2 * Copyright © 2015 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
26 #include "util/hash_table.h"
28 /* This file contains various little helpers for doing simple linking in
29 * NIR. Eventually, we'll probably want a full-blown varying packing
30 * implementation in here. Right now, it just deletes unused things.
34 * Returns the bits in the inputs_read, outputs_written, or
35 * system_values_read bitfield corresponding to this variable.
38 get_variable_io_mask(nir_variable
*var
, gl_shader_stage stage
)
40 if (var
->data
.location
< 0)
43 unsigned location
= var
->data
.patch
?
44 var
->data
.location
- VARYING_SLOT_PATCH0
: var
->data
.location
;
46 assert(var
->data
.mode
== nir_var_shader_in
||
47 var
->data
.mode
== nir_var_shader_out
||
48 var
->data
.mode
== nir_var_system_value
);
49 assert(var
->data
.location
>= 0);
51 const struct glsl_type
*type
= var
->type
;
52 if (nir_is_per_vertex_io(var
, stage
)) {
53 assert(glsl_type_is_array(type
));
54 type
= glsl_get_array_element(type
);
57 unsigned slots
= glsl_count_attribute_slots(type
, false);
58 return ((1ull << slots
) - 1) << location
;
62 tcs_add_output_reads(nir_shader
*shader
, uint64_t *read
, uint64_t *patches_read
)
64 nir_foreach_function(function
, shader
) {
68 nir_foreach_block(block
, function
->impl
) {
69 nir_foreach_instr(instr
, block
) {
70 if (instr
->type
!= nir_instr_type_intrinsic
)
73 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
74 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
)
78 nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[0]));
80 if (var
->data
.mode
!= nir_var_shader_out
)
83 if (var
->data
.patch
) {
84 patches_read
[var
->data
.location_frac
] |=
85 get_variable_io_mask(var
, shader
->info
.stage
);
87 read
[var
->data
.location_frac
] |=
88 get_variable_io_mask(var
, shader
->info
.stage
);
96 remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
97 uint64_t *used_by_other_stage
,
98 uint64_t *used_by_other_stage_patches
)
100 bool progress
= false;
103 nir_foreach_variable_safe(var
, var_list
) {
105 used
= used_by_other_stage_patches
;
107 used
= used_by_other_stage
;
109 if (var
->data
.location
< VARYING_SLOT_VAR0
&& var
->data
.location
>= 0)
112 if (var
->data
.always_active_io
)
115 uint64_t other_stage
= used
[var
->data
.location_frac
];
117 if (!(other_stage
& get_variable_io_mask(var
, shader
->info
.stage
))) {
118 /* This one is invalid, make it a global variable instead */
119 var
->data
.location
= 0;
120 var
->data
.mode
= nir_var_global
;
122 exec_node_remove(&var
->node
);
123 exec_list_push_tail(&shader
->globals
, &var
->node
);
133 nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
)
135 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
136 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
137 nir_assert_unlowered_derefs(producer
, nir_lower_load_store_derefs
);
138 nir_assert_unlowered_derefs(consumer
, nir_lower_load_store_derefs
);
140 uint64_t read
[4] = { 0 }, written
[4] = { 0 };
141 uint64_t patches_read
[4] = { 0 }, patches_written
[4] = { 0 };
143 nir_foreach_variable(var
, &producer
->outputs
) {
144 if (var
->data
.patch
) {
145 patches_written
[var
->data
.location_frac
] |=
146 get_variable_io_mask(var
, producer
->info
.stage
);
148 written
[var
->data
.location_frac
] |=
149 get_variable_io_mask(var
, producer
->info
.stage
);
153 nir_foreach_variable(var
, &consumer
->inputs
) {
154 if (var
->data
.patch
) {
155 patches_read
[var
->data
.location_frac
] |=
156 get_variable_io_mask(var
, consumer
->info
.stage
);
158 read
[var
->data
.location_frac
] |=
159 get_variable_io_mask(var
, consumer
->info
.stage
);
163 /* Each TCS invocation can read data written by other TCS invocations,
164 * so even if the outputs are not used by the TES we must also make
165 * sure they are not read by the TCS before demoting them to globals.
167 if (producer
->info
.stage
== MESA_SHADER_TESS_CTRL
)
168 tcs_add_output_reads(producer
, read
, patches_read
);
170 bool progress
= false;
171 progress
= remove_unused_io_vars(producer
, &producer
->outputs
, read
,
174 progress
= remove_unused_io_vars(consumer
, &consumer
->inputs
, written
,
175 patches_written
) || progress
;
181 get_interp_type(nir_variable
*var
, bool default_to_smooth_interp
)
183 if (var
->data
.interpolation
!= INTERP_MODE_NONE
)
184 return var
->data
.interpolation
;
185 else if (default_to_smooth_interp
)
186 return INTERP_MODE_SMOOTH
;
188 return INTERP_MODE_NONE
;
191 #define INTERPOLATE_LOC_SAMPLE 0
192 #define INTERPOLATE_LOC_CENTROID 1
193 #define INTERPOLATE_LOC_CENTER 2
196 get_interp_loc(nir_variable
*var
)
198 if (var
->data
.sample
)
199 return INTERPOLATE_LOC_SAMPLE
;
200 else if (var
->data
.centroid
)
201 return INTERPOLATE_LOC_CENTROID
;
203 return INTERPOLATE_LOC_CENTER
;
207 get_slot_component_masks_and_interp_types(struct exec_list
*var_list
,
209 uint8_t *interp_type
,
211 gl_shader_stage stage
,
212 bool default_to_smooth_interp
)
214 nir_foreach_variable_safe(var
, var_list
) {
215 assert(var
->data
.location
>= 0);
217 /* Only remap things that aren't built-ins.
218 * TODO: add TES patch support.
220 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
221 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
223 const struct glsl_type
*type
= var
->type
;
224 if (nir_is_per_vertex_io(var
, stage
)) {
225 assert(glsl_type_is_array(type
));
226 type
= glsl_get_array_element(type
);
229 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
231 glsl_get_vector_elements(glsl_without_array(type
));
233 bool dual_slot
= glsl_type_is_dual_slot(glsl_without_array(type
));
234 unsigned slots
= glsl_count_attribute_slots(type
, false);
235 unsigned comps_slot2
= 0;
236 for (unsigned i
= 0; i
< slots
; i
++) {
237 interp_type
[location
+ i
] =
238 get_interp_type(var
, default_to_smooth_interp
);
239 interp_loc
[location
+ i
] = get_interp_loc(var
);
243 comps
[location
+ i
] |= ((1 << comps_slot2
) - 1);
245 unsigned num_comps
= 4 - var
->data
.location_frac
;
246 comps_slot2
= (elements
* 2) - num_comps
;
248 /* Assume ARB_enhanced_layouts packing rules for doubles */
249 assert(var
->data
.location_frac
== 0 ||
250 var
->data
.location_frac
== 2);
251 assert(comps_slot2
<= 4);
253 comps
[location
+ i
] |=
254 ((1 << num_comps
) - 1) << var
->data
.location_frac
;
257 comps
[location
+ i
] |=
258 ((1 << elements
) - 1) << var
->data
.location_frac
;
272 remap_slots_and_components(struct exec_list
*var_list
, gl_shader_stage stage
,
273 struct varying_loc (*remap
)[4],
274 uint64_t *slots_used
, uint64_t *out_slots_read
)
276 uint64_t out_slots_read_tmp
= 0;
278 /* We don't touch builtins so just copy the bitmask */
279 uint64_t slots_used_tmp
=
280 *slots_used
& (((uint64_t)1 << (VARYING_SLOT_VAR0
- 1)) - 1);
282 nir_foreach_variable(var
, var_list
) {
283 assert(var
->data
.location
>= 0);
285 /* Only remap things that aren't built-ins */
286 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
287 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
288 assert(var
->data
.location
- VARYING_SLOT_VAR0
< 32);
290 const struct glsl_type
*type
= var
->type
;
291 if (nir_is_per_vertex_io(var
, stage
)) {
292 assert(glsl_type_is_array(type
));
293 type
= glsl_get_array_element(type
);
296 unsigned num_slots
= glsl_count_attribute_slots(type
, false);
297 bool used_across_stages
= false;
298 bool outputs_read
= false;
300 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
301 struct varying_loc
*new_loc
= &remap
[location
][var
->data
.location_frac
];
303 uint64_t slots
= (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
304 if (slots
& *slots_used
)
305 used_across_stages
= true;
307 if (slots
& *out_slots_read
)
310 if (new_loc
->location
) {
311 var
->data
.location
= new_loc
->location
;
312 var
->data
.location_frac
= new_loc
->component
;
315 if (var
->data
.always_active_io
) {
316 /* We can't apply link time optimisations (specifically array
317 * splitting) to these so we need to copy the existing mask
318 * otherwise we will mess up the mask for things like partially
321 if (used_across_stages
) {
323 *slots_used
& (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
327 out_slots_read_tmp
|=
328 *out_slots_read
& (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
332 for (unsigned i
= 0; i
< num_slots
; i
++) {
333 if (used_across_stages
)
334 slots_used_tmp
|= (uint64_t)1 << (var
->data
.location
+ i
);
337 out_slots_read_tmp
|= (uint64_t)1 << (var
->data
.location
+ i
);
343 *slots_used
= slots_used_tmp
;
344 *out_slots_read
= out_slots_read_tmp
;
347 /* If there are empty components in the slot compact the remaining components
348 * as close to component 0 as possible. This will make it easier to fill the
349 * empty components with components from a different slot in a following pass.
352 compact_components(nir_shader
*producer
, nir_shader
*consumer
, uint8_t *comps
,
353 uint8_t *interp_type
, uint8_t *interp_loc
,
354 bool default_to_smooth_interp
)
356 struct exec_list
*input_list
= &consumer
->inputs
;
357 struct exec_list
*output_list
= &producer
->outputs
;
358 struct varying_loc remap
[32][4] = {{{0}, {0}}};
360 /* Create a cursor for each interpolation type */
361 unsigned cursor
[4] = {0};
363 /* We only need to pass over one stage and we choose the consumer as it seems
364 * to cause a larger reduction in instruction counts (tested on i965).
366 nir_foreach_variable(var
, input_list
) {
368 /* Only remap things that aren't builtins.
369 * TODO: add TES patch support.
371 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
372 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
374 /* We can't repack xfb varyings. */
375 if (var
->data
.always_active_io
)
378 const struct glsl_type
*type
= var
->type
;
379 if (nir_is_per_vertex_io(var
, consumer
->info
.stage
)) {
380 assert(glsl_type_is_array(type
));
381 type
= glsl_get_array_element(type
);
384 /* Skip types that require more complex packing handling.
385 * TODO: add support for these types.
387 if (glsl_type_is_array(type
) ||
388 glsl_type_is_dual_slot(type
) ||
389 glsl_type_is_matrix(type
) ||
390 glsl_type_is_struct(type
) ||
391 glsl_type_is_64bit(type
))
394 /* We ignore complex types above and all other vector types should
395 * have been split into scalar variables by the lower_io_to_scalar
396 * pass. The only exeption should by OpenGL xfb varyings.
398 if (glsl_get_vector_elements(type
) != 1)
401 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
402 uint8_t used_comps
= comps
[location
];
404 /* If there are no empty components there is nothing more for us to do.
406 if (used_comps
== 0xf)
409 bool found_new_offset
= false;
410 uint8_t interp
= get_interp_type(var
, default_to_smooth_interp
);
411 for (; cursor
[interp
] < 32; cursor
[interp
]++) {
412 uint8_t cursor_used_comps
= comps
[cursor
[interp
]];
414 /* We couldn't find anywhere to pack the varying continue on. */
415 if (cursor
[interp
] == location
&&
416 (var
->data
.location_frac
== 0 ||
417 cursor_used_comps
& ((1 << (var
->data
.location_frac
)) - 1)))
420 /* We can only pack varyings with matching interpolation types */
421 if (interp_type
[cursor
[interp
]] != interp
)
424 /* Interpolation loc must match also.
425 * TODO: i965 can handle these if they don't match, but the
426 * radeonsi nir backend handles everything as vec4s and so expects
427 * this to be the same for all components. We could make this
428 * check driver specfific or drop it if NIR ever become the only
431 if (interp_loc
[cursor
[interp
]] != get_interp_loc(var
))
434 /* If the slot is empty just skip it for now, compact_var_list()
435 * can be called after this function to remove empty slots for us.
436 * TODO: finish implementing compact_var_list() requires array and
439 if (!cursor_used_comps
)
442 uint8_t unused_comps
= ~cursor_used_comps
;
444 for (unsigned i
= 0; i
< 4; i
++) {
445 uint8_t new_var_comps
= 1 << i
;
446 if (unused_comps
& new_var_comps
) {
447 remap
[location
][var
->data
.location_frac
].component
= i
;
448 remap
[location
][var
->data
.location_frac
].location
=
449 cursor
[interp
] + VARYING_SLOT_VAR0
;
451 found_new_offset
= true;
453 /* Turn off the mask for the component we are remapping */
454 if (comps
[location
] & 1 << var
->data
.location_frac
) {
455 comps
[location
] ^= 1 << var
->data
.location_frac
;
456 comps
[cursor
[interp
]] |= new_var_comps
;
462 if (found_new_offset
)
469 remap_slots_and_components(input_list
, consumer
->info
.stage
, remap
,
470 &consumer
->info
.inputs_read
, &zero
);
471 remap_slots_and_components(output_list
, producer
->info
.stage
, remap
,
472 &producer
->info
.outputs_written
,
473 &producer
->info
.outputs_read
);
476 /* We assume that this has been called more-or-less directly after
477 * remove_unused_varyings. At this point, all of the varyings that we
478 * aren't going to be using have been completely removed and the
479 * inputs_read and outputs_written fields in nir_shader_info reflect
480 * this. Therefore, the total set of valid slots is the OR of the two
481 * sets of varyings; this accounts for varyings which one side may need
482 * to read/write even if the other doesn't. This can happen if, for
483 * instance, an array is used indirectly from one side causing it to be
484 * unsplittable but directly from the other.
487 nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
488 bool default_to_smooth_interp
)
490 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
491 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
493 uint8_t comps
[32] = {0};
494 uint8_t interp_type
[32] = {0};
495 uint8_t interp_loc
[32] = {0};
497 get_slot_component_masks_and_interp_types(&producer
->outputs
, comps
,
498 interp_type
, interp_loc
,
499 producer
->info
.stage
,
500 default_to_smooth_interp
);
501 get_slot_component_masks_and_interp_types(&consumer
->inputs
, comps
,
502 interp_type
, interp_loc
,
503 consumer
->info
.stage
,
504 default_to_smooth_interp
);
506 compact_components(producer
, consumer
, comps
, interp_type
, interp_loc
,
507 default_to_smooth_interp
);