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