1 /*
2 * Copyright © 2017 Timothy Arceri
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "nir.h"
25 #include "nir_builder.h"
26
27 /** @file nir_lower_io_arrays_to_elements.c
28 *
29 * Split arrays/matrices with direct indexing into individual elements. This
30 * will allow optimisation passes to better clean up unused elements.
31 *
32 */
33
34 static unsigned
get_io_offset(nir_builder * b,nir_deref_var * deref,nir_variable * var,unsigned * element_index)35 get_io_offset(nir_builder *b, nir_deref_var *deref, nir_variable *var,
36 unsigned *element_index)
37 {
38 bool vs_in = (b->shader->info.stage == MESA_SHADER_VERTEX) &&
39 (var->data.mode == nir_var_shader_in);
40
41 nir_deref *tail = &deref->deref;
42
43 /* For per-vertex input arrays (i.e. geometry shader inputs), skip the
44 * outermost array index. Process the rest normally.
45 */
46 if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
47 tail = tail->child;
48 }
49
50 unsigned offset = 0;
51 while (tail->child != NULL) {
52 tail = tail->child;
53
54 if (tail->deref_type == nir_deref_type_array) {
55 nir_deref_array *deref_array = nir_deref_as_array(tail);
56 assert(deref_array->deref_array_type != nir_deref_array_type_indirect);
57
58 unsigned size = glsl_count_attribute_slots(tail->type, vs_in);
59 offset += size * deref_array->base_offset;
60
61 unsigned num_elements = glsl_type_is_array(tail->type) ?
62 glsl_get_aoa_size(tail->type) : 1;
63
64 num_elements *= glsl_type_is_matrix(glsl_without_array(tail->type)) ?
65 glsl_get_matrix_columns(glsl_without_array(tail->type)) : 1;
66
67 *element_index += num_elements * deref_array->base_offset;
68 } else if (tail->deref_type == nir_deref_type_struct) {
69 /* TODO: we could also add struct splitting support to this pass */
70 break;
71 }
72 }
73
74 return offset;
75 }
76
77 static nir_variable **
get_array_elements(struct hash_table * ht,nir_variable * var,gl_shader_stage stage)78 get_array_elements(struct hash_table *ht, nir_variable *var,
79 gl_shader_stage stage)
80 {
81 nir_variable **elements;
82 struct hash_entry *entry = _mesa_hash_table_search(ht, var);
83 if (!entry) {
84 const struct glsl_type *type = var->type;
85 if (nir_is_per_vertex_io(var, stage)) {
86 assert(glsl_type_is_array(type));
87 type = glsl_get_array_element(type);
88 }
89
90 unsigned num_elements = glsl_type_is_array(type) ?
91 glsl_get_aoa_size(type) : 1;
92
93 num_elements *= glsl_type_is_matrix(glsl_without_array(type)) ?
94 glsl_get_matrix_columns(glsl_without_array(type)) : 1;
95
96 elements = (nir_variable **) calloc(num_elements, sizeof(nir_variable *));
97 _mesa_hash_table_insert(ht, var, elements);
98 } else {
99 elements = (nir_variable **) entry->data;
100 }
101
102 return elements;
103 }
104
105 static void
create_array_deref(nir_intrinsic_instr * arr_intr,nir_intrinsic_instr * element_intr)106 create_array_deref(nir_intrinsic_instr *arr_intr,
107 nir_intrinsic_instr *element_intr)
108 {
109 assert(arr_intr->variables[0]->deref.child);
110
111 nir_deref *parent = &element_intr->variables[0]->deref;
112 nir_deref_array *darr =
113 nir_deref_as_array(arr_intr->variables[0]->deref.child);
114 nir_deref_array *ndarr = nir_deref_array_create(parent);
115
116 ndarr->deref.type = glsl_get_array_element(parent->type);
117 ndarr->deref_array_type = darr->deref_array_type;
118 ndarr->base_offset = darr->base_offset;
119
120 if (ndarr->deref_array_type == nir_deref_array_type_indirect)
121 nir_src_copy(&ndarr->indirect, &darr->indirect, parent);
122
123 element_intr->variables[0]->deref.child = &ndarr->deref;
124 }
125
126 static void
lower_array(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var,struct hash_table * varyings)127 lower_array(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var,
128 struct hash_table *varyings)
129 {
130 b->cursor = nir_before_instr(&intr->instr);
131
132 nir_variable **elements =
133 get_array_elements(varyings, var, b->shader->info.stage);
134
135 unsigned elements_index = 0;
136 unsigned io_offset = get_io_offset(b, intr->variables[0], var,
137 &elements_index);
138
139 nir_variable *element = elements[elements_index];
140 if (!element) {
141 element = nir_variable_clone(var, b->shader);
142 element->data.location = var->data.location + io_offset;
143
144 const struct glsl_type *type = glsl_without_array(element->type);
145
146 /* This pass also splits matrices so we need give them a new type. */
147 if (glsl_type_is_matrix(type)) {
148 type = glsl_vector_type(glsl_get_base_type(type),
149 glsl_get_vector_elements(type));
150 }
151
152 if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
153 type = glsl_get_array_instance(type,
154 glsl_get_length(element->type));
155 }
156
157 element->type = type;
158 elements[elements_index] = element;
159
160 nir_shader_add_variable(b->shader, element);
161 }
162
163 nir_intrinsic_instr *element_intr =
164 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
165 element_intr->num_components = intr->num_components;
166 element_intr->variables[0] = nir_deref_var_create(element_intr, element);
167
168 if (intr->intrinsic != nir_intrinsic_store_var) {
169 nir_ssa_dest_init(&element_intr->instr, &element_intr->dest,
170 intr->num_components, intr->dest.ssa.bit_size, NULL);
171
172 if (intr->intrinsic == nir_intrinsic_interp_var_at_offset ||
173 intr->intrinsic == nir_intrinsic_interp_var_at_sample) {
174 nir_src_copy(&element_intr->src[0], &intr->src[0],
175 &element_intr->instr);
176 }
177
178 nir_ssa_def_rewrite_uses(&intr->dest.ssa,
179 nir_src_for_ssa(&element_intr->dest.ssa));
180 } else {
181 nir_intrinsic_set_write_mask(element_intr,
182 nir_intrinsic_write_mask(intr));
183 nir_src_copy(&element_intr->src[0], &intr->src[0],
184 &element_intr->instr);
185 }
186
187 if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
188 create_array_deref(intr, element_intr);
189 }
190
191 nir_builder_instr_insert(b, &element_intr->instr);
192
193 /* Remove the old load intrinsic */
194 nir_instr_remove(&intr->instr);
195 }
196
197 static bool
deref_has_indirect(nir_builder * b,nir_variable * var,nir_deref_var * deref)198 deref_has_indirect(nir_builder *b, nir_variable *var, nir_deref_var *deref)
199 {
200 nir_deref *tail = &deref->deref;
201
202 if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
203 tail = tail->child;
204 }
205
206 for (tail = tail->child; tail; tail = tail->child) {
207 if (tail->deref_type != nir_deref_type_array)
208 continue;
209
210 nir_deref_array *arr = nir_deref_as_array(tail);
211 if (arr->deref_array_type == nir_deref_array_type_indirect)
212 return true;
213 }
214
215 return false;
216 }
217
218 /* Creates a mask of locations that contains arrays that are indexed via
219 * indirect indexing.
220 */
221 static void
create_indirects_mask(nir_shader * shader,uint64_t * indirects,uint64_t * patch_indirects,nir_variable_mode mode)222 create_indirects_mask(nir_shader *shader, uint64_t *indirects,
223 uint64_t *patch_indirects, nir_variable_mode mode)
224 {
225 nir_foreach_function(function, shader) {
226 if (function->impl) {
227 nir_builder b;
228 nir_builder_init(&b, function->impl);
229
230 nir_foreach_block(block, function->impl) {
231 nir_foreach_instr_safe(instr, block) {
232
233 if (instr->type != nir_instr_type_intrinsic)
234 continue;
235
236 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
237
238 if (intr->intrinsic != nir_intrinsic_load_var &&
239 intr->intrinsic != nir_intrinsic_store_var &&
240 intr->intrinsic != nir_intrinsic_interp_var_at_centroid &&
241 intr->intrinsic != nir_intrinsic_interp_var_at_sample &&
242 intr->intrinsic != nir_intrinsic_interp_var_at_offset)
243 continue;
244
245 nir_variable *var = intr->variables[0]->var;
246
247 if (var->data.mode != mode)
248 continue;
249
250 uint64_t loc_mask = ((uint64_t)1) << var->data.location;
251 if (var->data.patch) {
252 if (deref_has_indirect(&b, var, intr->variables[0]))
253 patch_indirects[var->data.location_frac] |= loc_mask;
254 } else {
255 if (deref_has_indirect(&b, var, intr->variables[0]))
256 indirects[var->data.location_frac] |= loc_mask;
257 }
258 }
259 }
260 }
261 }
262 }
263
264 static void
lower_io_arrays_to_elements(nir_shader * shader,nir_variable_mode mask,uint64_t * indirects,uint64_t * patch_indirects,struct hash_table * varyings,bool after_cross_stage_opts)265 lower_io_arrays_to_elements(nir_shader *shader, nir_variable_mode mask,
266 uint64_t *indirects, uint64_t *patch_indirects,
267 struct hash_table *varyings,
268 bool after_cross_stage_opts)
269 {
270 nir_foreach_function(function, shader) {
271 if (function->impl) {
272 nir_builder b;
273 nir_builder_init(&b, function->impl);
274
275 nir_foreach_block(block, function->impl) {
276 nir_foreach_instr_safe(instr, block) {
277 if (instr->type != nir_instr_type_intrinsic)
278 continue;
279
280 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
281
282 if (intr->intrinsic != nir_intrinsic_load_var &&
283 intr->intrinsic != nir_intrinsic_store_var &&
284 intr->intrinsic != nir_intrinsic_interp_var_at_centroid &&
285 intr->intrinsic != nir_intrinsic_interp_var_at_sample &&
286 intr->intrinsic != nir_intrinsic_interp_var_at_offset)
287 continue;
288
289 nir_variable *var = intr->variables[0]->var;
290
291 /* Skip indirects */
292 uint64_t loc_mask = ((uint64_t)1) << var->data.location;
293 if (var->data.patch) {
294 if (patch_indirects[var->data.location_frac] & loc_mask)
295 continue;
296 } else {
297 if (indirects[var->data.location_frac] & loc_mask)
298 continue;
299 }
300
301 nir_variable_mode mode = var->data.mode;
302
303 const struct glsl_type *type = var->type;
304 if (nir_is_per_vertex_io(var, b.shader->info.stage)) {
305 assert(glsl_type_is_array(type));
306 type = glsl_get_array_element(type);
307 }
308
309 /* Skip types we cannot split.
310 *
311 * TODO: Add support for struct splitting.
312 */
313 if ((!glsl_type_is_array(type) && !glsl_type_is_matrix(type))||
314 glsl_type_is_struct(glsl_without_array(type)))
315 continue;
316
317 /* Skip builtins */
318 if (!after_cross_stage_opts &&
319 var->data.location < VARYING_SLOT_VAR0 &&
320 var->data.location >= 0)
321 continue;
322
323 /* Don't bother splitting if we can't opt away any unused
324 * elements.
325 */
326 if (!after_cross_stage_opts && var->data.always_active_io)
327 continue;
328
329 switch (intr->intrinsic) {
330 case nir_intrinsic_interp_var_at_centroid:
331 case nir_intrinsic_interp_var_at_sample:
332 case nir_intrinsic_interp_var_at_offset:
333 case nir_intrinsic_load_var:
334 case nir_intrinsic_store_var:
335 if ((mask & nir_var_shader_in && mode == nir_var_shader_in) ||
336 (mask & nir_var_shader_out && mode == nir_var_shader_out))
337 lower_array(&b, intr, var, varyings);
338 break;
339 default:
340 break;
341 }
342 }
343 }
344 }
345 }
346 }
347
348 void
nir_lower_io_arrays_to_elements_no_indirects(nir_shader * shader)349 nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader)
350 {
351 struct hash_table *split_inputs =
352 _mesa_hash_table_create(NULL, _mesa_hash_pointer,
353 _mesa_key_pointer_equal);
354 struct hash_table *split_outputs =
355 _mesa_hash_table_create(NULL, _mesa_hash_pointer,
356 _mesa_key_pointer_equal);
357
358 uint64_t indirects[4] = {0}, patch_indirects[4] = {0};
359
360 lower_io_arrays_to_elements(shader, nir_var_shader_out, indirects,
361 patch_indirects, split_outputs, true);
362
363 lower_io_arrays_to_elements(shader, nir_var_shader_in, indirects,
364 patch_indirects, split_inputs, true);
365
366 /* Remove old input from the shaders inputs list */
367 struct hash_entry *entry;
368 hash_table_foreach(split_inputs, entry) {
369 nir_variable *var = (nir_variable *) entry->key;
370 exec_node_remove(&var->node);
371
372 free(entry->data);
373 }
374
375 /* Remove old output from the shaders outputs list */
376 hash_table_foreach(split_outputs, entry) {
377 nir_variable *var = (nir_variable *) entry->key;
378 exec_node_remove(&var->node);
379
380 free(entry->data);
381 }
382
383 _mesa_hash_table_destroy(split_inputs, NULL);
384 _mesa_hash_table_destroy(split_outputs, NULL);
385 }
386
387 void
nir_lower_io_arrays_to_elements(nir_shader * producer,nir_shader * consumer)388 nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader *consumer)
389 {
390 struct hash_table *split_inputs =
391 _mesa_hash_table_create(NULL, _mesa_hash_pointer,
392 _mesa_key_pointer_equal);
393 struct hash_table *split_outputs =
394 _mesa_hash_table_create(NULL, _mesa_hash_pointer,
395 _mesa_key_pointer_equal);
396
397 uint64_t indirects[4] = {0}, patch_indirects[4] = {0};
398 create_indirects_mask(producer, indirects, patch_indirects,
399 nir_var_shader_out);
400 create_indirects_mask(consumer, indirects, patch_indirects,
401 nir_var_shader_in);
402
403 lower_io_arrays_to_elements(producer, nir_var_shader_out, indirects,
404 patch_indirects, split_outputs, false);
405
406 lower_io_arrays_to_elements(consumer, nir_var_shader_in, indirects,
407 patch_indirects, split_inputs, false);
408
409 /* Remove old input from the shaders inputs list */
410 struct hash_entry *entry;
411 hash_table_foreach(split_inputs, entry) {
412 nir_variable *var = (nir_variable *) entry->key;
413 exec_node_remove(&var->node);
414
415 free(entry->data);
416 }
417
418 /* Remove old output from the shaders outputs list */
419 hash_table_foreach(split_outputs, entry) {
420 nir_variable *var = (nir_variable *) entry->key;
421 exec_node_remove(&var->node);
422
423 free(entry->data);
424 }
425
426 _mesa_hash_table_destroy(split_inputs, NULL);
427 _mesa_hash_table_destroy(split_outputs, NULL);
428 }
429