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/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 blorp_surf
*surf
,
234 struct pipe_resource
*p_res
,
235 enum isl_aux_usage aux_usage
,
237 bool is_render_target
)
239 struct iris_resource
*res
= (void *) p_res
;
241 if (aux_usage
== ISL_AUX_USAGE_HIZ
&&
242 !iris_resource_level_has_hiz(res
, level
))
243 aux_usage
= ISL_AUX_USAGE_NONE
;
245 *surf
= (struct blorp_surf
) {
247 .addr
= (struct blorp_address
) {
249 .offset
= 0, // XXX: ???
250 .reloc_flags
= is_render_target
? EXEC_OBJECT_WRITE
: 0,
251 .mocs
= vtbl
->mocs(res
->bo
),
253 .aux_usage
= aux_usage
,
256 if (aux_usage
!= ISL_AUX_USAGE_NONE
) {
257 surf
->aux_surf
= &res
->aux
.surf
;
258 surf
->aux_addr
= (struct blorp_address
) {
259 .buffer
= res
->aux
.bo
,
260 .offset
= res
->aux
.offset
,
261 .reloc_flags
= is_render_target
? EXEC_OBJECT_WRITE
: 0,
262 .mocs
= vtbl
->mocs(res
->bo
),
265 iris_resource_get_clear_color(res
, NULL
, NULL
);
266 surf
->clear_color_addr
= (struct blorp_address
) {
267 .buffer
= res
->aux
.clear_color_bo
,
268 .offset
= res
->aux
.clear_color_offset
,
270 .mocs
= vtbl
->mocs(res
->aux
.clear_color_bo
),
278 * The pipe->blit() driver hook.
280 * This performs a blit between two surfaces, which copies data but may
281 * also perform format conversion, scaling, flipping, and so on.
284 iris_blit(struct pipe_context
*ctx
, const struct pipe_blit_info
*info
)
286 struct iris_context
*ice
= (void *) ctx
;
287 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
288 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
289 struct iris_batch
*batch
= &ice
->batches
[IRIS_BATCH_RENDER
];
290 enum blorp_batch_flags blorp_flags
= 0;
291 struct iris_resource
*src_res
= (void *) info
->src
.resource
;
292 struct iris_resource
*dst_res
= (void *) info
->dst
.resource
;
294 /* We don't support color masking. */
295 assert((info
->mask
& PIPE_MASK_RGBA
) == PIPE_MASK_RGBA
||
296 (info
->mask
& PIPE_MASK_RGBA
) == 0);
298 if (info
->render_condition_enable
) {
299 if (ice
->state
.predicate
== IRIS_PREDICATE_STATE_DONT_RENDER
)
302 if (ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
)
303 blorp_flags
|= BLORP_BATCH_PREDICATE_ENABLE
;
306 struct iris_format_info src_fmt
=
307 iris_format_for_usage(devinfo
, info
->src
.format
,
308 ISL_SURF_USAGE_TEXTURE_BIT
);
309 enum isl_aux_usage src_aux_usage
=
310 iris_resource_texture_aux_usage(ice
, src_res
, src_fmt
.fmt
, 0);
312 if (src_aux_usage
== ISL_AUX_USAGE_HIZ
)
313 src_aux_usage
= ISL_AUX_USAGE_NONE
;
315 bool src_clear_supported
= src_aux_usage
!= ISL_AUX_USAGE_NONE
&&
316 src_res
->surf
.format
== src_fmt
.fmt
;
318 iris_resource_prepare_access(ice
, batch
, src_res
, info
->src
.level
, 1,
319 info
->src
.box
.z
, info
->src
.box
.depth
,
320 src_aux_usage
, src_clear_supported
);
322 struct iris_format_info dst_fmt
=
323 iris_format_for_usage(devinfo
, info
->dst
.format
,
324 ISL_SURF_USAGE_RENDER_TARGET_BIT
);
325 enum isl_aux_usage dst_aux_usage
=
326 iris_resource_render_aux_usage(ice
, dst_res
, dst_fmt
.fmt
, false, false);
327 bool dst_clear_supported
= dst_aux_usage
!= ISL_AUX_USAGE_NONE
;
329 struct blorp_surf src_surf
, dst_surf
;
330 iris_blorp_surf_for_resource(&ice
->vtbl
, &src_surf
, info
->src
.resource
,
331 src_aux_usage
, info
->src
.level
, false);
332 iris_blorp_surf_for_resource(&ice
->vtbl
, &dst_surf
, info
->dst
.resource
,
333 dst_aux_usage
, info
->dst
.level
, true);
335 iris_resource_prepare_access(ice
, batch
, dst_res
, info
->dst
.level
, 1,
336 info
->dst
.box
.z
, info
->dst
.box
.depth
,
337 dst_aux_usage
, dst_clear_supported
);
339 float src_x0
= info
->src
.box
.x
;
340 float src_x1
= info
->src
.box
.x
+ info
->src
.box
.width
;
341 float src_y0
= info
->src
.box
.y
;
342 float src_y1
= info
->src
.box
.y
+ info
->src
.box
.height
;
343 float dst_x0
= info
->dst
.box
.x
;
344 float dst_x1
= info
->dst
.box
.x
+ info
->dst
.box
.width
;
345 float dst_y0
= info
->dst
.box
.y
;
346 float dst_y1
= info
->dst
.box
.y
+ info
->dst
.box
.height
;
347 bool mirror_x
= apply_mirror(&src_x0
, &src_x1
);
348 bool mirror_y
= apply_mirror(&src_y0
, &src_y1
);
349 enum blorp_filter filter
;
351 if (info
->scissor_enable
) {
352 bool noop
= apply_blit_scissor(&info
->scissor
,
353 &src_x0
, &src_y0
, &src_x1
, &src_y1
,
354 &dst_x0
, &dst_y0
, &dst_x1
, &dst_y1
,
360 if (abs(info
->dst
.box
.width
) == abs(info
->src
.box
.width
) &&
361 abs(info
->dst
.box
.height
) == abs(info
->src
.box
.height
)) {
362 if (src_surf
.surf
->samples
> 1 && dst_surf
.surf
->samples
<= 1) {
363 /* The OpenGL ES 3.2 specification, section 16.2.1, says:
365 * "If the read framebuffer is multisampled (its effective
366 * value of SAMPLE_BUFFERS is one) and the draw framebuffer
367 * is not (its value of SAMPLE_BUFFERS is zero), the samples
368 * corresponding to each pixel location in the source are
369 * converted to a single sample before being written to the
370 * destination. The filter parameter is ignored. If the
371 * source formats are integer types or stencil values, a
372 * single sample’s value is selected for each pixel. If the
373 * source formats are floating-point or normalized types,
374 * the sample values for each pixel are resolved in an
375 * implementation-dependent manner. If the source formats
376 * are depth values, sample values are resolved in an
377 * implementation-dependent manner where the result will be
378 * between the minimum and maximum depth values in the pixel."
380 * When selecting a single sample, we always choose sample 0.
382 if (util_format_is_depth_or_stencil(info
->src
.format
) ||
383 util_format_is_pure_integer(info
->src
.format
)) {
384 filter
= BLORP_FILTER_SAMPLE_0
;
386 filter
= BLORP_FILTER_AVERAGE
;
389 /* The OpenGL 4.6 specification, section 18.3.1, says:
391 * "If the source and destination dimensions are identical,
392 * no filtering is applied."
394 * Using BLORP_FILTER_NONE will also handle the upsample case by
395 * replicating the one value in the source to all values in the
398 filter
= BLORP_FILTER_NONE
;
400 } else if (info
->filter
== PIPE_TEX_FILTER_LINEAR
) {
401 filter
= BLORP_FILTER_BILINEAR
;
403 filter
= BLORP_FILTER_NEAREST
;
406 struct blorp_batch blorp_batch
;
407 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, blorp_flags
);
410 if (util_format_is_depth_or_stencil(info
->dst
.format
))
411 main_mask
= PIPE_MASK_Z
;
413 main_mask
= PIPE_MASK_RGBA
;
415 if (info
->mask
& main_mask
) {
416 for (int slice
= 0; slice
< info
->dst
.box
.depth
; slice
++) {
417 iris_batch_maybe_flush(batch
, 1500);
419 blorp_blit(&blorp_batch
,
420 &src_surf
, info
->src
.level
, info
->src
.box
.z
+ slice
,
421 src_fmt
.fmt
, src_fmt
.swizzle
,
422 &dst_surf
, info
->dst
.level
, info
->dst
.box
.z
+ slice
,
423 dst_fmt
.fmt
, dst_fmt
.swizzle
,
424 src_x0
, src_y0
, src_x1
, src_y1
,
425 dst_x0
, dst_y0
, dst_x1
, dst_y1
,
426 filter
, mirror_x
, mirror_y
);
430 if ((info
->mask
& PIPE_MASK_S
) &&
431 util_format_has_stencil(util_format_description(info
->dst
.format
)) &&
432 util_format_has_stencil(util_format_description(info
->src
.format
))) {
433 struct iris_resource
*src_res
, *dst_res
, *junk
;
434 iris_get_depth_stencil_resources(info
->src
.resource
, &junk
, &src_res
);
435 iris_get_depth_stencil_resources(info
->dst
.resource
, &junk
, &dst_res
);
436 iris_blorp_surf_for_resource(&ice
->vtbl
, &src_surf
, &src_res
->base
,
437 ISL_AUX_USAGE_NONE
, info
->src
.level
, false);
438 iris_blorp_surf_for_resource(&ice
->vtbl
, &dst_surf
, &dst_res
->base
,
439 ISL_AUX_USAGE_NONE
, info
->dst
.level
, true);
441 for (int slice
= 0; slice
< info
->dst
.box
.depth
; slice
++) {
442 iris_batch_maybe_flush(batch
, 1500);
444 blorp_blit(&blorp_batch
,
445 &src_surf
, info
->src
.level
, info
->src
.box
.z
+ slice
,
446 ISL_FORMAT_R8_UINT
, ISL_SWIZZLE_IDENTITY
,
447 &dst_surf
, info
->dst
.level
, info
->dst
.box
.z
+ slice
,
448 ISL_FORMAT_R8_UINT
, ISL_SWIZZLE_IDENTITY
,
449 src_x0
, src_y0
, src_x1
, src_y1
,
450 dst_x0
, dst_y0
, dst_x1
, dst_y1
,
451 filter
, mirror_x
, mirror_y
);
455 blorp_batch_finish(&blorp_batch
);
457 iris_resource_finish_write(ice
, dst_res
, info
->dst
.level
, info
->dst
.box
.z
,
458 info
->dst
.box
.depth
, dst_aux_usage
);
460 iris_flush_and_dirty_for_history(ice
, batch
, (struct iris_resource
*)
465 get_copy_region_aux_settings(const struct gen_device_info
*devinfo
,
466 struct iris_resource
*res
,
467 enum isl_aux_usage
*out_aux_usage
,
468 bool *out_clear_supported
)
470 switch (res
->aux
.usage
) {
471 case ISL_AUX_USAGE_MCS
:
472 case ISL_AUX_USAGE_CCS_E
:
473 *out_aux_usage
= res
->aux
.usage
;
474 /* Prior to Gen9, fast-clear only supported 0/1 clear colors. Since
475 * we're going to re-interpret the format as an integer format possibly
476 * with a different number of components, we can't handle clear colors
479 *out_clear_supported
= devinfo
->gen
>= 9;
482 *out_aux_usage
= ISL_AUX_USAGE_NONE
;
483 *out_clear_supported
= false;
489 * Perform a GPU-based raw memory copy between compatible view classes.
491 * Does not perform any flushing - the new data may still be left in the
492 * render cache, and old data may remain in other caches.
494 * Wraps blorp_copy() and blorp_buffer_copy().
497 iris_copy_region(struct blorp_context
*blorp
,
498 struct iris_batch
*batch
,
499 struct pipe_resource
*dst
,
501 unsigned dstx
, unsigned dsty
, unsigned dstz
,
502 struct pipe_resource
*src
,
504 const struct pipe_box
*src_box
)
506 struct blorp_batch blorp_batch
;
507 struct iris_context
*ice
= blorp
->driver_ctx
;
508 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
509 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
510 struct iris_resource
*src_res
= (void *) src
;
511 struct iris_resource
*dst_res
= (void *) dst
;
513 enum isl_aux_usage src_aux_usage
, dst_aux_usage
;
514 bool src_clear_supported
, dst_clear_supported
;
515 get_copy_region_aux_settings(devinfo
, src_res
, &src_aux_usage
,
516 &src_clear_supported
);
517 get_copy_region_aux_settings(devinfo
, dst_res
, &dst_aux_usage
,
518 &dst_clear_supported
);
520 if (dst
->target
== PIPE_BUFFER
&& src
->target
== PIPE_BUFFER
) {
521 struct blorp_address src_addr
= {
522 .buffer
= iris_resource_bo(src
), .offset
= src_box
->x
,
524 struct blorp_address dst_addr
= {
525 .buffer
= iris_resource_bo(dst
), .offset
= dstx
,
528 iris_batch_maybe_flush(batch
, 1500);
530 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, 0);
531 blorp_buffer_copy(&blorp_batch
, src_addr
, dst_addr
, src_box
->width
);
532 blorp_batch_finish(&blorp_batch
);
534 iris_flush_and_dirty_for_history(ice
, batch
,
535 (struct iris_resource
*) dst
);
537 // XXX: what about one surface being a buffer and not the other?
539 struct blorp_surf src_surf
, dst_surf
;
540 iris_blorp_surf_for_resource(&ice
->vtbl
, &src_surf
, src
, src_aux_usage
,
542 iris_blorp_surf_for_resource(&ice
->vtbl
, &dst_surf
, dst
, dst_aux_usage
,
545 iris_resource_prepare_access(ice
, batch
, src_res
, src_level
, 1,
546 src_box
->z
, src_box
->depth
,
547 src_aux_usage
, src_clear_supported
);
548 iris_resource_prepare_access(ice
, batch
, dst_res
, dst_level
, 1,
549 dstz
, src_box
->depth
,
550 dst_aux_usage
, dst_clear_supported
);
552 blorp_batch_init(&ice
->blorp
, &blorp_batch
, batch
, 0);
554 for (int slice
= 0; slice
< src_box
->depth
; slice
++) {
555 iris_batch_maybe_flush(batch
, 1500);
557 blorp_copy(&blorp_batch
, &src_surf
, src_level
, src_box
->z
+ slice
,
558 &dst_surf
, dst_level
, dstz
+ slice
,
559 src_box
->x
, src_box
->y
, dstx
, dsty
,
560 src_box
->width
, src_box
->height
);
562 blorp_batch_finish(&blorp_batch
);
564 iris_resource_finish_write(ice
, dst_res
, dst_level
, dstz
,
565 src_box
->depth
, dst_aux_usage
);
571 * The pipe->resource_copy_region() driver hook.
573 * This implements ARB_copy_image semantics - a raw memory copy between
574 * compatible view classes.
577 iris_resource_copy_region(struct pipe_context
*ctx
,
578 struct pipe_resource
*dst
,
580 unsigned dstx
, unsigned dsty
, unsigned dstz
,
581 struct pipe_resource
*src
,
583 const struct pipe_box
*src_box
)
585 struct iris_context
*ice
= (void *) ctx
;
586 struct iris_batch
*batch
= &ice
->batches
[IRIS_BATCH_RENDER
];
588 iris_copy_region(&ice
->blorp
, batch
, dst
, dst_level
, dstx
, dsty
, dstz
,
589 src
, src_level
, src_box
);
593 iris_init_blit_functions(struct pipe_context
*ctx
)
595 ctx
->blit
= iris_blit
;
596 ctx
->resource_copy_region
= iris_resource_copy_region
;