2 * Copyright (C) 2019 Alyssa Rosenzweig
3 * Copyright (C) 2014-2017 Broadcom
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
28 #include "drm-uapi/panfrost_drm.h"
31 #include "pan_context.h"
32 #include "util/hash_table.h"
33 #include "util/ralloc.h"
34 #include "util/u_format.h"
35 #include "util/u_pack_color.h"
37 #include "pandecode/decode.h"
39 static struct panfrost_batch
*
40 panfrost_create_batch(struct panfrost_context
*ctx
,
41 const struct pipe_framebuffer_state
*key
)
43 struct panfrost_batch
*batch
= rzalloc(ctx
, struct panfrost_batch
);
47 batch
->bos
= _mesa_set_create(batch
,
49 _mesa_key_pointer_equal
);
51 batch
->minx
= batch
->miny
= ~0;
52 batch
->maxx
= batch
->maxy
= 0;
53 batch
->transient_offset
= 0;
55 util_dynarray_init(&batch
->headers
, batch
);
56 util_dynarray_init(&batch
->gpu_headers
, batch
);
57 util_copy_framebuffer_state(&batch
->key
, key
);
63 panfrost_free_batch(struct panfrost_batch
*batch
)
68 struct panfrost_context
*ctx
= batch
->ctx
;
70 set_foreach(batch
->bos
, entry
) {
71 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
72 panfrost_bo_unreference(bo
);
75 _mesa_hash_table_remove_key(ctx
->batches
, &batch
->key
);
77 if (ctx
->batch
== batch
)
80 util_unreference_framebuffer_state(&batch
->key
);
84 static struct panfrost_batch
*
85 panfrost_get_batch(struct panfrost_context
*ctx
,
86 const struct pipe_framebuffer_state
*key
)
88 /* Lookup the job first */
89 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->batches
, key
);
94 /* Otherwise, let's create a job */
96 struct panfrost_batch
*batch
= panfrost_create_batch(ctx
, key
);
98 /* Save the created job */
99 _mesa_hash_table_insert(ctx
->batches
, &batch
->key
, batch
);
104 /* Get the job corresponding to the FBO we're currently rendering into */
106 struct panfrost_batch
*
107 panfrost_get_batch_for_fbo(struct panfrost_context
*ctx
)
109 /* If we're wallpapering, we special case to workaround
112 if (ctx
->wallpaper_batch
)
113 return ctx
->wallpaper_batch
;
115 /* If we already began rendering, use that */
118 assert(util_framebuffer_state_equal(&ctx
->batch
->key
,
119 &ctx
->pipe_framebuffer
));
123 /* If not, look up the job */
124 struct panfrost_batch
*batch
= panfrost_get_batch(ctx
,
125 &ctx
->pipe_framebuffer
);
127 /* Set this job as the current FBO job. Will be reset when updating the
128 * FB state and when submitting or releasing a job.
135 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
141 if (_mesa_set_search(batch
->bos
, bo
))
144 panfrost_bo_reference(bo
);
145 _mesa_set_add(batch
->bos
, bo
);
148 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
150 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
151 PAN_BO_ACCESS_VERTEX_TILER
|
152 PAN_BO_ACCESS_FRAGMENT
;
154 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
155 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
156 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
159 if (batch
->key
.zsbuf
) {
160 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
161 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
166 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
167 uint32_t create_flags
, uint32_t access_flags
)
169 struct panfrost_bo
*bo
;
171 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
173 panfrost_batch_add_bo(batch
, bo
, access_flags
);
175 /* panfrost_batch_add_bo() has retained a reference and
176 * panfrost_bo_create() initialize the refcnt to 1, so let's
177 * unreference the BO here so it gets released when the batch is
178 * destroyed (unless it's retained by someone else in the meantime).
180 panfrost_bo_unreference(bo
);
184 /* Returns the polygon list's GPU address if available, or otherwise allocates
185 * the polygon list. It's perfectly fast to use allocate/free BO directly,
186 * since we'll hit the BO cache and this is one-per-batch anyway. */
189 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
191 if (batch
->polygon_list
) {
192 assert(batch
->polygon_list
->size
>= size
);
194 /* Create the BO as invisible, as there's no reason to map */
196 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
198 PAN_BO_ACCESS_PRIVATE
|
200 PAN_BO_ACCESS_VERTEX_TILER
|
201 PAN_BO_ACCESS_FRAGMENT
);
204 return batch
->polygon_list
->gpu
;
208 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
)
210 if (batch
->scratchpad
)
211 return batch
->scratchpad
;
213 batch
->scratchpad
= panfrost_batch_create_bo(batch
, 64 * 4 * 4096,
215 PAN_BO_ACCESS_PRIVATE
|
217 PAN_BO_ACCESS_VERTEX_TILER
|
218 PAN_BO_ACCESS_FRAGMENT
);
219 assert(batch
->scratchpad
);
220 return batch
->scratchpad
;
224 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
226 if (batch
->tiler_heap
)
227 return batch
->tiler_heap
;
229 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
232 PAN_BO_ACCESS_PRIVATE
|
234 PAN_BO_ACCESS_VERTEX_TILER
|
235 PAN_BO_ACCESS_FRAGMENT
);
236 assert(batch
->tiler_heap
);
237 return batch
->tiler_heap
;
241 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
243 if (batch
->tiler_dummy
)
244 return batch
->tiler_dummy
;
246 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
248 PAN_BO_ACCESS_PRIVATE
|
250 PAN_BO_ACCESS_VERTEX_TILER
|
251 PAN_BO_ACCESS_FRAGMENT
);
252 assert(batch
->tiler_dummy
);
253 return batch
->tiler_dummy
;
257 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
259 /* Nothing to reload? TODO: MRT wallpapers */
260 if (batch
->key
.cbufs
[0] == NULL
)
263 /* Check if the buffer has any content on it worth preserving */
265 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
266 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
267 unsigned level
= surf
->u
.tex
.level
;
269 if (!rsrc
->slices
[level
].initialized
)
272 batch
->ctx
->wallpaper_batch
= batch
;
274 /* Clamp the rendering area to the damage extent. The
275 * KHR_partial_update() spec states that trying to render outside of
276 * the damage region is "undefined behavior", so we should be safe.
278 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
279 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
281 if (damage_width
&& damage_height
) {
282 panfrost_batch_intersection_scissor(batch
,
283 rsrc
->damage
.extent
.minx
,
284 rsrc
->damage
.extent
.miny
,
285 rsrc
->damage
.extent
.maxx
,
286 rsrc
->damage
.extent
.maxy
);
289 /* FIXME: Looks like aligning on a tile is not enough, but
290 * aligning on twice the tile size seems to works. We don't
291 * know exactly what happens here but this deserves extra
292 * investigation to figure it out.
294 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
295 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
296 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
298 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
301 struct pipe_scissor_state damage
;
302 struct pipe_box rects
[4];
304 /* Clamp the damage box to the rendering area. */
305 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
306 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
307 damage
.maxx
= MIN2(batch
->maxx
,
308 rsrc
->damage
.biggest_rect
.x
+
309 rsrc
->damage
.biggest_rect
.width
);
310 damage
.maxy
= MIN2(batch
->maxy
,
311 rsrc
->damage
.biggest_rect
.y
+
312 rsrc
->damage
.biggest_rect
.height
);
314 /* One damage rectangle means we can end up with at most 4 reload
316 * 1: left region, only exists if damage.x > 0
317 * 2: right region, only exists if damage.x + damage.width < fb->width
318 * 3: top region, only exists if damage.y > 0. The intersection with
319 * the left and right regions are dropped
320 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
321 * The intersection with the left and right regions are dropped
323 * ____________________________
330 * |_______|___________|______|
332 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
333 batch
->maxy
- batch
->miny
, &rects
[0]);
334 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
335 batch
->maxy
- batch
->miny
, &rects
[1]);
336 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
337 damage
.miny
- batch
->miny
, &rects
[2]);
338 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
339 batch
->maxy
- damage
.maxy
, &rects
[3]);
341 for (unsigned i
= 0; i
< 4; i
++) {
342 /* Width and height are always >= 0 even if width is declared as a
343 * signed integer: u_box_2d() helper takes unsigned args and
344 * panfrost_set_damage_region() is taking care of clamping
347 if (!rects
[i
].width
|| !rects
[i
].height
)
350 /* Blit the wallpaper in */
351 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
353 batch
->ctx
->wallpaper_batch
= NULL
;
357 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
358 mali_ptr first_job_desc
,
361 struct panfrost_context
*ctx
= batch
->ctx
;
362 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
363 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
364 struct drm_panfrost_submit submit
= {0,};
365 uint32_t *bo_handles
;
368 submit
.in_syncs
= (u64
) (uintptr_t) &ctx
->out_sync
;
369 submit
.in_sync_count
= 1;
371 submit
.out_sync
= ctx
->out_sync
;
373 submit
.jc
= first_job_desc
;
374 submit
.requirements
= reqs
;
376 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
379 set_foreach(batch
->bos
, entry
) {
380 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
381 assert(bo
->gem_handle
> 0);
382 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
385 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
386 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
389 fprintf(stderr
, "Error submitting: %m\n");
393 /* Trace the job if we're doing that */
394 if (pan_debug
& PAN_DBG_TRACE
) {
395 /* Wait so we can get errors reported back */
396 drmSyncobjWait(screen
->fd
, &ctx
->out_sync
, 1, INT64_MAX
, 0, NULL
);
397 pandecode_jc(submit
.jc
, FALSE
);
404 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
406 bool has_draws
= batch
->first_job
.gpu
;
410 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0);
414 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
415 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
417 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
425 panfrost_batch_submit(struct panfrost_batch
*batch
)
429 struct panfrost_context
*ctx
= batch
->ctx
;
433 if (!batch
->last_job
.gpu
&& !batch
->clear
)
436 if (!batch
->clear
&& batch
->last_tiler
.gpu
)
437 panfrost_batch_draw_wallpaper(batch
);
439 panfrost_scoreboard_link_batch(batch
);
441 ret
= panfrost_batch_submit_jobs(batch
);
444 fprintf(stderr
, "panfrost_batch_submit failed: %d\n", ret
);
447 /* If this is the bound batch, the panfrost_context parameters are
448 * relevant so submitting it invalidates those paramaters, but if it's
449 * not bound, the context parameters are for some other batch so we
450 * can't invalidate them.
452 if (ctx
->batch
== batch
)
453 panfrost_invalidate_frame(ctx
);
455 /* The job has been submitted, let's invalidate the current FBO job
458 assert(!ctx
->batch
|| batch
== ctx
->batch
);
461 /* We always stall the pipeline for correct results since pipelined
462 * rendering is quite broken right now (to be fixed by the panfrost_job
463 * refactor, just take the perf hit for correctness)
465 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
, &ctx
->out_sync
, 1,
467 panfrost_free_batch(batch
);
471 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
473 struct panfrost_context
*ctx
= batch
->ctx
;
475 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
476 batch
->requirements
|= PAN_REQ_MSAA
;
478 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
479 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
482 /* Helper to smear a 32-bit color across 128-bit components */
485 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
487 for (unsigned i
= 0; i
< 4; ++i
)
492 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
494 for (unsigned i
= 0; i
< 4; i
+= 2) {
501 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
503 /* Alpha magicked to 1.0 if there is no alpha */
505 bool has_alpha
= util_format_has_alpha(format
);
506 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
508 /* Packed color depends on the framebuffer format */
510 const struct util_format_description
*desc
=
511 util_format_description(format
);
513 if (util_format_is_rgba8_variant(desc
)) {
514 pan_pack_color_32(packed
,
515 (float_to_ubyte(clear_alpha
) << 24) |
516 (float_to_ubyte(color
->f
[2]) << 16) |
517 (float_to_ubyte(color
->f
[1]) << 8) |
518 (float_to_ubyte(color
->f
[0]) << 0));
519 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
520 /* First, we convert the components to R5, G6, B5 separately */
521 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
522 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
523 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
525 /* Then we pack into a sparse u32. TODO: Why these shifts? */
526 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
527 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
528 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
529 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
530 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
531 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
532 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
534 /* Pack on *byte* intervals */
535 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
536 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
537 /* Scale as expected but shift oddly */
538 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
539 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
540 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
541 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
543 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
545 /* Try Gallium's generic default path. Doesn't work for all
546 * formats but it's a good guess. */
548 union util_color out
;
550 if (util_format_is_pure_integer(format
)) {
551 memcpy(out
.ui
, color
->ui
, 16);
553 util_pack_color(color
->f
, format
, &out
);
556 unsigned size
= util_format_get_blocksize(format
);
559 unsigned b
= out
.ui
[0];
560 unsigned s
= b
| (b
<< 8);
561 pan_pack_color_32(packed
, s
| (s
<< 16));
562 } else if (size
== 2)
563 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
565 pan_pack_color_32(packed
, out
.ui
[0]);
567 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
569 memcpy(packed
, out
.ui
, 16);
571 unreachable("Unknown generic format size packing clear colour");
576 panfrost_batch_clear(struct panfrost_batch
*batch
,
578 const union pipe_color_union
*color
,
579 double depth
, unsigned stencil
)
581 struct panfrost_context
*ctx
= batch
->ctx
;
583 if (buffers
& PIPE_CLEAR_COLOR
) {
584 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
585 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
588 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
589 pan_pack_color(batch
->clear_color
[i
], color
, format
);
593 if (buffers
& PIPE_CLEAR_DEPTH
) {
594 batch
->clear_depth
= depth
;
597 if (buffers
& PIPE_CLEAR_STENCIL
) {
598 batch
->clear_stencil
= stencil
;
601 batch
->clear
|= buffers
;
603 /* Clearing affects the entire framebuffer (by definition -- this is
604 * the Gallium clear callback, which clears the whole framebuffer. If
605 * the scissor test were enabled from the GL side, the state tracker
606 * would emit a quad instead and we wouldn't go down this code path) */
608 panfrost_batch_union_scissor(batch
, 0, 0,
609 ctx
->pipe_framebuffer
.width
,
610 ctx
->pipe_framebuffer
.height
);
614 panfrost_batch_compare(const void *a
, const void *b
)
616 return util_framebuffer_state_equal(a
, b
);
620 panfrost_batch_hash(const void *key
)
622 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
625 /* Given a new bounding rectangle (scissor), let the job cover the union of the
626 * new and old bounding rectangles */
629 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
630 unsigned minx
, unsigned miny
,
631 unsigned maxx
, unsigned maxy
)
633 batch
->minx
= MIN2(batch
->minx
, minx
);
634 batch
->miny
= MIN2(batch
->miny
, miny
);
635 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
636 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
640 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
641 unsigned minx
, unsigned miny
,
642 unsigned maxx
, unsigned maxy
)
644 batch
->minx
= MAX2(batch
->minx
, minx
);
645 batch
->miny
= MAX2(batch
->miny
, miny
);
646 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
647 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
650 /* Are we currently rendering to the screen (rather than an FBO)? */
653 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
655 /* If there is no color buffer, it's an FBO */
656 if (batch
->key
.nr_cbufs
!= 1)
659 /* If we're too early that no framebuffer was sent, it's scanout */
660 if (!batch
->key
.cbufs
[0])
663 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
664 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
665 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
669 panfrost_batch_init(struct panfrost_context
*ctx
)
671 ctx
->batches
= _mesa_hash_table_create(ctx
,
673 panfrost_batch_compare
);