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
) {
66 nir_foreach_block(block
, function
->impl
) {
67 nir_foreach_instr(instr
, block
) {
68 if (instr
->type
!= nir_instr_type_intrinsic
)
71 nir_intrinsic_instr
*intrin_instr
=
72 nir_instr_as_intrinsic(instr
);
73 if (intrin_instr
->intrinsic
== nir_intrinsic_load_var
&&
74 intrin_instr
->variables
[0]->var
->data
.mode
==
77 nir_variable
*var
= intrin_instr
->variables
[0]->var
;
78 if (var
->data
.patch
) {
79 patches_read
[var
->data
.location_frac
] |=
80 get_variable_io_mask(intrin_instr
->variables
[0]->var
,
83 read
[var
->data
.location_frac
] |=
84 get_variable_io_mask(intrin_instr
->variables
[0]->var
,
95 remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
96 uint64_t *used_by_other_stage
,
97 uint64_t *used_by_other_stage_patches
)
99 bool progress
= false;
102 nir_foreach_variable_safe(var
, var_list
) {
104 used
= used_by_other_stage_patches
;
106 used
= used_by_other_stage
;
108 if (var
->data
.location
< VARYING_SLOT_VAR0
&& var
->data
.location
>= 0)
111 if (var
->data
.always_active_io
)
114 uint64_t other_stage
= used
[var
->data
.location_frac
];
116 if (!(other_stage
& get_variable_io_mask(var
, shader
->info
.stage
))) {
117 /* This one is invalid, make it a global variable instead */
118 var
->data
.location
= 0;
119 var
->data
.mode
= nir_var_global
;
121 exec_node_remove(&var
->node
);
122 exec_list_push_tail(&shader
->globals
, &var
->node
);
132 nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
)
134 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
135 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
137 uint64_t read
[4] = { 0 }, written
[4] = { 0 };
138 uint64_t patches_read
[4] = { 0 }, patches_written
[4] = { 0 };
140 nir_foreach_variable(var
, &producer
->outputs
) {
141 if (var
->data
.patch
) {
142 patches_written
[var
->data
.location_frac
] |=
143 get_variable_io_mask(var
, producer
->info
.stage
);
145 written
[var
->data
.location_frac
] |=
146 get_variable_io_mask(var
, producer
->info
.stage
);
150 nir_foreach_variable(var
, &consumer
->inputs
) {
151 if (var
->data
.patch
) {
152 patches_read
[var
->data
.location_frac
] |=
153 get_variable_io_mask(var
, consumer
->info
.stage
);
155 read
[var
->data
.location_frac
] |=
156 get_variable_io_mask(var
, consumer
->info
.stage
);
160 /* Each TCS invocation can read data written by other TCS invocations,
161 * so even if the outputs are not used by the TES we must also make
162 * sure they are not read by the TCS before demoting them to globals.
164 if (producer
->info
.stage
== MESA_SHADER_TESS_CTRL
)
165 tcs_add_output_reads(producer
, read
, patches_read
);
167 bool progress
= false;
168 progress
= remove_unused_io_vars(producer
, &producer
->outputs
, read
,
171 progress
= remove_unused_io_vars(consumer
, &consumer
->inputs
, written
,
172 patches_written
) || progress
;
178 get_interp_type(nir_variable
*var
, bool default_to_smooth_interp
)
180 if (var
->data
.interpolation
!= INTERP_MODE_NONE
)
181 return var
->data
.interpolation
;
182 else if (default_to_smooth_interp
)
183 return INTERP_MODE_SMOOTH
;
185 return INTERP_MODE_NONE
;
188 #define INTERPOLATE_LOC_SAMPLE 0
189 #define INTERPOLATE_LOC_CENTROID 1
190 #define INTERPOLATE_LOC_CENTER 2
193 get_interp_loc(nir_variable
*var
)
195 if (var
->data
.sample
)
196 return INTERPOLATE_LOC_SAMPLE
;
197 else if (var
->data
.centroid
)
198 return INTERPOLATE_LOC_CENTROID
;
200 return INTERPOLATE_LOC_CENTER
;
204 get_slot_component_masks_and_interp_types(struct exec_list
*var_list
,
206 uint8_t *interp_type
,
208 gl_shader_stage stage
,
209 bool default_to_smooth_interp
)
211 nir_foreach_variable_safe(var
, var_list
) {
212 assert(var
->data
.location
>= 0);
214 /* Only remap things that aren't built-ins.
215 * TODO: add TES patch support.
217 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
218 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
220 const struct glsl_type
*type
= var
->type
;
221 if (nir_is_per_vertex_io(var
, stage
)) {
222 assert(glsl_type_is_array(type
));
223 type
= glsl_get_array_element(type
);
226 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
228 glsl_get_vector_elements(glsl_without_array(type
));
230 bool dual_slot
= glsl_type_is_dual_slot(glsl_without_array(type
));
231 unsigned slots
= glsl_count_attribute_slots(type
, false);
232 unsigned comps_slot2
= 0;
233 for (unsigned i
= 0; i
< slots
; i
++) {
234 interp_type
[location
+ i
] =
235 get_interp_type(var
, default_to_smooth_interp
);
236 interp_loc
[location
+ i
] = get_interp_loc(var
);
240 comps
[location
+ i
] |= ((1 << comps_slot2
) - 1);
242 unsigned num_comps
= 4 - var
->data
.location_frac
;
243 comps_slot2
= (elements
* 2) - num_comps
;
245 /* Assume ARB_enhanced_layouts packing rules for doubles */
246 assert(var
->data
.location_frac
== 0 ||
247 var
->data
.location_frac
== 2);
248 assert(comps_slot2
<= 4);
250 comps
[location
+ i
] |=
251 ((1 << num_comps
) - 1) << var
->data
.location_frac
;
254 comps
[location
+ i
] |=
255 ((1 << elements
) - 1) << var
->data
.location_frac
;
269 remap_slots_and_components(struct exec_list
*var_list
, gl_shader_stage stage
,
270 struct varying_loc (*remap
)[4],
271 uint64_t *slots_used
, uint64_t *out_slots_read
)
273 uint64_t out_slots_read_tmp
= 0;
275 /* We don't touch builtins so just copy the bitmask */
276 uint64_t slots_used_tmp
=
277 *slots_used
& (((uint64_t)1 << (VARYING_SLOT_VAR0
- 1)) - 1);
279 nir_foreach_variable(var
, var_list
) {
280 assert(var
->data
.location
>= 0);
282 /* Only remap things that aren't built-ins */
283 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
284 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
285 assert(var
->data
.location
- VARYING_SLOT_VAR0
< 32);
286 assert(remap
[var
->data
.location
- VARYING_SLOT_VAR0
] >= 0);
288 const struct glsl_type
*type
= var
->type
;
289 if (nir_is_per_vertex_io(var
, stage
)) {
290 assert(glsl_type_is_array(type
));
291 type
= glsl_get_array_element(type
);
294 unsigned num_slots
= glsl_count_attribute_slots(type
, false);
295 bool used_across_stages
= false;
296 bool outputs_read
= false;
298 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
299 struct varying_loc
*new_loc
= &remap
[location
][var
->data
.location_frac
];
300 if (new_loc
->location
) {
301 uint64_t slots
= (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
302 if (slots
& *slots_used
)
303 used_across_stages
= true;
305 if (slots
& *out_slots_read
)
308 var
->data
.location
= new_loc
->location
;
309 var
->data
.location_frac
= new_loc
->component
;
312 if (var
->data
.always_active_io
) {
313 /* We can't apply link time optimisations (specifically array
314 * splitting) to these so we need to copy the existing mask
315 * otherwise we will mess up the mask for things like partially
318 if (used_across_stages
) {
320 *slots_used
& (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
324 out_slots_read_tmp
|=
325 *out_slots_read
& (((uint64_t)1 << num_slots
) - 1) << var
->data
.location
;
329 for (unsigned i
= 0; i
< num_slots
; i
++) {
330 if (used_across_stages
)
331 slots_used_tmp
|= (uint64_t)1 << (var
->data
.location
+ i
);
334 out_slots_read_tmp
|= (uint64_t)1 << (var
->data
.location
+ i
);
340 *slots_used
= slots_used_tmp
;
341 *out_slots_read
= out_slots_read_tmp
;
344 /* If there are empty components in the slot compact the remaining components
345 * as close to component 0 as possible. This will make it easier to fill the
346 * empty components with components from a different slot in a following pass.
349 compact_components(nir_shader
*producer
, nir_shader
*consumer
, uint8_t *comps
,
350 uint8_t *interp_type
, uint8_t *interp_loc
,
351 bool default_to_smooth_interp
)
353 struct exec_list
*input_list
= &consumer
->inputs
;
354 struct exec_list
*output_list
= &producer
->outputs
;
355 struct varying_loc remap
[32][4] = {{{0}, {0}}};
357 /* Create a cursor for each interpolation type */
358 unsigned cursor
[4] = {0};
360 /* We only need to pass over one stage and we choose the consumer as it seems
361 * to cause a larger reduction in instruction counts (tested on i965).
363 nir_foreach_variable(var
, input_list
) {
365 /* Only remap things that aren't builtins.
366 * TODO: add TES patch support.
368 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
369 var
->data
.location
- VARYING_SLOT_VAR0
< 32) {
371 /* We can't repack xfb varyings. */
372 if (var
->data
.always_active_io
)
375 const struct glsl_type
*type
= var
->type
;
376 if (nir_is_per_vertex_io(var
, consumer
->info
.stage
)) {
377 assert(glsl_type_is_array(type
));
378 type
= glsl_get_array_element(type
);
381 /* Skip types that require more complex packing handling.
382 * TODO: add support for these types.
384 if (glsl_type_is_array(type
) ||
385 glsl_type_is_dual_slot(type
) ||
386 glsl_type_is_matrix(type
) ||
387 glsl_type_is_struct(type
) ||
388 glsl_type_is_64bit(type
))
391 /* We ignore complex types above and all other vector types should
392 * have been split into scalar variables by the lower_io_to_scalar
393 * pass. The only exeption should by OpenGL xfb varyings.
395 if (glsl_get_vector_elements(type
) != 1)
398 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
399 uint8_t used_comps
= comps
[location
];
401 /* If there are no empty components there is nothing more for us to do.
403 if (used_comps
== 0xf)
406 bool found_new_offset
= false;
407 uint8_t interp
= get_interp_type(var
, default_to_smooth_interp
);
408 for (; cursor
[interp
] < 32; cursor
[interp
]++) {
409 uint8_t cursor_used_comps
= comps
[cursor
[interp
]];
411 /* We couldn't find anywhere to pack the varying continue on. */
412 if (cursor
[interp
] == location
&&
413 (var
->data
.location_frac
== 0 ||
414 cursor_used_comps
& ((1 << (var
->data
.location_frac
)) - 1)))
417 /* We can only pack varyings with matching interpolation types */
418 if (interp_type
[cursor
[interp
]] != interp
)
421 /* Interpolation loc must match also.
422 * TODO: i965 can handle these if they don't match, but the
423 * radeonsi nir backend handles everything as vec4s and so expects
424 * this to be the same for all components. We could make this
425 * check driver specfific or drop it if NIR ever become the only
428 if (interp_loc
[cursor
[interp
]] != get_interp_loc(var
))
431 /* If the slot is empty just skip it for now, compact_var_list()
432 * can be called after this function to remove empty slots for us.
433 * TODO: finish implementing compact_var_list() requires array and
436 if (!cursor_used_comps
)
439 uint8_t unused_comps
= ~cursor_used_comps
;
441 for (unsigned i
= 0; i
< 4; i
++) {
442 uint8_t new_var_comps
= 1 << i
;
443 if (unused_comps
& new_var_comps
) {
444 remap
[location
][var
->data
.location_frac
].component
= i
;
445 remap
[location
][var
->data
.location_frac
].location
=
446 cursor
[interp
] + VARYING_SLOT_VAR0
;
448 found_new_offset
= true;
450 /* Turn off the mask for the component we are remapping */
451 if (comps
[location
] & 1 << var
->data
.location_frac
) {
452 comps
[location
] ^= 1 << var
->data
.location_frac
;
453 comps
[cursor
[interp
]] |= new_var_comps
;
459 if (found_new_offset
)
466 remap_slots_and_components(input_list
, consumer
->info
.stage
, remap
,
467 &consumer
->info
.inputs_read
, &zero
);
468 remap_slots_and_components(output_list
, producer
->info
.stage
, remap
,
469 &producer
->info
.outputs_written
,
470 &producer
->info
.outputs_read
);
473 /* We assume that this has been called more-or-less directly after
474 * remove_unused_varyings. At this point, all of the varyings that we
475 * aren't going to be using have been completely removed and the
476 * inputs_read and outputs_written fields in nir_shader_info reflect
477 * this. Therefore, the total set of valid slots is the OR of the two
478 * sets of varyings; this accounts for varyings which one side may need
479 * to read/write even if the other doesn't. This can happen if, for
480 * instance, an array is used indirectly from one side causing it to be
481 * unsplittable but directly from the other.
484 nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
485 bool default_to_smooth_interp
)
487 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
488 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
490 uint8_t comps
[32] = {0};
491 uint8_t interp_type
[32] = {0};
492 uint8_t interp_loc
[32] = {0};
494 get_slot_component_masks_and_interp_types(&producer
->outputs
, comps
,
495 interp_type
, interp_loc
,
496 producer
->info
.stage
,
497 default_to_smooth_interp
);
498 get_slot_component_masks_and_interp_types(&consumer
->inputs
, comps
,
499 interp_type
, interp_loc
,
500 consumer
->info
.stage
,
501 default_to_smooth_interp
);
503 compact_components(producer
, consumer
, comps
, interp_type
, interp_loc
,
504 default_to_smooth_interp
);