2 * Copyright © 2017 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included
12 * in all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
24 #include "pipe/p_defines.h"
25 #include "pipe/p_state.h"
26 #include "pipe/p_context.h"
27 #include "pipe/p_screen.h"
28 #include "util/format/u_format.h"
29 #include "util/u_inlines.h"
30 #include "util/ralloc.h"
31 #include "intel/blorp/blorp.h"
32 #include "iris_context.h"
33 #include "iris_resource.h"
34 #include "iris_screen.h"
37 * Helper function for handling mirror image blits.
39 * If coord0 > coord1, swap them and return "true" (mirrored).
42 apply_mirror(float *coord0
, float *coord1
)
44 if (*coord0
> *coord1
) {
54 * Compute the number of pixels to clip for each side of a rect
56 * \param x0 The rect's left coordinate
57 * \param y0 The rect's bottom coordinate
58 * \param x1 The rect's right coordinate
59 * \param y1 The rect's top coordinate
60 * \param min_x The clipping region's left coordinate
61 * \param min_y The clipping region's bottom coordinate
62 * \param max_x The clipping region's right coordinate
63 * \param max_y The clipping region's top coordinate
64 * \param clipped_x0 The number of pixels to clip from the left side
65 * \param clipped_y0 The number of pixels to clip from the bottom side
66 * \param clipped_x1 The number of pixels to clip from the right side
67 * \param clipped_y1 The number of pixels to clip from the top side
69 * \return false if we clip everything away, true otherwise
72 compute_pixels_clipped(float x0
, float y0
, float x1
, float y1
,
73 float min_x
, float min_y
, float max_x
, float max_y
,
74 float *clipped_x0
, float *clipped_y0
,
75 float *clipped_x1
, float *clipped_y1
)
77 /* If we are going to clip everything away, stop. */
78 if (!(min_x
<= max_x
&&
90 *clipped_x0
= min_x
- x0
;
94 *clipped_x1
= x1
- max_x
;
99 *clipped_y0
= min_y
- y0
;
103 *clipped_y1
= y1
- max_y
;
111 * Clips a coordinate (left, right, top or bottom) for the src or dst rect
112 * (whichever requires the largest clip) and adjusts the coordinate
113 * for the other rect accordingly.
115 * \param mirror true if mirroring is required
116 * \param src the source rect coordinate (for example src_x0)
117 * \param dst0 the dst rect coordinate (for example dst_x0)
118 * \param dst1 the opposite dst rect coordinate (for example dst_x1)
119 * \param clipped_dst0 number of pixels to clip from the dst coordinate
120 * \param clipped_dst1 number of pixels to clip from the opposite dst coordinate
121 * \param scale the src vs dst scale involved for that coordinate
122 * \param is_left_or_bottom true if we are clipping the left or bottom sides
126 clip_coordinates(bool mirror
,
127 float *src
, float *dst0
, float *dst1
,
131 bool is_left_or_bottom
)
133 /* When clipping we need to add or subtract pixels from the original
134 * coordinates depending on whether we are acting on the left/bottom
135 * or right/top sides of the rect respectively. We assume we have to
136 * add them in the code below, and multiply by -1 when we should
139 int mult
= is_left_or_bottom
? 1 : -1;
142 *dst0
+= clipped_dst0
* mult
;
143 *src
+= clipped_dst0
* scale
* mult
;
145 *dst1
-= clipped_dst1
* mult
;
146 *src
+= clipped_dst1
* scale
* mult
;
151 * Apply a scissor rectangle to blit coordinates.
153 * Returns true if the blit was entirely scissored away.
156 apply_blit_scissor(const struct pipe_scissor_state
*scissor
,
157 float *src_x0
, float *src_y0
,
158 float *src_x1
, float *src_y1
,
159 float *dst_x0
, float *dst_y0
,
160 float *dst_x1
, float *dst_y1
,
161 bool mirror_x
, bool mirror_y
)
163 float clip_dst_x0
, clip_dst_x1
, clip_dst_y0
, clip_dst_y1
;
165 /* Compute number of pixels to scissor away. */
166 if (!compute_pixels_clipped(*dst_x0
, *dst_y0
, *dst_x1
, *dst_y1
,
167 scissor
->minx
, scissor
->miny
,
168 scissor
->maxx
, scissor
->maxy
,
169 &clip_dst_x0
, &clip_dst_y0
,
170 &clip_dst_x1
, &clip_dst_y1
))
173 // XXX: comments assume source clipping, which we don't do
175 /* When clipping any of the two rects we need to adjust the coordinates
176 * in the other rect considering the scaling factor involved. To obtain
177 * the best precision we want to make sure that we only clip once per
178 * side to avoid accumulating errors due to the scaling adjustment.
180 * For example, if src_x0 and dst_x0 need both to be clipped we want to
181 * avoid the situation where we clip src_x0 first, then adjust dst_x0
182 * accordingly but then we realize that the resulting dst_x0 still needs
183 * to be clipped, so we clip dst_x0 and adjust src_x0 again. Because we are
184 * applying scaling factors to adjust the coordinates in each clipping
185 * pass we lose some precision and that can affect the results of the
186 * blorp blit operation slightly. What we want to do here is detect the
187 * rect that we should clip first for each side so that when we adjust
188 * the other rect we ensure the resulting coordinate does not need to be
191 * The code below implements this by comparing the number of pixels that
192 * we need to clip for each side of both rects considering the scales
193 * involved. For example, clip_src_x0 represents the number of pixels
194 * to be clipped for the src rect's left side, so if clip_src_x0 = 5,
195 * clip_dst_x0 = 4 and scale_x = 2 it means that we are clipping more
196 * from the dst rect so we should clip dst_x0 only and adjust src_x0.
197 * This is because clipping 4 pixels in the dst is equivalent to
198 * clipping 4 * 2 = 8 > 5 in the src.
201 if (*src_x0
== *src_x1
|| *src_y0
== *src_y1
202 || *dst_x0
== *dst_x1
|| *dst_y0
== *dst_y1
)
205 float scale_x
= (float) (*src_x1
- *src_x0
) / (*dst_x1
- *dst_x0
);
206 float scale_y
= (float) (*src_y1
- *src_y0
) / (*dst_y1
- *dst_y0
);
209 clip_coordinates(mirror_x
, src_x0
, dst_x0
, dst_x1
,
210 clip_dst_x0
, clip_dst_x1
, scale_x
, true);
212 /* Clip right side */
213 clip_coordinates(mirror_x
, src_x1
, dst_x1
, dst_x0
,
214 clip_dst_x1
, clip_dst_x0
, scale_x
, false);
216 /* Clip bottom side */
217 clip_coordinates(mirror_y
, src_y0
, dst_y0
, dst_y1
,
218 clip_dst_y0
, clip_dst_y1
, scale_y
, true);
221 clip_coordinates(mirror_y
, src_y1
, dst_y1
, dst_y0
,
222 clip_dst_y1
, clip_dst_y0
, scale_y
, false);
224 /* Check for invalid bounds
225 * Can't blit for 0-dimensions
227 return *src_x0
== *src_x1
|| *src_y0
== *src_y1
228 || *dst_x0
== *dst_x1
|| *dst_y0
== *dst_y1
;
232 iris_blorp_surf_for_resource(struct iris_vtable
*vtbl
,
233 struct isl_device
*isl_dev
,
234 struct blorp_surf
*surf
,
235 struct pipe_resource
*p_res
,
236 enum isl_aux_usage aux_usage
,
238 bool is_render_target
)
240 struct iris_resource
*res
= (void *) p_res
;
242 assert(!iris_resource_unfinished_aux_import(res
));
244 if (isl_aux_usage_has_hiz(aux_usage
) &&
245 !iris_resource_level_has_hiz(res
, level
))
246 aux_usage
= ISL_AUX_USAGE_NONE
;
248 *surf
= (struct blorp_surf
) {
250 .addr
= (struct blorp_address
) {
252 .offset
= res
->offset
,
253 .reloc_flags
= is_render_target
? EXEC_OBJECT_WRITE
: 0,
254 .mocs
= vtbl
->mocs(res
->bo
, isl_dev
),
256 .aux_usage
= aux_usage
,
259 if (aux_usage
!= ISL_AUX_USAGE_NONE
) {
260 surf
->aux_surf
= &res
->aux
.surf
;
261 surf
->aux_addr
= (struct blorp_address
) {
262 .buffer
= res
->aux
.bo
,
263 .offset
= res
->aux
.offset
,
264 .reloc_flags
= is_render_target
? EXEC_OBJECT_WRITE
: 0,
265 .mocs
= vtbl
->mocs(res
->bo
, isl_dev
),
268 iris_resource_get_clear_color(res
, NULL
, NULL
);
269 surf
->clear_color_addr
= (struct blorp_address
) {
270 .buffer
= res
->aux
.clear_color_bo
,
271 .offset
= res
->aux
.clear_color_offset
,
273 .mocs
= vtbl
->mocs(res
->aux
.clear_color_bo
, isl_dev
),
281 is_astc(enum isl_format format
)
283 return format
!= ISL_FORMAT_UNSUPPORTED
&&
284 isl_format_get_layout(format
)->txc
== ISL_TXC_ASTC
;
288 tex_cache_flush_hack(struct iris_batch
*batch
,
289 enum isl_format view_format
,
290 enum isl_format surf_format
)
292 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
294 /* The WaSamplerCacheFlushBetweenRedescribedSurfaceReads workaround says:
296 * "Currently Sampler assumes that a surface would not have two
297 * different format associate with it. It will not properly cache
298 * the different views in the MT cache, causing a data corruption."
300 * We may need to handle this for texture views in general someday, but
301 * for now we handle it here, as it hurts copies and blits particularly
302 * badly because they ofter reinterpret formats.
304 * If the BO hasn't been referenced yet this batch, we assume that the
305 * texture cache doesn't contain any relevant data nor need flushing.
307 * Icelake (Gen11+) claims to fix this issue, but seems to still have
308 * issues with ASTC formats.
310 bool need_flush
= devinfo
->gen
>= 11 ?
311 is_astc(surf_format
) != is_astc(view_format
) :
312 view_format
!= surf_format
;
317 "workaround: WaSamplerCacheFlushBetweenRedescribedSurfaceReads";
319 iris_emit_pipe_control_flush(batch
, reason
, PIPE_CONTROL_CS_STALL
);
320 iris_emit_pipe_control_flush(batch
, reason
,
321 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
325 * The pipe->blit() driver hook.
327 * This performs a blit between two surfaces, which copies data but may
328 * also perform format conversion, scaling, flipping, and so on.
331 iris_blit(struct pipe_context
*ctx
, const struct pipe_blit_info
*info
)
333 struct iris_context
*ice
= (void *) ctx
;
334 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
335 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
336 struct iris_batch
*batch
= &ice
->batches
[IRIS_BATCH_RENDER
];
337 enum blorp_batch_flags blorp_flags
= 0;
338 struct iris_resource
*src_res
= (void *) info
->src
.resource
;
339 struct iris_resource
*dst_res
= (void *) info
->dst
.resource
;
341 /* We don't support color masking. */
342 assert((info
->mask
& PIPE_MASK_RGBA
) == PIPE_MASK_RGBA
||
343 (info
->mask
& PIPE_MASK_RGBA
) == 0);
345 if (info
->render_condition_enable
) {
346 if (ice
->state
.predicate
== IRIS_PREDICATE_STATE_DONT_RENDER
)
349 if (ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
)
350 blorp_flags
|= BLORP_BATCH_PREDICATE_ENABLE
;
353 struct iris_format_info src_fmt
=
354 iris_format_for_usage(devinfo
, info
->src
.format
,
355 ISL_SURF_USAGE_TEXTURE_BIT
);
356 enum isl_aux_usage src_aux_usage
=
357 iris_resource_texture_aux_usage(ice
, src_res
, src_fmt
.fmt
, 0);
359 if (iris_resource_level_has_hiz(src_res
, info
->src
.level
))
360 src_aux_usage
= ISL_AUX_USAGE_NONE
;
362 bool src_clear_supported
= src_aux_usage
!= ISL_AUX_USAGE_NONE
&&
363 src_res
->surf
.format
== src_fmt
.fmt
;
365 iris_resource_prepare_access(ice
, batch
, src_res
, info
->src
.level
, 1,
366 info
->src
.box
.z
, info
->src
.box
.depth
,
367 src_aux_usage
, src_clear_supported
);
369 struct iris_format_info dst_fmt
=
370 iris_format_for_usage(devinfo
, info
->dst
.format
,
371 ISL_SURF_USAGE_RENDER_TARGET_BIT
);
372 enum isl_aux_usage dst_aux_usage
=
373 iris_resource_render_aux_usage(ice
, dst_res
, dst_fmt
.fmt
, false, false);
374 bool dst_clear_supported
= dst_aux_usage
!= ISL_AUX_USAGE_NONE
;
376 struct blorp_surf src_surf
, dst_surf
;
377 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &src_surf
,
378 info
->src
.resource
, src_aux_usage
,
379 info
->src
.level
, false);
380 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &dst_surf
,
381 info
->dst
.resource
, dst_aux_usage
,
382 info
->dst
.level
, true);
384 iris_resource_prepare_access(ice
, batch
, dst_res
, info
->dst
.level
, 1,
385 info
->dst
.box
.z
, info
->dst
.box
.depth
,
386 dst_aux_usage
, dst_clear_supported
);
388 float src_x0
= info
->src
.box
.x
;
389 float src_x1
= info
->src
.box
.x
+ info
->src
.box
.width
;
390 float src_y0
= info
->src
.box
.y
;
391 float src_y1
= info
->src
.box
.y
+ info
->src
.box
.height
;
392 float dst_x0
= info
->dst
.box
.x
;
393 float dst_x1
= info
->dst
.box
.x
+ info
->dst
.box
.width
;
394 float dst_y0
= info
->dst
.box
.y
;
395 float dst_y1
= info
->dst
.box
.y
+ info
->dst
.box
.height
;
396 bool mirror_x
= apply_mirror(&src_x0
, &src_x1
);
397 bool mirror_y
= apply_mirror(&src_y0
, &src_y1
);
398 enum blorp_filter filter
;
400 if (info
->scissor_enable
) {
401 bool noop
= apply_blit_scissor(&info
->scissor
,
402 &src_x0
, &src_y0
, &src_x1
, &src_y1
,
403 &dst_x0
, &dst_y0
, &dst_x1
, &dst_y1
,
409 if (abs(info
->dst
.box
.width
) == abs(info
->src
.box
.width
) &&
410 abs(info
->dst
.box
.height
) == abs(info
->src
.box
.height
)) {
411 if (src_surf
.surf
->samples
> 1 && dst_surf
.surf
->samples
<= 1) {
412 /* The OpenGL ES 3.2 specification, section 16.2.1, says:
414 * "If the read framebuffer is multisampled (its effective
415 * value of SAMPLE_BUFFERS is one) and the draw framebuffer
416 * is not (its value of SAMPLE_BUFFERS is zero), the samples
417 * corresponding to each pixel location in the source are
418 * converted to a single sample before being written to the
419 * destination. The filter parameter is ignored. If the
420 * source formats are integer types or stencil values, a
421 * single sample’s value is selected for each pixel. If the
422 * source formats are floating-point or normalized types,
423 * the sample values for each pixel are resolved in an
424 * implementation-dependent manner. If the source formats
425 * are depth values, sample values are resolved in an
426 * implementation-dependent manner where the result will be
427 * between the minimum and maximum depth values in the pixel."
429 * When selecting a single sample, we always choose sample 0.
431 if (util_format_is_depth_or_stencil(info
->src
.format
) ||
432 util_format_is_pure_integer(info
->src
.format
)) {
433 filter
= BLORP_FILTER_SAMPLE_0
;
435 filter
= BLORP_FILTER_AVERAGE
;
438 /* The OpenGL 4.6 specification, section 18.3.1, says:
440 * "If the source and destination dimensions are identical,
441 * no filtering is applied."
443 * Using BLORP_FILTER_NONE will also handle the upsample case by
444 * replicating the one value in the source to all values in the
447 filter
= BLORP_FILTER_NONE
;
449 } else if (info
->filter
== PIPE_TEX_FILTER_LINEAR
) {
450 filter
= BLORP_FILTER_BILINEAR
;
452 filter
= BLORP_FILTER_NEAREST
;
455 if (iris_batch_references(batch
, src_res
->bo
))
456 tex_cache_flush_hack(batch
, src_fmt
.fmt
, src_res
->surf
.format
);
458 if (dst_res
->base
.target
== PIPE_BUFFER
)
459 util_range_add(&dst_res
->base
, &dst_res
->valid_buffer_range
, dst_x0
, dst_x1
);
461 struct blorp_batch blorp_batch
;
462 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, blorp_flags
);
465 if (util_format_is_depth_or_stencil(info
->dst
.format
))
466 main_mask
= PIPE_MASK_Z
;
468 main_mask
= PIPE_MASK_RGBA
;
470 if (info
->mask
& main_mask
) {
471 for (int slice
= 0; slice
< info
->dst
.box
.depth
; slice
++) {
472 iris_batch_maybe_flush(batch
, 1500);
474 blorp_blit(&blorp_batch
,
475 &src_surf
, info
->src
.level
, info
->src
.box
.z
+ slice
,
476 src_fmt
.fmt
, src_fmt
.swizzle
,
477 &dst_surf
, info
->dst
.level
, info
->dst
.box
.z
+ slice
,
478 dst_fmt
.fmt
, dst_fmt
.swizzle
,
479 src_x0
, src_y0
, src_x1
, src_y1
,
480 dst_x0
, dst_y0
, dst_x1
, dst_y1
,
481 filter
, mirror_x
, mirror_y
);
485 struct iris_resource
*stc_dst
= NULL
;
486 enum isl_aux_usage stc_src_aux_usage
, stc_dst_aux_usage
;
487 if ((info
->mask
& PIPE_MASK_S
) &&
488 util_format_has_stencil(util_format_description(info
->dst
.format
)) &&
489 util_format_has_stencil(util_format_description(info
->src
.format
))) {
490 struct iris_resource
*src_res
, *junk
;
491 struct blorp_surf src_surf
, dst_surf
;
492 iris_get_depth_stencil_resources(info
->src
.resource
, &junk
, &src_res
);
493 iris_get_depth_stencil_resources(info
->dst
.resource
, &junk
, &stc_dst
);
495 struct iris_format_info src_fmt
=
496 iris_format_for_usage(devinfo
, src_res
->base
.format
,
497 ISL_SURF_USAGE_TEXTURE_BIT
);
499 iris_resource_texture_aux_usage(ice
, src_res
, src_fmt
.fmt
, 0);
501 struct iris_format_info dst_fmt
=
502 iris_format_for_usage(devinfo
, stc_dst
->base
.format
,
503 ISL_SURF_USAGE_RENDER_TARGET_BIT
);
505 iris_resource_render_aux_usage(ice
, stc_dst
, dst_fmt
.fmt
, false, false);
507 /* Resolve destination surface before blit because :
508 * 1. when we try to blit from the same surface, we can't read and
509 * write to the same surfaces at the same time when we have
510 * compression enabled so it's safe to resolve surface first and then
512 * 2. While bliting from one surface to another surface, we might be
513 * mixing compression formats, Our experiments shows that if after
514 * blit if we set DepthStencilResource flag to 0, blit passes but
517 * XXX: In second case by destructing the compression, we might lose
520 if (devinfo
->gen
>= 12)
521 stc_dst_aux_usage
= ISL_AUX_USAGE_NONE
;
523 iris_resource_prepare_access(ice
, batch
, src_res
, info
->src
.level
, 1,
524 info
->src
.box
.z
, info
->src
.box
.depth
,
525 stc_src_aux_usage
, false);
526 iris_resource_prepare_access(ice
, batch
, stc_dst
, info
->dst
.level
, 1,
527 info
->dst
.box
.z
, info
->dst
.box
.depth
,
528 stc_dst_aux_usage
, false);
529 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &src_surf
,
530 &src_res
->base
, stc_src_aux_usage
,
531 info
->src
.level
, false);
532 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &dst_surf
,
533 &stc_dst
->base
, stc_dst_aux_usage
,
534 info
->dst
.level
, true);
536 for (int slice
= 0; slice
< info
->dst
.box
.depth
; slice
++) {
537 iris_batch_maybe_flush(batch
, 1500);
539 blorp_blit(&blorp_batch
,
540 &src_surf
, info
->src
.level
, info
->src
.box
.z
+ slice
,
541 ISL_FORMAT_R8_UINT
, ISL_SWIZZLE_IDENTITY
,
542 &dst_surf
, info
->dst
.level
, info
->dst
.box
.z
+ slice
,
543 ISL_FORMAT_R8_UINT
, ISL_SWIZZLE_IDENTITY
,
544 src_x0
, src_y0
, src_x1
, src_y1
,
545 dst_x0
, dst_y0
, dst_x1
, dst_y1
,
546 filter
, mirror_x
, mirror_y
);
550 blorp_batch_finish(&blorp_batch
);
552 tex_cache_flush_hack(batch
, src_fmt
.fmt
, src_res
->surf
.format
);
554 if (info
->mask
& main_mask
) {
555 iris_resource_finish_write(ice
, dst_res
, info
->dst
.level
, info
->dst
.box
.z
,
556 info
->dst
.box
.depth
, dst_aux_usage
);
560 iris_resource_finish_write(ice
, stc_dst
, info
->dst
.level
, info
->dst
.box
.z
,
561 info
->dst
.box
.depth
, stc_dst_aux_usage
);
564 iris_flush_and_dirty_for_history(ice
, batch
, (struct iris_resource
*)
566 PIPE_CONTROL_RENDER_TARGET_FLUSH
,
567 "cache history: post-blit");
571 get_copy_region_aux_settings(const struct gen_device_info
*devinfo
,
572 struct iris_resource
*res
,
573 enum isl_aux_usage
*out_aux_usage
,
574 bool *out_clear_supported
,
575 bool is_render_target
)
577 switch (res
->aux
.usage
) {
578 case ISL_AUX_USAGE_MCS
:
579 case ISL_AUX_USAGE_MCS_CCS
:
580 case ISL_AUX_USAGE_CCS_E
:
581 /* A stencil resolve operation must be performed prior to doing resource
582 * copies or used by CPU.
583 * (see HSD 1209978162)
585 if (is_render_target
&& isl_surf_usage_is_stencil(res
->surf
.usage
)) {
586 *out_aux_usage
= ISL_AUX_USAGE_NONE
;
587 *out_clear_supported
= false;
589 *out_aux_usage
= res
->aux
.usage
;
590 /* Prior to Gen9, fast-clear only supported 0/1 clear colors. Since
591 * we're going to re-interpret the format as an integer format possibly
592 * with a different number of components, we can't handle clear colors
595 *out_clear_supported
= devinfo
->gen
>= 9;
599 *out_aux_usage
= ISL_AUX_USAGE_NONE
;
600 *out_clear_supported
= false;
606 * Perform a GPU-based raw memory copy between compatible view classes.
608 * Does not perform any flushing - the new data may still be left in the
609 * render cache, and old data may remain in other caches.
611 * Wraps blorp_copy() and blorp_buffer_copy().
614 iris_copy_region(struct blorp_context
*blorp
,
615 struct iris_batch
*batch
,
616 struct pipe_resource
*dst
,
618 unsigned dstx
, unsigned dsty
, unsigned dstz
,
619 struct pipe_resource
*src
,
621 const struct pipe_box
*src_box
)
623 struct blorp_batch blorp_batch
;
624 struct iris_context
*ice
= blorp
->driver_ctx
;
625 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
626 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
627 struct iris_resource
*src_res
= (void *) src
;
628 struct iris_resource
*dst_res
= (void *) dst
;
630 enum isl_aux_usage src_aux_usage
, dst_aux_usage
;
631 bool src_clear_supported
, dst_clear_supported
;
632 get_copy_region_aux_settings(devinfo
, src_res
, &src_aux_usage
,
633 &src_clear_supported
, false);
634 get_copy_region_aux_settings(devinfo
, dst_res
, &dst_aux_usage
,
635 &dst_clear_supported
, true);
637 if (iris_batch_references(batch
, src_res
->bo
))
638 tex_cache_flush_hack(batch
, ISL_FORMAT_UNSUPPORTED
, src_res
->surf
.format
);
640 if (dst
->target
== PIPE_BUFFER
)
641 util_range_add(&dst_res
->base
, &dst_res
->valid_buffer_range
, dstx
, dstx
+ src_box
->width
);
643 if (dst
->target
== PIPE_BUFFER
&& src
->target
== PIPE_BUFFER
) {
644 struct blorp_address src_addr
= {
645 .buffer
= iris_resource_bo(src
), .offset
= src_box
->x
,
647 struct blorp_address dst_addr
= {
648 .buffer
= iris_resource_bo(dst
), .offset
= dstx
,
649 .reloc_flags
= EXEC_OBJECT_WRITE
,
652 iris_batch_maybe_flush(batch
, 1500);
654 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, 0);
655 blorp_buffer_copy(&blorp_batch
, src_addr
, dst_addr
, src_box
->width
);
656 blorp_batch_finish(&blorp_batch
);
658 // XXX: what about one surface being a buffer and not the other?
660 struct blorp_surf src_surf
, dst_surf
;
661 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &src_surf
,
662 src
, src_aux_usage
, src_level
, false);
663 iris_blorp_surf_for_resource(&ice
->vtbl
, &screen
->isl_dev
, &dst_surf
,
664 dst
, dst_aux_usage
, dst_level
, true);
666 iris_resource_prepare_access(ice
, batch
, src_res
, src_level
, 1,
667 src_box
->z
, src_box
->depth
,
668 src_aux_usage
, src_clear_supported
);
669 iris_resource_prepare_access(ice
, batch
, dst_res
, dst_level
, 1,
670 dstz
, src_box
->depth
,
671 dst_aux_usage
, dst_clear_supported
);
673 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, 0);
675 for (int slice
= 0; slice
< src_box
->depth
; slice
++) {
676 iris_batch_maybe_flush(batch
, 1500);
678 blorp_copy(&blorp_batch
, &src_surf
, src_level
, src_box
->z
+ slice
,
679 &dst_surf
, dst_level
, dstz
+ slice
,
680 src_box
->x
, src_box
->y
, dstx
, dsty
,
681 src_box
->width
, src_box
->height
);
683 blorp_batch_finish(&blorp_batch
);
685 iris_resource_finish_write(ice
, dst_res
, dst_level
, dstz
,
686 src_box
->depth
, dst_aux_usage
);
689 tex_cache_flush_hack(batch
, ISL_FORMAT_UNSUPPORTED
, src_res
->surf
.format
);
692 static struct iris_batch
*
693 get_preferred_batch(struct iris_context
*ice
, struct iris_bo
*bo
)
695 /* If the compute batch is already using this buffer, we'd prefer to
696 * continue queueing in the compute batch.
698 if (iris_batch_references(&ice
->batches
[IRIS_BATCH_COMPUTE
], bo
))
699 return &ice
->batches
[IRIS_BATCH_COMPUTE
];
701 /* Otherwise default to the render batch. */
702 return &ice
->batches
[IRIS_BATCH_RENDER
];
707 * The pipe->resource_copy_region() driver hook.
709 * This implements ARB_copy_image semantics - a raw memory copy between
710 * compatible view classes.
713 iris_resource_copy_region(struct pipe_context
*ctx
,
714 struct pipe_resource
*dst
,
716 unsigned dstx
, unsigned dsty
, unsigned dstz
,
717 struct pipe_resource
*src
,
719 const struct pipe_box
*src_box
)
721 struct iris_context
*ice
= (void *) ctx
;
722 struct iris_batch
*batch
= &ice
->batches
[IRIS_BATCH_RENDER
];
724 /* Use MI_COPY_MEM_MEM for tiny (<= 16 byte, % 4) buffer copies. */
725 if (src
->target
== PIPE_BUFFER
&& dst
->target
== PIPE_BUFFER
&&
726 (src_box
->width
% 4 == 0) && src_box
->width
<= 16) {
727 struct iris_bo
*dst_bo
= iris_resource_bo(dst
);
728 batch
= get_preferred_batch(ice
, dst_bo
);
729 iris_batch_maybe_flush(batch
, 24 + 5 * (src_box
->width
/ 4));
730 iris_emit_pipe_control_flush(batch
,
731 "stall for MI_COPY_MEM_MEM copy_region",
732 PIPE_CONTROL_CS_STALL
);
733 ice
->vtbl
.copy_mem_mem(batch
, dst_bo
, dstx
, iris_resource_bo(src
),
734 src_box
->x
, src_box
->width
);
738 iris_copy_region(&ice
->blorp
, batch
, dst
, dst_level
, dstx
, dsty
, dstz
,
739 src
, src_level
, src_box
);
741 if (util_format_is_depth_and_stencil(dst
->format
) &&
742 util_format_has_stencil(util_format_description(src
->format
))) {
743 struct iris_resource
*junk
, *s_src_res
, *s_dst_res
;
744 iris_get_depth_stencil_resources(src
, &junk
, &s_src_res
);
745 iris_get_depth_stencil_resources(dst
, &junk
, &s_dst_res
);
747 iris_copy_region(&ice
->blorp
, batch
, &s_dst_res
->base
, dst_level
, dstx
,
748 dsty
, dstz
, &s_src_res
->base
, src_level
, src_box
);
751 iris_flush_and_dirty_for_history(ice
, batch
, (struct iris_resource
*) dst
,
752 PIPE_CONTROL_RENDER_TARGET_FLUSH
,
753 "cache history: post copy_region");
757 iris_init_blit_functions(struct pipe_context
*ctx
)
759 ctx
->blit
= iris_blit
;
760 ctx
->resource_copy_region
= iris_resource_copy_region
;