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
)
140 if (_mesa_set_search(batch
->bos
, bo
))
143 panfrost_bo_reference(bo
);
144 _mesa_set_add(batch
->bos
, bo
);
147 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
149 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
150 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
151 panfrost_batch_add_bo(batch
, rsrc
->bo
);
154 if (batch
->key
.zsbuf
) {
155 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
156 panfrost_batch_add_bo(batch
, rsrc
->bo
);
161 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
162 uint32_t create_flags
)
164 struct panfrost_bo
*bo
;
166 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
168 panfrost_batch_add_bo(batch
, bo
);
170 /* panfrost_batch_add_bo() has retained a reference and
171 * panfrost_bo_create() initialize the refcnt to 1, so let's
172 * unreference the BO here so it gets released when the batch is
173 * destroyed (unless it's retained by someone else in the meantime).
175 panfrost_bo_unreference(bo
);
179 /* Returns the polygon list's GPU address if available, or otherwise allocates
180 * the polygon list. It's perfectly fast to use allocate/free BO directly,
181 * since we'll hit the BO cache and this is one-per-batch anyway. */
184 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
186 if (batch
->polygon_list
) {
187 assert(batch
->polygon_list
->size
>= size
);
189 /* Create the BO as invisible, as there's no reason to map */
191 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
195 return batch
->polygon_list
->gpu
;
199 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
)
201 if (batch
->scratchpad
)
202 return batch
->scratchpad
;
204 batch
->scratchpad
= panfrost_batch_create_bo(batch
, 64 * 4 * 4096,
206 assert(batch
->scratchpad
);
207 return batch
->scratchpad
;
211 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
213 if (batch
->tiler_heap
)
214 return batch
->tiler_heap
;
216 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
219 assert(batch
->tiler_heap
);
220 return batch
->tiler_heap
;
224 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
226 if (batch
->tiler_dummy
)
227 return batch
->tiler_dummy
;
229 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
231 assert(batch
->tiler_dummy
);
232 return batch
->tiler_dummy
;
236 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
238 /* Nothing to reload? TODO: MRT wallpapers */
239 if (batch
->key
.cbufs
[0] == NULL
)
242 /* Check if the buffer has any content on it worth preserving */
244 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
245 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
246 unsigned level
= surf
->u
.tex
.level
;
248 if (!rsrc
->slices
[level
].initialized
)
251 batch
->ctx
->wallpaper_batch
= batch
;
253 /* Clamp the rendering area to the damage extent. The
254 * KHR_partial_update() spec states that trying to render outside of
255 * the damage region is "undefined behavior", so we should be safe.
257 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
258 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
260 if (damage_width
&& damage_height
) {
261 panfrost_batch_intersection_scissor(batch
,
262 rsrc
->damage
.extent
.minx
,
263 rsrc
->damage
.extent
.miny
,
264 rsrc
->damage
.extent
.maxx
,
265 rsrc
->damage
.extent
.maxy
);
268 /* FIXME: Looks like aligning on a tile is not enough, but
269 * aligning on twice the tile size seems to works. We don't
270 * know exactly what happens here but this deserves extra
271 * investigation to figure it out.
273 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
274 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
275 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
277 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
280 struct pipe_scissor_state damage
;
281 struct pipe_box rects
[4];
283 /* Clamp the damage box to the rendering area. */
284 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
285 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
286 damage
.maxx
= MIN2(batch
->maxx
,
287 rsrc
->damage
.biggest_rect
.x
+
288 rsrc
->damage
.biggest_rect
.width
);
289 damage
.maxy
= MIN2(batch
->maxy
,
290 rsrc
->damage
.biggest_rect
.y
+
291 rsrc
->damage
.biggest_rect
.height
);
293 /* One damage rectangle means we can end up with at most 4 reload
295 * 1: left region, only exists if damage.x > 0
296 * 2: right region, only exists if damage.x + damage.width < fb->width
297 * 3: top region, only exists if damage.y > 0. The intersection with
298 * the left and right regions are dropped
299 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
300 * The intersection with the left and right regions are dropped
302 * ____________________________
309 * |_______|___________|______|
311 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
312 batch
->maxy
- batch
->miny
, &rects
[0]);
313 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
314 batch
->maxy
- batch
->miny
, &rects
[1]);
315 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
316 damage
.miny
- batch
->miny
, &rects
[2]);
317 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
318 batch
->maxy
- damage
.maxy
, &rects
[3]);
320 for (unsigned i
= 0; i
< 4; i
++) {
321 /* Width and height are always >= 0 even if width is declared as a
322 * signed integer: u_box_2d() helper takes unsigned args and
323 * panfrost_set_damage_region() is taking care of clamping
326 if (!rects
[i
].width
|| !rects
[i
].height
)
329 /* Blit the wallpaper in */
330 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
332 batch
->ctx
->wallpaper_batch
= NULL
;
336 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
337 mali_ptr first_job_desc
,
340 struct panfrost_context
*ctx
= batch
->ctx
;
341 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
342 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
343 struct drm_panfrost_submit submit
= {0,};
344 uint32_t *bo_handles
;
347 submit
.in_syncs
= (u64
) (uintptr_t) &ctx
->out_sync
;
348 submit
.in_sync_count
= 1;
350 submit
.out_sync
= ctx
->out_sync
;
352 submit
.jc
= first_job_desc
;
353 submit
.requirements
= reqs
;
355 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
358 set_foreach(batch
->bos
, entry
) {
359 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
360 assert(bo
->gem_handle
> 0);
361 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
364 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
365 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
368 fprintf(stderr
, "Error submitting: %m\n");
372 /* Trace the job if we're doing that */
373 if (pan_debug
& PAN_DBG_TRACE
) {
374 /* Wait so we can get errors reported back */
375 drmSyncobjWait(screen
->fd
, &ctx
->out_sync
, 1, INT64_MAX
, 0, NULL
);
376 pandecode_jc(submit
.jc
, FALSE
);
383 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
385 bool has_draws
= batch
->first_job
.gpu
;
389 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0);
393 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
394 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
396 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
404 panfrost_batch_submit(struct panfrost_batch
*batch
)
408 struct panfrost_context
*ctx
= batch
->ctx
;
412 if (!batch
->last_job
.gpu
&& !batch
->clear
)
415 if (!batch
->clear
&& batch
->last_tiler
.gpu
)
416 panfrost_batch_draw_wallpaper(batch
);
418 panfrost_scoreboard_link_batch(batch
);
420 ret
= panfrost_batch_submit_jobs(batch
);
423 fprintf(stderr
, "panfrost_batch_submit failed: %d\n", ret
);
426 /* If this is the bound batch, the panfrost_context parameters are
427 * relevant so submitting it invalidates those paramaters, but if it's
428 * not bound, the context parameters are for some other batch so we
429 * can't invalidate them.
431 if (ctx
->batch
== batch
)
432 panfrost_invalidate_frame(ctx
);
434 /* The job has been submitted, let's invalidate the current FBO job
437 assert(!ctx
->batch
|| batch
== ctx
->batch
);
440 /* We always stall the pipeline for correct results since pipelined
441 * rendering is quite broken right now (to be fixed by the panfrost_job
442 * refactor, just take the perf hit for correctness)
444 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
, &ctx
->out_sync
, 1,
446 panfrost_free_batch(batch
);
450 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
452 struct panfrost_context
*ctx
= batch
->ctx
;
454 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
455 batch
->requirements
|= PAN_REQ_MSAA
;
457 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
458 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
461 /* Helper to smear a 32-bit color across 128-bit components */
464 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
466 for (unsigned i
= 0; i
< 4; ++i
)
471 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
473 for (unsigned i
= 0; i
< 4; i
+= 2) {
480 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
482 /* Alpha magicked to 1.0 if there is no alpha */
484 bool has_alpha
= util_format_has_alpha(format
);
485 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
487 /* Packed color depends on the framebuffer format */
489 const struct util_format_description
*desc
=
490 util_format_description(format
);
492 if (util_format_is_rgba8_variant(desc
)) {
493 pan_pack_color_32(packed
,
494 (float_to_ubyte(clear_alpha
) << 24) |
495 (float_to_ubyte(color
->f
[2]) << 16) |
496 (float_to_ubyte(color
->f
[1]) << 8) |
497 (float_to_ubyte(color
->f
[0]) << 0));
498 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
499 /* First, we convert the components to R5, G6, B5 separately */
500 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
501 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
502 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
504 /* Then we pack into a sparse u32. TODO: Why these shifts? */
505 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
506 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
507 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
508 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
509 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
510 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
511 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
513 /* Pack on *byte* intervals */
514 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
515 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
516 /* Scale as expected but shift oddly */
517 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
518 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
519 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
520 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
522 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
524 /* Try Gallium's generic default path. Doesn't work for all
525 * formats but it's a good guess. */
527 union util_color out
;
529 if (util_format_is_pure_integer(format
)) {
530 memcpy(out
.ui
, color
->ui
, 16);
532 util_pack_color(color
->f
, format
, &out
);
535 unsigned size
= util_format_get_blocksize(format
);
538 unsigned b
= out
.ui
[0];
539 unsigned s
= b
| (b
<< 8);
540 pan_pack_color_32(packed
, s
| (s
<< 16));
541 } else if (size
== 2)
542 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
544 pan_pack_color_32(packed
, out
.ui
[0]);
546 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
548 memcpy(packed
, out
.ui
, 16);
550 unreachable("Unknown generic format size packing clear colour");
555 panfrost_batch_clear(struct panfrost_batch
*batch
,
557 const union pipe_color_union
*color
,
558 double depth
, unsigned stencil
)
560 struct panfrost_context
*ctx
= batch
->ctx
;
562 if (buffers
& PIPE_CLEAR_COLOR
) {
563 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
564 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
567 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
568 pan_pack_color(batch
->clear_color
[i
], color
, format
);
572 if (buffers
& PIPE_CLEAR_DEPTH
) {
573 batch
->clear_depth
= depth
;
576 if (buffers
& PIPE_CLEAR_STENCIL
) {
577 batch
->clear_stencil
= stencil
;
580 batch
->clear
|= buffers
;
582 /* Clearing affects the entire framebuffer (by definition -- this is
583 * the Gallium clear callback, which clears the whole framebuffer. If
584 * the scissor test were enabled from the GL side, the state tracker
585 * would emit a quad instead and we wouldn't go down this code path) */
587 panfrost_batch_union_scissor(batch
, 0, 0,
588 ctx
->pipe_framebuffer
.width
,
589 ctx
->pipe_framebuffer
.height
);
593 panfrost_batch_compare(const void *a
, const void *b
)
595 return util_framebuffer_state_equal(a
, b
);
599 panfrost_batch_hash(const void *key
)
601 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
604 /* Given a new bounding rectangle (scissor), let the job cover the union of the
605 * new and old bounding rectangles */
608 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
609 unsigned minx
, unsigned miny
,
610 unsigned maxx
, unsigned maxy
)
612 batch
->minx
= MIN2(batch
->minx
, minx
);
613 batch
->miny
= MIN2(batch
->miny
, miny
);
614 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
615 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
619 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
620 unsigned minx
, unsigned miny
,
621 unsigned maxx
, unsigned maxy
)
623 batch
->minx
= MAX2(batch
->minx
, minx
);
624 batch
->miny
= MAX2(batch
->miny
, miny
);
625 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
626 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
629 /* Are we currently rendering to the screen (rather than an FBO)? */
632 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
634 /* If there is no color buffer, it's an FBO */
635 if (batch
->key
.nr_cbufs
!= 1)
638 /* If we're too early that no framebuffer was sent, it's scanout */
639 if (!batch
->key
.cbufs
[0])
642 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
643 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
644 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
648 panfrost_batch_init(struct panfrost_context
*ctx
)
650 ctx
->batches
= _mesa_hash_table_create(ctx
,
652 panfrost_batch_compare
);