1 /*
2 * Copyright (c) 2017-2019 Lima Project
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, sub license,
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
12 * next paragraph) shall be included in all copies or substantial portions
13 * of the 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 NON-INFRINGEMENT. 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
21 * DEALINGS IN THE SOFTWARE.
22 *
23 */
24
25 #include "util/u_memory.h"
26 #include "util/ralloc.h"
27 #include "util/u_debug.h"
28
29 #include "tgsi/tgsi_dump.h"
30 #include "compiler/nir/nir.h"
31 #include "compiler/nir/nir_serialize.h"
32 #include "nir/tgsi_to_nir.h"
33
34 #include "pipe/p_state.h"
35
36 #include "lima_screen.h"
37 #include "lima_context.h"
38 #include "lima_job.h"
39 #include "lima_program.h"
40 #include "lima_bo.h"
41 #include "lima_disk_cache.h"
42
43 #include "ir/lima_ir.h"
44
45 static const nir_shader_compiler_options vs_nir_options = {
46 .lower_ffma16 = true,
47 .lower_ffma32 = true,
48 .lower_ffma64 = true,
49 .lower_fpow = true,
50 .lower_ffract = true,
51 .lower_fdiv = true,
52 .lower_fmod = true,
53 .lower_fsqrt = true,
54 .lower_flrp32 = true,
55 .lower_flrp64 = true,
56 /* could be implemented by clamp */
57 .lower_fsat = true,
58 .lower_bitops = true,
59 .lower_sincos = true,
60 .lower_fceil = true,
61 .lower_insert_byte = true,
62 .lower_insert_word = true,
63 .force_indirect_unrolling = nir_var_all,
64 .force_indirect_unrolling_sampler = true,
65 .max_unroll_iterations = 32,
66 .no_integers = true,
67 .support_indirect_inputs = (uint8_t)BITFIELD_MASK(PIPE_SHADER_TYPES),
68 .max_varying_expression_cost = 2,
69 };
70
71 static const nir_shader_compiler_options fs_nir_options = {
72 .lower_ffma16 = true,
73 .lower_ffma32 = true,
74 .lower_ffma64 = true,
75 .lower_fpow = true,
76 .lower_fdiv = true,
77 .lower_fmod = true,
78 .lower_flrp32 = true,
79 .lower_flrp64 = true,
80 .lower_fsign = true,
81 .lower_fdot = true,
82 .lower_fdph = true,
83 .lower_insert_byte = true,
84 .lower_insert_word = true,
85 .lower_bitops = true,
86 .lower_vector_cmp = true,
87 .force_indirect_unrolling = (nir_var_shader_out | nir_var_function_temp),
88 .force_indirect_unrolling_sampler = true,
89 .max_unroll_iterations = 32,
90 .no_integers = true,
91 .support_indirect_inputs = (uint8_t)BITFIELD_MASK(PIPE_SHADER_TYPES),
92 .max_varying_expression_cost = 2,
93 };
94
95 const void *
lima_program_get_compiler_options(enum pipe_shader_type shader)96 lima_program_get_compiler_options(enum pipe_shader_type shader)
97 {
98 switch (shader) {
99 case PIPE_SHADER_VERTEX:
100 return &vs_nir_options;
101 case PIPE_SHADER_FRAGMENT:
102 return &fs_nir_options;
103 default:
104 return NULL;
105 }
106 }
107
108 static int
type_size(const struct glsl_type * type,bool bindless)109 type_size(const struct glsl_type *type, bool bindless)
110 {
111 return glsl_count_attribute_slots(type, false);
112 }
113
114 static void
lima_program_optimize_vs_nir(struct nir_shader * s)115 lima_program_optimize_vs_nir(struct nir_shader *s)
116 {
117 bool progress;
118
119 NIR_PASS_V(s, nir_lower_viewport_transform);
120 NIR_PASS_V(s, nir_lower_point_size, 1.0f, 100.0f);
121 NIR_PASS_V(s, nir_lower_io,
122 nir_var_shader_in | nir_var_shader_out, type_size, 0);
123 NIR_PASS_V(s, nir_lower_load_const_to_scalar);
124 NIR_PASS_V(s, lima_nir_lower_uniform_to_scalar);
125 NIR_PASS_V(s, nir_lower_io_to_scalar,
126 nir_var_shader_in|nir_var_shader_out, NULL, NULL);
127
128 do {
129 progress = false;
130
131 NIR_PASS_V(s, nir_lower_vars_to_ssa);
132 NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL, NULL);
133 NIR_PASS(progress, s, nir_lower_phis_to_scalar, false);
134 NIR_PASS(progress, s, nir_copy_prop);
135 NIR_PASS(progress, s, nir_opt_remove_phis);
136 NIR_PASS(progress, s, nir_opt_dce);
137 NIR_PASS(progress, s, nir_opt_dead_cf);
138 NIR_PASS(progress, s, nir_opt_cse);
139 NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
140 NIR_PASS(progress, s, nir_opt_algebraic);
141 NIR_PASS(progress, s, lima_nir_lower_ftrunc);
142 NIR_PASS(progress, s, nir_opt_constant_folding);
143 NIR_PASS(progress, s, nir_opt_undef);
144 NIR_PASS(progress, s, nir_lower_undef_to_zero);
145 NIR_PASS(progress, s, nir_opt_loop_unroll);
146 NIR_PASS(progress, s, nir_lower_undef_to_zero);
147 } while (progress);
148
149 NIR_PASS_V(s, nir_lower_int_to_float);
150 /* int_to_float pass generates ftrunc, so lower it */
151 NIR_PASS(progress, s, lima_nir_lower_ftrunc);
152 NIR_PASS_V(s, nir_lower_bool_to_float, true);
153
154 NIR_PASS_V(s, nir_copy_prop);
155 NIR_PASS_V(s, nir_opt_dce);
156 NIR_PASS_V(s, lima_nir_split_loads);
157 NIR_PASS_V(s, nir_convert_from_ssa, true);
158 NIR_PASS_V(s, nir_opt_dce);
159 NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
160 nir_sweep(s);
161 }
162
163 static bool
lima_alu_to_scalar_filter_cb(const nir_instr * instr,const void * data)164 lima_alu_to_scalar_filter_cb(const nir_instr *instr, const void *data)
165 {
166 if (instr->type != nir_instr_type_alu)
167 return false;
168
169 nir_alu_instr *alu = nir_instr_as_alu(instr);
170 switch (alu->op) {
171 case nir_op_frcp:
172 /* nir_op_idiv is lowered to frcp by lower_int_to_floats which
173 * will be run later, so lower idiv here
174 */
175 case nir_op_idiv:
176 case nir_op_frsq:
177 case nir_op_flog2:
178 case nir_op_fexp2:
179 case nir_op_fsqrt:
180 case nir_op_fsin:
181 case nir_op_fcos:
182 return true;
183 default:
184 break;
185 }
186
187 /* nir vec4 fcsel assumes that each component of the condition will be
188 * used to select the same component from the two options, but Utgard PP
189 * has only 1 component condition. If all condition components are not the
190 * same we need to lower it to scalar.
191 */
192 switch (alu->op) {
193 case nir_op_bcsel:
194 case nir_op_fcsel:
195 break;
196 default:
197 return false;
198 }
199
200 int num_components = alu->def.num_components;
201
202 uint8_t swizzle = alu->src[0].swizzle[0];
203
204 for (int i = 1; i < num_components; i++)
205 if (alu->src[0].swizzle[i] != swizzle)
206 return true;
207
208 return false;
209 }
210
211 static bool
lima_vec_to_regs_filter_cb(const nir_instr * instr,unsigned writemask,const void * data)212 lima_vec_to_regs_filter_cb(const nir_instr *instr, unsigned writemask,
213 const void *data)
214 {
215 assert(writemask > 0);
216 if (util_bitcount(writemask) == 1)
217 return true;
218
219 return !lima_alu_to_scalar_filter_cb(instr, data);
220 }
221
222 static void
lima_program_optimize_fs_nir(struct nir_shader * s,struct nir_lower_tex_options * tex_options)223 lima_program_optimize_fs_nir(struct nir_shader *s,
224 struct nir_lower_tex_options *tex_options)
225 {
226 bool progress;
227
228 NIR_PASS_V(s, nir_lower_fragcoord_wtrans);
229 NIR_PASS_V(s, nir_lower_io,
230 nir_var_shader_in | nir_var_shader_out, type_size, 0);
231 NIR_PASS_V(s, nir_lower_tex, tex_options);
232 NIR_PASS_V(s, lima_nir_lower_txp);
233
234 do {
235 progress = false;
236 NIR_PASS(progress, s, nir_opt_vectorize, NULL, NULL);
237 } while (progress);
238
239 do {
240 progress = false;
241
242 NIR_PASS_V(s, nir_lower_vars_to_ssa);
243 NIR_PASS(progress, s, nir_lower_alu_to_scalar, lima_alu_to_scalar_filter_cb, NULL);
244 NIR_PASS(progress, s, nir_copy_prop);
245 NIR_PASS(progress, s, nir_opt_remove_phis);
246 NIR_PASS(progress, s, nir_opt_dce);
247 NIR_PASS(progress, s, nir_opt_dead_cf);
248 NIR_PASS(progress, s, nir_opt_cse);
249 NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
250 NIR_PASS(progress, s, nir_opt_algebraic);
251 NIR_PASS(progress, s, nir_opt_constant_folding);
252 NIR_PASS(progress, s, nir_opt_undef);
253 NIR_PASS(progress, s, nir_opt_loop_unroll);
254 NIR_PASS(progress, s, lima_nir_split_load_input);
255 } while (progress);
256
257 NIR_PASS_V(s, nir_lower_int_to_float);
258 NIR_PASS_V(s, nir_lower_bool_to_float, true);
259
260 /* Some ops must be lowered after being converted from int ops,
261 * so re-run nir_opt_algebraic after int lowering. */
262 do {
263 progress = false;
264 NIR_PASS(progress, s, nir_opt_algebraic);
265 } while (progress);
266
267 /* Must be run after optimization loop */
268 NIR_PASS_V(s, lima_nir_scale_trig);
269 NIR_PASS_V(s, lima_nir_ppir_algebraic_late);
270
271 NIR_PASS_V(s, nir_copy_prop);
272 NIR_PASS_V(s, nir_opt_dce);
273
274 NIR_PASS_V(s, nir_convert_from_ssa, true);
275 NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
276
277 NIR_PASS_V(s, nir_move_vec_src_uses_to_dest, false);
278 NIR_PASS_V(s, nir_lower_vec_to_regs, lima_vec_to_regs_filter_cb, NULL);
279
280 NIR_PASS_V(s, nir_opt_dce); /* clean up any new dead code from vec to movs */
281
282 NIR_PASS_V(s, lima_nir_duplicate_load_uniforms);
283 NIR_PASS_V(s, lima_nir_duplicate_load_inputs);
284 NIR_PASS_V(s, lima_nir_duplicate_load_consts);
285
286 NIR_PASS_V(s, nir_trivialize_registers);
287
288 nir_sweep(s);
289 }
290
291 static bool
lima_fs_compile_shader(struct lima_context * ctx,struct lima_fs_key * key,struct lima_fs_uncompiled_shader * ufs,struct lima_fs_compiled_shader * fs)292 lima_fs_compile_shader(struct lima_context *ctx,
293 struct lima_fs_key *key,
294 struct lima_fs_uncompiled_shader *ufs,
295 struct lima_fs_compiled_shader *fs)
296 {
297 struct lima_screen *screen = lima_screen(ctx->base.screen);
298 nir_shader *nir = nir_shader_clone(fs, ufs->base.ir.nir);
299
300 struct nir_lower_tex_options tex_options = {
301 .swizzle_result = ~0u,
302 .lower_invalid_implicit_lod = true,
303 };
304
305 for (int i = 0; i < ARRAY_SIZE(key->tex); i++) {
306 for (int j = 0; j < 4; j++)
307 tex_options.swizzles[i][j] = key->tex[i].swizzle[j];
308 }
309
310 lima_program_optimize_fs_nir(nir, &tex_options);
311
312 if (lima_debug & LIMA_DEBUG_PP)
313 nir_print_shader(nir, stdout);
314
315 if (!ppir_compile_nir(fs, nir, screen->pp_ra, &ctx->base.debug)) {
316 ralloc_free(nir);
317 return false;
318 }
319
320 fs->state.uses_discard = nir->info.fs.uses_discard;
321 ralloc_free(nir);
322
323 return true;
324 }
325
326 static bool
lima_fs_upload_shader(struct lima_context * ctx,struct lima_fs_compiled_shader * fs)327 lima_fs_upload_shader(struct lima_context *ctx,
328 struct lima_fs_compiled_shader *fs)
329 {
330 static const uint32_t pp_clear_program[] = {
331 PP_CLEAR_PROGRAM
332 };
333 int shader_size = sizeof(pp_clear_program);
334 void *shader = (void *)pp_clear_program;
335 struct lima_screen *screen = lima_screen(ctx->base.screen);
336
337 if (fs->state.shader_size) {
338 shader_size = fs->state.shader_size;
339 shader = fs->shader;
340 }
341
342 fs->bo = lima_bo_create(screen, shader_size, 0);
343 if (!fs->bo) {
344 fprintf(stderr, "lima: create fs shader bo fail\n");
345 return false;
346 }
347
348 memcpy(lima_bo_map(fs->bo), shader, shader_size);
349
350 return true;
351 }
352
353 static struct lima_fs_compiled_shader *
lima_get_compiled_fs(struct lima_context * ctx,struct lima_fs_uncompiled_shader * ufs,struct lima_fs_key * key)354 lima_get_compiled_fs(struct lima_context *ctx,
355 struct lima_fs_uncompiled_shader *ufs,
356 struct lima_fs_key *key)
357 {
358 struct lima_screen *screen = lima_screen(ctx->base.screen);
359 struct hash_table *ht;
360 uint32_t key_size;
361
362 ht = ctx->fs_cache;
363 key_size = sizeof(struct lima_fs_key);
364
365 struct hash_entry *entry = _mesa_hash_table_search(ht, key);
366 if (entry)
367 return entry->data;
368
369 /* Not on memory cache, try disk cache */
370 struct lima_fs_compiled_shader *fs =
371 lima_fs_disk_cache_retrieve(screen->disk_cache, key);
372
373 if (!fs) {
374 /* Not on disk cache, compile and insert into disk cache*/
375 fs = rzalloc(NULL, struct lima_fs_compiled_shader);
376 if (!fs)
377 return NULL;
378
379 if (!lima_fs_compile_shader(ctx, key, ufs, fs))
380 goto err;
381
382 lima_fs_disk_cache_store(screen->disk_cache, key, fs);
383 }
384
385 if (!lima_fs_upload_shader(ctx, fs))
386 goto err;
387
388 ralloc_free(fs->shader);
389 fs->shader = NULL;
390
391 /* Insert into memory cache */
392 struct lima_key *dup_key;
393 dup_key = rzalloc_size(fs, key_size);
394 memcpy(dup_key, key, key_size);
395 _mesa_hash_table_insert(ht, dup_key, fs);
396
397 return fs;
398
399 err:
400 ralloc_free(fs);
401 return NULL;
402 }
403
404 static void *
lima_create_fs_state(struct pipe_context * pctx,const struct pipe_shader_state * cso)405 lima_create_fs_state(struct pipe_context *pctx,
406 const struct pipe_shader_state *cso)
407 {
408 struct lima_context *ctx = lima_context(pctx);
409 struct lima_fs_uncompiled_shader *so = rzalloc(NULL, struct lima_fs_uncompiled_shader);
410
411 if (!so)
412 return NULL;
413
414 nir_shader *nir;
415 if (cso->type == PIPE_SHADER_IR_NIR)
416 /* The backend takes ownership of the NIR shader on state
417 * creation. */
418 nir = cso->ir.nir;
419 else {
420 assert(cso->type == PIPE_SHADER_IR_TGSI);
421
422 nir = tgsi_to_nir(cso->tokens, pctx->screen, false);
423 }
424
425 so->base.type = PIPE_SHADER_IR_NIR;
426 so->base.ir.nir = nir;
427
428 /* Serialize the NIR to a binary blob that we can hash for the disk
429 * cache. Drop unnecessary information (like variable names)
430 * so the serialized NIR is smaller, and also to let us detect more
431 * isomorphic shaders when hashing, increasing cache hits.
432 */
433 struct blob blob;
434 blob_init(&blob);
435 nir_serialize(&blob, nir, true);
436 _mesa_sha1_compute(blob.data, blob.size, so->nir_sha1);
437 blob_finish(&blob);
438
439 if (lima_debug & LIMA_DEBUG_PRECOMPILE) {
440 /* Trigger initial compilation with default settings */
441 struct lima_fs_key key;
442 memset(&key, 0, sizeof(key));
443 memcpy(key.nir_sha1, so->nir_sha1, sizeof(so->nir_sha1));
444 for (int i = 0; i < ARRAY_SIZE(key.tex); i++) {
445 for (int j = 0; j < 4; j++)
446 key.tex[i].swizzle[j] = j;
447 }
448 lima_get_compiled_fs(ctx, so, &key);
449 }
450
451 return so;
452 }
453
454 static void
lima_bind_fs_state(struct pipe_context * pctx,void * hwcso)455 lima_bind_fs_state(struct pipe_context *pctx, void *hwcso)
456 {
457 struct lima_context *ctx = lima_context(pctx);
458
459 ctx->uncomp_fs = hwcso;
460 ctx->dirty |= LIMA_CONTEXT_DIRTY_UNCOMPILED_FS;
461 }
462
463 static void
lima_delete_fs_state(struct pipe_context * pctx,void * hwcso)464 lima_delete_fs_state(struct pipe_context *pctx, void *hwcso)
465 {
466 struct lima_context *ctx = lima_context(pctx);
467 struct lima_fs_uncompiled_shader *so = hwcso;
468
469 hash_table_foreach(ctx->fs_cache, entry) {
470 const struct lima_fs_key *key = entry->key;
471 if (!memcmp(key->nir_sha1, so->nir_sha1, sizeof(so->nir_sha1))) {
472 struct lima_fs_compiled_shader *fs = entry->data;
473 _mesa_hash_table_remove(ctx->fs_cache, entry);
474 if (fs->bo)
475 lima_bo_unreference(fs->bo);
476
477 if (fs == ctx->fs)
478 ctx->fs = NULL;
479
480 ralloc_free(fs);
481 }
482 }
483
484 ralloc_free(so->base.ir.nir);
485 ralloc_free(so);
486 }
487
488 static bool
lima_vs_compile_shader(struct lima_context * ctx,struct lima_vs_key * key,struct lima_vs_uncompiled_shader * uvs,struct lima_vs_compiled_shader * vs)489 lima_vs_compile_shader(struct lima_context *ctx,
490 struct lima_vs_key *key,
491 struct lima_vs_uncompiled_shader *uvs,
492 struct lima_vs_compiled_shader *vs)
493 {
494 nir_shader *nir = nir_shader_clone(vs, uvs->base.ir.nir);
495
496 lima_program_optimize_vs_nir(nir);
497
498 if (lima_debug & LIMA_DEBUG_GP)
499 nir_print_shader(nir, stdout);
500
501 if (!gpir_compile_nir(vs, nir, &ctx->base.debug)) {
502 ralloc_free(nir);
503 return false;
504 }
505
506 ralloc_free(nir);
507
508 return true;
509 }
510
511 static bool
lima_vs_upload_shader(struct lima_context * ctx,struct lima_vs_compiled_shader * vs)512 lima_vs_upload_shader(struct lima_context *ctx,
513 struct lima_vs_compiled_shader *vs)
514 {
515 struct lima_screen *screen = lima_screen(ctx->base.screen);
516 vs->bo = lima_bo_create(screen, vs->state.shader_size, 0);
517 if (!vs->bo) {
518 fprintf(stderr, "lima: create vs shader bo fail\n");
519 return false;
520 }
521
522 memcpy(lima_bo_map(vs->bo), vs->shader, vs->state.shader_size);
523
524 return true;
525 }
526
527 static struct lima_vs_compiled_shader *
lima_get_compiled_vs(struct lima_context * ctx,struct lima_vs_uncompiled_shader * uvs,struct lima_vs_key * key)528 lima_get_compiled_vs(struct lima_context *ctx,
529 struct lima_vs_uncompiled_shader *uvs,
530 struct lima_vs_key *key)
531 {
532 struct lima_screen *screen = lima_screen(ctx->base.screen);
533 struct hash_table *ht;
534 uint32_t key_size;
535
536 ht = ctx->vs_cache;
537 key_size = sizeof(struct lima_vs_key);
538
539 struct hash_entry *entry = _mesa_hash_table_search(ht, key);
540 if (entry)
541 return entry->data;
542
543 /* Not on memory cache, try disk cache */
544 struct lima_vs_compiled_shader *vs =
545 lima_vs_disk_cache_retrieve(screen->disk_cache, key);
546
547 if (!vs) {
548 /* Not on disk cache, compile and insert into disk cache */
549 vs = rzalloc(NULL, struct lima_vs_compiled_shader);
550 if (!vs)
551 return NULL;
552 if (!lima_vs_compile_shader(ctx, key, uvs, vs))
553 goto err;
554
555 lima_vs_disk_cache_store(screen->disk_cache, key, vs);
556 }
557
558 if (!lima_vs_upload_shader(ctx, vs))
559 goto err;
560
561 ralloc_free(vs->shader);
562 vs->shader = NULL;
563
564 struct lima_key *dup_key;
565 dup_key = rzalloc_size(vs, key_size);
566 memcpy(dup_key, key, key_size);
567 _mesa_hash_table_insert(ht, dup_key, vs);
568
569 return vs;
570
571 err:
572 ralloc_free(vs);
573 return NULL;
574 }
575
576 bool
lima_update_vs_state(struct lima_context * ctx)577 lima_update_vs_state(struct lima_context *ctx)
578 {
579 if (!(ctx->dirty & LIMA_CONTEXT_DIRTY_UNCOMPILED_VS)) {
580 return true;
581 }
582
583 struct lima_vs_key local_key;
584 struct lima_vs_key *key = &local_key;
585 memset(key, 0, sizeof(*key));
586 memcpy(key->nir_sha1, ctx->uncomp_vs->nir_sha1,
587 sizeof(ctx->uncomp_vs->nir_sha1));
588
589 struct lima_vs_compiled_shader *old_vs = ctx->vs;
590 struct lima_vs_compiled_shader *vs = lima_get_compiled_vs(ctx,
591 ctx->uncomp_vs,
592 key);
593 if (!vs)
594 return false;
595
596 ctx->vs = vs;
597
598 if (ctx->vs != old_vs)
599 ctx->dirty |= LIMA_CONTEXT_DIRTY_COMPILED_VS;
600
601 return true;
602 }
603
604 bool
lima_update_fs_state(struct lima_context * ctx)605 lima_update_fs_state(struct lima_context *ctx)
606 {
607 if (!(ctx->dirty & (LIMA_CONTEXT_DIRTY_UNCOMPILED_FS |
608 LIMA_CONTEXT_DIRTY_TEXTURES))) {
609 return true;
610 }
611
612 struct lima_texture_stateobj *lima_tex = &ctx->tex_stateobj;
613 struct lima_fs_key local_key;
614 struct lima_fs_key *key = &local_key;
615 memset(key, 0, sizeof(*key));
616 memcpy(key->nir_sha1, ctx->uncomp_fs->nir_sha1,
617 sizeof(ctx->uncomp_fs->nir_sha1));
618
619 uint8_t identity[4] = { PIPE_SWIZZLE_X, PIPE_SWIZZLE_Y,
620 PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W };
621 for (int i = 0; i < lima_tex->num_textures; i++) {
622 struct lima_sampler_view *sampler = lima_sampler_view(lima_tex->textures[i]);
623 if (!sampler) {
624 memcpy(key->tex[i].swizzle, identity, 4);
625 continue;
626 }
627 for (int j = 0; j < 4; j++)
628 key->tex[i].swizzle[j] = sampler->swizzle[j];
629 }
630
631 /* Fill rest with identity swizzle */
632 for (int i = lima_tex->num_textures; i < ARRAY_SIZE(key->tex); i++)
633 memcpy(key->tex[i].swizzle, identity, 4);
634
635 struct lima_fs_compiled_shader *old_fs = ctx->fs;
636
637 struct lima_fs_compiled_shader *fs = lima_get_compiled_fs(ctx,
638 ctx->uncomp_fs,
639 key);
640 if (!fs)
641 return false;
642
643 ctx->fs = fs;
644
645 if (ctx->fs != old_fs)
646 ctx->dirty |= LIMA_CONTEXT_DIRTY_COMPILED_FS;
647
648 return true;
649 }
650
651 static void *
lima_create_vs_state(struct pipe_context * pctx,const struct pipe_shader_state * cso)652 lima_create_vs_state(struct pipe_context *pctx,
653 const struct pipe_shader_state *cso)
654 {
655 struct lima_context *ctx = lima_context(pctx);
656 struct lima_vs_uncompiled_shader *so = rzalloc(NULL, struct lima_vs_uncompiled_shader);
657
658 if (!so)
659 return NULL;
660
661 nir_shader *nir;
662 if (cso->type == PIPE_SHADER_IR_NIR)
663 /* The backend takes ownership of the NIR shader on state
664 * creation. */
665 nir = cso->ir.nir;
666 else {
667 assert(cso->type == PIPE_SHADER_IR_TGSI);
668
669 nir = tgsi_to_nir(cso->tokens, pctx->screen, false);
670 }
671
672 so->base.type = PIPE_SHADER_IR_NIR;
673 so->base.ir.nir = nir;
674
675 /* Serialize the NIR to a binary blob that we can hash for the disk
676 * cache. Drop unnecessary information (like variable names)
677 * so the serialized NIR is smaller, and also to let us detect more
678 * isomorphic shaders when hashing, increasing cache hits.
679 */
680 struct blob blob;
681 blob_init(&blob);
682 nir_serialize(&blob, nir, true);
683 _mesa_sha1_compute(blob.data, blob.size, so->nir_sha1);
684 blob_finish(&blob);
685
686 if (lima_debug & LIMA_DEBUG_PRECOMPILE) {
687 /* Trigger initial compilation with default settings */
688 struct lima_vs_key key;
689 memset(&key, 0, sizeof(key));
690 memcpy(key.nir_sha1, so->nir_sha1, sizeof(so->nir_sha1));
691 lima_get_compiled_vs(ctx, so, &key);
692 }
693
694 return so;
695 }
696
697 static void
lima_bind_vs_state(struct pipe_context * pctx,void * hwcso)698 lima_bind_vs_state(struct pipe_context *pctx, void *hwcso)
699 {
700 struct lima_context *ctx = lima_context(pctx);
701
702 ctx->uncomp_vs = hwcso;
703 ctx->dirty |= LIMA_CONTEXT_DIRTY_UNCOMPILED_VS;
704 }
705
706 static void
lima_delete_vs_state(struct pipe_context * pctx,void * hwcso)707 lima_delete_vs_state(struct pipe_context *pctx, void *hwcso)
708 {
709 struct lima_context *ctx = lima_context(pctx);
710 struct lima_vs_uncompiled_shader *so = hwcso;
711
712 hash_table_foreach(ctx->vs_cache, entry) {
713 const struct lima_vs_key *key = entry->key;
714 if (!memcmp(key->nir_sha1, so->nir_sha1, sizeof(so->nir_sha1))) {
715 struct lima_vs_compiled_shader *vs = entry->data;
716 _mesa_hash_table_remove(ctx->vs_cache, entry);
717 if (vs->bo)
718 lima_bo_unreference(vs->bo);
719
720 if (vs == ctx->vs)
721 ctx->vs = NULL;
722
723 ralloc_free(vs);
724 }
725 }
726
727 ralloc_free(so->base.ir.nir);
728 ralloc_free(so);
729 }
730
731 static uint32_t
lima_fs_cache_hash(const void * key)732 lima_fs_cache_hash(const void *key)
733 {
734 return _mesa_hash_data(key, sizeof(struct lima_fs_key));
735 }
736
737 static uint32_t
lima_vs_cache_hash(const void * key)738 lima_vs_cache_hash(const void *key)
739 {
740 return _mesa_hash_data(key, sizeof(struct lima_vs_key));
741 }
742
743 static bool
lima_fs_cache_compare(const void * key1,const void * key2)744 lima_fs_cache_compare(const void *key1, const void *key2)
745 {
746 return memcmp(key1, key2, sizeof(struct lima_fs_key)) == 0;
747 }
748
749 static bool
lima_vs_cache_compare(const void * key1,const void * key2)750 lima_vs_cache_compare(const void *key1, const void *key2)
751 {
752 return memcmp(key1, key2, sizeof(struct lima_vs_key)) == 0;
753 }
754
755 void
lima_program_init(struct lima_context * ctx)756 lima_program_init(struct lima_context *ctx)
757 {
758 ctx->base.create_fs_state = lima_create_fs_state;
759 ctx->base.bind_fs_state = lima_bind_fs_state;
760 ctx->base.delete_fs_state = lima_delete_fs_state;
761
762 ctx->base.create_vs_state = lima_create_vs_state;
763 ctx->base.bind_vs_state = lima_bind_vs_state;
764 ctx->base.delete_vs_state = lima_delete_vs_state;
765
766 ctx->fs_cache = _mesa_hash_table_create(ctx, lima_fs_cache_hash,
767 lima_fs_cache_compare);
768 ctx->vs_cache = _mesa_hash_table_create(ctx, lima_vs_cache_hash,
769 lima_vs_cache_compare);
770 }
771
772 void
lima_program_fini(struct lima_context * ctx)773 lima_program_fini(struct lima_context *ctx)
774 {
775 hash_table_foreach(ctx->vs_cache, entry) {
776 struct lima_vs_compiled_shader *vs = entry->data;
777 if (vs->bo)
778 lima_bo_unreference(vs->bo);
779 ralloc_free(vs);
780 _mesa_hash_table_remove(ctx->vs_cache, entry);
781 }
782
783 hash_table_foreach(ctx->fs_cache, entry) {
784 struct lima_fs_compiled_shader *fs = entry->data;
785 if (fs->bo)
786 lima_bo_unreference(fs->bo);
787 ralloc_free(fs);
788 _mesa_hash_table_remove(ctx->fs_cache, entry);
789 }
790 }
791