1 /**************************************************************************
3 * Copyright 2006 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
29 #include <GL/internal/dri_interface.h>
31 #include "intel_batchbuffer.h"
32 #include "intel_chipset.h"
33 #include "intel_mipmap_tree.h"
34 #include "intel_regions.h"
35 #include "intel_resolve_map.h"
36 #include "intel_tex.h"
37 #include "intel_blit.h"
38 #include "intel_fbo.h"
40 #include "brw_blorp.h"
41 #include "brw_context.h"
43 #include "main/enums.h"
44 #include "main/fbobject.h"
45 #include "main/formats.h"
46 #include "main/glformats.h"
47 #include "main/texcompress_etc.h"
48 #include "main/teximage.h"
49 #include "main/streaming-load-memcpy.h"
51 #define FILE_DEBUG_FLAG DEBUG_MIPTREE
54 * Determine which MSAA layout should be used by the MSAA surface being
55 * created, based on the chip generation and the surface type.
57 static enum intel_msaa_layout
58 compute_msaa_layout(struct brw_context
*brw
, mesa_format format
, GLenum target
)
60 /* Prior to Gen7, all MSAA surfaces used IMS layout. */
62 return INTEL_MSAA_LAYOUT_IMS
;
64 /* In Gen7, IMS layout is only used for depth and stencil buffers. */
65 switch (_mesa_get_format_base_format(format
)) {
66 case GL_DEPTH_COMPONENT
:
67 case GL_STENCIL_INDEX
:
68 case GL_DEPTH_STENCIL
:
69 return INTEL_MSAA_LAYOUT_IMS
;
71 /* Disable MCS on Broadwell for now. We can enable it once things
72 * are working without it.
75 perf_debug("Missing CMS support on Broadwell.\n");
76 return INTEL_MSAA_LAYOUT_UMS
;
79 /* From the Ivy Bridge PRM, Vol4 Part1 p77 ("MCS Enable"):
81 * This field must be set to 0 for all SINT MSRTs when all RT channels
84 * In practice this means that we have to disable MCS for all signed
85 * integer MSAA buffers. The alternative, to disable MCS only when one
86 * of the render target channels is disabled, is impractical because it
87 * would require converting between CMS and UMS MSAA layouts on the fly,
90 if (_mesa_get_format_datatype(format
) == GL_INT
) {
91 /* TODO: is this workaround needed for future chipsets? */
92 assert(brw
->gen
== 7);
93 return INTEL_MSAA_LAYOUT_UMS
;
95 return INTEL_MSAA_LAYOUT_CMS
;
102 * For single-sampled render targets ("non-MSRT"), the MCS buffer is a
103 * scaled-down bitfield representation of the color buffer which is capable of
104 * recording when blocks of the color buffer are equal to the clear value.
105 * This function returns the block size that will be used by the MCS buffer
106 * corresponding to a certain color miptree.
108 * From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render Target(s)",
109 * beneath the "Fast Color Clear" bullet (p327):
111 * The following table describes the RT alignment
125 * This alignment has the following uses:
127 * - For figuring out the size of the MCS buffer. Each 4k tile in the MCS
128 * buffer contains 128 blocks horizontally and 256 blocks vertically.
130 * - For figuring out alignment restrictions for a fast clear operation. Fast
131 * clear operations must always clear aligned multiples of 16 blocks
132 * horizontally and 32 blocks vertically.
134 * - For scaling down the coordinates sent through the render pipeline during
135 * a fast clear. X coordinates must be scaled down by 8 times the block
136 * width, and Y coordinates by 16 times the block height.
138 * - For scaling down the coordinates sent through the render pipeline during
139 * a "Render Target Resolve" operation. X coordinates must be scaled down
140 * by half the block width, and Y coordinates by half the block height.
143 intel_get_non_msrt_mcs_alignment(struct brw_context
*brw
,
144 struct intel_mipmap_tree
*mt
,
145 unsigned *width_px
, unsigned *height
)
147 switch (mt
->region
->tiling
) {
149 assert(!"Non-MSRT MCS requires X or Y tiling");
150 /* In release builds, fall through */
152 *width_px
= 32 / mt
->cpp
;
156 *width_px
= 64 / mt
->cpp
;
163 * For a single-sampled render target ("non-MSRT"), determine if an MCS buffer
166 * From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render Target(s)",
167 * beneath the "Fast Color Clear" bullet (p326):
169 * - Support is limited to tiled render targets.
170 * - Support is for non-mip-mapped and non-array surface types only.
172 * And then later, on p327:
174 * - MCS buffer for non-MSRT is supported only for RT formats 32bpp,
178 intel_is_non_msrt_mcs_buffer_supported(struct brw_context
*brw
,
179 struct intel_mipmap_tree
*mt
)
181 /* MCS support does not exist prior to Gen7 */
182 if (brw
->gen
< 7 || brw
->gen
>= 8)
185 /* MCS is only supported for color buffers */
186 switch (_mesa_get_format_base_format(mt
->format
)) {
187 case GL_DEPTH_COMPONENT
:
188 case GL_DEPTH_STENCIL
:
189 case GL_STENCIL_INDEX
:
193 if (mt
->region
->tiling
!= I915_TILING_X
&&
194 mt
->region
->tiling
!= I915_TILING_Y
)
196 if (mt
->cpp
!= 4 && mt
->cpp
!= 8 && mt
->cpp
!= 16)
198 if (mt
->first_level
!= 0 || mt
->last_level
!= 0)
200 if (mt
->physical_depth0
!= 1)
203 /* There's no point in using an MCS buffer if the surface isn't in a
206 if (!brw
->format_supported_as_render_target
[mt
->format
])
214 * Determine depth format corresponding to a depth+stencil format,
215 * for separate stencil.
218 intel_depth_format_for_depthstencil_format(mesa_format format
) {
220 case MESA_FORMAT_Z24_UNORM_S8_UINT
:
221 return MESA_FORMAT_Z24_UNORM_X8_UINT
;
222 case MESA_FORMAT_Z32_FLOAT_S8X24_UINT
:
223 return MESA_FORMAT_Z_FLOAT32
;
231 * @param for_bo Indicates that the caller is
232 * intel_miptree_create_for_bo(). If true, then do not create
235 struct intel_mipmap_tree
*
236 intel_miptree_create_layout(struct brw_context
*brw
,
247 struct intel_mipmap_tree
*mt
= calloc(sizeof(*mt
), 1);
251 DBG("%s target %s format %s level %d..%d slices %d <-- %p\n", __FUNCTION__
,
252 _mesa_lookup_enum_by_nr(target
),
253 _mesa_get_format_name(format
),
254 first_level
, last_level
, depth0
, mt
);
258 mt
->first_level
= first_level
;
259 mt
->last_level
= last_level
;
260 mt
->logical_width0
= width0
;
261 mt
->logical_height0
= height0
;
262 mt
->logical_depth0
= depth0
;
263 mt
->fast_clear_state
= INTEL_FAST_CLEAR_STATE_NO_MCS
;
265 /* The cpp is bytes per (1, blockheight)-sized block for compressed
266 * textures. This is why you'll see divides by blockheight all over
269 _mesa_get_format_block_size(format
, &bw
, &bh
);
270 assert(_mesa_get_format_bytes(mt
->format
) % bw
== 0);
271 mt
->cpp
= _mesa_get_format_bytes(mt
->format
) / bw
;
273 mt
->num_samples
= num_samples
;
274 mt
->compressed
= _mesa_is_format_compressed(format
);
275 mt
->msaa_layout
= INTEL_MSAA_LAYOUT_NONE
;
278 if (num_samples
> 1) {
279 /* Adjust width/height/depth for MSAA */
280 mt
->msaa_layout
= compute_msaa_layout(brw
, format
, mt
->target
);
281 if (mt
->msaa_layout
== INTEL_MSAA_LAYOUT_IMS
) {
282 /* In the Sandy Bridge PRM, volume 4, part 1, page 31, it says:
284 * "Any of the other messages (sample*, LOD, load4) used with a
285 * (4x) multisampled surface will in-effect sample a surface with
286 * double the height and width as that indicated in the surface
287 * state. Each pixel position on the original-sized surface is
288 * replaced with a 2x2 of samples with the following arrangement:
293 * Thus, when sampling from a multisampled texture, it behaves as
294 * though the layout in memory for (x,y,sample) is:
296 * (0,0,0) (0,0,2) (1,0,0) (1,0,2)
297 * (0,0,1) (0,0,3) (1,0,1) (1,0,3)
299 * (0,1,0) (0,1,2) (1,1,0) (1,1,2)
300 * (0,1,1) (0,1,3) (1,1,1) (1,1,3)
302 * However, the actual layout of multisampled data in memory is:
304 * (0,0,0) (1,0,0) (0,0,1) (1,0,1)
305 * (0,1,0) (1,1,0) (0,1,1) (1,1,1)
307 * (0,0,2) (1,0,2) (0,0,3) (1,0,3)
308 * (0,1,2) (1,1,2) (0,1,3) (1,1,3)
310 * This pattern repeats for each 2x2 pixel block.
312 * As a result, when calculating the size of our 4-sample buffer for
313 * an odd width or height, we have to align before scaling up because
314 * sample 3 is in that bottom right 2x2 block.
316 switch (num_samples
) {
318 assert(brw
->gen
>= 8);
319 width0
= ALIGN(width0
, 2) * 2;
320 height0
= ALIGN(height0
, 2);
323 width0
= ALIGN(width0
, 2) * 2;
324 height0
= ALIGN(height0
, 2) * 2;
327 width0
= ALIGN(width0
, 2) * 4;
328 height0
= ALIGN(height0
, 2) * 2;
331 /* num_samples should already have been quantized to 0, 1, 2, 4, or
337 /* Non-interleaved */
338 depth0
*= num_samples
;
342 /* array_spacing_lod0 is only used for non-IMS MSAA surfaces. TODO: can we
345 switch (mt
->msaa_layout
) {
346 case INTEL_MSAA_LAYOUT_NONE
:
347 case INTEL_MSAA_LAYOUT_IMS
:
348 mt
->array_spacing_lod0
= false;
350 case INTEL_MSAA_LAYOUT_UMS
:
351 case INTEL_MSAA_LAYOUT_CMS
:
352 mt
->array_spacing_lod0
= true;
356 if (target
== GL_TEXTURE_CUBE_MAP
) {
361 mt
->physical_width0
= width0
;
362 mt
->physical_height0
= height0
;
363 mt
->physical_depth0
= depth0
;
366 _mesa_get_format_base_format(format
) == GL_DEPTH_STENCIL
&&
367 (brw
->must_use_separate_stencil
||
368 (brw
->has_separate_stencil
&& brw_is_hiz_depth_format(brw
, format
)))) {
369 mt
->stencil_mt
= intel_miptree_create(brw
,
379 INTEL_MIPTREE_TILING_ANY
);
380 if (!mt
->stencil_mt
) {
381 intel_miptree_release(&mt
);
385 /* Fix up the Z miptree format for how we're splitting out separate
386 * stencil. Gen7 expects there to be no stencil bits in its depth buffer.
388 mt
->format
= intel_depth_format_for_depthstencil_format(mt
->format
);
391 if (format
== mt
->format
) {
392 _mesa_problem(NULL
, "Unknown format %s in separate stencil mt\n",
393 _mesa_get_format_name(mt
->format
));
397 brw_miptree_layout(brw
, mt
);
403 * \brief Helper function for intel_miptree_create().
406 intel_miptree_choose_tiling(struct brw_context
*brw
,
409 uint32_t num_samples
,
410 enum intel_miptree_tiling_mode requested
,
411 struct intel_mipmap_tree
*mt
)
413 if (format
== MESA_FORMAT_S_UINT8
) {
414 /* The stencil buffer is W tiled. However, we request from the kernel a
415 * non-tiled buffer because the GTT is incapable of W fencing.
417 return I915_TILING_NONE
;
420 /* Some usages may want only one type of tiling, like depth miptrees (Y
421 * tiled), or temporary BOs for uploading data once (linear).
424 case INTEL_MIPTREE_TILING_ANY
:
426 case INTEL_MIPTREE_TILING_Y
:
427 return I915_TILING_Y
;
428 case INTEL_MIPTREE_TILING_NONE
:
429 return I915_TILING_NONE
;
432 if (num_samples
> 1) {
433 /* From p82 of the Sandy Bridge PRM, dw3[1] of SURFACE_STATE ("Tiled
436 * [DevSNB+]: For multi-sample render targets, this field must be
437 * 1. MSRTs can only be tiled.
439 * Our usual reason for preferring X tiling (fast blits using the
440 * blitting engine) doesn't apply to MSAA, since we'll generally be
441 * downsampling or upsampling when blitting between the MSAA buffer
442 * and another buffer, and the blitting engine doesn't support that.
443 * So use Y tiling, since it makes better use of the cache.
445 return I915_TILING_Y
;
448 GLenum base_format
= _mesa_get_format_base_format(format
);
449 if (base_format
== GL_DEPTH_COMPONENT
||
450 base_format
== GL_DEPTH_STENCIL_EXT
)
451 return I915_TILING_Y
;
453 int minimum_pitch
= mt
->total_width
* mt
->cpp
;
455 /* If the width is much smaller than a tile, don't bother tiling. */
456 if (minimum_pitch
< 64)
457 return I915_TILING_NONE
;
459 if (ALIGN(minimum_pitch
, 512) >= 32768 ||
460 mt
->total_width
>= 32768 || mt
->total_height
>= 32768) {
461 perf_debug("%dx%d miptree too large to blit, falling back to untiled",
462 mt
->total_width
, mt
->total_height
);
463 return I915_TILING_NONE
;
466 /* Pre-gen6 doesn't have BLORP to handle Y-tiling, so use X-tiling. */
468 return I915_TILING_X
;
470 /* From the Sandybridge PRM, Volume 1, Part 2, page 32:
471 * "NOTE: 128BPE Format Color Buffer ( render target ) MUST be either TileX
473 * 128 bits per pixel translates to 16 bytes per pixel. This is necessary
474 * all the way back to 965, but is explicitly permitted on Gen7.
476 if (brw
->gen
!= 7 && mt
->cpp
>= 16)
477 return I915_TILING_X
;
479 /* From the Ivy Bridge PRM, Vol4 Part1 2.12.2.1 (SURFACE_STATE for most
480 * messages), on p64, under the heading "Surface Vertical Alignment":
482 * This field must be set to VALIGN_4 for all tiled Y Render Target
485 * So if the surface is renderable and uses a vertical alignment of 2,
486 * force it to be X tiled. This is somewhat conservative (it's possible
487 * that the client won't ever render to this surface), but it's difficult
488 * to know that ahead of time. And besides, since we use a vertical
489 * alignment of 4 as often as we can, this shouldn't happen very often.
491 if (brw
->gen
== 7 && mt
->align_h
== 2 &&
492 brw
->format_supported_as_render_target
[format
]) {
493 return I915_TILING_X
;
496 return I915_TILING_Y
| I915_TILING_X
;
501 * Choose an appropriate uncompressed format for a requested
502 * compressed format, if unsupported.
505 intel_lower_compressed_format(struct brw_context
*brw
, mesa_format format
)
507 /* No need to lower ETC formats on these platforms,
508 * they are supported natively.
510 if (brw
->gen
>= 8 || brw
->is_baytrail
)
514 case MESA_FORMAT_ETC1_RGB8
:
515 return MESA_FORMAT_R8G8B8X8_UNORM
;
516 case MESA_FORMAT_ETC2_RGB8
:
517 return MESA_FORMAT_R8G8B8X8_UNORM
;
518 case MESA_FORMAT_ETC2_SRGB8
:
519 case MESA_FORMAT_ETC2_SRGB8_ALPHA8_EAC
:
520 case MESA_FORMAT_ETC2_SRGB8_PUNCHTHROUGH_ALPHA1
:
521 return MESA_FORMAT_B8G8R8A8_SRGB
;
522 case MESA_FORMAT_ETC2_RGBA8_EAC
:
523 case MESA_FORMAT_ETC2_RGB8_PUNCHTHROUGH_ALPHA1
:
524 return MESA_FORMAT_R8G8B8A8_UNORM
;
525 case MESA_FORMAT_ETC2_R11_EAC
:
526 return MESA_FORMAT_R_UNORM16
;
527 case MESA_FORMAT_ETC2_SIGNED_R11_EAC
:
528 return MESA_FORMAT_R_SNORM16
;
529 case MESA_FORMAT_ETC2_RG11_EAC
:
530 return MESA_FORMAT_R16G16_UNORM
;
531 case MESA_FORMAT_ETC2_SIGNED_RG11_EAC
:
532 return MESA_FORMAT_R16G16_SNORM
;
534 /* Non ETC1 / ETC2 format */
540 struct intel_mipmap_tree
*
541 intel_miptree_create(struct brw_context
*brw
,
549 bool expect_accelerated_upload
,
551 enum intel_miptree_tiling_mode requested_tiling
)
553 struct intel_mipmap_tree
*mt
;
554 mesa_format tex_format
= format
;
555 mesa_format etc_format
= MESA_FORMAT_NONE
;
556 GLuint total_width
, total_height
;
558 format
= intel_lower_compressed_format(brw
, format
);
560 etc_format
= (format
!= tex_format
) ? tex_format
: MESA_FORMAT_NONE
;
562 mt
= intel_miptree_create_layout(brw
, target
, format
,
563 first_level
, last_level
, width0
,
567 * pitch == 0 || height == 0 indicates the null texture
569 if (!mt
|| !mt
->total_width
|| !mt
->total_height
) {
570 intel_miptree_release(&mt
);
574 total_width
= mt
->total_width
;
575 total_height
= mt
->total_height
;
577 if (format
== MESA_FORMAT_S_UINT8
) {
578 /* Align to size of W tile, 64x64. */
579 total_width
= ALIGN(total_width
, 64);
580 total_height
= ALIGN(total_height
, 64);
583 uint32_t tiling
= intel_miptree_choose_tiling(brw
, format
, width0
,
584 num_samples
, requested_tiling
,
586 bool y_or_x
= tiling
== (I915_TILING_Y
| I915_TILING_X
);
588 mt
->etc_format
= etc_format
;
589 mt
->region
= intel_region_alloc(brw
->intelScreen
,
590 y_or_x
? I915_TILING_Y
: tiling
,
594 expect_accelerated_upload
);
596 /* If the region is too large to fit in the aperture, we need to use the
597 * BLT engine to support it. The BLT paths can't currently handle Y-tiling,
598 * so we need to fall back to X.
600 if (y_or_x
&& mt
->region
->bo
->size
>= brw
->max_gtt_map_object_size
) {
601 perf_debug("%dx%d miptree larger than aperture; falling back to X-tiled\n",
602 mt
->total_width
, mt
->total_height
);
603 intel_region_release(&mt
->region
);
605 mt
->region
= intel_region_alloc(brw
->intelScreen
,
610 expect_accelerated_upload
);
616 intel_miptree_release(&mt
);
621 if (mt
->msaa_layout
== INTEL_MSAA_LAYOUT_CMS
) {
622 if (!intel_miptree_alloc_mcs(brw
, mt
, num_samples
)) {
623 intel_miptree_release(&mt
);
628 /* If this miptree is capable of supporting fast color clears, set
629 * fast_clear_state appropriately to ensure that fast clears will occur.
630 * Allocation of the MCS miptree will be deferred until the first fast
631 * clear actually occurs.
633 if (intel_is_non_msrt_mcs_buffer_supported(brw
, mt
))
634 mt
->fast_clear_state
= INTEL_FAST_CLEAR_STATE_RESOLVED
;
639 struct intel_mipmap_tree
*
640 intel_miptree_create_for_bo(struct brw_context
*brw
,
649 struct intel_mipmap_tree
*mt
;
651 struct intel_region
*region
= calloc(1, sizeof(*region
));
655 /* Nothing will be able to use this miptree with the BO if the offset isn't
658 if (tiling
!= I915_TILING_NONE
)
659 assert(offset
% 4096 == 0);
661 /* miptrees can't handle negative pitch. If you need flipping of images,
662 * that's outside of the scope of the mt.
666 mt
= intel_miptree_create_layout(brw
, GL_TEXTURE_2D
, format
,
669 true, 0 /* num_samples */);
675 region
->cpp
= mt
->cpp
;
676 region
->width
= width
;
677 region
->height
= height
;
678 region
->pitch
= pitch
;
679 region
->refcount
= 1;
680 drm_intel_bo_reference(bo
);
682 region
->tiling
= tiling
;
691 * For a singlesample image buffer, this simply wraps the given region with a miptree.
693 * For a multisample image buffer, this wraps the given region with
694 * a singlesample miptree, then creates a multisample miptree into which the
695 * singlesample miptree is embedded as a child.
698 intel_update_winsys_renderbuffer_miptree(struct brw_context
*intel
,
699 struct intel_renderbuffer
*irb
,
700 struct intel_region
*region
)
702 struct intel_mipmap_tree
*singlesample_mt
= NULL
;
703 struct intel_mipmap_tree
*multisample_mt
= NULL
;
704 struct gl_renderbuffer
*rb
= &irb
->Base
.Base
;
705 mesa_format format
= rb
->Format
;
706 int num_samples
= rb
->NumSamples
;
708 /* Only the front and back buffers, which are color buffers, are allocated
709 * through the image loader.
711 assert(_mesa_get_format_base_format(format
) == GL_RGB
||
712 _mesa_get_format_base_format(format
) == GL_RGBA
);
714 singlesample_mt
= intel_miptree_create_for_bo(intel
,
722 if (!singlesample_mt
)
725 singlesample_mt
->region
->name
= region
->name
;
727 /* If this miptree is capable of supporting fast color clears, set
728 * mcs_state appropriately to ensure that fast clears will occur.
729 * Allocation of the MCS miptree will be deferred until the first fast
730 * clear actually occurs.
732 if (intel_is_non_msrt_mcs_buffer_supported(intel
, singlesample_mt
))
733 singlesample_mt
->fast_clear_state
= INTEL_FAST_CLEAR_STATE_RESOLVED
;
735 if (num_samples
== 0) {
736 intel_miptree_release(&irb
->mt
);
737 irb
->mt
= singlesample_mt
;
739 assert(!irb
->singlesample_mt
);
741 intel_miptree_release(&irb
->singlesample_mt
);
742 irb
->singlesample_mt
= singlesample_mt
;
745 irb
->mt
->logical_width0
!= region
->width
||
746 irb
->mt
->logical_height0
!= region
->height
) {
747 multisample_mt
= intel_miptree_create_for_renderbuffer(intel
,
755 irb
->need_downsample
= false;
756 intel_miptree_release(&irb
->mt
);
757 irb
->mt
= multisample_mt
;
763 intel_miptree_release(&irb
->singlesample_mt
);
764 intel_miptree_release(&irb
->mt
);
768 struct intel_mipmap_tree
*
769 intel_miptree_create_for_renderbuffer(struct brw_context
*brw
,
773 uint32_t num_samples
)
775 struct intel_mipmap_tree
*mt
;
778 GLenum target
= num_samples
> 1 ? GL_TEXTURE_2D_MULTISAMPLE
: GL_TEXTURE_2D
;
780 mt
= intel_miptree_create(brw
, target
, format
, 0, 0,
781 width
, height
, depth
, true, num_samples
,
782 INTEL_MIPTREE_TILING_ANY
);
786 if (brw_is_hiz_depth_format(brw
, format
)) {
787 ok
= intel_miptree_alloc_hiz(brw
, mt
);
795 intel_miptree_release(&mt
);
800 intel_miptree_reference(struct intel_mipmap_tree
**dst
,
801 struct intel_mipmap_tree
*src
)
806 intel_miptree_release(dst
);
810 DBG("%s %p refcount now %d\n", __FUNCTION__
, src
, src
->refcount
);
818 intel_miptree_release(struct intel_mipmap_tree
**mt
)
823 DBG("%s %p refcount will be %d\n", __FUNCTION__
, *mt
, (*mt
)->refcount
- 1);
824 if (--(*mt
)->refcount
<= 0) {
827 DBG("%s deleting %p\n", __FUNCTION__
, *mt
);
829 intel_region_release(&((*mt
)->region
));
830 intel_miptree_release(&(*mt
)->stencil_mt
);
831 intel_miptree_release(&(*mt
)->hiz_mt
);
832 intel_miptree_release(&(*mt
)->mcs_mt
);
833 intel_resolve_map_clear(&(*mt
)->hiz_map
);
835 for (i
= 0; i
< MAX_TEXTURE_LEVELS
; i
++) {
836 free((*mt
)->level
[i
].slice
);
845 intel_miptree_get_dimensions_for_image(struct gl_texture_image
*image
,
846 int *width
, int *height
, int *depth
)
848 switch (image
->TexObject
->Target
) {
849 case GL_TEXTURE_1D_ARRAY
:
850 *width
= image
->Width
;
852 *depth
= image
->Height
;
855 *width
= image
->Width
;
856 *height
= image
->Height
;
857 *depth
= image
->Depth
;
863 * Can the image be pulled into a unified mipmap tree? This mirrors
864 * the completeness test in a lot of ways.
866 * Not sure whether I want to pass gl_texture_image here.
869 intel_miptree_match_image(struct intel_mipmap_tree
*mt
,
870 struct gl_texture_image
*image
)
872 struct intel_texture_image
*intelImage
= intel_texture_image(image
);
873 GLuint level
= intelImage
->base
.Base
.Level
;
874 int width
, height
, depth
;
876 /* glTexImage* choose the texture object based on the target passed in, and
877 * objects can't change targets over their lifetimes, so this should be
880 assert(image
->TexObject
->Target
== mt
->target
);
882 mesa_format mt_format
= mt
->format
;
883 if (mt
->format
== MESA_FORMAT_Z24_UNORM_X8_UINT
&& mt
->stencil_mt
)
884 mt_format
= MESA_FORMAT_Z24_UNORM_S8_UINT
;
885 if (mt
->format
== MESA_FORMAT_Z_FLOAT32
&& mt
->stencil_mt
)
886 mt_format
= MESA_FORMAT_Z32_FLOAT_S8X24_UINT
;
887 if (mt
->etc_format
!= MESA_FORMAT_NONE
)
888 mt_format
= mt
->etc_format
;
890 if (image
->TexFormat
!= mt_format
)
893 intel_miptree_get_dimensions_for_image(image
, &width
, &height
, &depth
);
895 if (mt
->target
== GL_TEXTURE_CUBE_MAP
)
898 int level_depth
= mt
->level
[level
].depth
;
899 if (mt
->num_samples
> 1) {
900 switch (mt
->msaa_layout
) {
901 case INTEL_MSAA_LAYOUT_NONE
:
902 case INTEL_MSAA_LAYOUT_IMS
:
904 case INTEL_MSAA_LAYOUT_UMS
:
905 case INTEL_MSAA_LAYOUT_CMS
:
906 level_depth
/= mt
->num_samples
;
911 /* Test image dimensions against the base level image adjusted for
912 * minification. This will also catch images not present in the
913 * tree, changed targets, etc.
915 if (width
!= minify(mt
->logical_width0
, level
- mt
->first_level
) ||
916 height
!= minify(mt
->logical_height0
, level
- mt
->first_level
) ||
917 depth
!= level_depth
) {
921 if (image
->NumSamples
!= mt
->num_samples
)
929 intel_miptree_set_level_info(struct intel_mipmap_tree
*mt
,
931 GLuint x
, GLuint y
, GLuint d
)
933 mt
->level
[level
].depth
= d
;
934 mt
->level
[level
].level_x
= x
;
935 mt
->level
[level
].level_y
= y
;
937 DBG("%s level %d, depth %d, offset %d,%d\n", __FUNCTION__
,
940 assert(mt
->level
[level
].slice
== NULL
);
942 mt
->level
[level
].slice
= calloc(d
, sizeof(*mt
->level
[0].slice
));
943 mt
->level
[level
].slice
[0].x_offset
= mt
->level
[level
].level_x
;
944 mt
->level
[level
].slice
[0].y_offset
= mt
->level
[level
].level_y
;
949 intel_miptree_set_image_offset(struct intel_mipmap_tree
*mt
,
950 GLuint level
, GLuint img
,
953 if (img
== 0 && level
== 0)
954 assert(x
== 0 && y
== 0);
956 assert(img
< mt
->level
[level
].depth
);
958 mt
->level
[level
].slice
[img
].x_offset
= mt
->level
[level
].level_x
+ x
;
959 mt
->level
[level
].slice
[img
].y_offset
= mt
->level
[level
].level_y
+ y
;
961 DBG("%s level %d img %d pos %d,%d\n",
962 __FUNCTION__
, level
, img
,
963 mt
->level
[level
].slice
[img
].x_offset
,
964 mt
->level
[level
].slice
[img
].y_offset
);
968 intel_miptree_get_image_offset(const struct intel_mipmap_tree
*mt
,
969 GLuint level
, GLuint slice
,
970 GLuint
*x
, GLuint
*y
)
972 assert(slice
< mt
->level
[level
].depth
);
974 *x
= mt
->level
[level
].slice
[slice
].x_offset
;
975 *y
= mt
->level
[level
].slice
[slice
].y_offset
;
979 * Rendering with tiled buffers requires that the base address of the buffer
980 * be aligned to a page boundary. For renderbuffers, and sometimes with
981 * textures, we may want the surface to point at a texture image level that
982 * isn't at a page boundary.
984 * This function returns an appropriately-aligned base offset
985 * according to the tiling restrictions, plus any required x/y offset
989 intel_miptree_get_tile_offsets(const struct intel_mipmap_tree
*mt
,
990 GLuint level
, GLuint slice
,
994 const struct intel_region
*region
= mt
->region
;
996 uint32_t mask_x
, mask_y
;
998 intel_region_get_tile_masks(region
, &mask_x
, &mask_y
, false);
999 intel_miptree_get_image_offset(mt
, level
, slice
, &x
, &y
);
1001 *tile_x
= x
& mask_x
;
1002 *tile_y
= y
& mask_y
;
1004 return intel_region_get_aligned_offset(region
, x
& ~mask_x
, y
& ~mask_y
,
1009 intel_miptree_copy_slice_sw(struct brw_context
*brw
,
1010 struct intel_mipmap_tree
*dst_mt
,
1011 struct intel_mipmap_tree
*src_mt
,
1018 int src_stride
, dst_stride
;
1019 int cpp
= dst_mt
->cpp
;
1021 intel_miptree_map(brw
, src_mt
,
1025 GL_MAP_READ_BIT
| BRW_MAP_DIRECT_BIT
,
1028 intel_miptree_map(brw
, dst_mt
,
1032 GL_MAP_WRITE_BIT
| GL_MAP_INVALIDATE_RANGE_BIT
|
1036 DBG("sw blit %s mt %p %p/%d -> %s mt %p %p/%d (%dx%d)\n",
1037 _mesa_get_format_name(src_mt
->format
),
1038 src_mt
, src
, src_stride
,
1039 _mesa_get_format_name(dst_mt
->format
),
1040 dst_mt
, dst
, dst_stride
,
1043 int row_size
= cpp
* width
;
1044 if (src_stride
== row_size
&&
1045 dst_stride
== row_size
) {
1046 memcpy(dst
, src
, row_size
* height
);
1048 for (int i
= 0; i
< height
; i
++) {
1049 memcpy(dst
, src
, row_size
);
1055 intel_miptree_unmap(brw
, dst_mt
, level
, slice
);
1056 intel_miptree_unmap(brw
, src_mt
, level
, slice
);
1058 /* Don't forget to copy the stencil data over, too. We could have skipped
1059 * passing BRW_MAP_DIRECT_BIT, but that would have meant intel_miptree_map
1060 * shuffling the two data sources in/out of temporary storage instead of
1061 * the direct mapping we get this way.
1063 if (dst_mt
->stencil_mt
) {
1064 assert(src_mt
->stencil_mt
);
1065 intel_miptree_copy_slice_sw(brw
, dst_mt
->stencil_mt
, src_mt
->stencil_mt
,
1066 level
, slice
, width
, height
);
1071 intel_miptree_copy_slice(struct brw_context
*brw
,
1072 struct intel_mipmap_tree
*dst_mt
,
1073 struct intel_mipmap_tree
*src_mt
,
1079 mesa_format format
= src_mt
->format
;
1080 uint32_t width
= minify(src_mt
->physical_width0
, level
- src_mt
->first_level
);
1081 uint32_t height
= minify(src_mt
->physical_height0
, level
- src_mt
->first_level
);
1089 assert(depth
< src_mt
->level
[level
].depth
);
1090 assert(src_mt
->format
== dst_mt
->format
);
1092 if (dst_mt
->compressed
) {
1093 height
= ALIGN(height
, dst_mt
->align_h
) / dst_mt
->align_h
;
1094 width
= ALIGN(width
, dst_mt
->align_w
);
1097 /* If it's a packed depth/stencil buffer with separate stencil, the blit
1098 * below won't apply since we can't do the depth's Y tiling or the
1099 * stencil's W tiling in the blitter.
1101 if (src_mt
->stencil_mt
) {
1102 intel_miptree_copy_slice_sw(brw
,
1109 uint32_t dst_x
, dst_y
, src_x
, src_y
;
1110 intel_miptree_get_image_offset(dst_mt
, level
, slice
, &dst_x
, &dst_y
);
1111 intel_miptree_get_image_offset(src_mt
, level
, slice
, &src_x
, &src_y
);
1113 DBG("validate blit mt %s %p %d,%d/%d -> mt %s %p %d,%d/%d (%dx%d)\n",
1114 _mesa_get_format_name(src_mt
->format
),
1115 src_mt
, src_x
, src_y
, src_mt
->region
->pitch
,
1116 _mesa_get_format_name(dst_mt
->format
),
1117 dst_mt
, dst_x
, dst_y
, dst_mt
->region
->pitch
,
1120 if (!intel_miptree_blit(brw
,
1121 src_mt
, level
, slice
, 0, 0, false,
1122 dst_mt
, level
, slice
, 0, 0, false,
1123 width
, height
, GL_COPY
)) {
1124 perf_debug("miptree validate blit for %s failed\n",
1125 _mesa_get_format_name(format
));
1127 intel_miptree_copy_slice_sw(brw
, dst_mt
, src_mt
, level
, slice
,
1133 * Copies the image's current data to the given miptree, and associates that
1134 * miptree with the image.
1136 * If \c invalidate is true, then the actual image data does not need to be
1137 * copied, but the image still needs to be associated to the new miptree (this
1138 * is set to true if we're about to clear the image).
1141 intel_miptree_copy_teximage(struct brw_context
*brw
,
1142 struct intel_texture_image
*intelImage
,
1143 struct intel_mipmap_tree
*dst_mt
,
1146 struct intel_mipmap_tree
*src_mt
= intelImage
->mt
;
1147 struct intel_texture_object
*intel_obj
=
1148 intel_texture_object(intelImage
->base
.Base
.TexObject
);
1149 int level
= intelImage
->base
.Base
.Level
;
1150 int face
= intelImage
->base
.Base
.Face
;
1151 GLuint depth
= intelImage
->base
.Base
.Depth
;
1154 for (int slice
= 0; slice
< depth
; slice
++) {
1155 intel_miptree_copy_slice(brw
, dst_mt
, src_mt
, level
, face
, slice
);
1159 intel_miptree_reference(&intelImage
->mt
, dst_mt
);
1160 intel_obj
->needs_validate
= true;
1164 intel_miptree_alloc_mcs(struct brw_context
*brw
,
1165 struct intel_mipmap_tree
*mt
,
1168 assert(brw
->gen
>= 7); /* MCS only used on Gen7+ */
1169 assert(mt
->mcs_mt
== NULL
);
1171 /* Choose the correct format for the MCS buffer. All that really matters
1172 * is that we allocate the right buffer size, since we'll always be
1173 * accessing this miptree using MCS-specific hardware mechanisms, which
1174 * infer the correct format based on num_samples.
1177 switch (num_samples
) {
1179 /* 8 bits/pixel are required for MCS data when using 4x MSAA (2 bits for
1182 format
= MESA_FORMAT_R_UNORM8
;
1185 /* 32 bits/pixel are required for MCS data when using 8x MSAA (3 bits
1186 * for each sample, plus 8 padding bits).
1188 format
= MESA_FORMAT_R_UINT32
;
1191 assert(!"Unrecognized sample count in intel_miptree_alloc_mcs");
1195 /* From the Ivy Bridge PRM, Vol4 Part1 p76, "MCS Base Address":
1197 * "The MCS surface must be stored as Tile Y."
1199 mt
->mcs_mt
= intel_miptree_create(brw
,
1205 mt
->logical_height0
,
1208 0 /* num_samples */,
1209 INTEL_MIPTREE_TILING_Y
);
1211 /* From the Ivy Bridge PRM, Vol 2 Part 1 p326:
1213 * When MCS buffer is enabled and bound to MSRT, it is required that it
1214 * is cleared prior to any rendering.
1216 * Since we don't use the MCS buffer for any purpose other than rendering,
1217 * it makes sense to just clear it immediately upon allocation.
1219 * Note: the clear value for MCS buffers is all 1's, so we memset to 0xff.
1221 void *data
= intel_miptree_map_raw(brw
, mt
->mcs_mt
);
1222 memset(data
, 0xff, mt
->mcs_mt
->region
->bo
->size
);
1223 intel_miptree_unmap_raw(brw
, mt
->mcs_mt
);
1224 mt
->fast_clear_state
= INTEL_FAST_CLEAR_STATE_CLEAR
;
1231 intel_miptree_alloc_non_msrt_mcs(struct brw_context
*brw
,
1232 struct intel_mipmap_tree
*mt
)
1234 assert(mt
->mcs_mt
== NULL
);
1236 /* The format of the MCS buffer is opaque to the driver; all that matters
1237 * is that we get its size and pitch right. We'll pretend that the format
1238 * is R32. Since an MCS tile covers 128 blocks horizontally, and a Y-tiled
1239 * R32 buffer is 32 pixels across, we'll need to scale the width down by
1240 * the block width and then a further factor of 4. Since an MCS tile
1241 * covers 256 blocks vertically, and a Y-tiled R32 buffer is 32 rows high,
1242 * we'll need to scale the height down by the block height and then a
1243 * further factor of 8.
1245 const mesa_format format
= MESA_FORMAT_R_UINT32
;
1246 unsigned block_width_px
;
1247 unsigned block_height
;
1248 intel_get_non_msrt_mcs_alignment(brw
, mt
, &block_width_px
, &block_height
);
1249 unsigned width_divisor
= block_width_px
* 4;
1250 unsigned height_divisor
= block_height
* 8;
1251 unsigned mcs_width
=
1252 ALIGN(mt
->logical_width0
, width_divisor
) / width_divisor
;
1253 unsigned mcs_height
=
1254 ALIGN(mt
->logical_height0
, height_divisor
) / height_divisor
;
1255 assert(mt
->logical_depth0
== 1);
1256 mt
->mcs_mt
= intel_miptree_create(brw
,
1265 0 /* num_samples */,
1266 INTEL_MIPTREE_TILING_Y
);
1273 * Helper for intel_miptree_alloc_hiz() that sets
1274 * \c mt->level[level].slice[layer].has_hiz. Return true if and only if
1275 * \c has_hiz was set.
1278 intel_miptree_slice_enable_hiz(struct brw_context
*brw
,
1279 struct intel_mipmap_tree
*mt
,
1285 if (brw
->gen
>= 8 || brw
->is_haswell
) {
1286 uint32_t width
= minify(mt
->physical_width0
, level
);
1287 uint32_t height
= minify(mt
->physical_height0
, level
);
1289 /* Disable HiZ for LOD > 0 unless the width is 8 aligned
1290 * and the height is 4 aligned. This allows our HiZ support
1291 * to fulfill Haswell restrictions for HiZ ops. For LOD == 0,
1292 * we can grow the width & height to allow the HiZ op to
1293 * force the proper size alignments.
1295 if (level
> 0 && ((width
& 7) || (height
& 3))) {
1300 mt
->level
[level
].slice
[layer
].has_hiz
= true;
1307 intel_miptree_alloc_hiz(struct brw_context
*brw
,
1308 struct intel_mipmap_tree
*mt
)
1310 assert(mt
->hiz_mt
== NULL
);
1311 mt
->hiz_mt
= intel_miptree_create(brw
,
1317 mt
->logical_height0
,
1321 INTEL_MIPTREE_TILING_ANY
);
1326 /* Mark that all slices need a HiZ resolve. */
1327 struct intel_resolve_map
*head
= &mt
->hiz_map
;
1328 for (int level
= mt
->first_level
; level
<= mt
->last_level
; ++level
) {
1329 for (int layer
= 0; layer
< mt
->level
[level
].depth
; ++layer
) {
1330 if (!intel_miptree_slice_enable_hiz(brw
, mt
, level
, layer
))
1333 head
->next
= malloc(sizeof(*head
->next
));
1334 head
->next
->prev
= head
;
1335 head
->next
->next
= NULL
;
1338 head
->level
= level
;
1339 head
->layer
= layer
;
1340 head
->need
= GEN6_HIZ_OP_HIZ_RESOLVE
;
1348 * Does the miptree slice have hiz enabled?
1351 intel_miptree_slice_has_hiz(struct intel_mipmap_tree
*mt
,
1355 intel_miptree_check_level_layer(mt
, level
, layer
);
1356 return mt
->level
[level
].slice
[layer
].has_hiz
;
1360 intel_miptree_slice_set_needs_hiz_resolve(struct intel_mipmap_tree
*mt
,
1364 if (!intel_miptree_slice_has_hiz(mt
, level
, layer
))
1367 intel_resolve_map_set(&mt
->hiz_map
,
1368 level
, layer
, GEN6_HIZ_OP_HIZ_RESOLVE
);
1373 intel_miptree_slice_set_needs_depth_resolve(struct intel_mipmap_tree
*mt
,
1377 if (!intel_miptree_slice_has_hiz(mt
, level
, layer
))
1380 intel_resolve_map_set(&mt
->hiz_map
,
1381 level
, layer
, GEN6_HIZ_OP_DEPTH_RESOLVE
);
1385 intel_miptree_set_all_slices_need_depth_resolve(struct intel_mipmap_tree
*mt
,
1389 uint32_t end_layer
= mt
->level
[level
].depth
;
1391 for (layer
= 0; layer
< end_layer
; layer
++) {
1392 intel_miptree_slice_set_needs_depth_resolve(mt
, level
, layer
);
1397 intel_miptree_slice_resolve(struct brw_context
*brw
,
1398 struct intel_mipmap_tree
*mt
,
1401 enum gen6_hiz_op need
)
1403 intel_miptree_check_level_layer(mt
, level
, layer
);
1405 struct intel_resolve_map
*item
=
1406 intel_resolve_map_get(&mt
->hiz_map
, level
, layer
);
1408 if (!item
|| item
->need
!= need
)
1411 intel_hiz_exec(brw
, mt
, level
, layer
, need
);
1412 intel_resolve_map_remove(item
);
1417 intel_miptree_slice_resolve_hiz(struct brw_context
*brw
,
1418 struct intel_mipmap_tree
*mt
,
1422 return intel_miptree_slice_resolve(brw
, mt
, level
, layer
,
1423 GEN6_HIZ_OP_HIZ_RESOLVE
);
1427 intel_miptree_slice_resolve_depth(struct brw_context
*brw
,
1428 struct intel_mipmap_tree
*mt
,
1432 return intel_miptree_slice_resolve(brw
, mt
, level
, layer
,
1433 GEN6_HIZ_OP_DEPTH_RESOLVE
);
1437 intel_miptree_all_slices_resolve(struct brw_context
*brw
,
1438 struct intel_mipmap_tree
*mt
,
1439 enum gen6_hiz_op need
)
1441 bool did_resolve
= false;
1442 struct intel_resolve_map
*i
, *next
;
1444 for (i
= mt
->hiz_map
.next
; i
; i
= next
) {
1446 if (i
->need
!= need
)
1449 intel_hiz_exec(brw
, mt
, i
->level
, i
->layer
, need
);
1450 intel_resolve_map_remove(i
);
1458 intel_miptree_all_slices_resolve_hiz(struct brw_context
*brw
,
1459 struct intel_mipmap_tree
*mt
)
1461 return intel_miptree_all_slices_resolve(brw
, mt
,
1462 GEN6_HIZ_OP_HIZ_RESOLVE
);
1466 intel_miptree_all_slices_resolve_depth(struct brw_context
*brw
,
1467 struct intel_mipmap_tree
*mt
)
1469 return intel_miptree_all_slices_resolve(brw
, mt
,
1470 GEN6_HIZ_OP_DEPTH_RESOLVE
);
1475 intel_miptree_resolve_color(struct brw_context
*brw
,
1476 struct intel_mipmap_tree
*mt
)
1478 switch (mt
->fast_clear_state
) {
1479 case INTEL_FAST_CLEAR_STATE_NO_MCS
:
1480 case INTEL_FAST_CLEAR_STATE_RESOLVED
:
1481 /* No resolve needed */
1483 case INTEL_FAST_CLEAR_STATE_UNRESOLVED
:
1484 case INTEL_FAST_CLEAR_STATE_CLEAR
:
1485 /* Fast color clear resolves only make sense for non-MSAA buffers. */
1486 if (mt
->msaa_layout
== INTEL_MSAA_LAYOUT_NONE
)
1487 brw_blorp_resolve_color(brw
, mt
);
1494 * Make it possible to share the region backing the given miptree with another
1495 * process or another miptree.
1497 * Fast color clears are unsafe with shared buffers, so we need to resolve and
1498 * then discard the MCS buffer, if present. We also set the fast_clear_state
1499 * to INTEL_FAST_CLEAR_STATE_NO_MCS to ensure that no MCS buffer gets
1500 * allocated in the future.
1503 intel_miptree_make_shareable(struct brw_context
*brw
,
1504 struct intel_mipmap_tree
*mt
)
1506 /* MCS buffers are also used for multisample buffers, but we can't resolve
1507 * away a multisample MCS buffer because it's an integral part of how the
1508 * pixel data is stored. Fortunately this code path should never be
1509 * reached for multisample buffers.
1511 assert(mt
->msaa_layout
== INTEL_MSAA_LAYOUT_NONE
);
1514 intel_miptree_resolve_color(brw
, mt
);
1515 intel_miptree_release(&mt
->mcs_mt
);
1516 mt
->fast_clear_state
= INTEL_FAST_CLEAR_STATE_NO_MCS
;
1522 * \brief Get pointer offset into stencil buffer.
1524 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
1525 * must decode the tile's layout in software.
1528 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
1530 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
1532 * Even though the returned offset is always positive, the return type is
1534 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
1535 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
1538 intel_offset_S8(uint32_t stride
, uint32_t x
, uint32_t y
, bool swizzled
)
1540 uint32_t tile_size
= 4096;
1541 uint32_t tile_width
= 64;
1542 uint32_t tile_height
= 64;
1543 uint32_t row_size
= 64 * stride
;
1545 uint32_t tile_x
= x
/ tile_width
;
1546 uint32_t tile_y
= y
/ tile_height
;
1548 /* The byte's address relative to the tile's base addres. */
1549 uint32_t byte_x
= x
% tile_width
;
1550 uint32_t byte_y
= y
% tile_height
;
1552 uintptr_t u
= tile_y
* row_size
1553 + tile_x
* tile_size
1554 + 512 * (byte_x
/ 8)
1556 + 32 * ((byte_y
/ 4) % 2)
1557 + 16 * ((byte_x
/ 4) % 2)
1558 + 8 * ((byte_y
/ 2) % 2)
1559 + 4 * ((byte_x
/ 2) % 2)
1564 /* adjust for bit6 swizzling */
1565 if (((byte_x
/ 8) % 2) == 1) {
1566 if (((byte_y
/ 8) % 2) == 0) {
1578 intel_miptree_updownsample(struct brw_context
*brw
,
1579 struct intel_mipmap_tree
*src
,
1580 struct intel_mipmap_tree
*dst
)
1582 if (brw
->gen
< 8 || src
->format
== MESA_FORMAT_S_UINT8
) {
1583 brw_blorp_blit_miptrees(brw
,
1584 src
, 0 /* level */, 0 /* layer */,
1585 dst
, 0 /* level */, 0 /* layer */,
1587 src
->logical_width0
, src
->logical_height0
,
1589 dst
->logical_width0
, dst
->logical_height0
,
1590 GL_NEAREST
, false, false /*mirror x, y*/);
1592 brw_meta_updownsample(brw
, src
, dst
);
1595 if (src
->stencil_mt
) {
1596 brw_blorp_blit_miptrees(brw
,
1597 src
->stencil_mt
, 0 /* level */, 0 /* layer */,
1598 dst
->stencil_mt
, 0 /* level */, 0 /* layer */,
1600 src
->logical_width0
, src
->logical_height0
,
1602 dst
->logical_width0
, dst
->logical_height0
,
1603 GL_NEAREST
, false, false /*mirror x, y*/);
1608 intel_miptree_map_raw(struct brw_context
*brw
, struct intel_mipmap_tree
*mt
)
1610 /* CPU accesses to color buffers don't understand fast color clears, so
1611 * resolve any pending fast color clears before we map.
1613 intel_miptree_resolve_color(brw
, mt
);
1615 drm_intel_bo
*bo
= mt
->region
->bo
;
1617 intel_batchbuffer_flush(brw
);
1619 if (mt
->region
->tiling
!= I915_TILING_NONE
)
1620 brw_bo_map_gtt(brw
, bo
, "miptree");
1622 brw_bo_map(brw
, bo
, true, "miptree");
1628 intel_miptree_unmap_raw(struct brw_context
*brw
,
1629 struct intel_mipmap_tree
*mt
)
1631 drm_intel_bo_unmap(mt
->region
->bo
);
1635 intel_miptree_map_gtt(struct brw_context
*brw
,
1636 struct intel_mipmap_tree
*mt
,
1637 struct intel_miptree_map
*map
,
1638 unsigned int level
, unsigned int slice
)
1640 unsigned int bw
, bh
;
1642 unsigned int image_x
, image_y
;
1646 /* For compressed formats, the stride is the number of bytes per
1647 * row of blocks. intel_miptree_get_image_offset() already does
1650 _mesa_get_format_block_size(mt
->format
, &bw
, &bh
);
1651 assert(y
% bh
== 0);
1654 base
= intel_miptree_map_raw(brw
, mt
) + mt
->offset
;
1659 /* Note that in the case of cube maps, the caller must have passed the
1660 * slice number referencing the face.
1662 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
1666 map
->stride
= mt
->region
->pitch
;
1667 map
->ptr
= base
+ y
* map
->stride
+ x
* mt
->cpp
;
1670 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __FUNCTION__
,
1671 map
->x
, map
->y
, map
->w
, map
->h
,
1672 mt
, _mesa_get_format_name(mt
->format
),
1673 x
, y
, map
->ptr
, map
->stride
);
1677 intel_miptree_unmap_gtt(struct brw_context
*brw
,
1678 struct intel_mipmap_tree
*mt
,
1679 struct intel_miptree_map
*map
,
1683 intel_miptree_unmap_raw(brw
, mt
);
1687 intel_miptree_map_blit(struct brw_context
*brw
,
1688 struct intel_mipmap_tree
*mt
,
1689 struct intel_miptree_map
*map
,
1690 unsigned int level
, unsigned int slice
)
1692 map
->mt
= intel_miptree_create(brw
, GL_TEXTURE_2D
, mt
->format
,
1696 INTEL_MIPTREE_TILING_NONE
);
1698 fprintf(stderr
, "Failed to allocate blit temporary\n");
1701 map
->stride
= map
->mt
->region
->pitch
;
1703 if (!intel_miptree_blit(brw
,
1705 map
->x
, map
->y
, false,
1708 map
->w
, map
->h
, GL_COPY
)) {
1709 fprintf(stderr
, "Failed to blit\n");
1713 map
->ptr
= intel_miptree_map_raw(brw
, map
->mt
);
1715 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __FUNCTION__
,
1716 map
->x
, map
->y
, map
->w
, map
->h
,
1717 mt
, _mesa_get_format_name(mt
->format
),
1718 level
, slice
, map
->ptr
, map
->stride
);
1723 intel_miptree_release(&map
->mt
);
1729 intel_miptree_unmap_blit(struct brw_context
*brw
,
1730 struct intel_mipmap_tree
*mt
,
1731 struct intel_miptree_map
*map
,
1735 struct gl_context
*ctx
= &brw
->ctx
;
1737 intel_miptree_unmap_raw(brw
, map
->mt
);
1739 if (map
->mode
& GL_MAP_WRITE_BIT
) {
1740 bool ok
= intel_miptree_blit(brw
,
1744 map
->x
, map
->y
, false,
1745 map
->w
, map
->h
, GL_COPY
);
1746 WARN_ONCE(!ok
, "Failed to blit from linear temporary mapping");
1749 intel_miptree_release(&map
->mt
);
1754 * "Map" a buffer by copying it to an untiled temporary using MOVNTDQA.
1757 intel_miptree_map_movntdqa(struct brw_context
*brw
,
1758 struct intel_mipmap_tree
*mt
,
1759 struct intel_miptree_map
*map
,
1760 unsigned int level
, unsigned int slice
)
1762 assert(map
->mode
& GL_MAP_READ_BIT
);
1763 assert(!(map
->mode
& GL_MAP_WRITE_BIT
));
1765 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __FUNCTION__
,
1766 map
->x
, map
->y
, map
->w
, map
->h
,
1767 mt
, _mesa_get_format_name(mt
->format
),
1768 level
, slice
, map
->ptr
, map
->stride
);
1770 /* Map the original image */
1773 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
1777 void *src
= intel_miptree_map_raw(brw
, mt
);
1780 src
+= image_y
* mt
->region
->pitch
;
1781 src
+= image_x
* mt
->region
->cpp
;
1783 /* Due to the pixel offsets for the particular image being mapped, our
1784 * src pointer may not be 16-byte aligned. However, if the pitch is
1785 * divisible by 16, then the amount by which it's misaligned will remain
1786 * consistent from row to row.
1788 assert((mt
->region
->pitch
% 16) == 0);
1789 const int misalignment
= ((uintptr_t) src
) & 15;
1791 /* Create an untiled temporary buffer for the mapping. */
1792 const unsigned width_bytes
= _mesa_format_row_stride(mt
->format
, map
->w
);
1794 map
->stride
= ALIGN(misalignment
+ width_bytes
, 16);
1796 map
->buffer
= malloc(map
->stride
* map
->h
);
1797 /* Offset the destination so it has the same misalignment as src. */
1798 map
->ptr
= map
->buffer
+ misalignment
;
1800 assert((((uintptr_t) map
->ptr
) & 15) == misalignment
);
1802 for (uint32_t y
= 0; y
< map
->h
; y
++) {
1803 void *dst_ptr
= map
->ptr
+ y
* map
->stride
;
1804 void *src_ptr
= src
+ y
* mt
->region
->pitch
;
1806 _mesa_streaming_load_memcpy(dst_ptr
, src_ptr
, width_bytes
);
1809 intel_miptree_unmap_raw(brw
, mt
);
1813 intel_miptree_unmap_movntdqa(struct brw_context
*brw
,
1814 struct intel_mipmap_tree
*mt
,
1815 struct intel_miptree_map
*map
,
1826 intel_miptree_map_s8(struct brw_context
*brw
,
1827 struct intel_mipmap_tree
*mt
,
1828 struct intel_miptree_map
*map
,
1829 unsigned int level
, unsigned int slice
)
1831 map
->stride
= map
->w
;
1832 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
1836 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1837 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1838 * invalidate is set, since we'll be writing the whole rectangle from our
1839 * temporary buffer back out.
1841 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
1842 uint8_t *untiled_s8_map
= map
->ptr
;
1843 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
);
1844 unsigned int image_x
, image_y
;
1846 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
1848 for (uint32_t y
= 0; y
< map
->h
; y
++) {
1849 for (uint32_t x
= 0; x
< map
->w
; x
++) {
1850 ptrdiff_t offset
= intel_offset_S8(mt
->region
->pitch
,
1851 x
+ image_x
+ map
->x
,
1852 y
+ image_y
+ map
->y
,
1853 brw
->has_swizzling
);
1854 untiled_s8_map
[y
* map
->w
+ x
] = tiled_s8_map
[offset
];
1858 intel_miptree_unmap_raw(brw
, mt
);
1860 DBG("%s: %d,%d %dx%d from mt %p %d,%d = %p/%d\n", __FUNCTION__
,
1861 map
->x
, map
->y
, map
->w
, map
->h
,
1862 mt
, map
->x
+ image_x
, map
->y
+ image_y
, map
->ptr
, map
->stride
);
1864 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __FUNCTION__
,
1865 map
->x
, map
->y
, map
->w
, map
->h
,
1866 mt
, map
->ptr
, map
->stride
);
1871 intel_miptree_unmap_s8(struct brw_context
*brw
,
1872 struct intel_mipmap_tree
*mt
,
1873 struct intel_miptree_map
*map
,
1877 if (map
->mode
& GL_MAP_WRITE_BIT
) {
1878 unsigned int image_x
, image_y
;
1879 uint8_t *untiled_s8_map
= map
->ptr
;
1880 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
);
1882 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
1884 for (uint32_t y
= 0; y
< map
->h
; y
++) {
1885 for (uint32_t x
= 0; x
< map
->w
; x
++) {
1886 ptrdiff_t offset
= intel_offset_S8(mt
->region
->pitch
,
1889 brw
->has_swizzling
);
1890 tiled_s8_map
[offset
] = untiled_s8_map
[y
* map
->w
+ x
];
1894 intel_miptree_unmap_raw(brw
, mt
);
1901 intel_miptree_map_etc(struct brw_context
*brw
,
1902 struct intel_mipmap_tree
*mt
,
1903 struct intel_miptree_map
*map
,
1907 assert(mt
->etc_format
!= MESA_FORMAT_NONE
);
1908 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
) {
1909 assert(mt
->format
== MESA_FORMAT_R8G8B8X8_UNORM
);
1912 assert(map
->mode
& GL_MAP_WRITE_BIT
);
1913 assert(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
);
1915 map
->stride
= _mesa_format_row_stride(mt
->etc_format
, map
->w
);
1916 map
->buffer
= malloc(_mesa_format_image_size(mt
->etc_format
,
1917 map
->w
, map
->h
, 1));
1918 map
->ptr
= map
->buffer
;
1922 intel_miptree_unmap_etc(struct brw_context
*brw
,
1923 struct intel_mipmap_tree
*mt
,
1924 struct intel_miptree_map
*map
,
1930 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
1935 uint8_t *dst
= intel_miptree_map_raw(brw
, mt
)
1936 + image_y
* mt
->region
->pitch
1937 + image_x
* mt
->region
->cpp
;
1939 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
)
1940 _mesa_etc1_unpack_rgba8888(dst
, mt
->region
->pitch
,
1941 map
->ptr
, map
->stride
,
1944 _mesa_unpack_etc2_format(dst
, mt
->region
->pitch
,
1945 map
->ptr
, map
->stride
,
1946 map
->w
, map
->h
, mt
->etc_format
);
1948 intel_miptree_unmap_raw(brw
, mt
);
1953 * Mapping function for packed depth/stencil miptrees backed by real separate
1954 * miptrees for depth and stencil.
1956 * On gen7, and to support HiZ pre-gen7, we have to have the stencil buffer
1957 * separate from the depth buffer. Yet at the GL API level, we have to expose
1958 * packed depth/stencil textures and FBO attachments, and Mesa core expects to
1959 * be able to map that memory for texture storage and glReadPixels-type
1960 * operations. We give Mesa core that access by mallocing a temporary and
1961 * copying the data between the actual backing store and the temporary.
1964 intel_miptree_map_depthstencil(struct brw_context
*brw
,
1965 struct intel_mipmap_tree
*mt
,
1966 struct intel_miptree_map
*map
,
1967 unsigned int level
, unsigned int slice
)
1969 struct intel_mipmap_tree
*z_mt
= mt
;
1970 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
1971 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
1972 int packed_bpp
= map_z32f_x24s8
? 8 : 4;
1974 map
->stride
= map
->w
* packed_bpp
;
1975 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
1979 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1980 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1981 * invalidate is set, since we'll be writing the whole rectangle from our
1982 * temporary buffer back out.
1984 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
1985 uint32_t *packed_map
= map
->ptr
;
1986 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
);
1987 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
);
1988 unsigned int s_image_x
, s_image_y
;
1989 unsigned int z_image_x
, z_image_y
;
1991 intel_miptree_get_image_offset(s_mt
, level
, slice
,
1992 &s_image_x
, &s_image_y
);
1993 intel_miptree_get_image_offset(z_mt
, level
, slice
,
1994 &z_image_x
, &z_image_y
);
1996 for (uint32_t y
= 0; y
< map
->h
; y
++) {
1997 for (uint32_t x
= 0; x
< map
->w
; x
++) {
1998 int map_x
= map
->x
+ x
, map_y
= map
->y
+ y
;
1999 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->region
->pitch
,
2002 brw
->has_swizzling
);
2003 ptrdiff_t z_offset
= ((map_y
+ z_image_y
) *
2004 (z_mt
->region
->pitch
/ 4) +
2005 (map_x
+ z_image_x
));
2006 uint8_t s
= s_map
[s_offset
];
2007 uint32_t z
= z_map
[z_offset
];
2009 if (map_z32f_x24s8
) {
2010 packed_map
[(y
* map
->w
+ x
) * 2 + 0] = z
;
2011 packed_map
[(y
* map
->w
+ x
) * 2 + 1] = s
;
2013 packed_map
[y
* map
->w
+ x
] = (s
<< 24) | (z
& 0x00ffffff);
2018 intel_miptree_unmap_raw(brw
, s_mt
);
2019 intel_miptree_unmap_raw(brw
, z_mt
);
2021 DBG("%s: %d,%d %dx%d from z mt %p %d,%d, s mt %p %d,%d = %p/%d\n",
2023 map
->x
, map
->y
, map
->w
, map
->h
,
2024 z_mt
, map
->x
+ z_image_x
, map
->y
+ z_image_y
,
2025 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
2026 map
->ptr
, map
->stride
);
2028 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __FUNCTION__
,
2029 map
->x
, map
->y
, map
->w
, map
->h
,
2030 mt
, map
->ptr
, map
->stride
);
2035 intel_miptree_unmap_depthstencil(struct brw_context
*brw
,
2036 struct intel_mipmap_tree
*mt
,
2037 struct intel_miptree_map
*map
,
2041 struct intel_mipmap_tree
*z_mt
= mt
;
2042 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
2043 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
2045 if (map
->mode
& GL_MAP_WRITE_BIT
) {
2046 uint32_t *packed_map
= map
->ptr
;
2047 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
);
2048 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
);
2049 unsigned int s_image_x
, s_image_y
;
2050 unsigned int z_image_x
, z_image_y
;
2052 intel_miptree_get_image_offset(s_mt
, level
, slice
,
2053 &s_image_x
, &s_image_y
);
2054 intel_miptree_get_image_offset(z_mt
, level
, slice
,
2055 &z_image_x
, &z_image_y
);
2057 for (uint32_t y
= 0; y
< map
->h
; y
++) {
2058 for (uint32_t x
= 0; x
< map
->w
; x
++) {
2059 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->region
->pitch
,
2060 x
+ s_image_x
+ map
->x
,
2061 y
+ s_image_y
+ map
->y
,
2062 brw
->has_swizzling
);
2063 ptrdiff_t z_offset
= ((y
+ z_image_y
) *
2064 (z_mt
->region
->pitch
/ 4) +
2067 if (map_z32f_x24s8
) {
2068 z_map
[z_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 0];
2069 s_map
[s_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 1];
2071 uint32_t packed
= packed_map
[y
* map
->w
+ x
];
2072 s_map
[s_offset
] = packed
>> 24;
2073 z_map
[z_offset
] = packed
;
2078 intel_miptree_unmap_raw(brw
, s_mt
);
2079 intel_miptree_unmap_raw(brw
, z_mt
);
2081 DBG("%s: %d,%d %dx%d from z mt %p (%s) %d,%d, s mt %p %d,%d = %p/%d\n",
2083 map
->x
, map
->y
, map
->w
, map
->h
,
2084 z_mt
, _mesa_get_format_name(z_mt
->format
),
2085 map
->x
+ z_image_x
, map
->y
+ z_image_y
,
2086 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
2087 map
->ptr
, map
->stride
);
2094 * Create and attach a map to the miptree at (level, slice). Return the
2097 static struct intel_miptree_map
*
2098 intel_miptree_attach_map(struct intel_mipmap_tree
*mt
,
2107 struct intel_miptree_map
*map
= calloc(1, sizeof(*map
));
2112 assert(mt
->level
[level
].slice
[slice
].map
== NULL
);
2113 mt
->level
[level
].slice
[slice
].map
= map
;
2125 * Release the map at (level, slice).
2128 intel_miptree_release_map(struct intel_mipmap_tree
*mt
,
2132 struct intel_miptree_map
**map
;
2134 map
= &mt
->level
[level
].slice
[slice
].map
;
2140 can_blit_slice(struct intel_mipmap_tree
*mt
,
2141 unsigned int level
, unsigned int slice
)
2145 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
2146 if (image_x
>= 32768 || image_y
>= 32768)
2149 if (mt
->region
->pitch
>= 32768)
2156 intel_miptree_map(struct brw_context
*brw
,
2157 struct intel_mipmap_tree
*mt
,
2168 struct intel_miptree_map
*map
;
2170 assert(mt
->num_samples
<= 1);
2172 map
= intel_miptree_attach_map(mt
, level
, slice
, x
, y
, w
, h
, mode
);
2179 intel_miptree_slice_resolve_depth(brw
, mt
, level
, slice
);
2180 if (map
->mode
& GL_MAP_WRITE_BIT
) {
2181 intel_miptree_slice_set_needs_hiz_resolve(mt
, level
, slice
);
2184 if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2185 intel_miptree_map_s8(brw
, mt
, map
, level
, slice
);
2186 } else if (mt
->etc_format
!= MESA_FORMAT_NONE
&&
2187 !(mode
& BRW_MAP_DIRECT_BIT
)) {
2188 intel_miptree_map_etc(brw
, mt
, map
, level
, slice
);
2189 } else if (mt
->stencil_mt
&& !(mode
& BRW_MAP_DIRECT_BIT
)) {
2190 intel_miptree_map_depthstencil(brw
, mt
, map
, level
, slice
);
2192 /* See intel_miptree_blit() for details on the 32k pitch limit. */
2193 else if (brw
->has_llc
&&
2194 !(mode
& GL_MAP_WRITE_BIT
) &&
2196 (mt
->region
->tiling
== I915_TILING_X
||
2197 (brw
->gen
>= 6 && mt
->region
->tiling
== I915_TILING_Y
)) &&
2198 can_blit_slice(mt
, level
, slice
)) {
2199 intel_miptree_map_blit(brw
, mt
, map
, level
, slice
);
2200 } else if (mt
->region
->tiling
!= I915_TILING_NONE
&&
2201 mt
->region
->bo
->size
>= brw
->max_gtt_map_object_size
) {
2202 assert(can_blit_slice(mt
, level
, slice
));
2203 intel_miptree_map_blit(brw
, mt
, map
, level
, slice
);
2205 } else if (!(mode
& GL_MAP_WRITE_BIT
) && !mt
->compressed
) {
2206 intel_miptree_map_movntdqa(brw
, mt
, map
, level
, slice
);
2209 intel_miptree_map_gtt(brw
, mt
, map
, level
, slice
);
2212 *out_ptr
= map
->ptr
;
2213 *out_stride
= map
->stride
;
2215 if (map
->ptr
== NULL
)
2216 intel_miptree_release_map(mt
, level
, slice
);
2220 intel_miptree_unmap(struct brw_context
*brw
,
2221 struct intel_mipmap_tree
*mt
,
2225 struct intel_miptree_map
*map
= mt
->level
[level
].slice
[slice
].map
;
2227 assert(mt
->num_samples
<= 1);
2232 DBG("%s: mt %p (%s) level %d slice %d\n", __FUNCTION__
,
2233 mt
, _mesa_get_format_name(mt
->format
), level
, slice
);
2235 if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2236 intel_miptree_unmap_s8(brw
, mt
, map
, level
, slice
);
2237 } else if (mt
->etc_format
!= MESA_FORMAT_NONE
&&
2238 !(map
->mode
& BRW_MAP_DIRECT_BIT
)) {
2239 intel_miptree_unmap_etc(brw
, mt
, map
, level
, slice
);
2240 } else if (mt
->stencil_mt
&& !(map
->mode
& BRW_MAP_DIRECT_BIT
)) {
2241 intel_miptree_unmap_depthstencil(brw
, mt
, map
, level
, slice
);
2242 } else if (map
->mt
) {
2243 intel_miptree_unmap_blit(brw
, mt
, map
, level
, slice
);
2245 } else if (map
->buffer
) {
2246 intel_miptree_unmap_movntdqa(brw
, mt
, map
, level
, slice
);
2249 intel_miptree_unmap_gtt(brw
, mt
, map
, level
, slice
);
2252 intel_miptree_release_map(mt
, level
, slice
);