 for (unsigned i
= 0; i
< glsl_get_length(flat_type
); i
++) {
289 for (unsigned j
= 0; j
< 4; j
++)
290 new_vars
[loc
+ i
][j
] = var
;
291 flat_vars
[loc
+ i
] = true;
297 return merged_any_vars
;
300 static nir_deref_instr
*
301 build_array_deref_of_new_var(nir_builder
*b
, nir_variable
*new_var
,
302 nir_deref_instr
*leader
)
304 if (leader
->deref_type
== nir_deref_type_var
)
305 return nir_build_deref_var(b
, new_var
);
307 nir_deref_instr
*parent
=
308 build_array_deref_of_new_var(b
, new_var
, nir_deref_instr_parent(leader
));
310 return nir_build_deref_follower(b
, parent
, leader
);
314 build_array_index(nir_builder
*b
, nir_deref_instr
*deref
, nir_ssa_def
*base
,
317 switch (deref
->deref_type
) {
318 case nir_deref_type_var
:
320 case nir_deref_type_array
: {
321 nir_ssa_def
*index
= nir_i2i(b
, deref
->arr
.index
.ssa
,
322 deref
->dest
.ssa
.bit_size
);
324 b
, build_array_index(b
, nir_deref_instr_parent(deref
), base
, vs_in
),
325 nir_amul_imm(b
, index
, glsl_count_attribute_slots(deref
->type
, vs_in
)));
328 unreachable("Invalid deref instruction type");
332 static nir_deref_instr
*
333 build_array_deref_of_new_var_flat(nir_shader
*shader
,
334 nir_builder
*b
, nir_variable
*new_var
,
335 nir_deref_instr
*leader
, unsigned base
)
337 nir_deref_instr
*deref
= nir_build_deref_var(b
, new_var
);
339 if (nir_is_per_vertex_io(new_var
, shader
->info
.stage
)) {
340 assert(leader
->deref_type
== nir_deref_type_array
);
341 nir_ssa_def
*index
= leader
->arr
.index
.ssa
;
342 leader
= nir_deref_instr_parent(leader
);
343 deref
= nir_build_deref_array(b
, deref
, index
);
346 if (!glsl_type_is_array(deref
->type
))
349 bool vs_in
= shader
->info
.stage
== MESA_SHADER_VERTEX
&&
350 new_var
->data
.mode
== nir_var_shader_in
;
351 return nir_build_deref_array(
352 b
, deref
, build_array_index(b
, leader
, nir_imm_int(b
, base
), vs_in
));
356 nir_lower_io_to_vector_impl(nir_function_impl
*impl
, nir_variable_mode modes
)
358 assert(!(modes
& ~(nir_var_shader_in
| nir_var_shader_out
)));
361 nir_builder_init(&b
, impl
);
363 nir_metadata_require(impl
, nir_metadata_dominance
);
365 nir_shader
*shader
= impl
->function
->shader
;
366 nir_variable
*new_inputs
[MAX_SLOTS
][4] = {{0}};
367 nir_variable
*new_outputs
[MAX_SLOTS
][4] = {{0}};
368 bool flat_inputs
[MAX_SLOTS
] = {0};
369 bool flat_outputs
[MAX_SLOTS
] = {0};
371 if (modes
& nir_var_shader_in
) {
372 /* Vertex shaders support overlapping inputs. We don't do those */
373 assert(b
.shader
->info
.stage
!= MESA_SHADER_VERTEX
);
375 /* If we don't actually merge any variables, remove that bit from modes
376 * so we don't bother doing extra non-work.
378 if (!create_new_io_vars(shader
, &shader
->inputs
,
379 new_inputs
, flat_inputs
))
380 modes
&= ~nir_var_shader_in
;
383 if (modes
& nir_var_shader_out
) {
384 /* If we don't actually merge any variables, remove that bit from modes
385 * so we don't bother doing extra non-work.
387 if (!create_new_io_vars(shader
, &shader
->outputs
,
388 new_outputs
, flat_outputs
))
389 modes
&= ~nir_var_shader_out
;
395 bool progress
= false;
397 /* Actually lower all the IO load/store intrinsics. Load instructions are
398 * lowered to a vector load and an ALU instruction to grab the channels we
399 * want. Outputs are lowered to a write-masked store of the vector output.
400 * For non-TCS outputs, we then run nir_lower_io_to_temporaries at the end
401 * to clean up the partial writes.
403 nir_foreach_block(block
, impl
) {
404 nir_foreach_instr_safe(instr
, block
) {
405 if (instr
->type
!= nir_instr_type_intrinsic
)
408 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
410 switch (intrin
->intrinsic
) {
411 case nir_intrinsic_load_deref
:
412 case nir_intrinsic_interp_deref_at_centroid
:
413 case nir_intrinsic_interp_deref_at_sample
:
414 case nir_intrinsic_interp_deref_at_offset
: {
415 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
416 if (!(old_deref
->mode
& modes
))
419 if (old_deref
->mode
== nir_var_shader_out
)
420 assert(b
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
421 b
.shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
423 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
425 const unsigned loc
= get_slot(old_var
);
426 const unsigned old_frac
= old_var
->data
.location_frac
;
427 nir_variable
*new_var
= old_deref
->mode
== nir_var_shader_in
?
428 new_inputs
[loc
][old_frac
] :
429 new_outputs
[loc
][old_frac
];
430 bool flat
= old_deref
->mode
== nir_var_shader_in
?
431 flat_inputs
[loc
] : flat_outputs
[loc
];
435 const unsigned new_frac
= new_var
->data
.location_frac
;
437 nir_component_mask_t vec4_comp_mask
=
438 ((1 << intrin
->num_components
) - 1) << old_frac
;
440 b
.cursor
= nir_before_instr(&intrin
->instr
);
442 /* Rewrite the load to use the new variable and only select a
443 * portion of the result.
445 nir_deref_instr
*new_deref
;
447 new_deref
= build_array_deref_of_new_var_flat(
448 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
450 assert(get_slot(new_var
) == loc
);
451 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
452 assert(glsl_type_is_vector(new_deref
->type
));
454 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
455 nir_src_for_ssa(&new_deref
->dest
.ssa
));
457 intrin
->num_components
=
458 glsl_get_components(new_deref
->type
);
459 intrin
->dest
.ssa
.num_components
= intrin
->num_components
;
461 b
.cursor
= nir_after_instr(&intrin
->instr
);
463 nir_ssa_def
*new_vec
= nir_channels(&b
, &intrin
->dest
.ssa
,
464 vec4_comp_mask
>> new_frac
);
465 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
466 nir_src_for_ssa(new_vec
),
467 new_vec
->parent_instr
);
473 case nir_intrinsic_store_deref
: {
474 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
475 if (old_deref
->mode
!= nir_var_shader_out
)
478 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
480 const unsigned loc
= get_slot(old_var
);
481 const unsigned old_frac
= old_var
->data
.location_frac
;
482 nir_variable
*new_var
= new_outputs
[loc
][old_frac
];
483 bool flat
= flat_outputs
[loc
];
487 const unsigned new_frac
= new_var
->data
.location_frac
;
489 b
.cursor
= nir_before_instr(&intrin
->instr
);
491 /* Rewrite the store to be a masked store to the new variable */
492 nir_deref_instr
*new_deref
;
494 new_deref
= build_array_deref_of_new_var_flat(
495 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
497 assert(get_slot(new_var
) == loc
);
498 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
499 assert(glsl_type_is_vector(new_deref
->type
));
501 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
502 nir_src_for_ssa(&new_deref
->dest
.ssa
));
504 intrin
->num_components
=
505 glsl_get_components(new_deref
->type
);
507 nir_component_mask_t old_wrmask
= nir_intrinsic_write_mask(intrin
);
509 assert(intrin
->src
[1].is_ssa
);
510 nir_ssa_def
*old_value
= intrin
->src
[1].ssa
;
511 nir_ssa_def
*comps
[4];
512 for (unsigned c
= 0; c
< intrin
->num_components
; c
++) {
513 if (new_frac
+ c
>= old_frac
&&
514 (old_wrmask
& 1 << (new_frac
+ c
- old_frac
))) {
515 comps
[c
] = nir_channel(&b
, old_value
,
516 new_frac
+ c
- old_frac
);
518 comps
[c
] = nir_ssa_undef(&b
, old_value
->num_components
,
519 old_value
->bit_size
);
522 nir_ssa_def
*new_value
= nir_vec(&b
, comps
, intrin
->num_components
);
523 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
524 nir_src_for_ssa(new_value
));
526 nir_intrinsic_set_write_mask(intrin
,
527 old_wrmask
<< (old_frac
- new_frac
));
540 nir_metadata_preserve(impl
, nir_metadata_block_index
|
541 nir_metadata_dominance
);
548 nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode modes
)
550 bool progress
= false;
552 nir_foreach_function(function
, shader
) {
554 progress
|= nir_lower_io_to_vector_impl(function
->impl
, modes
);