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 "pan_context.h"
29 #include "util/hash_table.h"
30 #include "util/ralloc.h"
31 #include "util/u_format.h"
32 #include "util/u_pack_color.h"
35 panfrost_create_job(struct panfrost_context
*ctx
)
37 struct panfrost_job
*job
= rzalloc(ctx
, struct panfrost_job
);
41 job
->bos
= _mesa_set_create(job
,
43 _mesa_key_pointer_equal
);
45 job
->minx
= job
->miny
= ~0;
46 job
->maxx
= job
->maxy
= 0;
47 job
->transient_offset
= 0;
49 util_dynarray_init(&job
->headers
, job
);
50 util_dynarray_init(&job
->gpu_headers
, job
);
51 util_dynarray_init(&job
->transient_indices
, job
);
57 panfrost_free_job(struct panfrost_context
*ctx
, struct panfrost_job
*job
)
62 set_foreach(job
->bos
, entry
) {
63 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
64 panfrost_bo_unreference(ctx
->base
.screen
, bo
);
67 /* Free up the transient BOs we're sitting on */
68 struct panfrost_screen
*screen
= pan_screen(ctx
->base
.screen
);
70 util_dynarray_foreach(&job
->transient_indices
, unsigned, index
) {
72 BITSET_SET(screen
->free_transient
, *index
);
75 /* Unreference the polygon list */
76 panfrost_bo_unreference(ctx
->base
.screen
, job
->polygon_list
);
78 _mesa_hash_table_remove_key(ctx
->jobs
, &job
->key
);
87 panfrost_get_job(struct panfrost_context
*ctx
,
88 struct pipe_surface
**cbufs
, struct pipe_surface
*zsbuf
)
90 /* Lookup the job first */
92 struct panfrost_job_key key
= {
102 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->jobs
, &key
);
107 /* Otherwise, let's create a job */
109 struct panfrost_job
*job
= panfrost_create_job(ctx
);
111 /* Save the created job */
113 memcpy(&job
->key
, &key
, sizeof(key
));
114 _mesa_hash_table_insert(ctx
->jobs
, &job
->key
, job
);
119 /* Get the job corresponding to the FBO we're currently rendering into */
121 struct panfrost_job
*
122 panfrost_get_job_for_fbo(struct panfrost_context
*ctx
)
124 /* If we're wallpapering, we special case to workaround
127 if (ctx
->wallpaper_batch
)
128 return ctx
->wallpaper_batch
;
130 /* If we already began rendering, use that */
133 assert(ctx
->job
->key
.zsbuf
== ctx
->pipe_framebuffer
.zsbuf
&&
134 !memcmp(ctx
->job
->key
.cbufs
,
135 ctx
->pipe_framebuffer
.cbufs
,
136 sizeof(ctx
->job
->key
.cbufs
)));
140 /* If not, look up the job */
142 struct pipe_surface
**cbufs
= ctx
->pipe_framebuffer
.cbufs
;
143 struct pipe_surface
*zsbuf
= ctx
->pipe_framebuffer
.zsbuf
;
144 struct panfrost_job
*job
= panfrost_get_job(ctx
, cbufs
, zsbuf
);
146 /* Set this job as the current FBO job. Will be reset when updating the
147 * FB state and when submitting or releasing a job.
154 panfrost_job_add_bo(struct panfrost_job
*job
, struct panfrost_bo
*bo
)
159 if (_mesa_set_search(job
->bos
, bo
))
162 panfrost_bo_reference(bo
);
163 _mesa_set_add(job
->bos
, bo
);
166 /* Returns the polygon list's GPU address if available, or otherwise allocates
167 * the polygon list. It's perfectly fast to use allocate/free BO directly,
168 * since we'll hit the BO cache and this is one-per-batch anyway. */
171 panfrost_job_get_polygon_list(struct panfrost_job
*batch
, unsigned size
)
173 if (batch
->polygon_list
) {
174 assert(batch
->polygon_list
->size
>= size
);
176 struct panfrost_screen
*screen
= pan_screen(batch
->ctx
->base
.screen
);
178 /* Create the BO as invisible, as there's no reason to map */
180 batch
->polygon_list
= panfrost_drm_create_bo(screen
,
181 size
, PAN_ALLOCATE_INVISIBLE
);
184 return batch
->polygon_list
->gpu
;
188 panfrost_flush_jobs_writing_resource(struct panfrost_context
*panfrost
,
189 struct pipe_resource
*prsc
)
192 struct hash_entry
*entry
= _mesa_hash_table_search(panfrost
->write_jobs
,
195 struct panfrost_job
*job
= entry
->data
;
196 panfrost_job_submit(panfrost
, job
);
203 panfrost_job_submit(struct panfrost_context
*ctx
, struct panfrost_job
*job
)
208 panfrost_scoreboard_link_batch(job
);
210 bool has_draws
= job
->last_job
.gpu
;
211 bool is_scanout
= panfrost_is_scanout(ctx
);
213 ret
= panfrost_drm_submit_vs_fs_job(ctx
, has_draws
, is_scanout
);
216 fprintf(stderr
, "panfrost_job_submit failed: %d\n", ret
);
218 /* The job has been submitted, let's invalidate the current FBO job
221 assert(!ctx
->job
|| job
== ctx
->job
);
224 /* Remove the job from the ctx->jobs set so that future
225 * panfrost_get_job() calls don't see it.
226 * We must reset the job key to avoid removing another valid entry when
229 _mesa_hash_table_remove_key(ctx
->jobs
, &job
->key
);
230 memset(&job
->key
, 0, sizeof(job
->key
));
234 panfrost_job_set_requirements(struct panfrost_context
*ctx
,
235 struct panfrost_job
*job
)
237 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
238 job
->requirements
|= PAN_REQ_MSAA
;
240 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
241 job
->requirements
|= PAN_REQ_DEPTH_WRITE
;
244 /* Helper to smear a 32-bit color across 128-bit components */
247 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
249 for (unsigned i
= 0; i
< 4; ++i
)
254 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
256 for (unsigned i
= 0; i
< 4; i
+= 2) {
263 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
265 /* Alpha magicked to 1.0 if there is no alpha */
267 bool has_alpha
= util_format_has_alpha(format
);
268 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
270 /* Packed color depends on the framebuffer format */
272 const struct util_format_description
*desc
=
273 util_format_description(format
);
275 if (util_format_is_rgba8_variant(desc
)) {
276 pan_pack_color_32(packed
,
277 (float_to_ubyte(clear_alpha
) << 24) |
278 (float_to_ubyte(color
->f
[2]) << 16) |
279 (float_to_ubyte(color
->f
[1]) << 8) |
280 (float_to_ubyte(color
->f
[0]) << 0));
281 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
282 /* First, we convert the components to R5, G6, B5 separately */
283 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
284 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
285 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
287 /* Then we pack into a sparse u32. TODO: Why these shifts? */
288 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
289 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
290 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
291 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
292 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
293 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
294 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
296 /* Pack on *byte* intervals */
297 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
298 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
299 /* Scale as expected but shift oddly */
300 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
301 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
302 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
303 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
305 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
307 /* Try Gallium's generic default path. Doesn't work for all
308 * formats but it's a good guess. */
310 union util_color out
;
312 if (util_format_is_pure_integer(format
)) {
313 memcpy(out
.ui
, color
->ui
, 16);
315 util_pack_color(color
->f
, format
, &out
);
318 unsigned size
= util_format_get_blocksize(format
);
321 unsigned b
= out
.ui
[0];
322 unsigned s
= b
| (b
<< 8);
323 pan_pack_color_32(packed
, s
| (s
<< 16));
324 } else if (size
== 2)
325 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
327 pan_pack_color_32(packed
, out
.ui
[0]);
329 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
331 memcpy(packed
, out
.ui
, 16);
333 unreachable("Unknown generic format size packing clear colour");
338 panfrost_job_clear(struct panfrost_context
*ctx
,
339 struct panfrost_job
*job
,
341 const union pipe_color_union
*color
,
342 double depth
, unsigned stencil
)
345 if (buffers
& PIPE_CLEAR_COLOR
) {
346 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
347 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
350 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
351 pan_pack_color(job
->clear_color
[i
], color
, format
);
355 if (buffers
& PIPE_CLEAR_DEPTH
) {
356 job
->clear_depth
= depth
;
359 if (buffers
& PIPE_CLEAR_STENCIL
) {
360 job
->clear_stencil
= stencil
;
363 job
->clear
|= buffers
;
365 /* Clearing affects the entire framebuffer (by definition -- this is
366 * the Gallium clear callback, which clears the whole framebuffer. If
367 * the scissor test were enabled from the GL side, the state tracker
368 * would emit a quad instead and we wouldn't go down this code path) */
370 panfrost_job_union_scissor(job
, 0, 0,
371 ctx
->pipe_framebuffer
.width
,
372 ctx
->pipe_framebuffer
.height
);
376 panfrost_flush_jobs_reading_resource(struct panfrost_context
*panfrost
,
377 struct pipe_resource
*prsc
)
379 struct panfrost_resource
*rsc
= pan_resource(prsc
);
381 panfrost_flush_jobs_writing_resource(panfrost
, prsc
);
383 hash_table_foreach(panfrost
->jobs
, entry
) {
384 struct panfrost_job
*job
= entry
->data
;
386 if (_mesa_set_search(job
->bos
, rsc
->bo
)) {
387 printf("TODO: submit job for flush\n");
388 //panfrost_job_submit(panfrost, job);
395 panfrost_job_compare(const void *a
, const void *b
)
397 return memcmp(a
, b
, sizeof(struct panfrost_job_key
)) == 0;
401 panfrost_job_hash(const void *key
)
403 return _mesa_hash_data(key
, sizeof(struct panfrost_job_key
));
406 /* Given a new bounding rectangle (scissor), let the job cover the union of the
407 * new and old bounding rectangles */
410 panfrost_job_union_scissor(struct panfrost_job
*job
,
411 unsigned minx
, unsigned miny
,
412 unsigned maxx
, unsigned maxy
)
414 job
->minx
= MIN2(job
->minx
, minx
);
415 job
->miny
= MIN2(job
->miny
, miny
);
416 job
->maxx
= MAX2(job
->maxx
, maxx
);
417 job
->maxy
= MAX2(job
->maxy
, maxy
);
421 panfrost_job_intersection_scissor(struct panfrost_job
*job
,
422 unsigned minx
, unsigned miny
,
423 unsigned maxx
, unsigned maxy
)
425 job
->minx
= MAX2(job
->minx
, minx
);
426 job
->miny
= MAX2(job
->miny
, miny
);
427 job
->maxx
= MIN2(job
->maxx
, maxx
);
428 job
->maxy
= MIN2(job
->maxy
, maxy
);
432 panfrost_job_init(struct panfrost_context
*ctx
)
434 ctx
->jobs
= _mesa_hash_table_create(ctx
,
436 panfrost_job_compare
);
438 ctx
->write_jobs
= _mesa_hash_table_create(ctx
,
440 _mesa_key_pointer_equal
);