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 const struct glsl_type
*a_type_tail
= a
->type
;
85 const struct glsl_type
*b_type_tail
= b
->type
;
87 if (nir_is_per_vertex_io(a
, shader
->info
.stage
) !=
88 nir_is_per_vertex_io(b
, shader
->info
.stage
))
91 /* They must have the same array structure */
92 if (same_array_structure
) {
93 while (glsl_type_is_array(a_type_tail
)) {
94 if (!glsl_type_is_array(b_type_tail
))
97 if (glsl_get_length(a_type_tail
) != glsl_get_length(b_type_tail
))
100 a_type_tail
= glsl_get_array_element(a_type_tail
);
101 b_type_tail
= glsl_get_array_element(b_type_tail
);
103 if (glsl_type_is_array(b_type_tail
))
106 a_type_tail
= glsl_without_array(a_type_tail
);
107 b_type_tail
= glsl_without_array(b_type_tail
);
110 if (!glsl_type_is_vector_or_scalar(a_type_tail
) ||
111 !glsl_type_is_vector_or_scalar(b_type_tail
))
114 if (glsl_get_base_type(a_type_tail
) != glsl_get_base_type(b_type_tail
))
117 /* TODO: add 64/16bit support ? */
118 if (glsl_get_bit_size(a_type_tail
) != 32)
121 assert(a
->data
.mode
== b
->data
.mode
);
122 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
123 a
->data
.mode
== nir_var_shader_in
&&
124 a
->data
.interpolation
!= b
->data
.interpolation
)
127 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
128 a
->data
.mode
== nir_var_shader_out
&&
129 a
->data
.index
!= b
->data
.index
)
135 static const struct glsl_type
*
136 get_flat_type(const nir_shader
*shader
, nir_variable
*old_vars
[MAX_SLOTS
][4],
137 unsigned *loc
, nir_variable
**first_var
, unsigned *num_vertices
)
141 unsigned num_vars
= 0;
142 enum glsl_base_type base
;
147 assert(*loc
< MAX_SLOTS
);
148 for (unsigned frac
= 0; frac
< 4; frac
++) {
149 nir_variable
*var
= old_vars
[*loc
][frac
];
153 !variables_can_merge(shader
, var
, *first_var
, false)) ||
160 if (!glsl_type_is_vector_or_scalar(glsl_without_array(var
->type
))) {
165 base
= glsl_get_base_type(
166 glsl_without_array(get_per_vertex_type(shader
, var
, NULL
)));
169 bool vs_in
= shader
->info
.stage
== MESA_SHADER_VERTEX
&&
170 var
->data
.mode
== nir_var_shader_in
;
171 unsigned var_slots
= glsl_count_attribute_slots(
172 get_per_vertex_type(shader
, var
, num_vertices
), vs_in
);
173 todo
= MAX2(todo
, var_slots
);
184 return glsl_array_type(glsl_vector_type(base
, 4), slots
, 0);
188 create_new_io_vars(nir_shader
*shader
, struct exec_list
*io_list
,
189 nir_variable
*new_vars
[MAX_SLOTS
][4],
190 bool flat_vars
[MAX_SLOTS
])
192 if (exec_list_is_empty(io_list
))
195 nir_variable
*old_vars
[MAX_SLOTS
][4] = {{0}};
197 nir_foreach_variable(var
, io_list
) {
198 unsigned frac
= var
->data
.location_frac
;
199 old_vars
[get_slot(var
)][frac
] = var
;
202 bool merged_any_vars
= false;
204 for (unsigned loc
= 0; loc
< MAX_SLOTS
; loc
++) {
207 nir_variable
*first_var
= old_vars
[loc
][frac
];
214 bool found_merge
= false;
217 nir_variable
*var
= old_vars
[loc
][frac
];
221 if (var
!= first_var
) {
222 if (!variables_can_merge(shader
, first_var
, var
, true))
228 const unsigned num_components
=
229 glsl_get_components(glsl_without_array(var
->type
));
230 if (!num_components
) {
233 break; /* The type was a struct. */
236 /* We had better not have any overlapping vars */
237 for (unsigned i
= 1; i
< num_components
; i
++)
238 assert(old_vars
[loc
][frac
+ i
] == NULL
);
240 frac
+= num_components
;
246 merged_any_vars
= true;
248 nir_variable
*var
= nir_variable_clone(old_vars
[loc
][first
], shader
);
249 var
->data
.location_frac
= first
;
250 var
->type
= resize_array_vec_type(var
->type
, frac
- first
);
252 nir_shader_add_variable(shader
, var
);
253 for (unsigned i
= first
; i
< frac
; i
++) {
254 new_vars
[loc
][i
] = var
;
255 old_vars
[loc
][i
] = NULL
;
258 old_vars
[loc
][first
] = var
;
262 /* "flat" mode: tries to ensure there is at most one variable per slot by
263 * merging variables into vec4s
265 for (unsigned loc
= 0; loc
< MAX_SLOTS
;) {
266 nir_variable
*first_var
;
267 unsigned num_vertices
;
268 unsigned new_loc
= loc
;
269 const struct glsl_type
*flat_type
=
270 get_flat_type(shader
, old_vars
, &new_loc
, &first_var
, &num_vertices
);
272 merged_any_vars
= true;
274 nir_variable
*var
= nir_variable_clone(first_var
, shader
);
275 var
->data
.location_frac
= 0;
277 var
->type
= glsl_array_type(flat_type
, num_vertices
, 0);
279 var
->type
= flat_type
;
281 nir_shader_add_variable(shader
, var
);
282 for (unsigned i
= 0; i
< glsl_get_length(flat_type
); i
++) {
283 for (unsigned j
= 0; j
< 4; j
++)
284 new_vars
[loc
+ i
][j
] = var
;
285 flat_vars
[loc
+ i
] = true;
291 return merged_any_vars
;
294 static nir_deref_instr
*
295 build_array_deref_of_new_var(nir_builder
*b
, nir_variable
*new_var
,
296 nir_deref_instr
*leader
)
298 if (leader
->deref_type
== nir_deref_type_var
)
299 return nir_build_deref_var(b
, new_var
);
301 nir_deref_instr
*parent
=
302 build_array_deref_of_new_var(b
, new_var
, nir_deref_instr_parent(leader
));
304 return nir_build_deref_follower(b
, parent
, leader
);
308 build_array_index(nir_builder
*b
, nir_deref_instr
*deref
, nir_ssa_def
*base
,
311 switch (deref
->deref_type
) {
312 case nir_deref_type_var
:
314 case nir_deref_type_array
: {
315 nir_ssa_def
*index
= nir_i2i(b
, deref
->arr
.index
.ssa
,
316 deref
->dest
.ssa
.bit_size
);
318 b
, build_array_index(b
, nir_deref_instr_parent(deref
), base
, vs_in
),
319 nir_imul_imm(b
, index
, glsl_count_attribute_slots(deref
->type
, vs_in
)));
322 unreachable("Invalid deref instruction type");
326 static nir_deref_instr
*
327 build_array_deref_of_new_var_flat(nir_shader
*shader
,
328 nir_builder
*b
, nir_variable
*new_var
,
329 nir_deref_instr
*leader
, unsigned base
)
331 nir_deref_instr
*deref
= nir_build_deref_var(b
, new_var
);
333 if (nir_is_per_vertex_io(new_var
, shader
->info
.stage
)) {
334 assert(leader
->deref_type
== nir_deref_type_array
);
335 nir_ssa_def
*index
= leader
->arr
.index
.ssa
;
336 leader
= nir_deref_instr_parent(leader
);
337 deref
= nir_build_deref_array(b
, deref
, index
);
340 bool vs_in
= shader
->info
.stage
== MESA_SHADER_VERTEX
&&
341 new_var
->data
.mode
== nir_var_shader_in
;
342 return nir_build_deref_array(
343 b
, deref
, build_array_index(b
, leader
, nir_imm_int(b
, base
), vs_in
));
347 nir_lower_io_to_vector_impl(nir_function_impl
*impl
, nir_variable_mode modes
)
349 assert(!(modes
& ~(nir_var_shader_in
| nir_var_shader_out
)));
352 nir_builder_init(&b
, impl
);
354 nir_metadata_require(impl
, nir_metadata_dominance
);
356 nir_shader
*shader
= impl
->function
->shader
;
357 nir_variable
*new_inputs
[MAX_SLOTS
][4] = {{0}};
358 nir_variable
*new_outputs
[MAX_SLOTS
][4] = {{0}};
359 bool flat_inputs
[MAX_SLOTS
] = {0};
360 bool flat_outputs
[MAX_SLOTS
] = {0};
362 if (modes
& nir_var_shader_in
) {
363 /* Vertex shaders support overlapping inputs. We don't do those */
364 assert(b
.shader
->info
.stage
!= MESA_SHADER_VERTEX
);
366 /* If we don't actually merge any variables, remove that bit from modes
367 * so we don't bother doing extra non-work.
369 if (!create_new_io_vars(shader
, &shader
->inputs
,
370 new_inputs
, flat_inputs
))
371 modes
&= ~nir_var_shader_in
;
374 if (modes
& nir_var_shader_out
) {
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
->outputs
,
379 new_outputs
, flat_outputs
))
380 modes
&= ~nir_var_shader_out
;
386 bool progress
= false;
388 /* Actually lower all the IO load/store intrinsics. Load instructions are
389 * lowered to a vector load and an ALU instruction to grab the channels we
390 * want. Outputs are lowered to a write-masked store of the vector output.
391 * For non-TCS outputs, we then run nir_lower_io_to_temporaries at the end
392 * to clean up the partial writes.
394 nir_foreach_block(block
, impl
) {
395 nir_foreach_instr_safe(instr
, block
) {
396 if (instr
->type
!= nir_instr_type_intrinsic
)
399 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
401 switch (intrin
->intrinsic
) {
402 case nir_intrinsic_load_deref
:
403 case nir_intrinsic_interp_deref_at_centroid
:
404 case nir_intrinsic_interp_deref_at_sample
:
405 case nir_intrinsic_interp_deref_at_offset
: {
406 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
407 if (!(old_deref
->mode
& modes
))
410 if (old_deref
->mode
== nir_var_shader_out
)
411 assert(b
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
412 b
.shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
414 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
416 const unsigned loc
= get_slot(old_var
);
417 const unsigned old_frac
= old_var
->data
.location_frac
;
418 nir_variable
*new_var
= old_deref
->mode
== nir_var_shader_in
?
419 new_inputs
[loc
][old_frac
] :
420 new_outputs
[loc
][old_frac
];
421 bool flat
= old_deref
->mode
== nir_var_shader_in
?
422 flat_inputs
[loc
] : flat_outputs
[loc
];
426 const unsigned new_frac
= new_var
->data
.location_frac
;
428 nir_component_mask_t vec4_comp_mask
=
429 ((1 << intrin
->num_components
) - 1) << old_frac
;
431 b
.cursor
= nir_before_instr(&intrin
->instr
);
433 /* Rewrite the load to use the new variable and only select a
434 * portion of the result.
436 nir_deref_instr
*new_deref
;
438 new_deref
= build_array_deref_of_new_var_flat(
439 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
441 assert(get_slot(new_var
) == loc
);
442 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
443 assert(glsl_type_is_vector(new_deref
->type
));
445 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
446 nir_src_for_ssa(&new_deref
->dest
.ssa
));
448 intrin
->num_components
=
449 glsl_get_components(new_deref
->type
);
450 intrin
->dest
.ssa
.num_components
= intrin
->num_components
;
452 b
.cursor
= nir_after_instr(&intrin
->instr
);
454 nir_ssa_def
*new_vec
= nir_channels(&b
, &intrin
->dest
.ssa
,
455 vec4_comp_mask
>> new_frac
);
456 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
457 nir_src_for_ssa(new_vec
),
458 new_vec
->parent_instr
);
464 case nir_intrinsic_store_deref
: {
465 nir_deref_instr
*old_deref
= nir_src_as_deref(intrin
->src
[0]);
466 if (old_deref
->mode
!= nir_var_shader_out
)
469 nir_variable
*old_var
= nir_deref_instr_get_variable(old_deref
);
471 const unsigned loc
= get_slot(old_var
);
472 const unsigned old_frac
= old_var
->data
.location_frac
;
473 nir_variable
*new_var
= new_outputs
[loc
][old_frac
];
474 bool flat
= flat_outputs
[loc
];
478 const unsigned new_frac
= new_var
->data
.location_frac
;
480 b
.cursor
= nir_before_instr(&intrin
->instr
);
482 /* Rewrite the store to be a masked store to the new variable */
483 nir_deref_instr
*new_deref
;
485 new_deref
= build_array_deref_of_new_var_flat(
486 shader
, &b
, new_var
, old_deref
, loc
- get_slot(new_var
));
488 assert(get_slot(new_var
) == loc
);
489 new_deref
= build_array_deref_of_new_var(&b
, new_var
, old_deref
);
490 assert(glsl_type_is_vector(new_deref
->type
));
492 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
493 nir_src_for_ssa(&new_deref
->dest
.ssa
));
495 intrin
->num_components
=
496 glsl_get_components(new_deref
->type
);
498 nir_component_mask_t old_wrmask
= nir_intrinsic_write_mask(intrin
);
500 assert(intrin
->src
[1].is_ssa
);
501 nir_ssa_def
*old_value
= intrin
->src
[1].ssa
;
502 nir_ssa_def
*comps
[4];
503 for (unsigned c
= 0; c
< intrin
->num_components
; c
++) {
504 if (new_frac
+ c
>= old_frac
&&
505 (old_wrmask
& 1 << (new_frac
+ c
- old_frac
))) {
506 comps
[c
] = nir_channel(&b
, old_value
,
507 new_frac
+ c
- old_frac
);
509 comps
[c
] = nir_ssa_undef(&b
, old_value
->num_components
,
510 old_value
->bit_size
);
513 nir_ssa_def
*new_value
= nir_vec(&b
, comps
, intrin
->num_components
);
514 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
515 nir_src_for_ssa(new_value
));
517 nir_intrinsic_set_write_mask(intrin
,
518 old_wrmask
<< (old_frac
- new_frac
));
531 nir_metadata_preserve(impl
, nir_metadata_block_index
|
532 nir_metadata_dominance
);
539 nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode modes
)
541 bool progress
= false;
543 nir_foreach_function(function
, shader
) {
545 progress
|= nir_lower_io_to_vector_impl(function
->impl
, modes
);