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, 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 shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23
24#include <stdlib.h>
25#include <string.h>
26
27#include "xf86drm.h"
28#include "drm-uapi/lima_drm.h"
29
30#include "util/u_math.h"
31#include "util/ralloc.h"
32#include "util/os_time.h"
33#include "util/hash_table.h"
34#include "util/format/u_format.h"
35#include "util/u_upload_mgr.h"
36#include "util/u_inlines.h"
37#include "util/u_framebuffer.h"
38
39#include "lima_screen.h"
40#include "lima_context.h"
41#include "lima_job.h"
42#include "lima_bo.h"
43#include "lima_util.h"
44#include "lima_format.h"
45#include "lima_resource.h"
46#include "lima_texture.h"
47#include "lima_fence.h"
48#include "lima_gpu.h"
49#include "lima_blit.h"
50
51#define VOID2U64(x) ((uint64_t)(unsigned long)(x))
52
53static void
54lima_get_fb_info(struct lima_job *job)
55{
56   struct lima_context *ctx = job->ctx;
57   struct lima_job_fb_info *fb = &job->fb;
58   struct lima_surface *surf = lima_surface(job->key.cbuf);
59
60   if (!surf)
61      surf = lima_surface(job->key.zsbuf);
62
63   if (!surf) {
64      /* We don't have neither cbuf nor zsbuf, use dimensions from ctx */
65      fb->width = ctx->framebuffer.base.width;
66      fb->height =  ctx->framebuffer.base.height;
67   } else {
68      fb->width = surf->base.width;
69      fb->height = surf->base.height;
70   }
71
72   int width = align(fb->width, 16) >> 4;
73   int height = align(fb->height, 16) >> 4;
74
75   struct lima_screen *screen = lima_screen(ctx->base.screen);
76
77   fb->tiled_w = width;
78   fb->tiled_h = height;
79
80   fb->shift_h = 0;
81   fb->shift_w = 0;
82
83   int limit = screen->plb_max_blk;
84   while ((width * height) > limit) {
85      if (width >= height) {
86         width = (width + 1) >> 1;
87         fb->shift_w++;
88      } else {
89         height = (height + 1) >> 1;
90         fb->shift_h++;
91      }
92   }
93
94   fb->block_w = width;
95   fb->block_h = height;
96
97   fb->shift_min = MIN3(fb->shift_w, fb->shift_h, 2);
98}
99
100static struct lima_job *
101lima_job_create(struct lima_context *ctx,
102                struct pipe_surface *cbuf,
103                struct pipe_surface *zsbuf)
104{
105   struct lima_job *s;
106
107   s = rzalloc(ctx, struct lima_job);
108   if (!s)
109      return NULL;
110
111   s->fd = lima_screen(ctx->base.screen)->fd;
112   s->ctx = ctx;
113
114   s->damage_rect.minx = s->damage_rect.miny = 0xffff;
115   s->damage_rect.maxx = s->damage_rect.maxy = 0;
116   s->draws = 0;
117
118   s->clear.depth = 0x00ffffff;
119
120   for (int i = 0; i < 2; i++) {
121      util_dynarray_init(s->gem_bos + i, s);
122      util_dynarray_init(s->bos + i, s);
123   }
124
125   util_dynarray_init(&s->vs_cmd_array, s);
126   util_dynarray_init(&s->plbu_cmd_array, s);
127   util_dynarray_init(&s->plbu_cmd_head, s);
128
129   pipe_surface_reference(&s->key.cbuf, cbuf);
130   pipe_surface_reference(&s->key.zsbuf, zsbuf);
131
132   lima_get_fb_info(s);
133
134   s->dump = lima_dump_create();
135
136   return s;
137}
138
139static void
140lima_job_free(struct lima_job *job)
141{
142   struct lima_context *ctx = job->ctx;
143
144   _mesa_hash_table_remove_key(ctx->jobs, &job->key);
145
146   if (job->key.cbuf && (job->resolve & PIPE_CLEAR_COLOR0))
147      _mesa_hash_table_remove_key(ctx->write_jobs, job->key.cbuf->texture);
148   if (job->key.zsbuf && (job->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)))
149      _mesa_hash_table_remove_key(ctx->write_jobs, job->key.zsbuf->texture);
150
151   pipe_surface_reference(&job->key.cbuf, NULL);
152   pipe_surface_reference(&job->key.zsbuf, NULL);
153
154   lima_dump_free(job->dump);
155   job->dump = NULL;
156
157   /* TODO: do we need a cache for job? */
158   ralloc_free(job);
159}
160
161struct lima_job *
162lima_job_get_with_fb(struct lima_context *ctx,
163                      struct pipe_surface *cbuf,
164                      struct pipe_surface *zsbuf)
165{
166   struct lima_job_key local_key = {
167      .cbuf = cbuf,
168      .zsbuf = zsbuf,
169   };
170
171   struct hash_entry *entry = _mesa_hash_table_search(ctx->jobs, &local_key);
172   if (entry)
173      return entry->data;
174
175   struct lima_job *job = lima_job_create(ctx, cbuf, zsbuf);
176   if (!job)
177      return NULL;
178
179   _mesa_hash_table_insert(ctx->jobs, &job->key, job);
180
181   return job;
182}
183
184static struct lima_job *
185_lima_job_get(struct lima_context *ctx)
186{
187   struct lima_context_framebuffer *fb = &ctx->framebuffer;
188
189   return lima_job_get_with_fb(ctx, fb->base.cbufs[0], fb->base.zsbuf);
190}
191
192/*
193 * Note: this function can only be called in draw code path,
194 * must not exist in flush code path.
195 */
196struct lima_job *
197lima_job_get(struct lima_context *ctx)
198{
199   if (ctx->job)
200      return ctx->job;
201
202   ctx->job = _lima_job_get(ctx);
203   return ctx->job;
204}
205
206bool lima_job_add_bo(struct lima_job *job, int pipe,
207                     struct lima_bo *bo, uint32_t flags)
208{
209   util_dynarray_foreach(job->gem_bos + pipe, struct drm_lima_gem_submit_bo, gem_bo) {
210      if (bo->handle == gem_bo->handle) {
211         gem_bo->flags |= flags;
212         return true;
213      }
214   }
215
216   struct drm_lima_gem_submit_bo *job_bo =
217      util_dynarray_grow(job->gem_bos + pipe, struct drm_lima_gem_submit_bo, 1);
218   job_bo->handle = bo->handle;
219   job_bo->flags = flags;
220
221   struct lima_bo **jbo = util_dynarray_grow(job->bos + pipe, struct lima_bo *, 1);
222   *jbo = bo;
223
224   /* prevent bo from being freed when job start */
225   lima_bo_reference(bo);
226
227   return true;
228}
229
230static bool
231lima_job_start(struct lima_job *job, int pipe, void *frame, uint32_t size)
232{
233   struct lima_context *ctx = job->ctx;
234   struct drm_lima_gem_submit req = {
235      .ctx = ctx->id,
236      .pipe = pipe,
237      .nr_bos = job->gem_bos[pipe].size / sizeof(struct drm_lima_gem_submit_bo),
238      .bos = VOID2U64(util_dynarray_begin(job->gem_bos + pipe)),
239      .frame = VOID2U64(frame),
240      .frame_size = size,
241      .out_sync = ctx->out_sync[pipe],
242   };
243
244   if (ctx->in_sync_fd >= 0) {
245      int err = drmSyncobjImportSyncFile(job->fd, ctx->in_sync[pipe],
246                                         ctx->in_sync_fd);
247      if (err)
248         return false;
249
250      req.in_sync[0] = ctx->in_sync[pipe];
251      close(ctx->in_sync_fd);
252      ctx->in_sync_fd = -1;
253   }
254
255   bool ret = drmIoctl(job->fd, DRM_IOCTL_LIMA_GEM_SUBMIT, &req) == 0;
256
257   util_dynarray_foreach(job->bos + pipe, struct lima_bo *, bo) {
258      lima_bo_unreference(*bo);
259   }
260
261   return ret;
262}
263
264static bool
265lima_job_wait(struct lima_job *job, int pipe, uint64_t timeout_ns)
266{
267   int64_t abs_timeout = os_time_get_absolute_timeout(timeout_ns);
268   if (abs_timeout == OS_TIMEOUT_INFINITE)
269      abs_timeout = INT64_MAX;
270
271   struct lima_context *ctx = job->ctx;
272   return !drmSyncobjWait(job->fd, ctx->out_sync + pipe, 1, abs_timeout, 0, NULL);
273}
274
275static bool
276lima_job_has_bo(struct lima_job *job, struct lima_bo *bo, bool all)
277{
278   for (int i = 0; i < 2; i++) {
279      util_dynarray_foreach(job->gem_bos + i, struct drm_lima_gem_submit_bo, gem_bo) {
280         if (bo->handle == gem_bo->handle) {
281            if (all || gem_bo->flags & LIMA_SUBMIT_BO_WRITE)
282               return true;
283            else
284               break;
285         }
286      }
287   }
288
289   return false;
290}
291
292void *
293lima_job_create_stream_bo(struct lima_job *job, int pipe,
294                          unsigned size, uint32_t *va)
295{
296   struct lima_context *ctx = job->ctx;
297
298   void *cpu;
299   unsigned offset;
300   struct pipe_resource *pres = NULL;
301   u_upload_alloc(ctx->uploader, 0, size, 0x40, &offset, &pres, &cpu);
302
303   struct lima_resource *res = lima_resource(pres);
304   *va = res->bo->va + offset;
305
306   lima_job_add_bo(job, pipe, res->bo, LIMA_SUBMIT_BO_READ);
307
308   pipe_resource_reference(&pres, NULL);
309
310   return cpu;
311}
312
313static inline struct lima_damage_region *
314lima_job_get_damage(struct lima_job *job)
315{
316   if (!(job->key.cbuf && (job->resolve & PIPE_CLEAR_COLOR0)))
317      return NULL;
318
319   struct lima_surface *surf = lima_surface(job->key.cbuf);
320   struct lima_resource *res = lima_resource(surf->base.texture);
321   return &res->damage;
322}
323
324static bool
325lima_fb_cbuf_needs_reload(struct lima_job *job)
326{
327   if (!job->key.cbuf)
328      return false;
329
330   struct lima_surface *surf = lima_surface(job->key.cbuf);
331   struct lima_resource *res = lima_resource(surf->base.texture);
332   if (res->damage.region) {
333      /* for EGL_KHR_partial_update, when EGL_EXT_buffer_age is enabled,
334       * we need to reload damage region, otherwise just want to reload
335       * the region not aligned to tile boundary */
336      //if (!res->damage.aligned)
337      //   return true;
338      return true;
339   }
340   else if (surf->reload & PIPE_CLEAR_COLOR0)
341         return true;
342
343   return false;
344}
345
346static bool
347lima_fb_zsbuf_needs_reload(struct lima_job *job)
348{
349   if (!job->key.zsbuf)
350      return false;
351
352   struct lima_surface *surf = lima_surface(job->key.zsbuf);
353   if (surf->reload & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))
354         return true;
355
356   return false;
357}
358
359static void
360lima_pack_reload_plbu_cmd(struct lima_job *job, struct pipe_surface *psurf)
361{
362   struct lima_job_fb_info *fb = &job->fb;
363   struct lima_context *ctx = job->ctx;
364   struct pipe_box src = {
365      .x = 0,
366      .y = 0,
367      .width = fb->width,
368      .height = fb->height,
369   };
370
371   struct pipe_box dst = {
372      .x = 0,
373      .y = 0,
374      .width = fb->width,
375      .height = fb->height,
376   };
377
378   if (ctx->framebuffer.base.samples > 1) {
379      for (int i = 0; i < LIMA_MAX_SAMPLES; i++) {
380         lima_pack_blit_cmd(job, &job->plbu_cmd_head,
381                            psurf, &src, &dst,
382                            PIPE_TEX_FILTER_NEAREST, false,
383                            (1 << i), i);
384      }
385   } else {
386      lima_pack_blit_cmd(job, &job->plbu_cmd_head,
387                         psurf, &src, &dst,
388                         PIPE_TEX_FILTER_NEAREST, false,
389                         0xf, 0);
390   }
391}
392
393static void
394lima_pack_head_plbu_cmd(struct lima_job *job)
395{
396   struct lima_context *ctx = job->ctx;
397   struct lima_job_fb_info *fb = &job->fb;
398
399   PLBU_CMD_BEGIN(&job->plbu_cmd_head, 10);
400
401   PLBU_CMD_UNKNOWN2();
402   PLBU_CMD_BLOCK_STEP(fb->shift_min, fb->shift_h, fb->shift_w);
403   PLBU_CMD_TILED_DIMENSIONS(fb->tiled_w, fb->tiled_h);
404   PLBU_CMD_BLOCK_STRIDE(fb->block_w);
405
406   PLBU_CMD_ARRAY_ADDRESS(
407      ctx->plb_gp_stream->va + ctx->plb_index * ctx->plb_gp_size,
408      fb->block_w * fb->block_h);
409
410   PLBU_CMD_END();
411
412   if (lima_fb_cbuf_needs_reload(job)) {
413      lima_pack_reload_plbu_cmd(job, job->key.cbuf);
414   }
415
416   if (lima_fb_zsbuf_needs_reload(job))
417      lima_pack_reload_plbu_cmd(job, job->key.zsbuf);
418}
419
420static void
421hilbert_rotate(int n, int *x, int *y, int rx, int ry)
422{
423   if (ry == 0) {
424      if (rx == 1) {
425         *x = n-1 - *x;
426         *y = n-1 - *y;
427      }
428
429      /* Swap x and y */
430      int t  = *x;
431      *x = *y;
432      *y = t;
433   }
434}
435
436static void
437hilbert_coords(int n, int d, int *x, int *y)
438{
439   int rx, ry, i, t=d;
440
441   *x = *y = 0;
442
443   for (i = 0; (1 << i) < n; i++) {
444
445      rx = 1 & (t / 2);
446      ry = 1 & (t ^ rx);
447
448      hilbert_rotate(1 << i, x, y, rx, ry);
449
450      *x += rx << i;
451      *y += ry << i;
452
453      t /= 4;
454   }
455}
456
457static int
458lima_get_pp_stream_size(int num_pp, int tiled_w, int tiled_h, uint32_t *off)
459{
460   /* carefully calculate each stream start address:
461    * 1. overflow: each stream size may be different due to
462    *    fb->tiled_w * fb->tiled_h can't be divided by num_pp,
463    *    extra size should be added to the preceeding stream
464    * 2. alignment: each stream address should be 0x20 aligned
465    */
466   int delta = tiled_w * tiled_h / num_pp * 16 + 16;
467   int remain = tiled_w * tiled_h % num_pp;
468   int offset = 0;
469
470   for (int i = 0; i < num_pp; i++) {
471      off[i] = offset;
472
473      offset += delta;
474      if (remain) {
475         offset += 16;
476         remain--;
477      }
478      offset = align(offset, 0x20);
479   }
480
481   return offset;
482}
483
484static void
485lima_generate_pp_stream(struct lima_job *job, int off_x, int off_y,
486                        int tiled_w, int tiled_h)
487{
488   struct lima_context *ctx = job->ctx;
489   struct lima_pp_stream_state *ps = &ctx->pp_stream;
490   struct lima_job_fb_info *fb = &job->fb;
491   struct lima_screen *screen = lima_screen(ctx->base.screen);
492   int num_pp = screen->num_pp;
493   assert(num_pp > 0);
494
495   /* use hilbert_coords to generates 1D to 2D relationship.
496    * 1D for pp stream index and 2D for plb block x/y on framebuffer.
497    * if multi pp, interleave the 1D index to make each pp's render target
498    * close enough which should result close workload
499    */
500   int max = MAX2(tiled_w, tiled_h);
501   int index = 0;
502   uint32_t *stream[8];
503   int si[8] = {0};
504   int dim = 0;
505   int count = 0;
506
507   /* Don't update count if we get zero rect. We'll just generate
508    * PP stream with just terminators in it.
509    */
510   if ((tiled_w * tiled_h) != 0) {
511      dim = util_logbase2_ceil(max);
512      count = 1 << (dim + dim);
513   }
514
515   for (int i = 0; i < num_pp; i++)
516      stream[i] = ps->map + ps->offset[i];
517
518   for (int i = 0; i < count; i++) {
519      int x, y;
520      hilbert_coords(max, i, &x, &y);
521      if (x < tiled_w && y < tiled_h) {
522         x += off_x;
523         y += off_y;
524
525         int pp = index % num_pp;
526         int offset = ((y >> fb->shift_h) * fb->block_w +
527                       (x >> fb->shift_w)) * LIMA_CTX_PLB_BLK_SIZE;
528         int plb_va = ctx->plb[ctx->plb_index]->va + offset;
529
530         stream[pp][si[pp]++] = 0;
531         stream[pp][si[pp]++] = 0xB8000000 | x | (y << 8);
532         stream[pp][si[pp]++] = 0xE0000002 | ((plb_va >> 3) & ~0xE0000003);
533         stream[pp][si[pp]++] = 0xB0000000;
534
535         index++;
536      }
537   }
538
539   for (int i = 0; i < num_pp; i++) {
540      stream[i][si[i]++] = 0;
541      stream[i][si[i]++] = 0xBC000000;
542      stream[i][si[i]++] = 0;
543      stream[i][si[i]++] = 0;
544
545      lima_dump_command_stream_print(
546         job->dump, stream[i], si[i] * 4,
547         false, "pp plb stream %d at va %x\n",
548         i, ps->va + ps->offset[i]);
549   }
550}
551
552static void
553lima_free_stale_pp_stream_bo(struct lima_context *ctx)
554{
555   list_for_each_entry_safe(struct lima_ctx_plb_pp_stream, entry,
556                            &ctx->plb_pp_stream_lru_list, lru_list) {
557      if (ctx->plb_stream_cache_size <= lima_plb_pp_stream_cache_size)
558         break;
559
560      struct hash_entry *hash_entry =
561         _mesa_hash_table_search(ctx->plb_pp_stream, &entry->key);
562      if (hash_entry)
563         _mesa_hash_table_remove(ctx->plb_pp_stream, hash_entry);
564      list_del(&entry->lru_list);
565
566      ctx->plb_stream_cache_size -= entry->bo->size;
567      lima_bo_unreference(entry->bo);
568
569      ralloc_free(entry);
570   }
571}
572
573static void
574lima_update_damage_pp_stream(struct lima_job *job)
575{
576   struct lima_context *ctx = job->ctx;
577   struct lima_damage_region *ds = lima_job_get_damage(job);
578   struct lima_job_fb_info *fb = &job->fb;
579   struct pipe_scissor_state bound;
580   struct pipe_scissor_state *dr = &job->damage_rect;
581
582   if (ds && ds->region) {
583      struct pipe_scissor_state *dbound = &ds->bound;
584      bound.minx = MAX2(dbound->minx, dr->minx >> 4);
585      bound.miny = MAX2(dbound->miny, dr->miny >> 4);
586      bound.maxx = MIN2(dbound->maxx, (dr->maxx + 0xf) >> 4);
587      bound.maxy = MIN2(dbound->maxy, (dr->maxy + 0xf) >> 4);
588   } else {
589      bound.minx = dr->minx >> 4;
590      bound.miny = dr->miny >> 4;
591      bound.maxx = (dr->maxx + 0xf) >> 4;
592      bound.maxy = (dr->maxy + 0xf) >> 4;
593   }
594
595   /* Clamp to FB size */
596   bound.minx = MIN2(bound.minx, fb->tiled_w);
597   bound.miny = MIN2(bound.miny, fb->tiled_h);
598   bound.maxx = MIN2(bound.maxx, fb->tiled_w);
599   bound.maxy = MIN2(bound.maxy, fb->tiled_h);
600
601   struct lima_ctx_plb_pp_stream_key key = {
602      .plb_index = ctx->plb_index,
603      .minx = bound.minx,
604      .miny = bound.miny,
605      .maxx = bound.maxx,
606      .maxy = bound.maxy,
607      .shift_w = fb->shift_w,
608      .shift_h = fb->shift_h,
609      .block_w = fb->block_w,
610      .block_h = fb->block_h,
611   };
612
613   struct hash_entry *entry =
614      _mesa_hash_table_search(ctx->plb_pp_stream, &key);
615   if (entry) {
616      struct lima_ctx_plb_pp_stream *s = entry->data;
617
618      list_del(&s->lru_list);
619      list_addtail(&s->lru_list, &ctx->plb_pp_stream_lru_list);
620
621      ctx->pp_stream.map = lima_bo_map(s->bo);
622      ctx->pp_stream.va = s->bo->va;
623      memcpy(ctx->pp_stream.offset, s->offset, sizeof(s->offset));
624
625      lima_job_add_bo(job, LIMA_PIPE_PP, s->bo, LIMA_SUBMIT_BO_READ);
626
627      return;
628   }
629
630   lima_free_stale_pp_stream_bo(ctx);
631
632   struct lima_screen *screen = lima_screen(ctx->base.screen);
633   struct lima_ctx_plb_pp_stream *s =
634      rzalloc(ctx->plb_pp_stream, struct lima_ctx_plb_pp_stream);
635
636   list_inithead(&s->lru_list);
637   s->key.plb_index = ctx->plb_index;
638   s->key.minx = bound.minx;
639   s->key.maxx = bound.maxx;
640   s->key.miny = bound.miny;
641   s->key.maxy = bound.maxy;
642   s->key.shift_w = fb->shift_w;
643   s->key.shift_h = fb->shift_h;
644   s->key.block_w = fb->block_w;
645   s->key.block_h = fb->block_h;
646
647   int tiled_w = bound.maxx - bound.minx;
648   int tiled_h = bound.maxy - bound.miny;
649   int size = lima_get_pp_stream_size(
650      screen->num_pp, tiled_w, tiled_h, s->offset);
651
652   s->bo = lima_bo_create(screen, size, 0);
653
654   ctx->pp_stream.map = lima_bo_map(s->bo);
655   ctx->pp_stream.va = s->bo->va;
656   memcpy(ctx->pp_stream.offset, s->offset, sizeof(s->offset));
657
658   lima_generate_pp_stream(job, bound.minx, bound.miny, tiled_w, tiled_h);
659
660   ctx->plb_stream_cache_size += size;
661   list_addtail(&s->lru_list, &ctx->plb_pp_stream_lru_list);
662   _mesa_hash_table_insert(ctx->plb_pp_stream, &s->key, s);
663
664   lima_job_add_bo(job, LIMA_PIPE_PP, s->bo, LIMA_SUBMIT_BO_READ);
665}
666
667static bool
668lima_damage_fullscreen(struct lima_job *job)
669{
670   struct pipe_scissor_state *dr = &job->damage_rect;
671
672   return dr->minx == 0 &&
673          dr->miny == 0 &&
674          dr->maxx == job->fb.width &&
675          dr->maxy == job->fb.height;
676}
677
678static void
679lima_update_pp_stream(struct lima_job *job)
680{
681   struct lima_context *ctx = job->ctx;
682   struct lima_screen *screen = lima_screen(ctx->base.screen);
683   struct lima_damage_region *damage = lima_job_get_damage(job);
684   if ((screen->gpu_type == DRM_LIMA_PARAM_GPU_ID_MALI400) ||
685       (damage && damage->region) || !lima_damage_fullscreen(job))
686      lima_update_damage_pp_stream(job);
687   else
688      /* Mali450 doesn't need full PP stream */
689      ctx->pp_stream.map = NULL;
690}
691
692static void
693lima_update_job_bo(struct lima_job *job)
694{
695   struct lima_context *ctx = job->ctx;
696
697   lima_job_add_bo(job, LIMA_PIPE_GP, ctx->plb_gp_stream,
698                      LIMA_SUBMIT_BO_READ);
699   lima_job_add_bo(job, LIMA_PIPE_GP, ctx->plb[ctx->plb_index],
700                      LIMA_SUBMIT_BO_WRITE);
701   lima_job_add_bo(job, LIMA_PIPE_GP, ctx->gp_tile_heap[ctx->plb_index],
702                      LIMA_SUBMIT_BO_WRITE);
703
704   lima_dump_command_stream_print(
705      job->dump, ctx->plb_gp_stream->map + ctx->plb_index * ctx->plb_gp_size,
706      ctx->plb_gp_size, false, "gp plb stream at va %x\n",
707      ctx->plb_gp_stream->va + ctx->plb_index * ctx->plb_gp_size);
708
709   lima_job_add_bo(job, LIMA_PIPE_PP, ctx->plb[ctx->plb_index],
710                      LIMA_SUBMIT_BO_READ);
711   lima_job_add_bo(job, LIMA_PIPE_PP, ctx->gp_tile_heap[ctx->plb_index],
712                      LIMA_SUBMIT_BO_READ);
713
714   struct lima_screen *screen = lima_screen(ctx->base.screen);
715   lima_job_add_bo(job, LIMA_PIPE_PP, screen->pp_buffer, LIMA_SUBMIT_BO_READ);
716}
717
718static void
719lima_finish_plbu_cmd(struct util_dynarray *plbu_cmd_array)
720{
721   int i = 0;
722   uint32_t *plbu_cmd = util_dynarray_ensure_cap(plbu_cmd_array, plbu_cmd_array->size + 2 * 4);
723
724   plbu_cmd[i++] = 0x00000000;
725   plbu_cmd[i++] = 0x50000000; /* END */
726
727   plbu_cmd_array->size += i * 4;
728}
729
730static void
731lima_pack_wb_zsbuf_reg(struct lima_job *job, uint32_t *wb_reg, int wb_idx)
732{
733   struct lima_job_fb_info *fb = &job->fb;
734   struct pipe_surface *zsbuf = job->key.zsbuf;
735   struct lima_resource *res = lima_resource(zsbuf->texture);
736   int level = zsbuf->u.tex.level;
737   uint32_t format = lima_format_get_pixel(zsbuf->format);
738
739   struct lima_pp_wb_reg *wb = (void *)wb_reg;
740   wb[wb_idx].type = 0x01; /* 1 for depth, stencil */
741   wb[wb_idx].address = res->bo->va + res->levels[level].offset;
742   wb[wb_idx].pixel_format = format;
743   if (res->tiled) {
744      wb[wb_idx].pixel_layout = 0x2;
745      wb[wb_idx].pitch = fb->tiled_w;
746   } else {
747      wb[wb_idx].pixel_layout = 0x0;
748      wb[wb_idx].pitch = res->levels[level].stride / 8;
749   }
750   wb[wb_idx].flags = 0;
751   unsigned nr_samples = zsbuf->nr_samples ?
752                         zsbuf->nr_samples : MAX2(1, zsbuf->texture->nr_samples);
753   if (nr_samples > 1) {
754      wb[wb_idx].mrt_pitch = res->mrt_pitch;
755      wb[wb_idx].mrt_bits = u_bit_consecutive(0, nr_samples);
756   }
757}
758
759static void
760lima_pack_wb_cbuf_reg(struct lima_job *job, uint32_t *frame_reg,
761                      uint32_t *wb_reg, int wb_idx)
762{
763   struct lima_job_fb_info *fb = &job->fb;
764   struct pipe_surface *cbuf = job->key.cbuf;
765   struct lima_resource *res = lima_resource(cbuf->texture);
766   int level = cbuf->u.tex.level;
767   unsigned layer = cbuf->u.tex.first_layer;
768   uint32_t format = lima_format_get_pixel(cbuf->format);
769   bool swap_channels = lima_format_get_pixel_swap_rb(cbuf->format);
770
771   struct lima_pp_frame_reg *frame = (void *)frame_reg;
772   frame->channel_layout = lima_format_get_channel_layout(cbuf->format);
773
774   struct lima_pp_wb_reg *wb = (void *)wb_reg;
775   wb[wb_idx].type = 0x02; /* 2 for color buffer */
776   wb[wb_idx].address = res->bo->va + res->levels[level].offset + layer * res->levels[level].layer_stride;
777   wb[wb_idx].pixel_format = format;
778   if (res->tiled) {
779      wb[wb_idx].pixel_layout = 0x2;
780      wb[wb_idx].pitch = fb->tiled_w;
781   } else {
782      wb[wb_idx].pixel_layout = 0x0;
783      wb[wb_idx].pitch = res->levels[level].stride / 8;
784   }
785   wb[wb_idx].flags = swap_channels ? 0x4 : 0x0;
786   unsigned nr_samples = cbuf->nr_samples ?
787                         cbuf->nr_samples : MAX2(1, cbuf->texture->nr_samples);
788   if (nr_samples > 1) {
789      wb[wb_idx].mrt_pitch = res->mrt_pitch;
790      wb[wb_idx].mrt_bits = u_bit_consecutive(0, nr_samples);
791   }
792}
793
794static void
795lima_pack_pp_frame_reg(struct lima_job *job, uint32_t *frame_reg,
796                       uint32_t *wb_reg)
797{
798   struct lima_context *ctx = job->ctx;
799   struct lima_job_fb_info *fb = &job->fb;
800   struct pipe_surface *cbuf = job->key.cbuf;
801   struct lima_pp_frame_reg *frame = (void *)frame_reg;
802   struct lima_screen *screen = lima_screen(ctx->base.screen);
803   int wb_idx = 0;
804
805   frame->render_address = screen->pp_buffer->va + pp_frame_rsw_offset;
806   frame->flags = 0x02;
807   if (cbuf && util_format_is_float(cbuf->format)) {
808      frame->flags |= 0x01; /* enable fp16 */
809      frame->clear_value_color   = (uint32_t)(job->clear.color_16pc & 0xffffffffUL);
810      frame->clear_value_color_1 = (uint32_t)(job->clear.color_16pc >> 32);
811      frame->clear_value_color_2 = 0;
812      frame->clear_value_color_3 = 0;
813   }
814   else {
815      frame->clear_value_color   = job->clear.color_8pc;
816      frame->clear_value_color_1 = job->clear.color_8pc;
817      frame->clear_value_color_2 = job->clear.color_8pc;
818      frame->clear_value_color_3 = job->clear.color_8pc;
819   }
820
821   frame->clear_value_depth = job->clear.depth;
822   frame->clear_value_stencil = job->clear.stencil;
823   frame->one = 1;
824
825   frame->width = fb->width - 1;
826   frame->height = fb->height - 1;
827
828   /* frame->fragment_stack_address is overwritten per-pp in the kernel
829    * by the values of pp_frame.fragment_stack_address[i] */
830
831   /* These are "stack size" and "stack offset" shifted,
832    * here they are assumed to be always the same. */
833   frame->fragment_stack_size = job->pp_max_stack_size << 16 | job->pp_max_stack_size;
834
835   /* related with MSAA and different value when r4p0/r7p0 */
836   frame->supersampled_height = fb->height * 2 - 1;
837   frame->scale = 0xE0C;
838
839   frame->dubya = 0x77;
840   frame->onscreen = 1;
841   frame->blocking = (fb->shift_min << 28) | (fb->shift_h << 16) | fb->shift_w;
842
843   /* Set default layout to 8888 */
844   frame->channel_layout = 0x8888;
845
846   if (cbuf && (job->resolve & PIPE_CLEAR_COLOR0))
847      lima_pack_wb_cbuf_reg(job, frame_reg, wb_reg, wb_idx++);
848
849   if (job->key.zsbuf &&
850       (job->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)))
851      lima_pack_wb_zsbuf_reg(job, wb_reg, wb_idx++);
852}
853
854void
855lima_do_job(struct lima_job *job)
856{
857   #define pp_stack_pp_size 0x400
858
859   struct lima_context *ctx = job->ctx;
860
861   lima_pack_head_plbu_cmd(job);
862   lima_finish_plbu_cmd(&job->plbu_cmd_array);
863
864   lima_update_job_bo(job);
865
866   int vs_cmd_size = job->vs_cmd_array.size;
867   uint32_t vs_cmd_va = 0;
868
869   if (vs_cmd_size) {
870      void *vs_cmd = lima_job_create_stream_bo(
871         job, LIMA_PIPE_GP, vs_cmd_size, &vs_cmd_va);
872      memcpy(vs_cmd, util_dynarray_begin(&job->vs_cmd_array), vs_cmd_size);
873
874      lima_dump_command_stream_print(
875         job->dump, vs_cmd, vs_cmd_size, false, "flush vs cmd at va %x\n", vs_cmd_va);
876      lima_dump_vs_command_stream_print(job->dump, vs_cmd, vs_cmd_size, vs_cmd_va);
877   }
878
879   uint32_t plbu_cmd_va;
880   int plbu_cmd_size = job->plbu_cmd_array.size + job->plbu_cmd_head.size;
881   void *plbu_cmd = lima_job_create_stream_bo(
882      job, LIMA_PIPE_GP, plbu_cmd_size, &plbu_cmd_va);
883   memcpy(plbu_cmd,
884          util_dynarray_begin(&job->plbu_cmd_head),
885          job->plbu_cmd_head.size);
886   memcpy(plbu_cmd + job->plbu_cmd_head.size,
887          util_dynarray_begin(&job->plbu_cmd_array),
888          job->plbu_cmd_array.size);
889
890   lima_dump_command_stream_print(
891      job->dump, plbu_cmd, plbu_cmd_size, false, "flush plbu cmd at va %x\n", plbu_cmd_va);
892   lima_dump_plbu_command_stream_print(job->dump, plbu_cmd, plbu_cmd_size, plbu_cmd_va);
893
894   struct lima_screen *screen = lima_screen(ctx->base.screen);
895   struct drm_lima_gp_frame gp_frame;
896   struct lima_gp_frame_reg *gp_frame_reg = (void *)gp_frame.frame;
897   gp_frame_reg->vs_cmd_start = vs_cmd_va;
898   gp_frame_reg->vs_cmd_end = vs_cmd_va + vs_cmd_size;
899   gp_frame_reg->plbu_cmd_start = plbu_cmd_va;
900   gp_frame_reg->plbu_cmd_end = plbu_cmd_va + plbu_cmd_size;
901   gp_frame_reg->tile_heap_start = ctx->gp_tile_heap[ctx->plb_index]->va;
902   gp_frame_reg->tile_heap_end = ctx->gp_tile_heap[ctx->plb_index]->va + ctx->gp_tile_heap_size;
903
904   lima_dump_command_stream_print(
905      job->dump, &gp_frame, sizeof(gp_frame), false, "add gp frame\n");
906
907   if (!lima_job_start(job, LIMA_PIPE_GP, &gp_frame, sizeof(gp_frame)))
908      fprintf(stderr, "gp job error\n");
909
910   if (job->dump) {
911      if (lima_job_wait(job, LIMA_PIPE_GP, PIPE_TIMEOUT_INFINITE)) {
912         if (ctx->gp_output) {
913            float *pos = lima_bo_map(ctx->gp_output);
914            lima_dump_command_stream_print(
915               job->dump, pos, 4 * 4 * 16, true, "gl_pos dump at va %x\n",
916               ctx->gp_output->va);
917         }
918
919         uint32_t *plb = lima_bo_map(ctx->plb[ctx->plb_index]);
920         lima_dump_command_stream_print(
921            job->dump, plb, LIMA_CTX_PLB_BLK_SIZE, false, "plb dump at va %x\n",
922            ctx->plb[ctx->plb_index]->va);
923      }
924      else {
925         fprintf(stderr, "gp job wait error\n");
926         exit(1);
927      }
928   }
929
930   uint32_t pp_stack_va = 0;
931   if (job->pp_max_stack_size) {
932      lima_job_create_stream_bo(
933         job, LIMA_PIPE_PP,
934         screen->num_pp * job->pp_max_stack_size * pp_stack_pp_size,
935         &pp_stack_va);
936   }
937
938   lima_update_pp_stream(job);
939
940   struct lima_pp_stream_state *ps = &ctx->pp_stream;
941   if (screen->gpu_type == DRM_LIMA_PARAM_GPU_ID_MALI400) {
942      struct drm_lima_m400_pp_frame pp_frame = {0};
943      lima_pack_pp_frame_reg(job, pp_frame.frame, pp_frame.wb);
944      pp_frame.num_pp = screen->num_pp;
945
946      for (int i = 0; i < screen->num_pp; i++) {
947         pp_frame.plbu_array_address[i] = ps->va + ps->offset[i];
948         if (job->pp_max_stack_size)
949            pp_frame.fragment_stack_address[i] = pp_stack_va +
950               job->pp_max_stack_size * pp_stack_pp_size * i;
951      }
952
953      lima_dump_command_stream_print(
954         job->dump, &pp_frame, sizeof(pp_frame), false, "add pp frame\n");
955
956      if (!lima_job_start(job, LIMA_PIPE_PP, &pp_frame, sizeof(pp_frame)))
957         fprintf(stderr, "pp job error\n");
958   }
959   else {
960      struct drm_lima_m450_pp_frame pp_frame = {0};
961      lima_pack_pp_frame_reg(job, pp_frame.frame, pp_frame.wb);
962      pp_frame.num_pp = screen->num_pp;
963
964      if (job->pp_max_stack_size)
965         for (int i = 0; i < screen->num_pp; i++)
966            pp_frame.fragment_stack_address[i] = pp_stack_va +
967               job->pp_max_stack_size * pp_stack_pp_size * i;
968
969      if (ps->map) {
970         for (int i = 0; i < screen->num_pp; i++)
971            pp_frame.plbu_array_address[i] = ps->va + ps->offset[i];
972      }
973      else {
974         pp_frame.use_dlbu = true;
975
976         struct lima_job_fb_info *fb = &job->fb;
977         pp_frame.dlbu_regs[0] = ctx->plb[ctx->plb_index]->va;
978         pp_frame.dlbu_regs[1] = ((fb->tiled_h - 1) << 16) | (fb->tiled_w - 1);
979         unsigned s = util_logbase2(LIMA_CTX_PLB_BLK_SIZE) - 7;
980         pp_frame.dlbu_regs[2] = (s << 28) | (fb->shift_h << 16) | fb->shift_w;
981         pp_frame.dlbu_regs[3] = ((fb->tiled_h - 1) << 24) | ((fb->tiled_w - 1) << 16);
982      }
983
984      lima_dump_command_stream_print(
985         job->dump, &pp_frame, sizeof(pp_frame), false, "add pp frame\n");
986
987      if (!lima_job_start(job, LIMA_PIPE_PP, &pp_frame, sizeof(pp_frame)))
988         fprintf(stderr, "pp job error\n");
989   }
990
991   if (job->dump) {
992      if (!lima_job_wait(job, LIMA_PIPE_PP, PIPE_TIMEOUT_INFINITE)) {
993         fprintf(stderr, "pp wait error\n");
994         exit(1);
995      }
996   }
997
998   ctx->plb_index = (ctx->plb_index + 1) % lima_ctx_num_plb;
999
1000   /* Set reload flags for next draw. It'll be unset if buffer is cleared */
1001   if (job->key.cbuf && (job->resolve & PIPE_CLEAR_COLOR0)) {
1002      struct lima_surface *surf = lima_surface(job->key.cbuf);
1003      surf->reload = PIPE_CLEAR_COLOR0;
1004   }
1005
1006   if (job->key.zsbuf && (job->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))) {
1007      struct lima_surface *surf = lima_surface(job->key.zsbuf);
1008      surf->reload = (job->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL));
1009   }
1010
1011   if (ctx->job == job)
1012      ctx->job = NULL;
1013
1014   lima_job_free(job);
1015}
1016
1017void
1018lima_flush(struct lima_context *ctx)
1019{
1020   hash_table_foreach(ctx->jobs, entry) {
1021      struct lima_job *job = entry->data;
1022      lima_do_job(job);
1023   }
1024}
1025
1026void
1027lima_flush_job_accessing_bo(
1028   struct lima_context *ctx, struct lima_bo *bo, bool write)
1029{
1030   hash_table_foreach(ctx->jobs, entry) {
1031      struct lima_job *job = entry->data;
1032      if (lima_job_has_bo(job, bo, write))
1033         lima_do_job(job);
1034   }
1035}
1036
1037/*
1038 * This is for current job flush previous job which write to the resource it wants
1039 * to read. Tipical usage is flush the FBO which is used as current task's texture.
1040 */
1041void
1042lima_flush_previous_job_writing_resource(
1043   struct lima_context *ctx, struct pipe_resource *prsc)
1044{
1045   struct hash_entry *entry = _mesa_hash_table_search(ctx->write_jobs, prsc);
1046
1047   if (entry) {
1048      struct lima_job *job = entry->data;
1049
1050      /* do not flush current job */
1051      if (job != ctx->job)
1052         lima_do_job(job);
1053   }
1054}
1055
1056static void
1057lima_pipe_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
1058                unsigned flags)
1059{
1060   struct lima_context *ctx = lima_context(pctx);
1061
1062   lima_flush(ctx);
1063
1064   if (fence) {
1065      int drm_fd = lima_screen(ctx->base.screen)->fd;
1066      int fd;
1067
1068      if (!drmSyncobjExportSyncFile(drm_fd, ctx->out_sync[LIMA_PIPE_PP], &fd))
1069         *fence = lima_fence_create(fd);
1070   }
1071}
1072
1073static void
1074lima_texture_barrier(struct pipe_context *pctx, unsigned flags)
1075{
1076    struct lima_context *ctx = lima_context(pctx);
1077
1078    lima_flush(ctx);
1079}
1080
1081static bool
1082lima_job_compare(const void *s1, const void *s2)
1083{
1084   return memcmp(s1, s2, sizeof(struct lima_job_key)) == 0;
1085}
1086
1087static uint32_t
1088lima_job_hash(const void *key)
1089{
1090   return _mesa_hash_data(key, sizeof(struct lima_job_key));
1091}
1092
1093bool lima_job_init(struct lima_context *ctx)
1094{
1095   int fd = lima_screen(ctx->base.screen)->fd;
1096
1097   ctx->jobs = _mesa_hash_table_create(ctx, lima_job_hash, lima_job_compare);
1098   if (!ctx->jobs)
1099      return false;
1100
1101   ctx->write_jobs = _mesa_hash_table_create(
1102      ctx, _mesa_hash_pointer, _mesa_key_pointer_equal);
1103   if (!ctx->write_jobs)
1104      return false;
1105
1106   ctx->in_sync_fd = -1;
1107
1108   for (int i = 0; i < 2; i++) {
1109      if (drmSyncobjCreate(fd, DRM_SYNCOBJ_CREATE_SIGNALED, ctx->in_sync + i) ||
1110          drmSyncobjCreate(fd, DRM_SYNCOBJ_CREATE_SIGNALED, ctx->out_sync + i))
1111         return false;
1112   }
1113
1114   ctx->base.flush = lima_pipe_flush;
1115   ctx->base.texture_barrier = lima_texture_barrier;
1116
1117   return true;
1118}
1119
1120void lima_job_fini(struct lima_context *ctx)
1121{
1122   int fd = lima_screen(ctx->base.screen)->fd;
1123
1124   lima_flush(ctx);
1125
1126   for (int i = 0; i < 2; i++) {
1127      if (ctx->in_sync[i])
1128         drmSyncobjDestroy(fd, ctx->in_sync[i]);
1129      if (ctx->out_sync[i])
1130         drmSyncobjDestroy(fd, ctx->out_sync[i]);
1131   }
1132
1133   if (ctx->in_sync_fd >= 0)
1134      close(ctx->in_sync_fd);
1135}
1136