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_hash_table_create(batch
, _mesa_hash_pointer
,
48 _mesa_key_pointer_equal
);
50 batch
->minx
= batch
->miny
= ~0;
51 batch
->maxx
= batch
->maxy
= 0;
52 batch
->transient_offset
= 0;
54 util_dynarray_init(&batch
->headers
, batch
);
55 util_dynarray_init(&batch
->gpu_headers
, batch
);
56 util_copy_framebuffer_state(&batch
->key
, key
);
62 panfrost_free_batch(struct panfrost_batch
*batch
)
67 struct panfrost_context
*ctx
= batch
->ctx
;
69 hash_table_foreach(batch
->bos
, entry
)
70 panfrost_bo_unreference((struct panfrost_bo
*)entry
->key
);
72 _mesa_hash_table_remove_key(ctx
->batches
, &batch
->key
);
74 if (ctx
->batch
== batch
)
77 util_unreference_framebuffer_state(&batch
->key
);
81 static struct panfrost_batch
*
82 panfrost_get_batch(struct panfrost_context
*ctx
,
83 const struct pipe_framebuffer_state
*key
)
85 /* Lookup the job first */
86 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->batches
, key
);
91 /* Otherwise, let's create a job */
93 struct panfrost_batch
*batch
= panfrost_create_batch(ctx
, key
);
95 /* Save the created job */
96 _mesa_hash_table_insert(ctx
->batches
, &batch
->key
, batch
);
101 /* Get the job corresponding to the FBO we're currently rendering into */
103 struct panfrost_batch
*
104 panfrost_get_batch_for_fbo(struct panfrost_context
*ctx
)
106 /* If we're wallpapering, we special case to workaround
109 if (ctx
->wallpaper_batch
)
110 return ctx
->wallpaper_batch
;
112 /* If we already began rendering, use that */
115 assert(util_framebuffer_state_equal(&ctx
->batch
->key
,
116 &ctx
->pipe_framebuffer
));
120 /* If not, look up the job */
121 struct panfrost_batch
*batch
= panfrost_get_batch(ctx
,
122 &ctx
->pipe_framebuffer
);
124 /* Set this job as the current FBO job. Will be reset when updating the
125 * FB state and when submitting or releasing a job.
132 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
138 struct hash_entry
*entry
;
139 uint32_t old_flags
= 0;
141 entry
= _mesa_hash_table_search(batch
->bos
, bo
);
143 entry
= _mesa_hash_table_insert(batch
->bos
, bo
,
144 (void *)(uintptr_t)flags
);
145 panfrost_bo_reference(bo
);
147 old_flags
= (uintptr_t)entry
->data
;
152 if (old_flags
== flags
)
156 entry
->data
= (void *)(uintptr_t)flags
;
159 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
161 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
162 PAN_BO_ACCESS_VERTEX_TILER
|
163 PAN_BO_ACCESS_FRAGMENT
;
165 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
166 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
167 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
170 if (batch
->key
.zsbuf
) {
171 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
172 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
177 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
178 uint32_t create_flags
, uint32_t access_flags
)
180 struct panfrost_bo
*bo
;
182 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
184 panfrost_batch_add_bo(batch
, bo
, access_flags
);
186 /* panfrost_batch_add_bo() has retained a reference and
187 * panfrost_bo_create() initialize the refcnt to 1, so let's
188 * unreference the BO here so it gets released when the batch is
189 * destroyed (unless it's retained by someone else in the meantime).
191 panfrost_bo_unreference(bo
);
195 /* Returns the polygon list's GPU address if available, or otherwise allocates
196 * the polygon list. It's perfectly fast to use allocate/free BO directly,
197 * since we'll hit the BO cache and this is one-per-batch anyway. */
200 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
202 if (batch
->polygon_list
) {
203 assert(batch
->polygon_list
->size
>= size
);
205 /* Create the BO as invisible, as there's no reason to map */
207 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
209 PAN_BO_ACCESS_PRIVATE
|
211 PAN_BO_ACCESS_VERTEX_TILER
|
212 PAN_BO_ACCESS_FRAGMENT
);
215 return batch
->polygon_list
->gpu
;
219 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
)
221 if (batch
->scratchpad
)
222 return batch
->scratchpad
;
224 batch
->scratchpad
= panfrost_batch_create_bo(batch
, 64 * 4 * 4096,
226 PAN_BO_ACCESS_PRIVATE
|
228 PAN_BO_ACCESS_VERTEX_TILER
|
229 PAN_BO_ACCESS_FRAGMENT
);
230 assert(batch
->scratchpad
);
231 return batch
->scratchpad
;
235 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
237 if (batch
->tiler_heap
)
238 return batch
->tiler_heap
;
240 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
243 PAN_BO_ACCESS_PRIVATE
|
245 PAN_BO_ACCESS_VERTEX_TILER
|
246 PAN_BO_ACCESS_FRAGMENT
);
247 assert(batch
->tiler_heap
);
248 return batch
->tiler_heap
;
252 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
254 if (batch
->tiler_dummy
)
255 return batch
->tiler_dummy
;
257 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
259 PAN_BO_ACCESS_PRIVATE
|
261 PAN_BO_ACCESS_VERTEX_TILER
|
262 PAN_BO_ACCESS_FRAGMENT
);
263 assert(batch
->tiler_dummy
);
264 return batch
->tiler_dummy
;
268 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
270 /* Nothing to reload? TODO: MRT wallpapers */
271 if (batch
->key
.cbufs
[0] == NULL
)
274 /* Check if the buffer has any content on it worth preserving */
276 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
277 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
278 unsigned level
= surf
->u
.tex
.level
;
280 if (!rsrc
->slices
[level
].initialized
)
283 batch
->ctx
->wallpaper_batch
= batch
;
285 /* Clamp the rendering area to the damage extent. The
286 * KHR_partial_update() spec states that trying to render outside of
287 * the damage region is "undefined behavior", so we should be safe.
289 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
290 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
292 if (damage_width
&& damage_height
) {
293 panfrost_batch_intersection_scissor(batch
,
294 rsrc
->damage
.extent
.minx
,
295 rsrc
->damage
.extent
.miny
,
296 rsrc
->damage
.extent
.maxx
,
297 rsrc
->damage
.extent
.maxy
);
300 /* FIXME: Looks like aligning on a tile is not enough, but
301 * aligning on twice the tile size seems to works. We don't
302 * know exactly what happens here but this deserves extra
303 * investigation to figure it out.
305 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
306 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
307 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
309 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
312 struct pipe_scissor_state damage
;
313 struct pipe_box rects
[4];
315 /* Clamp the damage box to the rendering area. */
316 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
317 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
318 damage
.maxx
= MIN2(batch
->maxx
,
319 rsrc
->damage
.biggest_rect
.x
+
320 rsrc
->damage
.biggest_rect
.width
);
321 damage
.maxy
= MIN2(batch
->maxy
,
322 rsrc
->damage
.biggest_rect
.y
+
323 rsrc
->damage
.biggest_rect
.height
);
325 /* One damage rectangle means we can end up with at most 4 reload
327 * 1: left region, only exists if damage.x > 0
328 * 2: right region, only exists if damage.x + damage.width < fb->width
329 * 3: top region, only exists if damage.y > 0. The intersection with
330 * the left and right regions are dropped
331 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
332 * The intersection with the left and right regions are dropped
334 * ____________________________
341 * |_______|___________|______|
343 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
344 batch
->maxy
- batch
->miny
, &rects
[0]);
345 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
346 batch
->maxy
- batch
->miny
, &rects
[1]);
347 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
348 damage
.miny
- batch
->miny
, &rects
[2]);
349 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
350 batch
->maxy
- damage
.maxy
, &rects
[3]);
352 for (unsigned i
= 0; i
< 4; i
++) {
353 /* Width and height are always >= 0 even if width is declared as a
354 * signed integer: u_box_2d() helper takes unsigned args and
355 * panfrost_set_damage_region() is taking care of clamping
358 if (!rects
[i
].width
|| !rects
[i
].height
)
361 /* Blit the wallpaper in */
362 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
364 batch
->ctx
->wallpaper_batch
= NULL
;
368 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
369 mali_ptr first_job_desc
,
372 struct panfrost_context
*ctx
= batch
->ctx
;
373 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
374 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
375 struct drm_panfrost_submit submit
= {0,};
376 uint32_t *bo_handles
;
379 submit
.in_syncs
= (u64
) (uintptr_t) &ctx
->out_sync
;
380 submit
.in_sync_count
= 1;
382 submit
.out_sync
= ctx
->out_sync
;
384 submit
.jc
= first_job_desc
;
385 submit
.requirements
= reqs
;
387 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
390 hash_table_foreach(batch
->bos
, entry
) {
391 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
392 assert(bo
->gem_handle
> 0);
393 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
396 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
397 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
400 fprintf(stderr
, "Error submitting: %m\n");
404 /* Trace the job if we're doing that */
405 if (pan_debug
& PAN_DBG_TRACE
) {
406 /* Wait so we can get errors reported back */
407 drmSyncobjWait(screen
->fd
, &ctx
->out_sync
, 1, INT64_MAX
, 0, NULL
);
408 pandecode_jc(submit
.jc
, FALSE
);
415 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
417 bool has_draws
= batch
->first_job
.gpu
;
421 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0);
425 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
426 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
428 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
436 panfrost_batch_submit(struct panfrost_batch
*batch
)
440 struct panfrost_context
*ctx
= batch
->ctx
;
444 if (!batch
->last_job
.gpu
&& !batch
->clear
)
447 if (!batch
->clear
&& batch
->last_tiler
.gpu
)
448 panfrost_batch_draw_wallpaper(batch
);
450 panfrost_scoreboard_link_batch(batch
);
452 ret
= panfrost_batch_submit_jobs(batch
);
455 fprintf(stderr
, "panfrost_batch_submit failed: %d\n", ret
);
458 /* If this is the bound batch, the panfrost_context parameters are
459 * relevant so submitting it invalidates those paramaters, but if it's
460 * not bound, the context parameters are for some other batch so we
461 * can't invalidate them.
463 if (ctx
->batch
== batch
)
464 panfrost_invalidate_frame(ctx
);
466 /* The job has been submitted, let's invalidate the current FBO job
469 assert(!ctx
->batch
|| batch
== ctx
->batch
);
472 /* We always stall the pipeline for correct results since pipelined
473 * rendering is quite broken right now (to be fixed by the panfrost_job
474 * refactor, just take the perf hit for correctness)
476 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
, &ctx
->out_sync
, 1,
478 panfrost_free_batch(batch
);
482 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
484 struct panfrost_context
*ctx
= batch
->ctx
;
486 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
487 batch
->requirements
|= PAN_REQ_MSAA
;
489 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
490 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
493 /* Helper to smear a 32-bit color across 128-bit components */
496 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
498 for (unsigned i
= 0; i
< 4; ++i
)
503 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
505 for (unsigned i
= 0; i
< 4; i
+= 2) {
512 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
514 /* Alpha magicked to 1.0 if there is no alpha */
516 bool has_alpha
= util_format_has_alpha(format
);
517 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
519 /* Packed color depends on the framebuffer format */
521 const struct util_format_description
*desc
=
522 util_format_description(format
);
524 if (util_format_is_rgba8_variant(desc
)) {
525 pan_pack_color_32(packed
,
526 (float_to_ubyte(clear_alpha
) << 24) |
527 (float_to_ubyte(color
->f
[2]) << 16) |
528 (float_to_ubyte(color
->f
[1]) << 8) |
529 (float_to_ubyte(color
->f
[0]) << 0));
530 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
531 /* First, we convert the components to R5, G6, B5 separately */
532 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
533 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
534 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
536 /* Then we pack into a sparse u32. TODO: Why these shifts? */
537 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
538 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
539 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
540 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
541 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
542 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
543 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
545 /* Pack on *byte* intervals */
546 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
547 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
548 /* Scale as expected but shift oddly */
549 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
550 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
551 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
552 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
554 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
556 /* Try Gallium's generic default path. Doesn't work for all
557 * formats but it's a good guess. */
559 union util_color out
;
561 if (util_format_is_pure_integer(format
)) {
562 memcpy(out
.ui
, color
->ui
, 16);
564 util_pack_color(color
->f
, format
, &out
);
567 unsigned size
= util_format_get_blocksize(format
);
570 unsigned b
= out
.ui
[0];
571 unsigned s
= b
| (b
<< 8);
572 pan_pack_color_32(packed
, s
| (s
<< 16));
573 } else if (size
== 2)
574 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
576 pan_pack_color_32(packed
, out
.ui
[0]);
578 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
580 memcpy(packed
, out
.ui
, 16);
582 unreachable("Unknown generic format size packing clear colour");
587 panfrost_batch_clear(struct panfrost_batch
*batch
,
589 const union pipe_color_union
*color
,
590 double depth
, unsigned stencil
)
592 struct panfrost_context
*ctx
= batch
->ctx
;
594 if (buffers
& PIPE_CLEAR_COLOR
) {
595 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
596 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
599 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
600 pan_pack_color(batch
->clear_color
[i
], color
, format
);
604 if (buffers
& PIPE_CLEAR_DEPTH
) {
605 batch
->clear_depth
= depth
;
608 if (buffers
& PIPE_CLEAR_STENCIL
) {
609 batch
->clear_stencil
= stencil
;
612 batch
->clear
|= buffers
;
614 /* Clearing affects the entire framebuffer (by definition -- this is
615 * the Gallium clear callback, which clears the whole framebuffer. If
616 * the scissor test were enabled from the GL side, the state tracker
617 * would emit a quad instead and we wouldn't go down this code path) */
619 panfrost_batch_union_scissor(batch
, 0, 0,
620 ctx
->pipe_framebuffer
.width
,
621 ctx
->pipe_framebuffer
.height
);
625 panfrost_batch_compare(const void *a
, const void *b
)
627 return util_framebuffer_state_equal(a
, b
);
631 panfrost_batch_hash(const void *key
)
633 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
636 /* Given a new bounding rectangle (scissor), let the job cover the union of the
637 * new and old bounding rectangles */
640 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
641 unsigned minx
, unsigned miny
,
642 unsigned maxx
, unsigned maxy
)
644 batch
->minx
= MIN2(batch
->minx
, minx
);
645 batch
->miny
= MIN2(batch
->miny
, miny
);
646 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
647 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
651 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
652 unsigned minx
, unsigned miny
,
653 unsigned maxx
, unsigned maxy
)
655 batch
->minx
= MAX2(batch
->minx
, minx
);
656 batch
->miny
= MAX2(batch
->miny
, miny
);
657 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
658 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
661 /* Are we currently rendering to the screen (rather than an FBO)? */
664 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
666 /* If there is no color buffer, it's an FBO */
667 if (batch
->key
.nr_cbufs
!= 1)
670 /* If we're too early that no framebuffer was sent, it's scanout */
671 if (!batch
->key
.cbufs
[0])
674 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
675 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
676 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
680 panfrost_batch_init(struct panfrost_context
*ctx
)
682 ctx
->batches
= _mesa_hash_table_create(ctx
,
684 panfrost_batch_compare
);