1 /*
2 * Copyright (C) 2019-2020 Collabora, Ltd.
3 * Copyright (C) 2019 Alyssa Rosenzweig
4 * Copyright (C) 2014-2017 Broadcom
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 */
26
27 #include <assert.h>
28
29 #include "drm-uapi/panfrost_drm.h"
30
31 #include "pan_bo.h"
32 #include "pan_context.h"
33 #include "util/hash_table.h"
34 #include "util/ralloc.h"
35 #include "util/format/u_format.h"
36 #include "util/u_pack_color.h"
37 #include "util/rounding.h"
38 #include "pan_util.h"
39 #include "pan_blending.h"
40 #include "pan_cmdstream.h"
41 #include "decode.h"
42 #include "panfrost-quirks.h"
43
44 /* panfrost_bo_access is here to help us keep track of batch accesses to BOs
45 * and build a proper dependency graph such that batches can be pipelined for
46 * better GPU utilization.
47 *
48 * Each accessed BO has a corresponding entry in the ->accessed_bos hash table.
49 * A BO is either being written or read at any time (see last_is_write).
50 * When the last access is a write, the batch writing the BO might have read
51 * dependencies (readers that have not been executed yet and want to read the
52 * previous BO content), and when the last access is a read, all readers might
53 * depend on another batch to push its results to memory. That's what the
54 * readers/writers keep track off.
55 * There can only be one writer at any given time, if a new batch wants to
56 * write to the same BO, a dependency will be added between the new writer and
57 * the old writer (at the batch level), and panfrost_bo_access->writer will be
58 * updated to point to the new writer.
59 */
60 struct panfrost_bo_access {
61 struct util_dynarray readers;
62 struct panfrost_batch_fence *writer;
63 bool last_is_write;
64 };
65
66 static struct panfrost_batch_fence *
panfrost_create_batch_fence(struct panfrost_batch * batch)67 panfrost_create_batch_fence(struct panfrost_batch *batch)
68 {
69 struct panfrost_batch_fence *fence;
70
71 fence = rzalloc(NULL, struct panfrost_batch_fence);
72 assert(fence);
73 pipe_reference_init(&fence->reference, 1);
74 fence->batch = batch;
75
76 return fence;
77 }
78
79 static void
panfrost_free_batch_fence(struct panfrost_batch_fence * fence)80 panfrost_free_batch_fence(struct panfrost_batch_fence *fence)
81 {
82 ralloc_free(fence);
83 }
84
85 void
panfrost_batch_fence_unreference(struct panfrost_batch_fence * fence)86 panfrost_batch_fence_unreference(struct panfrost_batch_fence *fence)
87 {
88 if (pipe_reference(&fence->reference, NULL))
89 panfrost_free_batch_fence(fence);
90 }
91
92 void
panfrost_batch_fence_reference(struct panfrost_batch_fence * fence)93 panfrost_batch_fence_reference(struct panfrost_batch_fence *fence)
94 {
95 pipe_reference(NULL, &fence->reference);
96 }
97
98 static void
99 panfrost_batch_add_fbo_bos(struct panfrost_batch *batch);
100
101 static struct panfrost_batch *
panfrost_create_batch(struct panfrost_context * ctx,const struct pipe_framebuffer_state * key)102 panfrost_create_batch(struct panfrost_context *ctx,
103 const struct pipe_framebuffer_state *key)
104 {
105 struct panfrost_batch *batch = rzalloc(ctx, struct panfrost_batch);
106 struct panfrost_device *dev = pan_device(ctx->base.screen);
107
108 batch->ctx = ctx;
109
110 batch->bos = _mesa_hash_table_create(batch, _mesa_hash_pointer,
111 _mesa_key_pointer_equal);
112
113 batch->minx = batch->miny = ~0;
114 batch->maxx = batch->maxy = 0;
115
116 batch->out_sync = panfrost_create_batch_fence(batch);
117 util_copy_framebuffer_state(&batch->key, key);
118
119 /* Preallocate the main pool, since every batch has at least one job
120 * structure so it will be used */
121 panfrost_pool_init(&batch->pool, batch, dev, 0, true);
122
123 /* Don't preallocate the invisible pool, since not every batch will use
124 * the pre-allocation, particularly if the varyings are larger than the
125 * preallocation and a reallocation is needed after anyway. */
126 panfrost_pool_init(&batch->invisible_pool, batch, dev, PAN_BO_INVISIBLE, false);
127
128 panfrost_batch_add_fbo_bos(batch);
129
130 return batch;
131 }
132
133 static void
panfrost_freeze_batch(struct panfrost_batch * batch)134 panfrost_freeze_batch(struct panfrost_batch *batch)
135 {
136 struct panfrost_context *ctx = batch->ctx;
137 struct hash_entry *entry;
138
139 /* Remove the entry in the FBO -> batch hash table if the batch
140 * matches and drop the context reference. This way, next draws/clears
141 * targeting this FBO will trigger the creation of a new batch.
142 */
143 entry = _mesa_hash_table_search(ctx->batches, &batch->key);
144 if (entry && entry->data == batch)
145 _mesa_hash_table_remove(ctx->batches, entry);
146
147 if (ctx->batch == batch)
148 ctx->batch = NULL;
149 }
150
151 #ifdef PAN_BATCH_DEBUG
panfrost_batch_is_frozen(struct panfrost_batch * batch)152 static bool panfrost_batch_is_frozen(struct panfrost_batch *batch)
153 {
154 struct panfrost_context *ctx = batch->ctx;
155 struct hash_entry *entry;
156
157 entry = _mesa_hash_table_search(ctx->batches, &batch->key);
158 if (entry && entry->data == batch)
159 return false;
160
161 if (ctx->batch == batch)
162 return false;
163
164 return true;
165 }
166 #endif
167
168 static void
panfrost_free_batch(struct panfrost_batch * batch)169 panfrost_free_batch(struct panfrost_batch *batch)
170 {
171 if (!batch)
172 return;
173
174 #ifdef PAN_BATCH_DEBUG
175 assert(panfrost_batch_is_frozen(batch));
176 #endif
177
178 hash_table_foreach(batch->bos, entry)
179 panfrost_bo_unreference((struct panfrost_bo *)entry->key);
180
181 panfrost_pool_cleanup(&batch->pool);
182 panfrost_pool_cleanup(&batch->invisible_pool);
183
184 util_dynarray_foreach(&batch->dependencies,
185 struct panfrost_batch_fence *, dep) {
186 panfrost_batch_fence_unreference(*dep);
187 }
188
189 util_dynarray_fini(&batch->dependencies);
190
191 /* The out_sync fence lifetime is different from the the batch one
192 * since other batches might want to wait on a fence of already
193 * submitted/signaled batch. All we need to do here is make sure the
194 * fence does not point to an invalid batch, which the core will
195 * interpret as 'batch is already submitted'.
196 */
197 batch->out_sync->batch = NULL;
198 panfrost_batch_fence_unreference(batch->out_sync);
199
200 util_unreference_framebuffer_state(&batch->key);
201 ralloc_free(batch);
202 }
203
204 #ifdef PAN_BATCH_DEBUG
205 static bool
panfrost_dep_graph_contains_batch(struct panfrost_batch * root,struct panfrost_batch * batch)206 panfrost_dep_graph_contains_batch(struct panfrost_batch *root,
207 struct panfrost_batch *batch)
208 {
209 if (!root)
210 return false;
211
212 util_dynarray_foreach(&root->dependencies,
213 struct panfrost_batch_fence *, dep) {
214 if ((*dep)->batch == batch ||
215 panfrost_dep_graph_contains_batch((*dep)->batch, batch))
216 return true;
217 }
218
219 return false;
220 }
221 #endif
222
223 static void
panfrost_batch_add_dep(struct panfrost_batch * batch,struct panfrost_batch_fence * newdep)224 panfrost_batch_add_dep(struct panfrost_batch *batch,
225 struct panfrost_batch_fence *newdep)
226 {
227 if (batch == newdep->batch)
228 return;
229
230 /* We might want to turn ->dependencies into a set if the number of
231 * deps turns out to be big enough to make this 'is dep already there'
232 * search inefficient.
233 */
234 util_dynarray_foreach(&batch->dependencies,
235 struct panfrost_batch_fence *, dep) {
236 if (*dep == newdep)
237 return;
238 }
239
240 #ifdef PAN_BATCH_DEBUG
241 /* Make sure the dependency graph is acyclic. */
242 assert(!panfrost_dep_graph_contains_batch(newdep->batch, batch));
243 #endif
244
245 panfrost_batch_fence_reference(newdep);
246 util_dynarray_append(&batch->dependencies,
247 struct panfrost_batch_fence *, newdep);
248
249 /* We now have a batch depending on us, let's make sure new draw/clear
250 * calls targeting the same FBO use a new batch object.
251 */
252 if (newdep->batch)
253 panfrost_freeze_batch(newdep->batch);
254 }
255
256 static struct panfrost_batch *
panfrost_get_batch(struct panfrost_context * ctx,const struct pipe_framebuffer_state * key)257 panfrost_get_batch(struct panfrost_context *ctx,
258 const struct pipe_framebuffer_state *key)
259 {
260 /* Lookup the job first */
261 struct hash_entry *entry = _mesa_hash_table_search(ctx->batches, key);
262
263 if (entry)
264 return entry->data;
265
266 /* Otherwise, let's create a job */
267
268 struct panfrost_batch *batch = panfrost_create_batch(ctx, key);
269
270 /* Save the created job */
271 _mesa_hash_table_insert(ctx->batches, &batch->key, batch);
272
273 return batch;
274 }
275
276 /* Get the job corresponding to the FBO we're currently rendering into */
277
278 struct panfrost_batch *
panfrost_get_batch_for_fbo(struct panfrost_context * ctx)279 panfrost_get_batch_for_fbo(struct panfrost_context *ctx)
280 {
281 /* If we already began rendering, use that */
282
283 if (ctx->batch) {
284 assert(util_framebuffer_state_equal(&ctx->batch->key,
285 &ctx->pipe_framebuffer));
286 return ctx->batch;
287 }
288
289 /* If not, look up the job */
290 struct panfrost_batch *batch = panfrost_get_batch(ctx,
291 &ctx->pipe_framebuffer);
292
293 /* Set this job as the current FBO job. Will be reset when updating the
294 * FB state and when submitting or releasing a job.
295 */
296 ctx->batch = batch;
297 return batch;
298 }
299
300 struct panfrost_batch *
panfrost_get_fresh_batch_for_fbo(struct panfrost_context * ctx)301 panfrost_get_fresh_batch_for_fbo(struct panfrost_context *ctx)
302 {
303 struct panfrost_batch *batch;
304
305 batch = panfrost_get_batch(ctx, &ctx->pipe_framebuffer);
306
307 /* The batch has no draw/clear queued, let's return it directly.
308 * Note that it's perfectly fine to re-use a batch with an
309 * existing clear, we'll just update it with the new clear request.
310 */
311 if (!batch->scoreboard.first_job)
312 return batch;
313
314 /* Otherwise, we need to freeze the existing one and instantiate a new
315 * one.
316 */
317 panfrost_freeze_batch(batch);
318 return panfrost_get_batch(ctx, &ctx->pipe_framebuffer);
319 }
320
321 static void
panfrost_bo_access_gc_fences(struct panfrost_context * ctx,struct panfrost_bo_access * access,const struct panfrost_bo * bo)322 panfrost_bo_access_gc_fences(struct panfrost_context *ctx,
323 struct panfrost_bo_access *access,
324 const struct panfrost_bo *bo)
325 {
326 if (access->writer) {
327 panfrost_batch_fence_unreference(access->writer);
328 access->writer = NULL;
329 }
330
331 struct panfrost_batch_fence **readers_array = util_dynarray_begin(&access->readers);
332 struct panfrost_batch_fence **new_readers = readers_array;
333
334 util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
335 reader) {
336 if (!(*reader))
337 continue;
338
339 panfrost_batch_fence_unreference(*reader);
340 *reader = NULL;
341 }
342
343 if (!util_dynarray_resize(&access->readers, struct panfrost_batch_fence *,
344 new_readers - readers_array) &&
345 new_readers != readers_array)
346 unreachable("Invalid dynarray access->readers");
347 }
348
349 /* Collect signaled fences to keep the kernel-side syncobj-map small. The
350 * idea is to collect those signaled fences at the end of each flush_all
351 * call. This function is likely to collect only fences from previous
352 * batch flushes not the one that have just have just been submitted and
353 * are probably still in flight when we trigger the garbage collection.
354 * Anyway, we need to do this garbage collection at some point if we don't
355 * want the BO access map to keep invalid entries around and retain
356 * syncobjs forever.
357 */
358 static void
panfrost_gc_fences(struct panfrost_context * ctx)359 panfrost_gc_fences(struct panfrost_context *ctx)
360 {
361 hash_table_foreach(ctx->accessed_bos, entry) {
362 struct panfrost_bo_access *access = entry->data;
363
364 assert(access);
365 panfrost_bo_access_gc_fences(ctx, access, entry->key);
366 if (!util_dynarray_num_elements(&access->readers,
367 struct panfrost_batch_fence *) &&
368 !access->writer) {
369 ralloc_free(access);
370 _mesa_hash_table_remove(ctx->accessed_bos, entry);
371 }
372 }
373 }
374
375 #ifdef PAN_BATCH_DEBUG
376 static bool
panfrost_batch_in_readers(struct panfrost_batch * batch,struct panfrost_bo_access * access)377 panfrost_batch_in_readers(struct panfrost_batch *batch,
378 struct panfrost_bo_access *access)
379 {
380 util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
381 reader) {
382 if (*reader && (*reader)->batch == batch)
383 return true;
384 }
385
386 return false;
387 }
388 #endif
389
390 static void
panfrost_batch_update_bo_access(struct panfrost_batch * batch,struct panfrost_bo * bo,bool writes,bool already_accessed)391 panfrost_batch_update_bo_access(struct panfrost_batch *batch,
392 struct panfrost_bo *bo, bool writes,
393 bool already_accessed)
394 {
395 struct panfrost_context *ctx = batch->ctx;
396 struct panfrost_bo_access *access;
397 bool old_writes = false;
398 struct hash_entry *entry;
399
400 entry = _mesa_hash_table_search(ctx->accessed_bos, bo);
401 access = entry ? entry->data : NULL;
402 if (access) {
403 old_writes = access->last_is_write;
404 } else {
405 access = rzalloc(ctx, struct panfrost_bo_access);
406 util_dynarray_init(&access->readers, access);
407 _mesa_hash_table_insert(ctx->accessed_bos, bo, access);
408 /* We are the first to access this BO, let's initialize
409 * old_writes to our own access type in that case.
410 */
411 old_writes = writes;
412 }
413
414 assert(access);
415
416 if (writes && !old_writes) {
417 /* Previous access was a read and we want to write this BO.
418 * We first need to add explicit deps between our batch and
419 * the previous readers.
420 */
421 util_dynarray_foreach(&access->readers,
422 struct panfrost_batch_fence *, reader) {
423 /* We were already reading the BO, no need to add a dep
424 * on ourself (the acyclic check would complain about
425 * that).
426 */
427 if (!(*reader) || (*reader)->batch == batch)
428 continue;
429
430 panfrost_batch_add_dep(batch, *reader);
431 }
432 panfrost_batch_fence_reference(batch->out_sync);
433
434 if (access->writer)
435 panfrost_batch_fence_unreference(access->writer);
436
437 /* We now are the new writer. */
438 access->writer = batch->out_sync;
439
440 /* Release the previous readers and reset the readers array. */
441 util_dynarray_foreach(&access->readers,
442 struct panfrost_batch_fence *,
443 reader) {
444 if (!*reader)
445 continue;
446 panfrost_batch_fence_unreference(*reader);
447 }
448
449 util_dynarray_clear(&access->readers);
450 } else if (writes && old_writes) {
451 /* First check if we were the previous writer, in that case
452 * there's nothing to do. Otherwise we need to add a
453 * dependency between the new writer and the old one.
454 */
455 if (access->writer != batch->out_sync) {
456 if (access->writer) {
457 panfrost_batch_add_dep(batch, access->writer);
458 panfrost_batch_fence_unreference(access->writer);
459 }
460 panfrost_batch_fence_reference(batch->out_sync);
461 access->writer = batch->out_sync;
462 }
463 } else if (!writes && old_writes) {
464 /* First check if we were the previous writer, in that case
465 * we want to keep the access type unchanged, as a write is
466 * more constraining than a read.
467 */
468 if (access->writer != batch->out_sync) {
469 /* Add a dependency on the previous writer. */
470 panfrost_batch_add_dep(batch, access->writer);
471
472 /* The previous access was a write, there's no reason
473 * to have entries in the readers array.
474 */
475 assert(!util_dynarray_num_elements(&access->readers,
476 struct panfrost_batch_fence *));
477
478 /* Add ourselves to the readers array. */
479 panfrost_batch_fence_reference(batch->out_sync);
480 util_dynarray_append(&access->readers,
481 struct panfrost_batch_fence *,
482 batch->out_sync);
483 }
484 } else {
485 /* We already accessed this BO before, so we should already be
486 * in the reader array.
487 */
488 #ifdef PAN_BATCH_DEBUG
489 if (already_accessed) {
490 assert(panfrost_batch_in_readers(batch, access));
491 return;
492 }
493 #endif
494
495 /* Previous access was a read and we want to read this BO.
496 * Add ourselves to the readers array and add a dependency on
497 * the previous writer if any.
498 */
499 panfrost_batch_fence_reference(batch->out_sync);
500 util_dynarray_append(&access->readers,
501 struct panfrost_batch_fence *,
502 batch->out_sync);
503
504 if (access->writer)
505 panfrost_batch_add_dep(batch, access->writer);
506 }
507
508 access->last_is_write = writes;
509 }
510
511 void
panfrost_batch_add_bo(struct panfrost_batch * batch,struct panfrost_bo * bo,uint32_t flags)512 panfrost_batch_add_bo(struct panfrost_batch *batch, struct panfrost_bo *bo,
513 uint32_t flags)
514 {
515 if (!bo)
516 return;
517
518 struct hash_entry *entry;
519 uint32_t old_flags = 0;
520
521 entry = _mesa_hash_table_search(batch->bos, bo);
522 if (!entry) {
523 entry = _mesa_hash_table_insert(batch->bos, bo,
524 (void *)(uintptr_t)flags);
525 panfrost_bo_reference(bo);
526 } else {
527 old_flags = (uintptr_t)entry->data;
528
529 /* All batches have to agree on the shared flag. */
530 assert((old_flags & PAN_BO_ACCESS_SHARED) ==
531 (flags & PAN_BO_ACCESS_SHARED));
532 }
533
534 assert(entry);
535
536 if (old_flags == flags)
537 return;
538
539 flags |= old_flags;
540 entry->data = (void *)(uintptr_t)flags;
541
542 /* If this is not a shared BO, we don't really care about dependency
543 * tracking.
544 */
545 if (!(flags & PAN_BO_ACCESS_SHARED))
546 return;
547
548 assert(flags & PAN_BO_ACCESS_RW);
549 panfrost_batch_update_bo_access(batch, bo, flags & PAN_BO_ACCESS_WRITE,
550 old_flags != 0);
551 }
552
553 static void
panfrost_batch_add_resource_bos(struct panfrost_batch * batch,struct panfrost_resource * rsrc,uint32_t flags)554 panfrost_batch_add_resource_bos(struct panfrost_batch *batch,
555 struct panfrost_resource *rsrc,
556 uint32_t flags)
557 {
558 panfrost_batch_add_bo(batch, rsrc->bo, flags);
559
560 for (unsigned i = 0; i < MAX_MIP_LEVELS; i++)
561 if (rsrc->slices[i].checksum_bo)
562 panfrost_batch_add_bo(batch, rsrc->slices[i].checksum_bo, flags);
563
564 if (rsrc->separate_stencil)
565 panfrost_batch_add_bo(batch, rsrc->separate_stencil->bo, flags);
566 }
567
568 static void
panfrost_batch_add_fbo_bos(struct panfrost_batch * batch)569 panfrost_batch_add_fbo_bos(struct panfrost_batch *batch)
570 {
571 uint32_t flags = PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_WRITE |
572 PAN_BO_ACCESS_VERTEX_TILER |
573 PAN_BO_ACCESS_FRAGMENT;
574
575 for (unsigned i = 0; i < batch->key.nr_cbufs; ++i) {
576 struct panfrost_resource *rsrc = pan_resource(batch->key.cbufs[i]->texture);
577 panfrost_batch_add_resource_bos(batch, rsrc, flags);
578 }
579
580 if (batch->key.zsbuf) {
581 struct panfrost_resource *rsrc = pan_resource(batch->key.zsbuf->texture);
582 panfrost_batch_add_resource_bos(batch, rsrc, flags);
583 }
584 }
585
586 struct panfrost_bo *
panfrost_batch_create_bo(struct panfrost_batch * batch,size_t size,uint32_t create_flags,uint32_t access_flags)587 panfrost_batch_create_bo(struct panfrost_batch *batch, size_t size,
588 uint32_t create_flags, uint32_t access_flags)
589 {
590 struct panfrost_bo *bo;
591
592 bo = panfrost_bo_create(pan_device(batch->ctx->base.screen), size,
593 create_flags);
594 panfrost_batch_add_bo(batch, bo, access_flags);
595
596 /* panfrost_batch_add_bo() has retained a reference and
597 * panfrost_bo_create() initialize the refcnt to 1, so let's
598 * unreference the BO here so it gets released when the batch is
599 * destroyed (unless it's retained by someone else in the meantime).
600 */
601 panfrost_bo_unreference(bo);
602 return bo;
603 }
604
605 /* Returns the polygon list's GPU address if available, or otherwise allocates
606 * the polygon list. It's perfectly fast to use allocate/free BO directly,
607 * since we'll hit the BO cache and this is one-per-batch anyway. */
608
609 mali_ptr
panfrost_batch_get_polygon_list(struct panfrost_batch * batch,unsigned size)610 panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size)
611 {
612 if (batch->polygon_list) {
613 assert(batch->polygon_list->size >= size);
614 } else {
615 /* Create the BO as invisible, as there's no reason to map */
616 size = util_next_power_of_two(size);
617
618 batch->polygon_list = panfrost_batch_create_bo(batch, size,
619 PAN_BO_INVISIBLE,
620 PAN_BO_ACCESS_PRIVATE |
621 PAN_BO_ACCESS_RW |
622 PAN_BO_ACCESS_VERTEX_TILER |
623 PAN_BO_ACCESS_FRAGMENT);
624 }
625
626 return batch->polygon_list->ptr.gpu;
627 }
628
629 struct panfrost_bo *
panfrost_batch_get_scratchpad(struct panfrost_batch * batch,unsigned size_per_thread,unsigned thread_tls_alloc,unsigned core_count)630 panfrost_batch_get_scratchpad(struct panfrost_batch *batch,
631 unsigned size_per_thread,
632 unsigned thread_tls_alloc,
633 unsigned core_count)
634 {
635 unsigned size = panfrost_get_total_stack_size(size_per_thread,
636 thread_tls_alloc,
637 core_count);
638
639 if (batch->scratchpad) {
640 assert(batch->scratchpad->size >= size);
641 } else {
642 batch->scratchpad = panfrost_batch_create_bo(batch, size,
643 PAN_BO_INVISIBLE,
644 PAN_BO_ACCESS_PRIVATE |
645 PAN_BO_ACCESS_RW |
646 PAN_BO_ACCESS_VERTEX_TILER |
647 PAN_BO_ACCESS_FRAGMENT);
648 }
649
650 return batch->scratchpad;
651 }
652
653 struct panfrost_bo *
panfrost_batch_get_shared_memory(struct panfrost_batch * batch,unsigned size,unsigned workgroup_count)654 panfrost_batch_get_shared_memory(struct panfrost_batch *batch,
655 unsigned size,
656 unsigned workgroup_count)
657 {
658 if (batch->shared_memory) {
659 assert(batch->shared_memory->size >= size);
660 } else {
661 batch->shared_memory = panfrost_batch_create_bo(batch, size,
662 PAN_BO_INVISIBLE,
663 PAN_BO_ACCESS_PRIVATE |
664 PAN_BO_ACCESS_RW |
665 PAN_BO_ACCESS_VERTEX_TILER);
666 }
667
668 return batch->shared_memory;
669 }
670
671 mali_ptr
panfrost_batch_get_bifrost_tiler(struct panfrost_batch * batch,unsigned vertex_count)672 panfrost_batch_get_bifrost_tiler(struct panfrost_batch *batch, unsigned vertex_count)
673 {
674 if (!vertex_count)
675 return 0;
676
677 if (batch->tiler_meta)
678 return batch->tiler_meta;
679
680 struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
681 struct panfrost_ptr t =
682 panfrost_pool_alloc_aligned(&batch->pool, MALI_BIFROST_TILER_HEAP_LENGTH, 64);
683
684 pan_pack(t.cpu, BIFROST_TILER_HEAP, heap) {
685 heap.size = dev->tiler_heap->size;
686 heap.base = dev->tiler_heap->ptr.gpu;
687 heap.bottom = dev->tiler_heap->ptr.gpu;
688 heap.top = dev->tiler_heap->ptr.gpu + dev->tiler_heap->size;
689 }
690
691 mali_ptr heap = t.gpu;
692
693 t = panfrost_pool_alloc_aligned(&batch->pool, MALI_BIFROST_TILER_LENGTH, 64);
694 pan_pack(t.cpu, BIFROST_TILER, tiler) {
695 tiler.hierarchy_mask = 0x28;
696 tiler.fb_width = batch->key.width;
697 tiler.fb_height = batch->key.height;
698 tiler.heap = heap;
699 }
700
701 batch->tiler_meta = t.gpu;
702 return batch->tiler_meta;
703 }
704
705 struct panfrost_bo *
panfrost_batch_get_tiler_dummy(struct panfrost_batch * batch)706 panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch)
707 {
708 struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
709
710 uint32_t create_flags = 0;
711
712 if (batch->tiler_dummy)
713 return batch->tiler_dummy;
714
715 if (!(dev->quirks & MIDGARD_NO_HIER_TILING))
716 create_flags = PAN_BO_INVISIBLE;
717
718 batch->tiler_dummy = panfrost_batch_create_bo(batch, 4096,
719 create_flags,
720 PAN_BO_ACCESS_PRIVATE |
721 PAN_BO_ACCESS_RW |
722 PAN_BO_ACCESS_VERTEX_TILER |
723 PAN_BO_ACCESS_FRAGMENT);
724 assert(batch->tiler_dummy);
725 return batch->tiler_dummy;
726 }
727
728 mali_ptr
panfrost_batch_reserve_framebuffer(struct panfrost_batch * batch)729 panfrost_batch_reserve_framebuffer(struct panfrost_batch *batch)
730 {
731 struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
732
733 /* If we haven't, reserve space for the thread storage descriptor (or a
734 * full framebuffer descriptor on Midgard) */
735
736 if (!batch->framebuffer.gpu) {
737 unsigned size = (dev->quirks & IS_BIFROST) ?
738 MALI_LOCAL_STORAGE_LENGTH :
739 (dev->quirks & MIDGARD_SFBD) ?
740 MALI_SINGLE_TARGET_FRAMEBUFFER_LENGTH :
741 MALI_MULTI_TARGET_FRAMEBUFFER_LENGTH;
742
743 batch->framebuffer = panfrost_pool_alloc_aligned(&batch->pool, size, 64);
744
745 /* Tag the pointer */
746 if (!(dev->quirks & (MIDGARD_SFBD | IS_BIFROST)))
747 batch->framebuffer.gpu |= MALI_FBD_TAG_IS_MFBD;
748 }
749
750 return batch->framebuffer.gpu;
751 }
752
753 static void
panfrost_load_surface(struct panfrost_batch * batch,struct pipe_surface * surf,unsigned loc)754 panfrost_load_surface(struct panfrost_batch *batch, struct pipe_surface *surf, unsigned loc)
755 {
756 if (!surf)
757 return;
758
759 struct panfrost_resource *rsrc = pan_resource(surf->texture);
760 unsigned level = surf->u.tex.level;
761
762 if (!rsrc->slices[level].initialized)
763 return;
764
765 if (!rsrc->damage.inverted_len)
766 return;
767
768 /* Clamp the rendering area to the damage extent. The
769 * KHR_partial_update() spec states that trying to render outside of
770 * the damage region is "undefined behavior", so we should be safe.
771 */
772 unsigned damage_width = (rsrc->damage.extent.maxx - rsrc->damage.extent.minx);
773 unsigned damage_height = (rsrc->damage.extent.maxy - rsrc->damage.extent.miny);
774
775 if (damage_width && damage_height) {
776 panfrost_batch_intersection_scissor(batch,
777 rsrc->damage.extent.minx,
778 rsrc->damage.extent.miny,
779 rsrc->damage.extent.maxx,
780 rsrc->damage.extent.maxy);
781 }
782
783 enum pipe_format format = rsrc->base.format;
784
785 if (loc == FRAG_RESULT_DEPTH) {
786 if (!util_format_has_depth(util_format_description(format)))
787 return;
788
789 format = util_format_get_depth_only(format);
790 } else if (loc == FRAG_RESULT_STENCIL) {
791 if (!util_format_has_stencil(util_format_description(format)))
792 return;
793
794 if (rsrc->separate_stencil) {
795 rsrc = rsrc->separate_stencil;
796 format = rsrc->base.format;
797 }
798
799 format = util_format_stencil_only(format);
800 }
801
802 enum mali_texture_dimension dim =
803 panfrost_translate_texture_dimension(rsrc->base.target);
804
805 struct pan_image img = {
806 .width0 = rsrc->base.width0,
807 .height0 = rsrc->base.height0,
808 .depth0 = rsrc->base.depth0,
809 .format = format,
810 .dim = dim,
811 .modifier = rsrc->modifier,
812 .array_size = rsrc->base.array_size,
813 .first_level = level,
814 .last_level = level,
815 .first_layer = surf->u.tex.first_layer,
816 .last_layer = surf->u.tex.last_layer,
817 .nr_samples = rsrc->base.nr_samples,
818 .cubemap_stride = rsrc->cubemap_stride,
819 .bo = rsrc->bo,
820 .slices = rsrc->slices
821 };
822
823 mali_ptr blend_shader = 0;
824
825 if (loc >= FRAG_RESULT_DATA0 && !panfrost_can_fixed_blend(rsrc->base.format)) {
826 struct panfrost_blend_shader *b =
827 panfrost_get_blend_shader(batch->ctx, batch->ctx->blit_blend,
828 rsrc->base.format, loc - FRAG_RESULT_DATA0,
829 NULL);
830
831 struct panfrost_bo *bo = panfrost_batch_create_bo(batch, b->size,
832 PAN_BO_EXECUTE,
833 PAN_BO_ACCESS_PRIVATE |
834 PAN_BO_ACCESS_READ |
835 PAN_BO_ACCESS_FRAGMENT);
836
837 memcpy(bo->ptr.cpu, b->buffer, b->size);
838 assert(b->work_count <= 4);
839
840 blend_shader = bo->ptr.gpu | b->first_tag;
841 }
842
843 struct panfrost_ptr transfer = panfrost_pool_alloc_aligned(&batch->pool,
844 4 * 4 * 6 * rsrc->damage.inverted_len, 64);
845
846 for (unsigned i = 0; i < rsrc->damage.inverted_len; ++i) {
847 float *o = (float *) (transfer.cpu + (4 * 4 * 6 * i));
848 struct pan_rect r = rsrc->damage.inverted_rects[i];
849
850 float rect[] = {
851 r.minx, rsrc->base.height0 - r.miny, 0.0, 1.0,
852 r.maxx, rsrc->base.height0 - r.miny, 0.0, 1.0,
853 r.minx, rsrc->base.height0 - r.maxy, 0.0, 1.0,
854
855 r.maxx, rsrc->base.height0 - r.miny, 0.0, 1.0,
856 r.minx, rsrc->base.height0 - r.maxy, 0.0, 1.0,
857 r.maxx, rsrc->base.height0 - r.maxy, 0.0, 1.0,
858 };
859
860 assert(sizeof(rect) == 4 * 4 * 6);
861 memcpy(o, rect, sizeof(rect));
862 }
863
864 unsigned vertex_count = rsrc->damage.inverted_len * 6;
865 if (batch->pool.dev->quirks & IS_BIFROST) {
866 mali_ptr tiler =
867 panfrost_batch_get_bifrost_tiler(batch, vertex_count);
868 panfrost_load_bifrost(&batch->pool, &batch->scoreboard,
869 blend_shader,
870 batch->framebuffer.gpu,
871 tiler,
872 transfer.gpu, vertex_count,
873 &img, loc);
874 } else {
875 panfrost_load_midg(&batch->pool, &batch->scoreboard,
876 blend_shader,
877 batch->framebuffer.gpu,
878 transfer.gpu, vertex_count,
879 &img, loc);
880 }
881
882 panfrost_batch_add_bo(batch, batch->pool.dev->blit_shaders.bo,
883 PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ | PAN_BO_ACCESS_FRAGMENT);
884 }
885
886 static void
panfrost_batch_draw_wallpaper(struct panfrost_batch * batch)887 panfrost_batch_draw_wallpaper(struct panfrost_batch *batch)
888 {
889 panfrost_batch_reserve_framebuffer(batch);
890
891 /* Assume combined. If either depth or stencil is written, they will
892 * both be written so we need to be careful for reloading */
893
894 unsigned draws = batch->draws;
895
896 if (draws & PIPE_CLEAR_DEPTHSTENCIL)
897 draws |= PIPE_CLEAR_DEPTHSTENCIL;
898
899 /* Mask of buffers which need reload since they are not cleared and
900 * they are drawn. (If they are cleared, reload is useless; if they are
901 * not drawn and also not cleared, we can generally omit the attachment
902 * at the framebuffer descriptor level */
903
904 unsigned reload = ~batch->clear & draws;
905
906 for (unsigned i = 0; i < batch->key.nr_cbufs; ++i) {
907 if (reload & (PIPE_CLEAR_COLOR0 << i))
908 panfrost_load_surface(batch, batch->key.cbufs[i], FRAG_RESULT_DATA0 + i);
909 }
910
911 if (reload & PIPE_CLEAR_DEPTH)
912 panfrost_load_surface(batch, batch->key.zsbuf, FRAG_RESULT_DEPTH);
913
914 if (reload & PIPE_CLEAR_STENCIL)
915 panfrost_load_surface(batch, batch->key.zsbuf, FRAG_RESULT_STENCIL);
916 }
917
918 static void
panfrost_batch_record_bo(struct hash_entry * entry,unsigned * bo_handles,unsigned idx)919 panfrost_batch_record_bo(struct hash_entry *entry, unsigned *bo_handles, unsigned idx)
920 {
921 struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
922 uint32_t flags = (uintptr_t)entry->data;
923
924 assert(bo->gem_handle > 0);
925 bo_handles[idx] = bo->gem_handle;
926
927 /* Update the BO access flags so that panfrost_bo_wait() knows
928 * about all pending accesses.
929 * We only keep the READ/WRITE info since this is all the BO
930 * wait logic cares about.
931 * We also preserve existing flags as this batch might not
932 * be the first one to access the BO.
933 */
934 bo->gpu_access |= flags & (PAN_BO_ACCESS_RW);
935 }
936
937 static int
panfrost_batch_submit_ioctl(struct panfrost_batch * batch,mali_ptr first_job_desc,uint32_t reqs,uint32_t in_sync,uint32_t out_sync)938 panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
939 mali_ptr first_job_desc,
940 uint32_t reqs,
941 uint32_t in_sync,
942 uint32_t out_sync)
943 {
944 struct panfrost_context *ctx = batch->ctx;
945 struct pipe_context *gallium = (struct pipe_context *) ctx;
946 struct panfrost_device *dev = pan_device(gallium->screen);
947 struct drm_panfrost_submit submit = {0,};
948 uint32_t *bo_handles;
949 int ret;
950
951 /* If we trace, we always need a syncobj, so make one of our own if we
952 * weren't given one to use. Remember that we did so, so we can free it
953 * after we're done but preventing double-frees if we were given a
954 * syncobj */
955
956 if (!out_sync && dev->debug & (PAN_DBG_TRACE | PAN_DBG_SYNC))
957 out_sync = ctx->syncobj;
958
959 submit.out_sync = out_sync;
960 submit.jc = first_job_desc;
961 submit.requirements = reqs;
962 if (in_sync) {
963 submit.in_syncs = (u64)(uintptr_t)(&in_sync);
964 submit.in_sync_count = 1;
965 }
966
967 bo_handles = calloc(panfrost_pool_num_bos(&batch->pool) +
968 panfrost_pool_num_bos(&batch->invisible_pool) +
969 batch->bos->entries + 1,
970 sizeof(*bo_handles));
971 assert(bo_handles);
972
973 hash_table_foreach(batch->bos, entry)
974 panfrost_batch_record_bo(entry, bo_handles, submit.bo_handle_count++);
975
976 panfrost_pool_get_bo_handles(&batch->pool, bo_handles + submit.bo_handle_count);
977 submit.bo_handle_count += panfrost_pool_num_bos(&batch->pool);
978 panfrost_pool_get_bo_handles(&batch->invisible_pool, bo_handles + submit.bo_handle_count);
979 submit.bo_handle_count += panfrost_pool_num_bos(&batch->invisible_pool);
980
981 /* Used by all tiler jobs (XXX: skip for compute-only) */
982 if (!(reqs & PANFROST_JD_REQ_FS))
983 bo_handles[submit.bo_handle_count++] = dev->tiler_heap->gem_handle;
984
985 submit.bo_handles = (u64) (uintptr_t) bo_handles;
986 ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
987 free(bo_handles);
988
989 if (ret) {
990 if (dev->debug & PAN_DBG_MSGS)
991 fprintf(stderr, "Error submitting: %m\n");
992
993 return errno;
994 }
995
996 /* Trace the job if we're doing that */
997 if (dev->debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) {
998 /* Wait so we can get errors reported back */
999 drmSyncobjWait(dev->fd, &out_sync, 1,
1000 INT64_MAX, 0, NULL);
1001
1002 /* Trace gets priority over sync */
1003 bool minimal = !(dev->debug & PAN_DBG_TRACE);
1004 pandecode_jc(submit.jc, dev->quirks & IS_BIFROST, dev->gpu_id, minimal);
1005 }
1006
1007 return 0;
1008 }
1009
1010 /* Submit both vertex/tiler and fragment jobs for a batch, possibly with an
1011 * outsync corresponding to the later of the two (since there will be an
1012 * implicit dep between them) */
1013
1014 static int
panfrost_batch_submit_jobs(struct panfrost_batch * batch,uint32_t in_sync,uint32_t out_sync)1015 panfrost_batch_submit_jobs(struct panfrost_batch *batch, uint32_t in_sync, uint32_t out_sync)
1016 {
1017 bool has_draws = batch->scoreboard.first_job;
1018 bool has_frag = batch->scoreboard.tiler_dep || batch->clear;
1019 int ret = 0;
1020
1021 if (has_draws) {
1022 ret = panfrost_batch_submit_ioctl(batch, batch->scoreboard.first_job,
1023 0, in_sync, has_frag ? 0 : out_sync);
1024 assert(!ret);
1025 }
1026
1027 if (has_frag) {
1028 /* Whether we program the fragment job for draws or not depends
1029 * on whether there is any *tiler* activity (so fragment
1030 * shaders). If there are draws but entirely RASTERIZER_DISCARD
1031 * (say, for transform feedback), we want a fragment job that
1032 * *only* clears, since otherwise the tiler structures will be
1033 * uninitialized leading to faults (or state leaks) */
1034
1035 mali_ptr fragjob = panfrost_fragment_job(batch,
1036 batch->scoreboard.tiler_dep != 0);
1037 ret = panfrost_batch_submit_ioctl(batch, fragjob,
1038 PANFROST_JD_REQ_FS, 0,
1039 out_sync);
1040 assert(!ret);
1041 }
1042
1043 return ret;
1044 }
1045
1046 static void
panfrost_batch_submit(struct panfrost_batch * batch,uint32_t in_sync,uint32_t out_sync)1047 panfrost_batch_submit(struct panfrost_batch *batch,
1048 uint32_t in_sync, uint32_t out_sync)
1049 {
1050 assert(batch);
1051 struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
1052
1053 /* Submit the dependencies first. Don't pass along the out_sync since
1054 * they are guaranteed to terminate sooner */
1055 util_dynarray_foreach(&batch->dependencies,
1056 struct panfrost_batch_fence *, dep) {
1057 if ((*dep)->batch)
1058 panfrost_batch_submit((*dep)->batch, 0, 0);
1059 }
1060
1061 int ret;
1062
1063 /* Nothing to do! */
1064 if (!batch->scoreboard.first_job && !batch->clear)
1065 goto out;
1066
1067 panfrost_batch_draw_wallpaper(batch);
1068
1069 /* Now that all draws are in, we can finally prepare the
1070 * FBD for the batch */
1071
1072 if (batch->framebuffer.gpu && batch->scoreboard.first_job) {
1073 struct panfrost_context *ctx = batch->ctx;
1074 struct pipe_context *gallium = (struct pipe_context *) ctx;
1075 struct panfrost_device *dev = pan_device(gallium->screen);
1076
1077 if (dev->quirks & MIDGARD_SFBD)
1078 panfrost_attach_sfbd(batch, ~0);
1079 else
1080 panfrost_attach_mfbd(batch, ~0);
1081 }
1082
1083 mali_ptr polygon_list = panfrost_batch_get_polygon_list(batch,
1084 MALI_MIDGARD_TILER_MINIMUM_HEADER_SIZE);
1085
1086 panfrost_scoreboard_initialize_tiler(&batch->pool, &batch->scoreboard, polygon_list);
1087
1088 ret = panfrost_batch_submit_jobs(batch, in_sync, out_sync);
1089
1090 if (ret && dev->debug & PAN_DBG_MSGS)
1091 fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret);
1092
1093 /* We must reset the damage info of our render targets here even
1094 * though a damage reset normally happens when the DRI layer swaps
1095 * buffers. That's because there can be implicit flushes the GL
1096 * app is not aware of, and those might impact the damage region: if
1097 * part of the damaged portion is drawn during those implicit flushes,
1098 * you have to reload those areas before next draws are pushed, and
1099 * since the driver can't easily know what's been modified by the draws
1100 * it flushed, the easiest solution is to reload everything.
1101 */
1102 for (unsigned i = 0; i < batch->key.nr_cbufs; i++) {
1103 if (!batch->key.cbufs[i])
1104 continue;
1105
1106 panfrost_resource_set_damage_region(NULL,
1107 batch->key.cbufs[i]->texture, 0, NULL);
1108 }
1109
1110 out:
1111 panfrost_freeze_batch(batch);
1112 panfrost_free_batch(batch);
1113 }
1114
1115 /* Submit all batches, applying the out_sync to the currently bound batch */
1116
1117 void
panfrost_flush_all_batches(struct panfrost_context * ctx)1118 panfrost_flush_all_batches(struct panfrost_context *ctx)
1119 {
1120 struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
1121 panfrost_batch_submit(batch, ctx->syncobj, ctx->syncobj);
1122
1123 hash_table_foreach(ctx->batches, hentry) {
1124 struct panfrost_batch *batch = hentry->data;
1125 assert(batch);
1126
1127 panfrost_batch_submit(batch, ctx->syncobj, ctx->syncobj);
1128 }
1129
1130 assert(!ctx->batches->entries);
1131
1132 /* Collect batch fences before returning */
1133 panfrost_gc_fences(ctx);
1134 }
1135
1136 bool
panfrost_pending_batches_access_bo(struct panfrost_context * ctx,const struct panfrost_bo * bo)1137 panfrost_pending_batches_access_bo(struct panfrost_context *ctx,
1138 const struct panfrost_bo *bo)
1139 {
1140 struct panfrost_bo_access *access;
1141 struct hash_entry *hentry;
1142
1143 hentry = _mesa_hash_table_search(ctx->accessed_bos, bo);
1144 access = hentry ? hentry->data : NULL;
1145 if (!access)
1146 return false;
1147
1148 if (access->writer && access->writer->batch)
1149 return true;
1150
1151 util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
1152 reader) {
1153 if (*reader && (*reader)->batch)
1154 return true;
1155 }
1156
1157 return false;
1158 }
1159
1160 /* We always flush writers. We might also need to flush readers */
1161
1162 void
panfrost_flush_batches_accessing_bo(struct panfrost_context * ctx,struct panfrost_bo * bo,bool flush_readers)1163 panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx,
1164 struct panfrost_bo *bo,
1165 bool flush_readers)
1166 {
1167 struct panfrost_bo_access *access;
1168 struct hash_entry *hentry;
1169
1170 hentry = _mesa_hash_table_search(ctx->accessed_bos, bo);
1171 access = hentry ? hentry->data : NULL;
1172 if (!access)
1173 return;
1174
1175 if (access->writer && access->writer->batch)
1176 panfrost_batch_submit(access->writer->batch, ctx->syncobj, ctx->syncobj);
1177
1178 if (!flush_readers)
1179 return;
1180
1181 util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
1182 reader) {
1183 if (*reader && (*reader)->batch)
1184 panfrost_batch_submit((*reader)->batch, ctx->syncobj, ctx->syncobj);
1185 }
1186 }
1187
1188 void
panfrost_batch_set_requirements(struct panfrost_batch * batch)1189 panfrost_batch_set_requirements(struct panfrost_batch *batch)
1190 {
1191 struct panfrost_context *ctx = batch->ctx;
1192
1193 if (ctx->rasterizer->base.multisample)
1194 batch->requirements |= PAN_REQ_MSAA;
1195
1196 if (ctx->depth_stencil && ctx->depth_stencil->base.depth.writemask) {
1197 batch->requirements |= PAN_REQ_DEPTH_WRITE;
1198 batch->draws |= PIPE_CLEAR_DEPTH;
1199 }
1200
1201 if (ctx->depth_stencil && ctx->depth_stencil->base.stencil[0].enabled)
1202 batch->draws |= PIPE_CLEAR_STENCIL;
1203 }
1204
1205 void
panfrost_batch_adjust_stack_size(struct panfrost_batch * batch)1206 panfrost_batch_adjust_stack_size(struct panfrost_batch *batch)
1207 {
1208 struct panfrost_context *ctx = batch->ctx;
1209
1210 for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i) {
1211 struct panfrost_shader_state *ss;
1212
1213 ss = panfrost_get_shader_state(ctx, i);
1214 if (!ss)
1215 continue;
1216
1217 batch->stack_size = MAX2(batch->stack_size, ss->stack_size);
1218 }
1219 }
1220
1221 /* Helper to smear a 32-bit color across 128-bit components */
1222
1223 static void
pan_pack_color_32(uint32_t * packed,uint32_t v)1224 pan_pack_color_32(uint32_t *packed, uint32_t v)
1225 {
1226 for (unsigned i = 0; i < 4; ++i)
1227 packed[i] = v;
1228 }
1229
1230 static void
pan_pack_color_64(uint32_t * packed,uint32_t lo,uint32_t hi)1231 pan_pack_color_64(uint32_t *packed, uint32_t lo, uint32_t hi)
1232 {
1233 for (unsigned i = 0; i < 4; i += 2) {
1234 packed[i + 0] = lo;
1235 packed[i + 1] = hi;
1236 }
1237 }
1238
1239 static void
pan_pack_color(uint32_t * packed,const union pipe_color_union * color,enum pipe_format format)1240 pan_pack_color(uint32_t *packed, const union pipe_color_union *color, enum pipe_format format)
1241 {
1242 /* Alpha magicked to 1.0 if there is no alpha */
1243
1244 bool has_alpha = util_format_has_alpha(format);
1245 float clear_alpha = has_alpha ? color->f[3] : 1.0f;
1246
1247 /* Packed color depends on the framebuffer format */
1248
1249 const struct util_format_description *desc =
1250 util_format_description(format);
1251
1252 if (util_format_is_rgba8_variant(desc) && desc->colorspace != UTIL_FORMAT_COLORSPACE_SRGB) {
1253 pan_pack_color_32(packed,
1254 ((uint32_t) float_to_ubyte(clear_alpha) << 24) |
1255 ((uint32_t) float_to_ubyte(color->f[2]) << 16) |
1256 ((uint32_t) float_to_ubyte(color->f[1]) << 8) |
1257 ((uint32_t) float_to_ubyte(color->f[0]) << 0));
1258 } else if (format == PIPE_FORMAT_B5G6R5_UNORM) {
1259 /* First, we convert the components to R5, G6, B5 separately */
1260 unsigned r5 = _mesa_roundevenf(SATURATE(color->f[0]) * 31.0);
1261 unsigned g6 = _mesa_roundevenf(SATURATE(color->f[1]) * 63.0);
1262 unsigned b5 = _mesa_roundevenf(SATURATE(color->f[2]) * 31.0);
1263
1264 /* Then we pack into a sparse u32. TODO: Why these shifts? */
1265 pan_pack_color_32(packed, (b5 << 25) | (g6 << 14) | (r5 << 5));
1266 } else if (format == PIPE_FORMAT_B4G4R4A4_UNORM) {
1267 /* Convert to 4-bits */
1268 unsigned r4 = _mesa_roundevenf(SATURATE(color->f[0]) * 15.0);
1269 unsigned g4 = _mesa_roundevenf(SATURATE(color->f[1]) * 15.0);
1270 unsigned b4 = _mesa_roundevenf(SATURATE(color->f[2]) * 15.0);
1271 unsigned a4 = _mesa_roundevenf(SATURATE(clear_alpha) * 15.0);
1272
1273 /* Pack on *byte* intervals */
1274 pan_pack_color_32(packed, (a4 << 28) | (b4 << 20) | (g4 << 12) | (r4 << 4));
1275 } else if (format == PIPE_FORMAT_B5G5R5A1_UNORM) {
1276 /* Scale as expected but shift oddly */
1277 unsigned r5 = _mesa_roundevenf(SATURATE(color->f[0]) * 31.0);
1278 unsigned g5 = _mesa_roundevenf(SATURATE(color->f[1]) * 31.0);
1279 unsigned b5 = _mesa_roundevenf(SATURATE(color->f[2]) * 31.0);
1280 unsigned a1 = _mesa_roundevenf(SATURATE(clear_alpha) * 1.0);
1281
1282 pan_pack_color_32(packed, (a1 << 31) | (b5 << 25) | (g5 << 15) | (r5 << 5));
1283 } else {
1284 /* Otherwise, it's generic subject to replication */
1285
1286 union util_color out = { 0 };
1287 unsigned size = util_format_get_blocksize(format);
1288
1289 util_pack_color(color->f, format, &out);
1290
1291 if (size == 1) {
1292 unsigned b = out.ui[0];
1293 unsigned s = b | (b << 8);
1294 pan_pack_color_32(packed, s | (s << 16));
1295 } else if (size == 2)
1296 pan_pack_color_32(packed, out.ui[0] | (out.ui[0] << 16));
1297 else if (size == 3 || size == 4)
1298 pan_pack_color_32(packed, out.ui[0]);
1299 else if (size == 6)
1300 pan_pack_color_64(packed, out.ui[0], out.ui[1] | (out.ui[1] << 16)); /* RGB16F -- RGBB */
1301 else if (size == 8)
1302 pan_pack_color_64(packed, out.ui[0], out.ui[1]);
1303 else if (size == 16)
1304 memcpy(packed, out.ui, 16);
1305 else
1306 unreachable("Unknown generic format size packing clear colour");
1307 }
1308 }
1309
1310 void
panfrost_batch_clear(struct panfrost_batch * batch,unsigned buffers,const union pipe_color_union * color,double depth,unsigned stencil)1311 panfrost_batch_clear(struct panfrost_batch *batch,
1312 unsigned buffers,
1313 const union pipe_color_union *color,
1314 double depth, unsigned stencil)
1315 {
1316 struct panfrost_context *ctx = batch->ctx;
1317
1318 if (buffers & PIPE_CLEAR_COLOR) {
1319 for (unsigned i = 0; i < PIPE_MAX_COLOR_BUFS; ++i) {
1320 if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
1321 continue;
1322
1323 enum pipe_format format = ctx->pipe_framebuffer.cbufs[i]->format;
1324 pan_pack_color(batch->clear_color[i], color, format);
1325 }
1326 }
1327
1328 if (buffers & PIPE_CLEAR_DEPTH) {
1329 batch->clear_depth = depth;
1330 }
1331
1332 if (buffers & PIPE_CLEAR_STENCIL) {
1333 batch->clear_stencil = stencil;
1334 }
1335
1336 batch->clear |= buffers;
1337
1338 /* Clearing affects the entire framebuffer (by definition -- this is
1339 * the Gallium clear callback, which clears the whole framebuffer. If
1340 * the scissor test were enabled from the GL side, the gallium frontend
1341 * would emit a quad instead and we wouldn't go down this code path) */
1342
1343 panfrost_batch_union_scissor(batch, 0, 0,
1344 ctx->pipe_framebuffer.width,
1345 ctx->pipe_framebuffer.height);
1346 }
1347
1348 static bool
panfrost_batch_compare(const void * a,const void * b)1349 panfrost_batch_compare(const void *a, const void *b)
1350 {
1351 return util_framebuffer_state_equal(a, b);
1352 }
1353
1354 static uint32_t
panfrost_batch_hash(const void * key)1355 panfrost_batch_hash(const void *key)
1356 {
1357 return _mesa_hash_data(key, sizeof(struct pipe_framebuffer_state));
1358 }
1359
1360 /* Given a new bounding rectangle (scissor), let the job cover the union of the
1361 * new and old bounding rectangles */
1362
1363 void
panfrost_batch_union_scissor(struct panfrost_batch * batch,unsigned minx,unsigned miny,unsigned maxx,unsigned maxy)1364 panfrost_batch_union_scissor(struct panfrost_batch *batch,
1365 unsigned minx, unsigned miny,
1366 unsigned maxx, unsigned maxy)
1367 {
1368 batch->minx = MIN2(batch->minx, minx);
1369 batch->miny = MIN2(batch->miny, miny);
1370 batch->maxx = MAX2(batch->maxx, maxx);
1371 batch->maxy = MAX2(batch->maxy, maxy);
1372 }
1373
1374 void
panfrost_batch_intersection_scissor(struct panfrost_batch * batch,unsigned minx,unsigned miny,unsigned maxx,unsigned maxy)1375 panfrost_batch_intersection_scissor(struct panfrost_batch *batch,
1376 unsigned minx, unsigned miny,
1377 unsigned maxx, unsigned maxy)
1378 {
1379 batch->minx = MAX2(batch->minx, minx);
1380 batch->miny = MAX2(batch->miny, miny);
1381 batch->maxx = MIN2(batch->maxx, maxx);
1382 batch->maxy = MIN2(batch->maxy, maxy);
1383 }
1384
1385 /* Are we currently rendering to the dev (rather than an FBO)? */
1386
1387 bool
panfrost_batch_is_scanout(struct panfrost_batch * batch)1388 panfrost_batch_is_scanout(struct panfrost_batch *batch)
1389 {
1390 /* If there is no color buffer, it's an FBO */
1391 if (batch->key.nr_cbufs != 1)
1392 return false;
1393
1394 /* If we're too early that no framebuffer was sent, it's scanout */
1395 if (!batch->key.cbufs[0])
1396 return true;
1397
1398 return batch->key.cbufs[0]->texture->bind & PIPE_BIND_DISPLAY_TARGET ||
1399 batch->key.cbufs[0]->texture->bind & PIPE_BIND_SCANOUT ||
1400 batch->key.cbufs[0]->texture->bind & PIPE_BIND_SHARED;
1401 }
1402
1403 void
panfrost_batch_init(struct panfrost_context * ctx)1404 panfrost_batch_init(struct panfrost_context *ctx)
1405 {
1406 ctx->batches = _mesa_hash_table_create(ctx,
1407 panfrost_batch_hash,
1408 panfrost_batch_compare);
1409 ctx->accessed_bos = _mesa_hash_table_create(ctx, _mesa_hash_pointer,
1410 _mesa_key_pointer_equal);
1411 }
1412