2 * Copyright 2006 VMware, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
21 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
22 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
23 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 #include <GL/internal/dri_interface.h>
28 #include <drm_fourcc.h>
30 #include "intel_batchbuffer.h"
31 #include "intel_image.h"
32 #include "intel_mipmap_tree.h"
33 #include "intel_tex.h"
34 #include "intel_blit.h"
35 #include "intel_fbo.h"
37 #include "brw_blorp.h"
38 #include "brw_context.h"
39 #include "brw_state.h"
41 #include "main/enums.h"
42 #include "main/fbobject.h"
43 #include "main/formats.h"
44 #include "main/glformats.h"
45 #include "main/texcompress_etc.h"
46 #include "main/teximage.h"
47 #include "main/streaming-load-memcpy.h"
49 #include "util/format_srgb.h"
51 #include "x86/common_x86_asm.h"
53 #define FILE_DEBUG_FLAG DEBUG_MIPTREE
55 static void *intel_miptree_map_raw(struct brw_context
*brw
,
56 struct intel_mipmap_tree
*mt
,
59 static void intel_miptree_unmap_raw(struct intel_mipmap_tree
*mt
);
62 intel_miptree_supports_mcs(struct brw_context
*brw
,
63 const struct intel_mipmap_tree
*mt
)
65 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
67 /* MCS compression only applies to multisampled miptrees */
68 if (mt
->surf
.samples
<= 1)
71 /* Prior to Gen7, all MSAA surfaces used IMS layout. */
75 /* See isl_surf_get_mcs_surf for details. */
76 if (mt
->surf
.samples
== 16 && mt
->surf
.logical_level0_px
.width
> 8192)
79 /* In Gen7, IMS layout is only used for depth and stencil buffers. */
80 switch (_mesa_get_format_base_format(mt
->format
)) {
81 case GL_DEPTH_COMPONENT
:
82 case GL_STENCIL_INDEX
:
83 case GL_DEPTH_STENCIL
:
86 /* From the Ivy Bridge PRM, Vol4 Part1 p77 ("MCS Enable"):
88 * This field must be set to 0 for all SINT MSRTs when all RT channels
91 * In practice this means that we have to disable MCS for all signed
92 * integer MSAA buffers. The alternative, to disable MCS only when one
93 * of the render target channels is disabled, is impractical because it
94 * would require converting between CMS and UMS MSAA layouts on the fly,
97 if (devinfo
->gen
== 7 && _mesa_get_format_datatype(mt
->format
) == GL_INT
) {
106 intel_tiling_supports_ccs(const struct brw_context
*brw
,
107 enum isl_tiling tiling
)
109 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
111 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
112 * Target(s)", beneath the "Fast Color Clear" bullet (p326):
114 * - Support is limited to tiled render targets.
116 * Gen9 changes the restriction to Y-tile only.
118 if (devinfo
->gen
>= 9)
119 return tiling
== ISL_TILING_Y0
;
120 else if (devinfo
->gen
>= 7)
121 return tiling
!= ISL_TILING_LINEAR
;
127 * For a single-sampled render target ("non-MSRT"), determine if an MCS buffer
128 * can be used. This doesn't (and should not) inspect any of the properties of
131 * From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render Target(s)",
132 * beneath the "Fast Color Clear" bullet (p326):
134 * - Support is for non-mip-mapped and non-array surface types only.
136 * And then later, on p327:
138 * - MCS buffer for non-MSRT is supported only for RT formats 32bpp,
141 * From the Skylake documentation, it is made clear that X-tiling is no longer
144 * - MCS and Lossless compression is supported for TiledY/TileYs/TileYf
148 intel_miptree_supports_ccs(struct brw_context
*brw
,
149 const struct intel_mipmap_tree
*mt
)
151 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
153 /* MCS support does not exist prior to Gen7 */
154 if (devinfo
->gen
< 7)
157 /* This function applies only to non-multisampled render targets. */
158 if (mt
->surf
.samples
> 1)
161 /* MCS is only supported for color buffers */
162 if (!_mesa_is_format_color_format(mt
->format
))
165 if (mt
->cpp
!= 4 && mt
->cpp
!= 8 && mt
->cpp
!= 16)
168 const bool mip_mapped
= mt
->first_level
!= 0 || mt
->last_level
!= 0;
169 const bool arrayed
= mt
->surf
.logical_level0_px
.array_len
> 1 ||
170 mt
->surf
.logical_level0_px
.depth
> 1;
173 /* Multisample surfaces with the CMS layout are not layered surfaces,
174 * yet still have physical_depth0 > 1. Assert that we don't
175 * accidentally reject a multisampled surface here. We should have
176 * rejected it earlier by explicitly checking the sample count.
178 assert(mt
->surf
.samples
== 1);
181 /* Handle the hardware restrictions...
183 * All GENs have the following restriction: "MCS buffer for non-MSRT is
184 * supported only for RT formats 32bpp, 64bpp, and 128bpp."
186 * From the HSW PRM Volume 7: 3D-Media-GPGPU, page 652: (Color Clear of
187 * Non-MultiSampler Render Target Restrictions) Support is for
188 * non-mip-mapped and non-array surface types only.
190 * From the BDW PRM Volume 7: 3D-Media-GPGPU, page 649: (Color Clear of
191 * Non-MultiSampler Render Target Restriction). Mip-mapped and arrayed
192 * surfaces are supported with MCS buffer layout with these alignments in
193 * the RT space: Horizontal Alignment = 256 and Vertical Alignment = 128.
195 * From the SKL PRM Volume 7: 3D-Media-GPGPU, page 632: (Color Clear of
196 * Non-MultiSampler Render Target Restriction). Mip-mapped and arrayed
197 * surfaces are supported with MCS buffer layout with these alignments in
198 * the RT space: Horizontal Alignment = 128 and Vertical Alignment = 64.
200 if (devinfo
->gen
< 8 && (mip_mapped
|| arrayed
))
203 /* The PRM doesn't say this explicitly, but fast-clears don't appear to
204 * work for 3D textures until gen9 where the layout of 3D textures changes
205 * to match 2D array textures.
207 if (devinfo
->gen
<= 8 && mt
->surf
.dim
!= ISL_SURF_DIM_2D
)
210 /* There's no point in using an MCS buffer if the surface isn't in a
213 if (!brw
->mesa_format_supports_render
[mt
->format
])
220 intel_tiling_supports_hiz(const struct brw_context
*brw
,
221 enum isl_tiling tiling
)
223 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
225 if (devinfo
->gen
< 6)
228 return tiling
== ISL_TILING_Y0
;
232 intel_miptree_supports_hiz(const struct brw_context
*brw
,
233 const struct intel_mipmap_tree
*mt
)
238 switch (mt
->format
) {
239 case MESA_FORMAT_Z_FLOAT32
:
240 case MESA_FORMAT_Z32_FLOAT_S8X24_UINT
:
241 case MESA_FORMAT_Z24_UNORM_X8_UINT
:
242 case MESA_FORMAT_Z24_UNORM_S8_UINT
:
243 case MESA_FORMAT_Z_UNORM16
:
251 * Return true if the format that will be used to access the miptree is
252 * CCS_E-compatible with the miptree's linear/non-sRGB format.
254 * Why use the linear format? Well, although the miptree may be specified with
255 * an sRGB format, the usage of that color space/format can be toggled. Since
256 * our HW tends to support more linear formats than sRGB ones, we use this
257 * format variant for check for CCS_E compatibility.
260 format_ccs_e_compat_with_miptree(const struct gen_device_info
*devinfo
,
261 const struct intel_mipmap_tree
*mt
,
262 enum isl_format access_format
)
264 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
);
266 mesa_format linear_format
= _mesa_get_srgb_format_linear(mt
->format
);
267 enum isl_format isl_format
= brw_isl_format_for_mesa_format(linear_format
);
268 return isl_formats_are_ccs_e_compatible(devinfo
, isl_format
, access_format
);
272 intel_miptree_supports_ccs_e(struct brw_context
*brw
,
273 const struct intel_mipmap_tree
*mt
)
275 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
277 if (devinfo
->gen
< 9)
280 /* For now compression is only enabled for integer formats even though
281 * there exist supported floating point formats also. This is a heuristic
282 * decision based on current public benchmarks. In none of the cases these
283 * formats provided any improvement but a few cases were seen to regress.
284 * Hence these are left to to be enabled in the future when they are known
287 if (_mesa_get_format_datatype(mt
->format
) == GL_FLOAT
)
290 if (!intel_miptree_supports_ccs(brw
, mt
))
293 /* Many window system buffers are sRGB even if they are never rendered as
294 * sRGB. For those, we want CCS_E for when sRGBEncode is false. When the
295 * surface is used as sRGB, we fall back to CCS_D.
297 mesa_format linear_format
= _mesa_get_srgb_format_linear(mt
->format
);
298 enum isl_format isl_format
= brw_isl_format_for_mesa_format(linear_format
);
299 return isl_format_supports_ccs_e(&brw
->screen
->devinfo
, isl_format
);
303 * Determine depth format corresponding to a depth+stencil format,
304 * for separate stencil.
307 intel_depth_format_for_depthstencil_format(mesa_format format
) {
309 case MESA_FORMAT_Z24_UNORM_S8_UINT
:
310 return MESA_FORMAT_Z24_UNORM_X8_UINT
;
311 case MESA_FORMAT_Z32_FLOAT_S8X24_UINT
:
312 return MESA_FORMAT_Z_FLOAT32
;
319 create_mapping_table(GLenum target
, unsigned first_level
, unsigned last_level
,
320 unsigned depth0
, struct intel_mipmap_level
*table
)
322 for (unsigned level
= first_level
; level
<= last_level
; level
++) {
324 target
== GL_TEXTURE_3D
? minify(depth0
, level
) : depth0
;
326 table
[level
].slice
= calloc(d
, sizeof(*table
[0].slice
));
327 if (!table
[level
].slice
)
334 for (unsigned level
= first_level
; level
<= last_level
; level
++)
335 free(table
[level
].slice
);
341 needs_separate_stencil(const struct brw_context
*brw
,
342 struct intel_mipmap_tree
*mt
,
345 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
347 if (_mesa_get_format_base_format(format
) != GL_DEPTH_STENCIL
)
350 if (devinfo
->must_use_separate_stencil
)
353 return brw
->has_separate_stencil
&&
354 intel_miptree_supports_hiz(brw
, mt
);
358 * Choose the aux usage for this miptree. This function must be called fairly
359 * late in the miptree create process after we have a tiling.
362 intel_miptree_choose_aux_usage(struct brw_context
*brw
,
363 struct intel_mipmap_tree
*mt
)
365 assert(mt
->aux_usage
== ISL_AUX_USAGE_NONE
);
367 if (intel_miptree_supports_mcs(brw
, mt
)) {
368 assert(mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
369 mt
->aux_usage
= ISL_AUX_USAGE_MCS
;
370 } else if (intel_tiling_supports_ccs(brw
, mt
->surf
.tiling
) &&
371 intel_miptree_supports_ccs(brw
, mt
)) {
372 if (!unlikely(INTEL_DEBUG
& DEBUG_NO_RBC
) &&
373 intel_miptree_supports_ccs_e(brw
, mt
)) {
374 mt
->aux_usage
= ISL_AUX_USAGE_CCS_E
;
376 mt
->aux_usage
= ISL_AUX_USAGE_CCS_D
;
378 } else if (intel_tiling_supports_hiz(brw
, mt
->surf
.tiling
) &&
379 intel_miptree_supports_hiz(brw
, mt
)) {
380 mt
->aux_usage
= ISL_AUX_USAGE_HIZ
;
383 /* We can do fast-clear on all auxiliary surface types that are
384 * allocated through the normal texture creation paths.
386 if (mt
->aux_usage
!= ISL_AUX_USAGE_NONE
)
387 mt
->supports_fast_clear
= true;
392 * Choose an appropriate uncompressed format for a requested
393 * compressed format, if unsupported.
396 intel_lower_compressed_format(struct brw_context
*brw
, mesa_format format
)
398 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
400 /* No need to lower ETC formats on these platforms,
401 * they are supported natively.
403 if (devinfo
->gen
>= 8 || devinfo
->is_baytrail
)
407 case MESA_FORMAT_ETC1_RGB8
:
408 return MESA_FORMAT_R8G8B8X8_UNORM
;
409 case MESA_FORMAT_ETC2_RGB8
:
410 return MESA_FORMAT_R8G8B8X8_UNORM
;
411 case MESA_FORMAT_ETC2_SRGB8
:
412 case MESA_FORMAT_ETC2_SRGB8_ALPHA8_EAC
:
413 case MESA_FORMAT_ETC2_SRGB8_PUNCHTHROUGH_ALPHA1
:
414 return MESA_FORMAT_B8G8R8A8_SRGB
;
415 case MESA_FORMAT_ETC2_RGBA8_EAC
:
416 case MESA_FORMAT_ETC2_RGB8_PUNCHTHROUGH_ALPHA1
:
417 return MESA_FORMAT_R8G8B8A8_UNORM
;
418 case MESA_FORMAT_ETC2_R11_EAC
:
419 return MESA_FORMAT_R_UNORM16
;
420 case MESA_FORMAT_ETC2_SIGNED_R11_EAC
:
421 return MESA_FORMAT_R_SNORM16
;
422 case MESA_FORMAT_ETC2_RG11_EAC
:
423 return MESA_FORMAT_R16G16_UNORM
;
424 case MESA_FORMAT_ETC2_SIGNED_RG11_EAC
:
425 return MESA_FORMAT_R16G16_SNORM
;
427 /* Non ETC1 / ETC2 format */
433 brw_get_num_logical_layers(const struct intel_mipmap_tree
*mt
, unsigned level
)
435 if (mt
->surf
.dim
== ISL_SURF_DIM_3D
)
436 return minify(mt
->surf
.logical_level0_px
.depth
, level
);
438 return mt
->surf
.logical_level0_px
.array_len
;
441 UNUSED
static unsigned
442 get_num_phys_layers(const struct isl_surf
*surf
, unsigned level
)
444 /* In case of physical dimensions one needs to consider also the layout.
445 * See isl_calc_phys_level0_extent_sa().
447 if (surf
->dim
!= ISL_SURF_DIM_3D
)
448 return surf
->phys_level0_sa
.array_len
;
450 if (surf
->dim_layout
== ISL_DIM_LAYOUT_GEN4_2D
)
451 return minify(surf
->phys_level0_sa
.array_len
, level
);
453 return minify(surf
->phys_level0_sa
.depth
, level
);
456 /** \brief Assert that the level and layer are valid for the miptree. */
458 intel_miptree_check_level_layer(const struct intel_mipmap_tree
*mt
,
466 assert(level
>= mt
->first_level
);
467 assert(level
<= mt
->last_level
);
468 assert(layer
< get_num_phys_layers(&mt
->surf
, level
));
471 static enum isl_aux_state
**
472 create_aux_state_map(struct intel_mipmap_tree
*mt
,
473 enum isl_aux_state initial
)
475 const uint32_t levels
= mt
->last_level
+ 1;
477 uint32_t total_slices
= 0;
478 for (uint32_t level
= 0; level
< levels
; level
++)
479 total_slices
+= brw_get_num_logical_layers(mt
, level
);
481 const size_t per_level_array_size
= levels
* sizeof(enum isl_aux_state
*);
483 /* We're going to allocate a single chunk of data for both the per-level
484 * reference array and the arrays of aux_state. This makes cleanup
485 * significantly easier.
487 const size_t total_size
= per_level_array_size
+
488 total_slices
* sizeof(enum isl_aux_state
);
489 void *data
= malloc(total_size
);
493 enum isl_aux_state
**per_level_arr
= data
;
494 enum isl_aux_state
*s
= data
+ per_level_array_size
;
495 for (uint32_t level
= 0; level
< levels
; level
++) {
496 per_level_arr
[level
] = s
;
497 const unsigned level_layers
= brw_get_num_logical_layers(mt
, level
);
498 for (uint32_t a
= 0; a
< level_layers
; a
++)
501 assert((void *)s
== data
+ total_size
);
503 return per_level_arr
;
507 free_aux_state_map(enum isl_aux_state
**state
)
513 need_to_retile_as_linear(struct brw_context
*brw
, unsigned row_pitch
,
514 enum isl_tiling tiling
, unsigned samples
)
519 if (tiling
== ISL_TILING_LINEAR
)
522 /* If the width is much smaller than a tile, don't bother tiling. */
526 if (ALIGN(row_pitch
, 512) >= 32768) {
527 perf_debug("row pitch %u too large to blit, falling back to untiled",
536 need_to_retile_as_x(const struct brw_context
*brw
, uint64_t size
,
537 enum isl_tiling tiling
)
539 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
541 /* If the BO is too large to fit in the aperture, we need to use the
542 * BLT engine to support it. Prior to Sandybridge, the BLT paths can't
543 * handle Y-tiling, so we need to fall back to X.
545 if (devinfo
->gen
< 6 && size
>= brw
->max_gtt_map_object_size
&&
546 tiling
== ISL_TILING_Y0
)
552 static struct intel_mipmap_tree
*
553 make_surface(struct brw_context
*brw
, GLenum target
, mesa_format format
,
554 unsigned first_level
, unsigned last_level
,
555 unsigned width0
, unsigned height0
, unsigned depth0
,
556 unsigned num_samples
, isl_tiling_flags_t tiling_flags
,
557 isl_surf_usage_flags_t isl_usage_flags
, uint32_t alloc_flags
,
558 unsigned row_pitch
, struct brw_bo
*bo
)
560 struct intel_mipmap_tree
*mt
= calloc(sizeof(*mt
), 1);
564 if (!create_mapping_table(target
, first_level
, last_level
, depth0
,
572 if (target
== GL_TEXTURE_CUBE_MAP
||
573 target
== GL_TEXTURE_CUBE_MAP_ARRAY
)
574 isl_usage_flags
|= ISL_SURF_USAGE_CUBE_BIT
;
576 DBG("%s: %s %s %ux %u:%u:%u %d..%d <-- %p\n",
578 _mesa_enum_to_string(target
),
579 _mesa_get_format_name(format
),
580 num_samples
, width0
, height0
, depth0
,
581 first_level
, last_level
, mt
);
583 struct isl_surf_init_info init_info
= {
584 .dim
= get_isl_surf_dim(target
),
585 .format
= translate_tex_format(brw
, format
, false),
588 .depth
= target
== GL_TEXTURE_3D
? depth0
: 1,
589 .levels
= last_level
- first_level
+ 1,
590 .array_len
= target
== GL_TEXTURE_3D
? 1 : depth0
,
591 .samples
= num_samples
,
592 .row_pitch
= row_pitch
,
593 .usage
= isl_usage_flags
,
594 .tiling_flags
= tiling_flags
,
597 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
600 /* Depth surfaces are always Y-tiled and stencil is always W-tiled, although
601 * on gen7 platforms we also need to create Y-tiled copies of stencil for
602 * texturing since the hardware can't sample from W-tiled surfaces. For
603 * everything else, check for corner cases needing special treatment.
605 bool is_depth_stencil
=
606 mt
->surf
.usage
& (ISL_SURF_USAGE_STENCIL_BIT
| ISL_SURF_USAGE_DEPTH_BIT
);
607 if (!is_depth_stencil
) {
608 if (need_to_retile_as_linear(brw
, mt
->surf
.row_pitch
,
609 mt
->surf
.tiling
, mt
->surf
.samples
)) {
610 init_info
.tiling_flags
= 1u << ISL_TILING_LINEAR
;
611 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
613 } else if (need_to_retile_as_x(brw
, mt
->surf
.size
, mt
->surf
.tiling
)) {
614 init_info
.tiling_flags
= 1u << ISL_TILING_X
;
615 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
620 /* In case of linear the buffer gets padded by fixed 64 bytes and therefore
621 * the size may not be multiple of row_pitch.
622 * See isl_apply_surface_padding().
624 if (mt
->surf
.tiling
!= ISL_TILING_LINEAR
)
625 assert(mt
->surf
.size
% mt
->surf
.row_pitch
== 0);
628 mt
->bo
= brw_bo_alloc_tiled(brw
->bufmgr
, "isl-miptree",
631 isl_tiling_to_i915_tiling(
633 mt
->surf
.row_pitch
, alloc_flags
);
640 mt
->first_level
= first_level
;
641 mt
->last_level
= last_level
;
644 mt
->aux_state
= NULL
;
645 mt
->cpp
= isl_format_get_layout(mt
->surf
.format
)->bpb
/ 8;
646 mt
->compressed
= _mesa_is_format_compressed(format
);
647 mt
->drm_modifier
= DRM_FORMAT_MOD_INVALID
;
652 intel_miptree_release(&mt
);
657 make_separate_stencil_surface(struct brw_context
*brw
,
658 struct intel_mipmap_tree
*mt
)
660 mt
->stencil_mt
= make_surface(brw
, mt
->target
, MESA_FORMAT_S_UINT8
,
661 0, mt
->surf
.levels
- 1,
662 mt
->surf
.logical_level0_px
.width
,
663 mt
->surf
.logical_level0_px
.height
,
664 mt
->surf
.dim
== ISL_SURF_DIM_3D
?
665 mt
->surf
.logical_level0_px
.depth
:
666 mt
->surf
.logical_level0_px
.array_len
,
667 mt
->surf
.samples
, ISL_TILING_W_BIT
,
668 ISL_SURF_USAGE_STENCIL_BIT
|
669 ISL_SURF_USAGE_TEXTURE_BIT
,
670 BO_ALLOC_BUSY
, 0, NULL
);
675 mt
->stencil_mt
->r8stencil_needs_update
= true;
680 static struct intel_mipmap_tree
*
681 miptree_create(struct brw_context
*brw
,
690 enum intel_miptree_create_flags flags
)
692 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
694 if (format
== MESA_FORMAT_S_UINT8
)
695 return make_surface(brw
, target
, format
, first_level
, last_level
,
696 width0
, height0
, depth0
, num_samples
,
698 ISL_SURF_USAGE_STENCIL_BIT
|
699 ISL_SURF_USAGE_TEXTURE_BIT
,
704 const GLenum base_format
= _mesa_get_format_base_format(format
);
705 if ((base_format
== GL_DEPTH_COMPONENT
||
706 base_format
== GL_DEPTH_STENCIL
) &&
707 !(flags
& MIPTREE_CREATE_LINEAR
)) {
708 /* Fix up the Z miptree format for how we're splitting out separate
709 * stencil. Gen7 expects there to be no stencil bits in its depth buffer.
711 const mesa_format depth_only_format
=
712 intel_depth_format_for_depthstencil_format(format
);
713 struct intel_mipmap_tree
*mt
= make_surface(
714 brw
, target
, devinfo
->gen
>= 6 ? depth_only_format
: format
,
715 first_level
, last_level
,
716 width0
, height0
, depth0
, num_samples
, ISL_TILING_Y0_BIT
,
717 ISL_SURF_USAGE_DEPTH_BIT
| ISL_SURF_USAGE_TEXTURE_BIT
,
718 BO_ALLOC_BUSY
, 0, NULL
);
720 if (needs_separate_stencil(brw
, mt
, format
) &&
721 !make_separate_stencil_surface(brw
, mt
)) {
722 intel_miptree_release(&mt
);
726 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
727 intel_miptree_choose_aux_usage(brw
, mt
);
732 mesa_format tex_format
= format
;
733 mesa_format etc_format
= MESA_FORMAT_NONE
;
734 uint32_t alloc_flags
= 0;
736 format
= intel_lower_compressed_format(brw
, format
);
738 etc_format
= (format
!= tex_format
) ? tex_format
: MESA_FORMAT_NONE
;
740 if (flags
& MIPTREE_CREATE_BUSY
)
741 alloc_flags
|= BO_ALLOC_BUSY
;
743 isl_tiling_flags_t tiling_flags
= (flags
& MIPTREE_CREATE_LINEAR
) ?
744 ISL_TILING_LINEAR_BIT
: ISL_TILING_ANY_MASK
;
746 /* TODO: This used to be because there wasn't BLORP to handle Y-tiling. */
747 if (devinfo
->gen
< 6)
748 tiling_flags
&= ~ISL_TILING_Y0_BIT
;
750 struct intel_mipmap_tree
*mt
= make_surface(
752 first_level
, last_level
,
753 width0
, height0
, depth0
,
754 num_samples
, tiling_flags
,
755 ISL_SURF_USAGE_RENDER_TARGET_BIT
|
756 ISL_SURF_USAGE_TEXTURE_BIT
,
757 alloc_flags
, 0, NULL
);
761 mt
->etc_format
= etc_format
;
763 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
764 intel_miptree_choose_aux_usage(brw
, mt
);
769 struct intel_mipmap_tree
*
770 intel_miptree_create(struct brw_context
*brw
,
779 enum intel_miptree_create_flags flags
)
781 assert(num_samples
> 0);
783 struct intel_mipmap_tree
*mt
= miptree_create(
785 first_level
, last_level
,
786 width0
, height0
, depth0
, num_samples
,
793 /* Create the auxiliary surface up-front. CCS_D, on the other hand, can only
794 * compress clear color so we wait until an actual fast-clear to allocate
797 if (mt
->aux_usage
!= ISL_AUX_USAGE_CCS_D
&&
798 !intel_miptree_alloc_aux(brw
, mt
)) {
799 intel_miptree_release(&mt
);
806 struct intel_mipmap_tree
*
807 intel_miptree_create_for_bo(struct brw_context
*brw
,
815 enum isl_tiling tiling
,
816 enum intel_miptree_create_flags flags
)
818 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
819 struct intel_mipmap_tree
*mt
;
820 const GLenum target
= depth
> 1 ? GL_TEXTURE_2D_ARRAY
: GL_TEXTURE_2D
;
821 const GLenum base_format
= _mesa_get_format_base_format(format
);
823 if ((base_format
== GL_DEPTH_COMPONENT
||
824 base_format
== GL_DEPTH_STENCIL
)) {
825 const mesa_format depth_only_format
=
826 intel_depth_format_for_depthstencil_format(format
);
827 mt
= make_surface(brw
, target
,
828 devinfo
->gen
>= 6 ? depth_only_format
: format
,
829 0, 0, width
, height
, depth
, 1, ISL_TILING_Y0_BIT
,
830 ISL_SURF_USAGE_DEPTH_BIT
| ISL_SURF_USAGE_TEXTURE_BIT
,
835 brw_bo_reference(bo
);
837 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
838 intel_miptree_choose_aux_usage(brw
, mt
);
841 } else if (format
== MESA_FORMAT_S_UINT8
) {
842 mt
= make_surface(brw
, target
, MESA_FORMAT_S_UINT8
,
843 0, 0, width
, height
, depth
, 1,
845 ISL_SURF_USAGE_STENCIL_BIT
|
846 ISL_SURF_USAGE_TEXTURE_BIT
,
851 assert(bo
->size
>= mt
->surf
.size
);
853 brw_bo_reference(bo
);
857 /* Nothing will be able to use this miptree with the BO if the offset isn't
860 if (tiling
!= ISL_TILING_LINEAR
)
861 assert(offset
% 4096 == 0);
863 /* miptrees can't handle negative pitch. If you need flipping of images,
864 * that's outside of the scope of the mt.
868 /* The BO already has a tiling format and we shouldn't confuse the lower
869 * layers by making it try to find a tiling format again.
871 assert((flags
& MIPTREE_CREATE_LINEAR
) == 0);
873 mt
= make_surface(brw
, target
, format
,
874 0, 0, width
, height
, depth
, 1,
876 ISL_SURF_USAGE_RENDER_TARGET_BIT
|
877 ISL_SURF_USAGE_TEXTURE_BIT
,
882 brw_bo_reference(bo
);
886 if (!(flags
& MIPTREE_CREATE_NO_AUX
)) {
887 intel_miptree_choose_aux_usage(brw
, mt
);
889 /* Create the auxiliary surface up-front. CCS_D, on the other hand, can
890 * only compress clear color so we wait until an actual fast-clear to
893 if (mt
->aux_usage
!= ISL_AUX_USAGE_CCS_D
&&
894 !intel_miptree_alloc_aux(brw
, mt
)) {
895 intel_miptree_release(&mt
);
903 static struct intel_mipmap_tree
*
904 miptree_create_for_planar_image(struct brw_context
*brw
,
905 __DRIimage
*image
, GLenum target
,
906 enum isl_tiling tiling
)
908 const struct intel_image_format
*f
= image
->planar_format
;
909 struct intel_mipmap_tree
*planar_mt
= NULL
;
911 for (int i
= 0; i
< f
->nplanes
; i
++) {
912 const int index
= f
->planes
[i
].buffer_index
;
913 const uint32_t dri_format
= f
->planes
[i
].dri_format
;
914 const mesa_format format
= driImageFormatToGLFormat(dri_format
);
915 const uint32_t width
= image
->width
>> f
->planes
[i
].width_shift
;
916 const uint32_t height
= image
->height
>> f
->planes
[i
].height_shift
;
918 /* Disable creation of the texture's aux buffers because the driver
919 * exposes no EGL API to manage them. That is, there is no API for
920 * resolving the aux buffer's content to the main buffer nor for
921 * invalidating the aux buffer's content.
923 struct intel_mipmap_tree
*mt
=
924 intel_miptree_create_for_bo(brw
, image
->bo
, format
,
925 image
->offsets
[index
],
927 image
->strides
[index
],
929 MIPTREE_CREATE_NO_AUX
);
931 intel_miptree_release(&planar_mt
);
940 planar_mt
->plane
[i
- 1] = mt
;
943 planar_mt
->drm_modifier
= image
->modifier
;
949 create_ccs_buf_for_image(struct brw_context
*brw
,
951 struct intel_mipmap_tree
*mt
,
952 enum isl_aux_state initial_state
)
954 struct isl_surf temp_ccs_surf
;
956 /* CCS is only supported for very simple miptrees */
957 assert(image
->aux_offset
!= 0 && image
->aux_pitch
!= 0);
958 assert(image
->tile_x
== 0 && image
->tile_y
== 0);
959 assert(mt
->surf
.samples
== 1);
960 assert(mt
->surf
.levels
== 1);
961 assert(mt
->surf
.logical_level0_px
.depth
== 1);
962 assert(mt
->surf
.logical_level0_px
.array_len
== 1);
963 assert(mt
->first_level
== 0);
964 assert(mt
->last_level
== 0);
966 /* We shouldn't already have a CCS */
967 assert(!mt
->aux_buf
);
969 if (!isl_surf_get_ccs_surf(&brw
->isl_dev
, &mt
->surf
, &temp_ccs_surf
,
973 assert(image
->aux_offset
< image
->bo
->size
);
974 assert(temp_ccs_surf
.size
<= image
->bo
->size
- image
->aux_offset
);
976 mt
->aux_buf
= calloc(sizeof(*mt
->aux_buf
), 1);
977 if (mt
->aux_buf
== NULL
)
980 mt
->aux_state
= create_aux_state_map(mt
, initial_state
);
981 if (!mt
->aux_state
) {
987 /* On gen10+ we start using an extra space in the aux buffer to store the
988 * indirect clear color. However, if we imported an image from the window
989 * system with CCS, we don't have the extra space at the end of the aux
990 * buffer. So create a new bo here that will store that clear color.
992 if (brw
->isl_dev
.ss
.clear_color_state_size
> 0) {
993 mt
->aux_buf
->clear_color_bo
=
994 brw_bo_alloc_tiled(brw
->bufmgr
, "clear_color_bo",
995 brw
->isl_dev
.ss
.clear_color_state_size
,
996 BRW_MEMZONE_OTHER
, I915_TILING_NONE
, 0,
998 if (!mt
->aux_buf
->clear_color_bo
) {
1005 mt
->aux_buf
->bo
= image
->bo
;
1006 brw_bo_reference(image
->bo
);
1008 mt
->aux_buf
->offset
= image
->aux_offset
;
1009 mt
->aux_buf
->surf
= temp_ccs_surf
;
1014 struct intel_mipmap_tree
*
1015 intel_miptree_create_for_dri_image(struct brw_context
*brw
,
1016 __DRIimage
*image
, GLenum target
,
1018 bool allow_internal_aux
)
1020 uint32_t bo_tiling
, bo_swizzle
;
1021 brw_bo_get_tiling(image
->bo
, &bo_tiling
, &bo_swizzle
);
1023 const struct isl_drm_modifier_info
*mod_info
=
1024 isl_drm_modifier_get_info(image
->modifier
);
1026 const enum isl_tiling tiling
=
1027 mod_info
? mod_info
->tiling
: isl_tiling_from_i915_tiling(bo_tiling
);
1029 if (image
->planar_format
&& image
->planar_format
->nplanes
> 1)
1030 return miptree_create_for_planar_image(brw
, image
, target
, tiling
);
1032 if (image
->planar_format
)
1033 assert(image
->planar_format
->planes
[0].dri_format
== image
->dri_format
);
1035 if (!brw
->ctx
.TextureFormatSupported
[format
]) {
1036 /* The texture storage paths in core Mesa detect if the driver does not
1037 * support the user-requested format, and then searches for a
1038 * fallback format. The DRIimage code bypasses core Mesa, though. So we
1039 * do the fallbacks here for important formats.
1041 * We must support DRM_FOURCC_XBGR8888 textures because the Android
1042 * framework produces HAL_PIXEL_FORMAT_RGBX8888 winsys surfaces, which
1043 * the Chrome OS compositor consumes as dma_buf EGLImages.
1045 format
= _mesa_format_fallback_rgbx_to_rgba(format
);
1048 if (!brw
->ctx
.TextureFormatSupported
[format
])
1051 enum intel_miptree_create_flags mt_create_flags
= 0;
1053 /* If this image comes in from a window system, we have different
1054 * requirements than if it comes in via an EGL import operation. Window
1055 * system images can use any form of auxiliary compression we wish because
1056 * they get "flushed" before being handed off to the window system and we
1057 * have the opportunity to do resolves. Non window-system images, on the
1058 * other hand, have no resolve point so we can't have aux without a
1061 if (!allow_internal_aux
)
1062 mt_create_flags
|= MIPTREE_CREATE_NO_AUX
;
1064 /* If we have a modifier which specifies aux, don't create one yet */
1065 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
)
1066 mt_create_flags
|= MIPTREE_CREATE_NO_AUX
;
1068 /* Disable creation of the texture's aux buffers because the driver exposes
1069 * no EGL API to manage them. That is, there is no API for resolving the aux
1070 * buffer's content to the main buffer nor for invalidating the aux buffer's
1073 struct intel_mipmap_tree
*mt
=
1074 intel_miptree_create_for_bo(brw
, image
->bo
, format
,
1075 image
->offset
, image
->width
, image
->height
, 1,
1076 image
->pitch
, tiling
, mt_create_flags
);
1080 mt
->target
= target
;
1081 mt
->level
[0].level_x
= image
->tile_x
;
1082 mt
->level
[0].level_y
= image
->tile_y
;
1083 mt
->drm_modifier
= image
->modifier
;
1085 /* From "OES_EGL_image" error reporting. We report GL_INVALID_OPERATION
1086 * for EGL images from non-tile aligned sufaces in gen4 hw and earlier which has
1087 * trouble resolving back to destination image due to alignment issues.
1089 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1090 if (!devinfo
->has_surface_tile_offset
) {
1091 uint32_t draw_x
, draw_y
;
1092 intel_miptree_get_tile_offsets(mt
, 0, 0, &draw_x
, &draw_y
);
1094 if (draw_x
!= 0 || draw_y
!= 0) {
1095 _mesa_error(&brw
->ctx
, GL_INVALID_OPERATION
, __func__
);
1096 intel_miptree_release(&mt
);
1101 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
1102 assert(mod_info
->aux_usage
== ISL_AUX_USAGE_CCS_E
);
1104 mt
->aux_usage
= mod_info
->aux_usage
;
1105 /* If we are a window system buffer, then we can support fast-clears
1106 * even if the modifier doesn't support them by doing a partial resolve
1107 * as part of the flush operation.
1109 mt
->supports_fast_clear
=
1110 allow_internal_aux
|| mod_info
->supports_clear_color
;
1112 /* We don't know the actual state of the surface when we get it but we
1113 * can make a pretty good guess based on the modifier. What we do know
1114 * for sure is that it isn't in the AUX_INVALID state, so we just assume
1115 * a worst case of compression.
1117 enum isl_aux_state initial_state
=
1118 isl_drm_modifier_get_default_aux_state(image
->modifier
);
1120 if (!create_ccs_buf_for_image(brw
, image
, mt
, initial_state
)) {
1121 intel_miptree_release(&mt
);
1126 /* Don't assume coherency for imported EGLimages. We don't know what
1127 * external clients are going to do with it. They may scan it out.
1129 image
->bo
->cache_coherent
= false;
1135 * For a singlesample renderbuffer, this simply wraps the given BO with a
1138 * For a multisample renderbuffer, this wraps the window system's
1139 * (singlesample) BO with a singlesample miptree attached to the
1140 * intel_renderbuffer, then creates a multisample miptree attached to irb->mt
1141 * that will contain the actual rendering (which is lazily resolved to
1142 * irb->singlesample_mt).
1145 intel_update_winsys_renderbuffer_miptree(struct brw_context
*intel
,
1146 struct intel_renderbuffer
*irb
,
1147 struct intel_mipmap_tree
*singlesample_mt
,
1148 uint32_t width
, uint32_t height
,
1151 struct intel_mipmap_tree
*multisample_mt
= NULL
;
1152 struct gl_renderbuffer
*rb
= &irb
->Base
.Base
;
1153 mesa_format format
= rb
->Format
;
1154 const unsigned num_samples
= MAX2(rb
->NumSamples
, 1);
1156 /* Only the front and back buffers, which are color buffers, are allocated
1157 * through the image loader.
1159 assert(_mesa_get_format_base_format(format
) == GL_RGB
||
1160 _mesa_get_format_base_format(format
) == GL_RGBA
);
1162 assert(singlesample_mt
);
1164 if (num_samples
== 1) {
1165 intel_miptree_release(&irb
->mt
);
1166 irb
->mt
= singlesample_mt
;
1168 assert(!irb
->singlesample_mt
);
1170 intel_miptree_release(&irb
->singlesample_mt
);
1171 irb
->singlesample_mt
= singlesample_mt
;
1174 irb
->mt
->surf
.logical_level0_px
.width
!= width
||
1175 irb
->mt
->surf
.logical_level0_px
.height
!= height
) {
1176 multisample_mt
= intel_miptree_create_for_renderbuffer(intel
,
1181 if (!multisample_mt
)
1184 irb
->need_downsample
= false;
1185 intel_miptree_release(&irb
->mt
);
1186 irb
->mt
= multisample_mt
;
1192 intel_miptree_release(&irb
->mt
);
1196 struct intel_mipmap_tree
*
1197 intel_miptree_create_for_renderbuffer(struct brw_context
*brw
,
1201 uint32_t num_samples
)
1203 struct intel_mipmap_tree
*mt
;
1205 GLenum target
= num_samples
> 1 ? GL_TEXTURE_2D_MULTISAMPLE
: GL_TEXTURE_2D
;
1207 mt
= intel_miptree_create(brw
, target
, format
, 0, 0,
1208 width
, height
, depth
, num_samples
,
1209 MIPTREE_CREATE_BUSY
);
1216 intel_miptree_release(&mt
);
1221 intel_miptree_reference(struct intel_mipmap_tree
**dst
,
1222 struct intel_mipmap_tree
*src
)
1227 intel_miptree_release(dst
);
1231 DBG("%s %p refcount now %d\n", __func__
, src
, src
->refcount
);
1238 intel_miptree_aux_buffer_free(struct intel_miptree_aux_buffer
*aux_buf
)
1240 if (aux_buf
== NULL
)
1243 brw_bo_unreference(aux_buf
->bo
);
1244 brw_bo_unreference(aux_buf
->clear_color_bo
);
1250 intel_miptree_release(struct intel_mipmap_tree
**mt
)
1255 DBG("%s %p refcount will be %d\n", __func__
, *mt
, (*mt
)->refcount
- 1);
1256 if (--(*mt
)->refcount
<= 0) {
1259 DBG("%s deleting %p\n", __func__
, *mt
);
1261 brw_bo_unreference((*mt
)->bo
);
1262 intel_miptree_release(&(*mt
)->stencil_mt
);
1263 intel_miptree_release(&(*mt
)->r8stencil_mt
);
1264 intel_miptree_aux_buffer_free((*mt
)->aux_buf
);
1265 free_aux_state_map((*mt
)->aux_state
);
1267 intel_miptree_release(&(*mt
)->plane
[0]);
1268 intel_miptree_release(&(*mt
)->plane
[1]);
1270 for (i
= 0; i
< MAX_TEXTURE_LEVELS
; i
++) {
1271 free((*mt
)->level
[i
].slice
);
1281 intel_get_image_dims(struct gl_texture_image
*image
,
1282 int *width
, int *height
, int *depth
)
1284 switch (image
->TexObject
->Target
) {
1285 case GL_TEXTURE_1D_ARRAY
:
1286 /* For a 1D Array texture the OpenGL API will treat the image height as
1287 * the number of array slices. For Intel hardware, we treat the 1D array
1288 * as a 2D Array with a height of 1. So, here we want to swap image
1291 assert(image
->Depth
== 1);
1292 *width
= image
->Width
;
1294 *depth
= image
->Height
;
1296 case GL_TEXTURE_CUBE_MAP
:
1297 /* For Cube maps, the mesa/main api layer gives us a depth of 1 even
1298 * though we really have 6 slices.
1300 assert(image
->Depth
== 1);
1301 *width
= image
->Width
;
1302 *height
= image
->Height
;
1306 *width
= image
->Width
;
1307 *height
= image
->Height
;
1308 *depth
= image
->Depth
;
1314 * Can the image be pulled into a unified mipmap tree? This mirrors
1315 * the completeness test in a lot of ways.
1317 * Not sure whether I want to pass gl_texture_image here.
1320 intel_miptree_match_image(struct intel_mipmap_tree
*mt
,
1321 struct gl_texture_image
*image
)
1323 struct intel_texture_image
*intelImage
= intel_texture_image(image
);
1324 GLuint level
= intelImage
->base
.Base
.Level
;
1325 int width
, height
, depth
;
1327 /* glTexImage* choose the texture object based on the target passed in, and
1328 * objects can't change targets over their lifetimes, so this should be
1331 assert(image
->TexObject
->Target
== mt
->target
);
1333 mesa_format mt_format
= mt
->format
;
1334 if (mt
->format
== MESA_FORMAT_Z24_UNORM_X8_UINT
&& mt
->stencil_mt
)
1335 mt_format
= MESA_FORMAT_Z24_UNORM_S8_UINT
;
1336 if (mt
->format
== MESA_FORMAT_Z_FLOAT32
&& mt
->stencil_mt
)
1337 mt_format
= MESA_FORMAT_Z32_FLOAT_S8X24_UINT
;
1338 if (mt
->etc_format
!= MESA_FORMAT_NONE
)
1339 mt_format
= mt
->etc_format
;
1341 if (_mesa_get_srgb_format_linear(image
->TexFormat
) !=
1342 _mesa_get_srgb_format_linear(mt_format
))
1345 intel_get_image_dims(image
, &width
, &height
, &depth
);
1347 if (mt
->target
== GL_TEXTURE_CUBE_MAP
)
1350 if (level
>= mt
->surf
.levels
)
1353 const unsigned level_depth
=
1354 mt
->surf
.dim
== ISL_SURF_DIM_3D
?
1355 minify(mt
->surf
.logical_level0_px
.depth
, level
) :
1356 mt
->surf
.logical_level0_px
.array_len
;
1358 return width
== minify(mt
->surf
.logical_level0_px
.width
, level
) &&
1359 height
== minify(mt
->surf
.logical_level0_px
.height
, level
) &&
1360 depth
== level_depth
&&
1361 MAX2(image
->NumSamples
, 1) == mt
->surf
.samples
;
1365 intel_miptree_get_image_offset(const struct intel_mipmap_tree
*mt
,
1366 GLuint level
, GLuint slice
,
1367 GLuint
*x
, GLuint
*y
)
1369 if (level
== 0 && slice
== 0) {
1370 *x
= mt
->level
[0].level_x
;
1371 *y
= mt
->level
[0].level_y
;
1375 uint32_t x_offset_sa
, y_offset_sa
;
1377 /* Miptree itself can have an offset only if it represents a single
1378 * slice in an imported buffer object.
1379 * See intel_miptree_create_for_dri_image().
1381 assert(mt
->level
[0].level_x
== 0);
1382 assert(mt
->level
[0].level_y
== 0);
1384 /* Given level is relative to level zero while the miptree may be
1385 * represent just a subset of all levels starting from 'first_level'.
1387 assert(level
>= mt
->first_level
);
1388 level
-= mt
->first_level
;
1390 const unsigned z
= mt
->surf
.dim
== ISL_SURF_DIM_3D
? slice
: 0;
1391 slice
= mt
->surf
.dim
== ISL_SURF_DIM_3D
? 0 : slice
;
1392 isl_surf_get_image_offset_el(&mt
->surf
, level
, slice
, z
,
1393 &x_offset_sa
, &y_offset_sa
);
1401 * This function computes the tile_w (in bytes) and tile_h (in rows) of
1402 * different tiling patterns. If the BO is untiled, tile_w is set to cpp
1403 * and tile_h is set to 1.
1406 intel_get_tile_dims(enum isl_tiling tiling
, uint32_t cpp
,
1407 uint32_t *tile_w
, uint32_t *tile_h
)
1418 case ISL_TILING_LINEAR
:
1423 unreachable("not reached");
1429 * This function computes masks that may be used to select the bits of the X
1430 * and Y coordinates that indicate the offset within a tile. If the BO is
1431 * untiled, the masks are set to 0.
1434 intel_get_tile_masks(enum isl_tiling tiling
, uint32_t cpp
,
1435 uint32_t *mask_x
, uint32_t *mask_y
)
1437 uint32_t tile_w_bytes
, tile_h
;
1439 intel_get_tile_dims(tiling
, cpp
, &tile_w_bytes
, &tile_h
);
1441 *mask_x
= tile_w_bytes
/ cpp
- 1;
1442 *mask_y
= tile_h
- 1;
1446 * Compute the offset (in bytes) from the start of the BO to the given x
1447 * and y coordinate. For tiled BOs, caller must ensure that x and y are
1448 * multiples of the tile size.
1451 intel_miptree_get_aligned_offset(const struct intel_mipmap_tree
*mt
,
1452 uint32_t x
, uint32_t y
)
1455 uint32_t pitch
= mt
->surf
.row_pitch
;
1457 switch (mt
->surf
.tiling
) {
1459 unreachable("not reached");
1460 case ISL_TILING_LINEAR
:
1461 return y
* pitch
+ x
* cpp
;
1463 assert((x
% (512 / cpp
)) == 0);
1464 assert((y
% 8) == 0);
1465 return y
* pitch
+ x
/ (512 / cpp
) * 4096;
1467 assert((x
% (128 / cpp
)) == 0);
1468 assert((y
% 32) == 0);
1469 return y
* pitch
+ x
/ (128 / cpp
) * 4096;
1474 * Rendering with tiled buffers requires that the base address of the buffer
1475 * be aligned to a page boundary. For renderbuffers, and sometimes with
1476 * textures, we may want the surface to point at a texture image level that
1477 * isn't at a page boundary.
1479 * This function returns an appropriately-aligned base offset
1480 * according to the tiling restrictions, plus any required x/y offset
1484 intel_miptree_get_tile_offsets(const struct intel_mipmap_tree
*mt
,
1485 GLuint level
, GLuint slice
,
1490 uint32_t mask_x
, mask_y
;
1492 intel_get_tile_masks(mt
->surf
.tiling
, mt
->cpp
, &mask_x
, &mask_y
);
1493 intel_miptree_get_image_offset(mt
, level
, slice
, &x
, &y
);
1495 *tile_x
= x
& mask_x
;
1496 *tile_y
= y
& mask_y
;
1498 return intel_miptree_get_aligned_offset(mt
, x
& ~mask_x
, y
& ~mask_y
);
1502 intel_miptree_copy_slice_sw(struct brw_context
*brw
,
1503 struct intel_mipmap_tree
*src_mt
,
1504 unsigned src_level
, unsigned src_layer
,
1505 struct intel_mipmap_tree
*dst_mt
,
1506 unsigned dst_level
, unsigned dst_layer
,
1507 unsigned width
, unsigned height
)
1510 ptrdiff_t src_stride
, dst_stride
;
1511 const unsigned cpp
= (isl_format_get_layout(dst_mt
->surf
.format
)->bpb
/ 8);
1513 intel_miptree_map(brw
, src_mt
,
1514 src_level
, src_layer
,
1517 GL_MAP_READ_BIT
| BRW_MAP_DIRECT_BIT
,
1520 intel_miptree_map(brw
, dst_mt
,
1521 dst_level
, dst_layer
,
1524 GL_MAP_WRITE_BIT
| GL_MAP_INVALIDATE_RANGE_BIT
|
1528 DBG("sw blit %s mt %p %p/%"PRIdPTR
" -> %s mt %p %p/%"PRIdPTR
" (%dx%d)\n",
1529 _mesa_get_format_name(src_mt
->format
),
1530 src_mt
, src
, src_stride
,
1531 _mesa_get_format_name(dst_mt
->format
),
1532 dst_mt
, dst
, dst_stride
,
1535 int row_size
= cpp
* width
;
1536 if (src_stride
== row_size
&&
1537 dst_stride
== row_size
) {
1538 memcpy(dst
, src
, row_size
* height
);
1540 for (int i
= 0; i
< height
; i
++) {
1541 memcpy(dst
, src
, row_size
);
1547 intel_miptree_unmap(brw
, dst_mt
, dst_level
, dst_layer
);
1548 intel_miptree_unmap(brw
, src_mt
, src_level
, src_layer
);
1550 /* Don't forget to copy the stencil data over, too. We could have skipped
1551 * passing BRW_MAP_DIRECT_BIT, but that would have meant intel_miptree_map
1552 * shuffling the two data sources in/out of temporary storage instead of
1553 * the direct mapping we get this way.
1555 if (dst_mt
->stencil_mt
) {
1556 assert(src_mt
->stencil_mt
);
1557 intel_miptree_copy_slice_sw(brw
,
1558 src_mt
->stencil_mt
, src_level
, src_layer
,
1559 dst_mt
->stencil_mt
, dst_level
, dst_layer
,
1565 intel_miptree_copy_slice(struct brw_context
*brw
,
1566 struct intel_mipmap_tree
*src_mt
,
1567 unsigned src_level
, unsigned src_layer
,
1568 struct intel_mipmap_tree
*dst_mt
,
1569 unsigned dst_level
, unsigned dst_layer
)
1572 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1573 mesa_format format
= src_mt
->format
;
1574 unsigned width
= minify(src_mt
->surf
.phys_level0_sa
.width
,
1575 src_level
- src_mt
->first_level
);
1576 unsigned height
= minify(src_mt
->surf
.phys_level0_sa
.height
,
1577 src_level
- src_mt
->first_level
);
1579 assert(src_layer
< get_num_phys_layers(&src_mt
->surf
,
1580 src_level
- src_mt
->first_level
));
1582 assert(_mesa_get_srgb_format_linear(src_mt
->format
) ==
1583 _mesa_get_srgb_format_linear(dst_mt
->format
));
1585 DBG("validate blit mt %s %p %d,%d -> mt %s %p %d,%d (%dx%d)\n",
1586 _mesa_get_format_name(src_mt
->format
),
1587 src_mt
, src_level
, src_layer
,
1588 _mesa_get_format_name(dst_mt
->format
),
1589 dst_mt
, dst_level
, dst_layer
,
1592 if (devinfo
->gen
>= 6) {
1593 /* On gen6 and above, we just use blorp. It's faster than the blitter
1594 * and can handle everything without software fallbacks.
1596 brw_blorp_copy_miptrees(brw
,
1597 src_mt
, src_level
, src_layer
,
1598 dst_mt
, dst_level
, dst_layer
,
1599 0, 0, 0, 0, width
, height
);
1601 if (src_mt
->stencil_mt
) {
1602 assert(dst_mt
->stencil_mt
);
1603 brw_blorp_copy_miptrees(brw
,
1604 src_mt
->stencil_mt
, src_level
, src_layer
,
1605 dst_mt
->stencil_mt
, dst_level
, dst_layer
,
1606 0, 0, 0, 0, width
, height
);
1611 if (dst_mt
->compressed
) {
1613 _mesa_get_format_block_size(dst_mt
->format
, &i
, &j
);
1614 height
= ALIGN_NPOT(height
, j
) / j
;
1615 width
= ALIGN_NPOT(width
, i
) / i
;
1618 /* Gen4-5 doesn't support separate stencil */
1619 assert(!src_mt
->stencil_mt
);
1621 uint32_t dst_x
, dst_y
, src_x
, src_y
;
1622 intel_miptree_get_image_offset(dst_mt
, dst_level
, dst_layer
,
1624 intel_miptree_get_image_offset(src_mt
, src_level
, src_layer
,
1627 DBG("validate blit mt %s %p %d,%d/%d -> mt %s %p %d,%d/%d (%dx%d)\n",
1628 _mesa_get_format_name(src_mt
->format
),
1629 src_mt
, src_x
, src_y
, src_mt
->surf
.row_pitch
,
1630 _mesa_get_format_name(dst_mt
->format
),
1631 dst_mt
, dst_x
, dst_y
, dst_mt
->surf
.row_pitch
,
1634 if (!intel_miptree_blit(brw
,
1635 src_mt
, src_level
, src_layer
, 0, 0, false,
1636 dst_mt
, dst_level
, dst_layer
, 0, 0, false,
1637 width
, height
, COLOR_LOGICOP_COPY
)) {
1638 perf_debug("miptree validate blit for %s failed\n",
1639 _mesa_get_format_name(format
));
1641 intel_miptree_copy_slice_sw(brw
,
1642 src_mt
, src_level
, src_layer
,
1643 dst_mt
, dst_level
, dst_layer
,
1649 * Copies the image's current data to the given miptree, and associates that
1650 * miptree with the image.
1653 intel_miptree_copy_teximage(struct brw_context
*brw
,
1654 struct intel_texture_image
*intelImage
,
1655 struct intel_mipmap_tree
*dst_mt
)
1657 struct intel_mipmap_tree
*src_mt
= intelImage
->mt
;
1658 struct intel_texture_object
*intel_obj
=
1659 intel_texture_object(intelImage
->base
.Base
.TexObject
);
1660 int level
= intelImage
->base
.Base
.Level
;
1661 const unsigned face
= intelImage
->base
.Base
.Face
;
1662 unsigned start_layer
, end_layer
;
1664 if (intel_obj
->base
.Target
== GL_TEXTURE_1D_ARRAY
) {
1666 assert(intelImage
->base
.Base
.Height
);
1668 end_layer
= intelImage
->base
.Base
.Height
- 1;
1669 } else if (face
> 0) {
1673 assert(intelImage
->base
.Base
.Depth
);
1675 end_layer
= intelImage
->base
.Base
.Depth
- 1;
1678 for (unsigned i
= start_layer
; i
<= end_layer
; i
++) {
1679 intel_miptree_copy_slice(brw
,
1684 intel_miptree_reference(&intelImage
->mt
, dst_mt
);
1685 intel_obj
->needs_validate
= true;
1688 static struct intel_miptree_aux_buffer
*
1689 intel_alloc_aux_buffer(struct brw_context
*brw
,
1690 const struct isl_surf
*aux_surf
,
1692 uint8_t memset_value
)
1694 struct intel_miptree_aux_buffer
*buf
= calloc(sizeof(*buf
), 1);
1698 uint64_t size
= aux_surf
->size
;
1700 const bool has_indirect_clear
= brw
->isl_dev
.ss
.clear_color_state_size
> 0;
1701 if (has_indirect_clear
) {
1702 /* On CNL+, instead of setting the clear color in the SURFACE_STATE, we
1703 * will set a pointer to a dword somewhere that contains the color. So,
1704 * allocate the space for the clear color value here on the aux buffer.
1706 buf
->clear_color_offset
= size
;
1707 size
+= brw
->isl_dev
.ss
.clear_color_state_size
;
1710 /* If the buffer needs to be initialised (requiring the buffer to be
1711 * immediately mapped to cpu space for writing), do not use the gpu access
1712 * flag which can cause an unnecessary delay if the backing pages happened
1713 * to be just used by the GPU.
1715 const bool alloc_zeroed
= wants_memset
&& memset_value
== 0;
1716 const bool needs_memset
=
1717 !alloc_zeroed
&& (wants_memset
|| has_indirect_clear
);
1718 const uint32_t alloc_flags
=
1719 alloc_zeroed
? BO_ALLOC_ZEROED
: (needs_memset
? 0 : BO_ALLOC_BUSY
);
1721 /* ISL has stricter set of alignment rules then the drm allocator.
1722 * Therefore one can pass the ISL dimensions in terms of bytes instead of
1723 * trying to recalculate based on different format block sizes.
1725 buf
->bo
= brw_bo_alloc_tiled(brw
->bufmgr
, "aux-miptree", size
,
1726 BRW_MEMZONE_OTHER
, I915_TILING_Y
,
1727 aux_surf
->row_pitch
, alloc_flags
);
1733 /* Initialize the bo to the desired value */
1735 assert(!(alloc_flags
& BO_ALLOC_BUSY
));
1737 void *map
= brw_bo_map(brw
, buf
->bo
, MAP_WRITE
| MAP_RAW
);
1739 intel_miptree_aux_buffer_free(buf
);
1743 /* Memset the aux_surf portion of the BO. */
1745 memset(map
, memset_value
, aux_surf
->size
);
1747 /* Zero the indirect clear color to match ::fast_clear_color. */
1748 if (has_indirect_clear
) {
1749 memset((char *)map
+ buf
->clear_color_offset
, 0,
1750 brw
->isl_dev
.ss
.clear_color_state_size
);
1753 brw_bo_unmap(buf
->bo
);
1756 if (has_indirect_clear
) {
1757 buf
->clear_color_bo
= buf
->bo
;
1758 brw_bo_reference(buf
->clear_color_bo
);
1761 buf
->surf
= *aux_surf
;
1768 * Helper for intel_miptree_alloc_aux() that sets
1769 * \c mt->level[level].has_hiz. Return true if and only if
1770 * \c has_hiz was set.
1773 intel_miptree_level_enable_hiz(struct brw_context
*brw
,
1774 struct intel_mipmap_tree
*mt
,
1777 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1779 assert(mt
->aux_buf
);
1780 assert(mt
->surf
.size
> 0);
1782 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
1783 uint32_t width
= minify(mt
->surf
.phys_level0_sa
.width
, level
);
1784 uint32_t height
= minify(mt
->surf
.phys_level0_sa
.height
, level
);
1786 /* Disable HiZ for LOD > 0 unless the width is 8 aligned
1787 * and the height is 4 aligned. This allows our HiZ support
1788 * to fulfill Haswell restrictions for HiZ ops. For LOD == 0,
1789 * we can grow the width & height to allow the HiZ op to
1790 * force the proper size alignments.
1792 if (level
> 0 && ((width
& 7) || (height
& 3))) {
1793 DBG("mt %p level %d: HiZ DISABLED\n", mt
, level
);
1798 DBG("mt %p level %d: HiZ enabled\n", mt
, level
);
1799 mt
->level
[level
].has_hiz
= true;
1805 * Allocate the initial aux surface for a miptree based on mt->aux_usage
1807 * Since MCS, HiZ, and CCS_E can compress more than just clear color, we
1808 * create the auxiliary surfaces up-front. CCS_D, on the other hand, can only
1809 * compress clear color so we wait until an actual fast-clear to allocate it.
1812 intel_miptree_alloc_aux(struct brw_context
*brw
,
1813 struct intel_mipmap_tree
*mt
)
1815 assert(mt
->aux_buf
== NULL
);
1817 /* Get the aux buf allocation parameters for this miptree. */
1818 enum isl_aux_state initial_state
;
1819 uint8_t memset_value
;
1820 struct isl_surf aux_surf
;
1821 MAYBE_UNUSED
bool aux_surf_ok
;
1823 switch (mt
->aux_usage
) {
1824 case ISL_AUX_USAGE_NONE
:
1828 case ISL_AUX_USAGE_HIZ
:
1829 initial_state
= ISL_AUX_STATE_AUX_INVALID
;
1830 aux_surf_ok
= isl_surf_get_hiz_surf(&brw
->isl_dev
, &mt
->surf
, &aux_surf
);
1832 case ISL_AUX_USAGE_MCS
:
1833 /* From the Ivy Bridge PRM, Vol 2 Part 1 p326:
1835 * When MCS buffer is enabled and bound to MSRT, it is required that
1836 * it is cleared prior to any rendering.
1838 * Since we don't use the MCS buffer for any purpose other than
1839 * rendering, it makes sense to just clear it immediately upon
1842 * Note: the clear value for MCS buffers is all 1's, so we memset to
1845 initial_state
= ISL_AUX_STATE_CLEAR
;
1846 memset_value
= 0xFF;
1847 aux_surf_ok
= isl_surf_get_mcs_surf(&brw
->isl_dev
, &mt
->surf
, &aux_surf
);
1849 case ISL_AUX_USAGE_CCS_D
:
1850 case ISL_AUX_USAGE_CCS_E
:
1851 /* When CCS_E is used, we need to ensure that the CCS starts off in a
1852 * valid state. From the Sky Lake PRM, "MCS Buffer for Render
1855 * "If Software wants to enable Color Compression without Fast
1856 * clear, Software needs to initialize MCS with zeros."
1858 * A CCS value of 0 indicates that the corresponding block is in the
1859 * pass-through state which is what we want.
1861 * For CCS_D, do the same thing. On gen9+, this avoids having any
1862 * undefined bits in the aux buffer.
1864 initial_state
= ISL_AUX_STATE_PASS_THROUGH
;
1867 isl_surf_get_ccs_surf(&brw
->isl_dev
, &mt
->surf
, &aux_surf
, 0);
1871 /* We should have a valid aux_surf. */
1872 assert(aux_surf_ok
);
1874 /* No work is needed for a zero-sized auxiliary buffer. */
1875 if (aux_surf
.size
== 0)
1878 /* Create the aux_state for the auxiliary buffer. */
1879 mt
->aux_state
= create_aux_state_map(mt
, initial_state
);
1880 if (mt
->aux_state
== NULL
)
1883 /* Allocate the auxiliary buffer. */
1884 const bool needs_memset
= initial_state
!= ISL_AUX_STATE_AUX_INVALID
;
1885 mt
->aux_buf
= intel_alloc_aux_buffer(brw
, &aux_surf
, needs_memset
,
1887 if (mt
->aux_buf
== NULL
) {
1888 free_aux_state_map(mt
->aux_state
);
1889 mt
->aux_state
= NULL
;
1893 /* Perform aux_usage-specific initialization. */
1894 if (mt
->aux_usage
== ISL_AUX_USAGE_HIZ
) {
1895 for (unsigned level
= mt
->first_level
; level
<= mt
->last_level
; ++level
)
1896 intel_miptree_level_enable_hiz(brw
, mt
, level
);
1904 * Can the miptree sample using the hiz buffer?
1907 intel_miptree_sample_with_hiz(struct brw_context
*brw
,
1908 struct intel_mipmap_tree
*mt
)
1910 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1912 if (!devinfo
->has_sample_with_hiz
) {
1920 /* It seems the hardware won't fallback to the depth buffer if some of the
1921 * mipmap levels aren't available in the HiZ buffer. So we need all levels
1922 * of the texture to be HiZ enabled.
1924 for (unsigned level
= 0; level
< mt
->surf
.levels
; ++level
) {
1925 if (!intel_miptree_level_has_hiz(mt
, level
))
1929 /* If compressed multisampling is enabled, then we use it for the auxiliary
1932 * From the BDW PRM (Volume 2d: Command Reference: Structures
1933 * RENDER_SURFACE_STATE.AuxiliarySurfaceMode):
1935 * "If this field is set to AUX_HIZ, Number of Multisamples must be
1936 * MULTISAMPLECOUNT_1, and Surface Type cannot be SURFTYPE_3D.
1938 * There is no such blurb for 1D textures, but there is sufficient evidence
1939 * that this is broken on SKL+.
1941 return (mt
->surf
.samples
== 1 &&
1942 mt
->target
!= GL_TEXTURE_3D
&&
1943 mt
->target
!= GL_TEXTURE_1D
/* gen9+ restriction */);
1947 * Does the miptree slice have hiz enabled?
1950 intel_miptree_level_has_hiz(const struct intel_mipmap_tree
*mt
, uint32_t level
)
1952 intel_miptree_check_level_layer(mt
, level
, 0);
1953 return mt
->level
[level
].has_hiz
;
1956 static inline uint32_t
1957 miptree_level_range_length(const struct intel_mipmap_tree
*mt
,
1958 uint32_t start_level
, uint32_t num_levels
)
1960 assert(start_level
>= mt
->first_level
);
1961 assert(start_level
<= mt
->last_level
);
1963 if (num_levels
== INTEL_REMAINING_LAYERS
)
1964 num_levels
= mt
->last_level
- start_level
+ 1;
1965 /* Check for overflow */
1966 assert(start_level
+ num_levels
>= start_level
);
1967 assert(start_level
+ num_levels
<= mt
->last_level
+ 1);
1972 static inline uint32_t
1973 miptree_layer_range_length(const struct intel_mipmap_tree
*mt
, uint32_t level
,
1974 uint32_t start_layer
, uint32_t num_layers
)
1976 assert(level
<= mt
->last_level
);
1978 const uint32_t total_num_layers
= brw_get_num_logical_layers(mt
, level
);
1979 assert(start_layer
< total_num_layers
);
1980 if (num_layers
== INTEL_REMAINING_LAYERS
)
1981 num_layers
= total_num_layers
- start_layer
;
1982 /* Check for overflow */
1983 assert(start_layer
+ num_layers
>= start_layer
);
1984 assert(start_layer
+ num_layers
<= total_num_layers
);
1990 intel_miptree_has_color_unresolved(const struct intel_mipmap_tree
*mt
,
1991 unsigned start_level
, unsigned num_levels
,
1992 unsigned start_layer
, unsigned num_layers
)
1994 assert(_mesa_is_format_color_format(mt
->format
));
1999 /* Clamp the level range to fit the miptree */
2000 num_levels
= miptree_level_range_length(mt
, start_level
, num_levels
);
2002 for (uint32_t l
= 0; l
< num_levels
; l
++) {
2003 const uint32_t level
= start_level
+ l
;
2004 const uint32_t level_layers
=
2005 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2006 for (unsigned a
= 0; a
< level_layers
; a
++) {
2007 enum isl_aux_state aux_state
=
2008 intel_miptree_get_aux_state(mt
, level
, start_layer
+ a
);
2009 assert(aux_state
!= ISL_AUX_STATE_AUX_INVALID
);
2010 if (aux_state
!= ISL_AUX_STATE_PASS_THROUGH
)
2019 intel_miptree_check_color_resolve(const struct brw_context
*brw
,
2020 const struct intel_mipmap_tree
*mt
,
2021 unsigned level
, unsigned layer
)
2026 /* Fast color clear is supported for mipmapped surfaces only on Gen8+. */
2027 assert(brw
->screen
->devinfo
.gen
>= 8 ||
2028 (level
== 0 && mt
->first_level
== 0 && mt
->last_level
== 0));
2030 /* Compression of arrayed msaa surfaces is supported. */
2031 if (mt
->surf
.samples
> 1)
2034 /* Fast color clear is supported for non-msaa arrays only on Gen8+. */
2035 assert(brw
->screen
->devinfo
.gen
>= 8 ||
2037 mt
->surf
.logical_level0_px
.depth
== 1 &&
2038 mt
->surf
.logical_level0_px
.array_len
== 1));
2044 static enum isl_aux_op
2045 get_ccs_d_resolve_op(enum isl_aux_state aux_state
,
2046 enum isl_aux_usage aux_usage
,
2047 bool fast_clear_supported
)
2049 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_CCS_D
);
2051 const bool ccs_supported
= aux_usage
== ISL_AUX_USAGE_CCS_D
;
2053 assert(ccs_supported
== fast_clear_supported
);
2055 switch (aux_state
) {
2056 case ISL_AUX_STATE_CLEAR
:
2057 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2059 return ISL_AUX_OP_FULL_RESOLVE
;
2061 return ISL_AUX_OP_NONE
;
2063 case ISL_AUX_STATE_PASS_THROUGH
:
2064 return ISL_AUX_OP_NONE
;
2066 case ISL_AUX_STATE_RESOLVED
:
2067 case ISL_AUX_STATE_AUX_INVALID
:
2068 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2069 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2073 unreachable("Invalid aux state for CCS_D");
2076 static enum isl_aux_op
2077 get_ccs_e_resolve_op(enum isl_aux_state aux_state
,
2078 enum isl_aux_usage aux_usage
,
2079 bool fast_clear_supported
)
2081 /* CCS_E surfaces can be accessed as CCS_D if we're careful. */
2082 assert(aux_usage
== ISL_AUX_USAGE_NONE
||
2083 aux_usage
== ISL_AUX_USAGE_CCS_D
||
2084 aux_usage
== ISL_AUX_USAGE_CCS_E
);
2086 if (aux_usage
== ISL_AUX_USAGE_CCS_D
)
2087 assert(fast_clear_supported
);
2089 switch (aux_state
) {
2090 case ISL_AUX_STATE_CLEAR
:
2091 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2092 if (fast_clear_supported
)
2093 return ISL_AUX_OP_NONE
;
2094 else if (aux_usage
== ISL_AUX_USAGE_CCS_E
)
2095 return ISL_AUX_OP_PARTIAL_RESOLVE
;
2097 return ISL_AUX_OP_FULL_RESOLVE
;
2099 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2100 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2101 return ISL_AUX_OP_FULL_RESOLVE
;
2102 else if (!fast_clear_supported
)
2103 return ISL_AUX_OP_PARTIAL_RESOLVE
;
2105 return ISL_AUX_OP_NONE
;
2107 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2108 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2109 return ISL_AUX_OP_FULL_RESOLVE
;
2111 return ISL_AUX_OP_NONE
;
2113 case ISL_AUX_STATE_PASS_THROUGH
:
2114 return ISL_AUX_OP_NONE
;
2116 case ISL_AUX_STATE_RESOLVED
:
2117 case ISL_AUX_STATE_AUX_INVALID
:
2121 unreachable("Invalid aux state for CCS_E");
2125 intel_miptree_prepare_ccs_access(struct brw_context
*brw
,
2126 struct intel_mipmap_tree
*mt
,
2127 uint32_t level
, uint32_t layer
,
2128 enum isl_aux_usage aux_usage
,
2129 bool fast_clear_supported
)
2131 enum isl_aux_state aux_state
= intel_miptree_get_aux_state(mt
, level
, layer
);
2133 enum isl_aux_op resolve_op
;
2134 if (mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2135 resolve_op
= get_ccs_e_resolve_op(aux_state
, aux_usage
,
2136 fast_clear_supported
);
2138 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2139 resolve_op
= get_ccs_d_resolve_op(aux_state
, aux_usage
,
2140 fast_clear_supported
);
2143 if (resolve_op
!= ISL_AUX_OP_NONE
) {
2144 intel_miptree_check_color_resolve(brw
, mt
, level
, layer
);
2145 brw_blorp_resolve_color(brw
, mt
, level
, layer
, resolve_op
);
2147 switch (resolve_op
) {
2148 case ISL_AUX_OP_FULL_RESOLVE
:
2149 /* The CCS full resolve operation destroys the CCS and sets it to the
2150 * pass-through state. (You can also think of this as being both a
2151 * resolve and an ambiguate in one operation.)
2153 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2154 ISL_AUX_STATE_PASS_THROUGH
);
2157 case ISL_AUX_OP_PARTIAL_RESOLVE
:
2158 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2159 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2163 unreachable("Invalid resolve op");
2169 intel_miptree_finish_ccs_write(struct brw_context
*brw
,
2170 struct intel_mipmap_tree
*mt
,
2171 uint32_t level
, uint32_t layer
,
2172 enum isl_aux_usage aux_usage
)
2174 assert(aux_usage
== ISL_AUX_USAGE_NONE
||
2175 aux_usage
== ISL_AUX_USAGE_CCS_D
||
2176 aux_usage
== ISL_AUX_USAGE_CCS_E
);
2178 enum isl_aux_state aux_state
= intel_miptree_get_aux_state(mt
, level
, layer
);
2180 if (mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2181 switch (aux_state
) {
2182 case ISL_AUX_STATE_CLEAR
:
2183 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2184 assert(aux_usage
== ISL_AUX_USAGE_CCS_E
||
2185 aux_usage
== ISL_AUX_USAGE_CCS_D
);
2187 if (aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2188 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2189 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2190 } else if (aux_state
!= ISL_AUX_STATE_PARTIAL_CLEAR
) {
2191 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2192 ISL_AUX_STATE_PARTIAL_CLEAR
);
2196 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2197 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2198 assert(aux_usage
== ISL_AUX_USAGE_CCS_E
);
2199 break; /* Nothing to do */
2201 case ISL_AUX_STATE_PASS_THROUGH
:
2202 if (aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2203 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2204 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2210 case ISL_AUX_STATE_RESOLVED
:
2211 case ISL_AUX_STATE_AUX_INVALID
:
2212 unreachable("Invalid aux state for CCS_E");
2215 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2216 /* CCS_D is a bit simpler */
2217 switch (aux_state
) {
2218 case ISL_AUX_STATE_CLEAR
:
2219 assert(aux_usage
== ISL_AUX_USAGE_CCS_D
);
2220 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2221 ISL_AUX_STATE_PARTIAL_CLEAR
);
2224 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2225 assert(aux_usage
== ISL_AUX_USAGE_CCS_D
);
2226 break; /* Nothing to do */
2228 case ISL_AUX_STATE_PASS_THROUGH
:
2232 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2233 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2234 case ISL_AUX_STATE_RESOLVED
:
2235 case ISL_AUX_STATE_AUX_INVALID
:
2236 unreachable("Invalid aux state for CCS_D");
2242 intel_miptree_prepare_mcs_access(struct brw_context
*brw
,
2243 struct intel_mipmap_tree
*mt
,
2245 enum isl_aux_usage aux_usage
,
2246 bool fast_clear_supported
)
2248 assert(aux_usage
== ISL_AUX_USAGE_MCS
);
2250 switch (intel_miptree_get_aux_state(mt
, 0, layer
)) {
2251 case ISL_AUX_STATE_CLEAR
:
2252 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2253 if (!fast_clear_supported
) {
2254 brw_blorp_mcs_partial_resolve(brw
, mt
, layer
, 1);
2255 intel_miptree_set_aux_state(brw
, mt
, 0, layer
, 1,
2256 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2260 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2261 break; /* Nothing to do */
2263 case ISL_AUX_STATE_RESOLVED
:
2264 case ISL_AUX_STATE_PASS_THROUGH
:
2265 case ISL_AUX_STATE_AUX_INVALID
:
2266 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2267 unreachable("Invalid aux state for MCS");
2272 intel_miptree_finish_mcs_write(struct brw_context
*brw
,
2273 struct intel_mipmap_tree
*mt
,
2275 enum isl_aux_usage aux_usage
)
2277 assert(aux_usage
== ISL_AUX_USAGE_MCS
);
2279 switch (intel_miptree_get_aux_state(mt
, 0, layer
)) {
2280 case ISL_AUX_STATE_CLEAR
:
2281 intel_miptree_set_aux_state(brw
, mt
, 0, layer
, 1,
2282 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2285 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2286 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2287 break; /* Nothing to do */
2289 case ISL_AUX_STATE_RESOLVED
:
2290 case ISL_AUX_STATE_PASS_THROUGH
:
2291 case ISL_AUX_STATE_AUX_INVALID
:
2292 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2293 unreachable("Invalid aux state for MCS");
2298 intel_miptree_prepare_hiz_access(struct brw_context
*brw
,
2299 struct intel_mipmap_tree
*mt
,
2300 uint32_t level
, uint32_t layer
,
2301 enum isl_aux_usage aux_usage
,
2302 bool fast_clear_supported
)
2304 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_HIZ
);
2306 enum isl_aux_op hiz_op
= ISL_AUX_OP_NONE
;
2307 switch (intel_miptree_get_aux_state(mt
, level
, layer
)) {
2308 case ISL_AUX_STATE_CLEAR
:
2309 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2310 if (aux_usage
!= ISL_AUX_USAGE_HIZ
|| !fast_clear_supported
)
2311 hiz_op
= ISL_AUX_OP_FULL_RESOLVE
;
2314 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2315 if (aux_usage
!= ISL_AUX_USAGE_HIZ
)
2316 hiz_op
= ISL_AUX_OP_FULL_RESOLVE
;
2319 case ISL_AUX_STATE_PASS_THROUGH
:
2320 case ISL_AUX_STATE_RESOLVED
:
2323 case ISL_AUX_STATE_AUX_INVALID
:
2324 if (aux_usage
== ISL_AUX_USAGE_HIZ
)
2325 hiz_op
= ISL_AUX_OP_AMBIGUATE
;
2328 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2329 unreachable("Invalid HiZ state");
2332 if (hiz_op
!= ISL_AUX_OP_NONE
) {
2333 intel_hiz_exec(brw
, mt
, level
, layer
, 1, hiz_op
);
2336 case ISL_AUX_OP_FULL_RESOLVE
:
2337 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2338 ISL_AUX_STATE_RESOLVED
);
2341 case ISL_AUX_OP_AMBIGUATE
:
2342 /* The HiZ resolve operation is actually an ambiguate */
2343 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2344 ISL_AUX_STATE_PASS_THROUGH
);
2348 unreachable("Invalid HiZ op");
2354 intel_miptree_finish_hiz_write(struct brw_context
*brw
,
2355 struct intel_mipmap_tree
*mt
,
2356 uint32_t level
, uint32_t layer
,
2357 enum isl_aux_usage aux_usage
)
2359 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_HIZ
);
2361 switch (intel_miptree_get_aux_state(mt
, level
, layer
)) {
2362 case ISL_AUX_STATE_CLEAR
:
2363 assert(aux_usage
== ISL_AUX_USAGE_HIZ
);
2364 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2365 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2368 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2369 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2370 assert(aux_usage
== ISL_AUX_USAGE_HIZ
);
2371 break; /* Nothing to do */
2373 case ISL_AUX_STATE_RESOLVED
:
2374 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
2375 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2376 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2378 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2379 ISL_AUX_STATE_AUX_INVALID
);
2383 case ISL_AUX_STATE_PASS_THROUGH
:
2384 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
2385 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2386 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2390 case ISL_AUX_STATE_AUX_INVALID
:
2391 assert(aux_usage
!= ISL_AUX_USAGE_HIZ
);
2394 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2395 unreachable("Invalid HiZ state");
2400 intel_miptree_prepare_access(struct brw_context
*brw
,
2401 struct intel_mipmap_tree
*mt
,
2402 uint32_t start_level
, uint32_t num_levels
,
2403 uint32_t start_layer
, uint32_t num_layers
,
2404 enum isl_aux_usage aux_usage
,
2405 bool fast_clear_supported
)
2407 num_levels
= miptree_level_range_length(mt
, start_level
, num_levels
);
2409 switch (mt
->aux_usage
) {
2410 case ISL_AUX_USAGE_NONE
:
2414 case ISL_AUX_USAGE_MCS
:
2415 assert(mt
->aux_buf
);
2416 assert(start_level
== 0 && num_levels
== 1);
2417 const uint32_t level_layers
=
2418 miptree_layer_range_length(mt
, 0, start_layer
, num_layers
);
2419 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2420 intel_miptree_prepare_mcs_access(brw
, mt
, start_layer
+ a
,
2421 aux_usage
, fast_clear_supported
);
2425 case ISL_AUX_USAGE_CCS_D
:
2426 case ISL_AUX_USAGE_CCS_E
:
2430 for (uint32_t l
= 0; l
< num_levels
; l
++) {
2431 const uint32_t level
= start_level
+ l
;
2432 const uint32_t level_layers
=
2433 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2434 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2435 intel_miptree_prepare_ccs_access(brw
, mt
, level
,
2437 aux_usage
, fast_clear_supported
);
2442 case ISL_AUX_USAGE_HIZ
:
2443 assert(mt
->aux_buf
);
2444 for (uint32_t l
= 0; l
< num_levels
; l
++) {
2445 const uint32_t level
= start_level
+ l
;
2446 if (!intel_miptree_level_has_hiz(mt
, level
))
2449 const uint32_t level_layers
=
2450 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2451 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2452 intel_miptree_prepare_hiz_access(brw
, mt
, level
, start_layer
+ a
,
2453 aux_usage
, fast_clear_supported
);
2459 unreachable("Invalid aux usage");
2464 intel_miptree_finish_write(struct brw_context
*brw
,
2465 struct intel_mipmap_tree
*mt
, uint32_t level
,
2466 uint32_t start_layer
, uint32_t num_layers
,
2467 enum isl_aux_usage aux_usage
)
2469 num_layers
= miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2471 switch (mt
->aux_usage
) {
2472 case ISL_AUX_USAGE_NONE
:
2476 case ISL_AUX_USAGE_MCS
:
2477 assert(mt
->aux_buf
);
2478 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2479 intel_miptree_finish_mcs_write(brw
, mt
, start_layer
+ a
,
2484 case ISL_AUX_USAGE_CCS_D
:
2485 case ISL_AUX_USAGE_CCS_E
:
2489 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2490 intel_miptree_finish_ccs_write(brw
, mt
, level
, start_layer
+ a
,
2495 case ISL_AUX_USAGE_HIZ
:
2496 if (!intel_miptree_level_has_hiz(mt
, level
))
2499 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2500 intel_miptree_finish_hiz_write(brw
, mt
, level
, start_layer
+ a
,
2506 unreachable("Invavlid aux usage");
2511 intel_miptree_get_aux_state(const struct intel_mipmap_tree
*mt
,
2512 uint32_t level
, uint32_t layer
)
2514 intel_miptree_check_level_layer(mt
, level
, layer
);
2516 if (_mesa_is_format_color_format(mt
->format
)) {
2517 assert(mt
->aux_buf
!= NULL
);
2518 assert(mt
->surf
.samples
== 1 ||
2519 mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
2520 } else if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2521 unreachable("Cannot get aux state for stencil");
2523 assert(intel_miptree_level_has_hiz(mt
, level
));
2526 return mt
->aux_state
[level
][layer
];
2530 intel_miptree_set_aux_state(struct brw_context
*brw
,
2531 struct intel_mipmap_tree
*mt
, uint32_t level
,
2532 uint32_t start_layer
, uint32_t num_layers
,
2533 enum isl_aux_state aux_state
)
2535 num_layers
= miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2537 if (_mesa_is_format_color_format(mt
->format
)) {
2538 assert(mt
->aux_buf
!= NULL
);
2539 assert(mt
->surf
.samples
== 1 ||
2540 mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
2541 } else if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2542 unreachable("Cannot get aux state for stencil");
2544 assert(intel_miptree_level_has_hiz(mt
, level
));
2547 for (unsigned a
= 0; a
< num_layers
; a
++) {
2548 if (mt
->aux_state
[level
][start_layer
+ a
] != aux_state
) {
2549 mt
->aux_state
[level
][start_layer
+ a
] = aux_state
;
2550 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
2555 /* On Gen9 color buffers may be compressed by the hardware (lossless
2556 * compression). There are, however, format restrictions and care needs to be
2557 * taken that the sampler engine is capable for re-interpreting a buffer with
2558 * format different the buffer was originally written with.
2560 * For example, SRGB formats are not compressible and the sampler engine isn't
2561 * capable of treating RGBA_UNORM as SRGB_ALPHA. In such a case the underlying
2562 * color buffer needs to be resolved so that the sampling surface can be
2563 * sampled as non-compressed (i.e., without the auxiliary MCS buffer being
2567 can_texture_with_ccs(struct brw_context
*brw
,
2568 struct intel_mipmap_tree
*mt
,
2569 enum isl_format view_format
)
2571 if (mt
->aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2574 if (!format_ccs_e_compat_with_miptree(&brw
->screen
->devinfo
,
2576 perf_debug("Incompatible sampling format (%s) for rbc (%s)\n",
2577 isl_format_get_layout(view_format
)->name
,
2578 _mesa_get_format_name(mt
->format
));
2586 intel_miptree_texture_aux_usage(struct brw_context
*brw
,
2587 struct intel_mipmap_tree
*mt
,
2588 enum isl_format view_format
)
2590 switch (mt
->aux_usage
) {
2591 case ISL_AUX_USAGE_HIZ
:
2592 if (intel_miptree_sample_with_hiz(brw
, mt
))
2593 return ISL_AUX_USAGE_HIZ
;
2596 case ISL_AUX_USAGE_MCS
:
2597 return ISL_AUX_USAGE_MCS
;
2599 case ISL_AUX_USAGE_CCS_D
:
2600 case ISL_AUX_USAGE_CCS_E
:
2602 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2603 return ISL_AUX_USAGE_NONE
;
2606 /* If we don't have any unresolved color, report an aux usage of
2607 * ISL_AUX_USAGE_NONE. This way, texturing won't even look at the
2608 * aux surface and we can save some bandwidth.
2610 if (!intel_miptree_has_color_unresolved(mt
, 0, INTEL_REMAINING_LEVELS
,
2611 0, INTEL_REMAINING_LAYERS
))
2612 return ISL_AUX_USAGE_NONE
;
2614 if (can_texture_with_ccs(brw
, mt
, view_format
))
2615 return ISL_AUX_USAGE_CCS_E
;
2622 return ISL_AUX_USAGE_NONE
;
2626 isl_formats_are_fast_clear_compatible(enum isl_format a
, enum isl_format b
)
2628 /* On gen8 and earlier, the hardware was only capable of handling 0/1 clear
2629 * values so sRGB curve application was a no-op for all fast-clearable
2632 * On gen9+, the hardware supports arbitrary clear values. For sRGB clear
2633 * values, the hardware interprets the floats, not as what would be
2634 * returned from the sampler (or written by the shader), but as being
2635 * between format conversion and sRGB curve application. This means that
2636 * we can switch between sRGB and UNORM without having to whack the clear
2639 return isl_format_srgb_to_linear(a
) == isl_format_srgb_to_linear(b
);
2643 intel_miptree_prepare_texture(struct brw_context
*brw
,
2644 struct intel_mipmap_tree
*mt
,
2645 enum isl_format view_format
,
2646 uint32_t start_level
, uint32_t num_levels
,
2647 uint32_t start_layer
, uint32_t num_layers
)
2649 enum isl_aux_usage aux_usage
=
2650 intel_miptree_texture_aux_usage(brw
, mt
, view_format
);
2651 bool clear_supported
= aux_usage
!= ISL_AUX_USAGE_NONE
;
2653 /* Clear color is specified as ints or floats and the conversion is done by
2654 * the sampler. If we have a texture view, we would have to perform the
2655 * clear color conversion manually. Just disable clear color.
2657 if (!isl_formats_are_fast_clear_compatible(mt
->surf
.format
, view_format
))
2658 clear_supported
= false;
2660 intel_miptree_prepare_access(brw
, mt
, start_level
, num_levels
,
2661 start_layer
, num_layers
,
2662 aux_usage
, clear_supported
);
2666 intel_miptree_prepare_image(struct brw_context
*brw
,
2667 struct intel_mipmap_tree
*mt
)
2669 /* The data port doesn't understand any compression */
2670 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2671 0, INTEL_REMAINING_LAYERS
,
2672 ISL_AUX_USAGE_NONE
, false);
2676 intel_miptree_render_aux_usage(struct brw_context
*brw
,
2677 struct intel_mipmap_tree
*mt
,
2678 enum isl_format render_format
,
2680 bool draw_aux_disabled
)
2682 struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
2684 if (draw_aux_disabled
)
2685 return ISL_AUX_USAGE_NONE
;
2687 switch (mt
->aux_usage
) {
2688 case ISL_AUX_USAGE_MCS
:
2689 assert(mt
->aux_buf
);
2690 return ISL_AUX_USAGE_MCS
;
2692 case ISL_AUX_USAGE_CCS_D
:
2693 case ISL_AUX_USAGE_CCS_E
:
2695 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2696 return ISL_AUX_USAGE_NONE
;
2699 /* gen9+ hardware technically supports non-0/1 clear colors with sRGB
2700 * formats. However, there are issues with blending where it doesn't
2701 * properly apply the sRGB curve to the clear color when blending.
2703 if (devinfo
->gen
>= 9 && blend_enabled
&&
2704 isl_format_is_srgb(render_format
) &&
2705 !isl_color_value_is_zero_one(mt
->fast_clear_color
, render_format
))
2706 return ISL_AUX_USAGE_NONE
;
2708 if (mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
&&
2709 format_ccs_e_compat_with_miptree(&brw
->screen
->devinfo
,
2711 return ISL_AUX_USAGE_CCS_E
;
2713 /* Otherwise, we have to fall back to CCS_D */
2714 return ISL_AUX_USAGE_CCS_D
;
2717 return ISL_AUX_USAGE_NONE
;
2722 intel_miptree_prepare_render(struct brw_context
*brw
,
2723 struct intel_mipmap_tree
*mt
, uint32_t level
,
2724 uint32_t start_layer
, uint32_t layer_count
,
2725 enum isl_aux_usage aux_usage
)
2727 intel_miptree_prepare_access(brw
, mt
, level
, 1, start_layer
, layer_count
,
2728 aux_usage
, aux_usage
!= ISL_AUX_USAGE_NONE
);
2732 intel_miptree_finish_render(struct brw_context
*brw
,
2733 struct intel_mipmap_tree
*mt
, uint32_t level
,
2734 uint32_t start_layer
, uint32_t layer_count
,
2735 enum isl_aux_usage aux_usage
)
2737 assert(_mesa_is_format_color_format(mt
->format
));
2739 intel_miptree_finish_write(brw
, mt
, level
, start_layer
, layer_count
,
2744 intel_miptree_prepare_depth(struct brw_context
*brw
,
2745 struct intel_mipmap_tree
*mt
, uint32_t level
,
2746 uint32_t start_layer
, uint32_t layer_count
)
2748 intel_miptree_prepare_access(brw
, mt
, level
, 1, start_layer
, layer_count
,
2749 mt
->aux_usage
, mt
->aux_buf
!= NULL
);
2753 intel_miptree_finish_depth(struct brw_context
*brw
,
2754 struct intel_mipmap_tree
*mt
, uint32_t level
,
2755 uint32_t start_layer
, uint32_t layer_count
,
2758 if (depth_written
) {
2759 intel_miptree_finish_write(brw
, mt
, level
, start_layer
, layer_count
,
2760 mt
->aux_buf
!= NULL
);
2765 intel_miptree_prepare_external(struct brw_context
*brw
,
2766 struct intel_mipmap_tree
*mt
)
2768 enum isl_aux_usage aux_usage
= ISL_AUX_USAGE_NONE
;
2769 bool supports_fast_clear
= false;
2771 const struct isl_drm_modifier_info
*mod_info
=
2772 isl_drm_modifier_get_info(mt
->drm_modifier
);
2774 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
2775 /* CCS_E is the only supported aux for external images and it's only
2776 * supported on very simple images.
2778 assert(mod_info
->aux_usage
== ISL_AUX_USAGE_CCS_E
);
2779 assert(_mesa_is_format_color_format(mt
->format
));
2780 assert(mt
->first_level
== 0 && mt
->last_level
== 0);
2781 assert(mt
->surf
.logical_level0_px
.depth
== 1);
2782 assert(mt
->surf
.logical_level0_px
.array_len
== 1);
2783 assert(mt
->surf
.samples
== 1);
2784 assert(mt
->aux_buf
!= NULL
);
2786 aux_usage
= mod_info
->aux_usage
;
2787 supports_fast_clear
= mod_info
->supports_clear_color
;
2790 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2791 0, INTEL_REMAINING_LAYERS
,
2792 aux_usage
, supports_fast_clear
);
2796 intel_miptree_finish_external(struct brw_context
*brw
,
2797 struct intel_mipmap_tree
*mt
)
2802 /* We don't know the actual aux state of the aux surface. The previous
2803 * owner could have given it to us in a number of different states.
2804 * Because we don't know the aux state, we reset the aux state to the
2805 * least common denominator of possible valid states.
2807 enum isl_aux_state default_aux_state
=
2808 isl_drm_modifier_get_default_aux_state(mt
->drm_modifier
);
2809 assert(mt
->last_level
== mt
->first_level
);
2810 intel_miptree_set_aux_state(brw
, mt
, 0, 0, INTEL_REMAINING_LAYERS
,
2815 * Make it possible to share the BO backing the given miptree with another
2816 * process or another miptree.
2818 * Fast color clears are unsafe with shared buffers, so we need to resolve and
2819 * then discard the MCS buffer, if present. We also set the no_ccs flag to
2820 * ensure that no MCS buffer gets allocated in the future.
2822 * HiZ is similarly unsafe with shared buffers.
2825 intel_miptree_make_shareable(struct brw_context
*brw
,
2826 struct intel_mipmap_tree
*mt
)
2828 /* MCS buffers are also used for multisample buffers, but we can't resolve
2829 * away a multisample MCS buffer because it's an integral part of how the
2830 * pixel data is stored. Fortunately this code path should never be
2831 * reached for multisample buffers.
2833 assert(mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_NONE
||
2834 mt
->surf
.samples
== 1);
2836 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2837 0, INTEL_REMAINING_LAYERS
,
2838 ISL_AUX_USAGE_NONE
, false);
2841 intel_miptree_aux_buffer_free(mt
->aux_buf
);
2844 /* Make future calls of intel_miptree_level_has_hiz() return false. */
2845 for (uint32_t l
= mt
->first_level
; l
<= mt
->last_level
; ++l
) {
2846 mt
->level
[l
].has_hiz
= false;
2849 free(mt
->aux_state
);
2850 mt
->aux_state
= NULL
;
2851 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
2854 mt
->aux_usage
= ISL_AUX_USAGE_NONE
;
2855 mt
->supports_fast_clear
= false;
2860 * \brief Get pointer offset into stencil buffer.
2862 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
2863 * must decode the tile's layout in software.
2866 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
2868 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
2870 * Even though the returned offset is always positive, the return type is
2872 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
2873 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
2876 intel_offset_S8(uint32_t stride
, uint32_t x
, uint32_t y
, bool swizzled
)
2878 uint32_t tile_size
= 4096;
2879 uint32_t tile_width
= 64;
2880 uint32_t tile_height
= 64;
2881 uint32_t row_size
= 64 * stride
/ 2; /* Two rows are interleaved. */
2883 uint32_t tile_x
= x
/ tile_width
;
2884 uint32_t tile_y
= y
/ tile_height
;
2886 /* The byte's address relative to the tile's base addres. */
2887 uint32_t byte_x
= x
% tile_width
;
2888 uint32_t byte_y
= y
% tile_height
;
2890 uintptr_t u
= tile_y
* row_size
2891 + tile_x
* tile_size
2892 + 512 * (byte_x
/ 8)
2894 + 32 * ((byte_y
/ 4) % 2)
2895 + 16 * ((byte_x
/ 4) % 2)
2896 + 8 * ((byte_y
/ 2) % 2)
2897 + 4 * ((byte_x
/ 2) % 2)
2902 /* adjust for bit6 swizzling */
2903 if (((byte_x
/ 8) % 2) == 1) {
2904 if (((byte_y
/ 8) % 2) == 0) {
2916 intel_miptree_updownsample(struct brw_context
*brw
,
2917 struct intel_mipmap_tree
*src
,
2918 struct intel_mipmap_tree
*dst
)
2920 unsigned src_w
= src
->surf
.logical_level0_px
.width
;
2921 unsigned src_h
= src
->surf
.logical_level0_px
.height
;
2922 unsigned dst_w
= dst
->surf
.logical_level0_px
.width
;
2923 unsigned dst_h
= dst
->surf
.logical_level0_px
.height
;
2925 brw_blorp_blit_miptrees(brw
,
2926 src
, 0 /* level */, 0 /* layer */,
2927 src
->format
, SWIZZLE_XYZW
,
2928 dst
, 0 /* level */, 0 /* layer */, dst
->format
,
2931 GL_NEAREST
, false, false /*mirror x, y*/,
2934 if (src
->stencil_mt
) {
2935 src_w
= src
->stencil_mt
->surf
.logical_level0_px
.width
;
2936 src_h
= src
->stencil_mt
->surf
.logical_level0_px
.height
;
2937 dst_w
= dst
->stencil_mt
->surf
.logical_level0_px
.width
;
2938 dst_h
= dst
->stencil_mt
->surf
.logical_level0_px
.height
;
2940 brw_blorp_blit_miptrees(brw
,
2941 src
->stencil_mt
, 0 /* level */, 0 /* layer */,
2942 src
->stencil_mt
->format
, SWIZZLE_XYZW
,
2943 dst
->stencil_mt
, 0 /* level */, 0 /* layer */,
2944 dst
->stencil_mt
->format
,
2947 GL_NEAREST
, false, false /*mirror x, y*/,
2948 false, false /* decode/encode srgb */);
2953 intel_update_r8stencil(struct brw_context
*brw
,
2954 struct intel_mipmap_tree
*mt
)
2956 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
2958 assert(devinfo
->gen
>= 7);
2959 struct intel_mipmap_tree
*src
=
2960 mt
->format
== MESA_FORMAT_S_UINT8
? mt
: mt
->stencil_mt
;
2961 if (!src
|| devinfo
->gen
>= 8 || !src
->r8stencil_needs_update
)
2964 assert(src
->surf
.size
> 0);
2966 if (!mt
->r8stencil_mt
) {
2967 assert(devinfo
->gen
> 6); /* Handle MIPTREE_LAYOUT_GEN6_HIZ_STENCIL */
2968 mt
->r8stencil_mt
= make_surface(
2971 MESA_FORMAT_R_UINT8
,
2972 src
->first_level
, src
->last_level
,
2973 src
->surf
.logical_level0_px
.width
,
2974 src
->surf
.logical_level0_px
.height
,
2975 src
->surf
.dim
== ISL_SURF_DIM_3D
?
2976 src
->surf
.logical_level0_px
.depth
:
2977 src
->surf
.logical_level0_px
.array_len
,
2980 ISL_SURF_USAGE_TEXTURE_BIT
,
2981 BO_ALLOC_BUSY
, 0, NULL
);
2982 assert(mt
->r8stencil_mt
);
2985 struct intel_mipmap_tree
*dst
= mt
->r8stencil_mt
;
2987 for (int level
= src
->first_level
; level
<= src
->last_level
; level
++) {
2988 const unsigned depth
= src
->surf
.dim
== ISL_SURF_DIM_3D
?
2989 minify(src
->surf
.phys_level0_sa
.depth
, level
) :
2990 src
->surf
.phys_level0_sa
.array_len
;
2992 for (unsigned layer
= 0; layer
< depth
; layer
++) {
2993 brw_blorp_copy_miptrees(brw
,
2997 minify(src
->surf
.logical_level0_px
.width
,
2999 minify(src
->surf
.logical_level0_px
.height
,
3004 brw_cache_flush_for_read(brw
, dst
->bo
);
3005 src
->r8stencil_needs_update
= false;
3009 intel_miptree_map_raw(struct brw_context
*brw
,
3010 struct intel_mipmap_tree
*mt
,
3013 struct brw_bo
*bo
= mt
->bo
;
3015 if (brw_batch_references(&brw
->batch
, bo
))
3016 intel_batchbuffer_flush(brw
);
3018 return brw_bo_map(brw
, bo
, mode
);
3022 intel_miptree_unmap_raw(struct intel_mipmap_tree
*mt
)
3024 brw_bo_unmap(mt
->bo
);
3028 intel_miptree_unmap_gtt(struct brw_context
*brw
,
3029 struct intel_mipmap_tree
*mt
,
3030 struct intel_miptree_map
*map
,
3031 unsigned int level
, unsigned int slice
)
3033 intel_miptree_unmap_raw(mt
);
3037 intel_miptree_map_gtt(struct brw_context
*brw
,
3038 struct intel_mipmap_tree
*mt
,
3039 struct intel_miptree_map
*map
,
3040 unsigned int level
, unsigned int slice
)
3042 unsigned int bw
, bh
;
3044 unsigned int image_x
, image_y
;
3045 intptr_t x
= map
->x
;
3046 intptr_t y
= map
->y
;
3048 /* For compressed formats, the stride is the number of bytes per
3049 * row of blocks. intel_miptree_get_image_offset() already does
3052 _mesa_get_format_block_size(mt
->format
, &bw
, &bh
);
3053 assert(y
% bh
== 0);
3054 assert(x
% bw
== 0);
3058 intel_miptree_access_raw(brw
, mt
, level
, slice
,
3059 map
->mode
& GL_MAP_WRITE_BIT
);
3061 base
= intel_miptree_map_raw(brw
, mt
, map
->mode
);
3068 /* Note that in the case of cube maps, the caller must have passed the
3069 * slice number referencing the face.
3071 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3075 map
->stride
= mt
->surf
.row_pitch
;
3076 map
->ptr
= base
+ y
* map
->stride
+ x
* mt
->cpp
;
3079 DBG("%s: %d,%d %dx%d from mt %p (%s) "
3080 "%"PRIiPTR
",%"PRIiPTR
" = %p/%d\n", __func__
,
3081 map
->x
, map
->y
, map
->w
, map
->h
,
3082 mt
, _mesa_get_format_name(mt
->format
),
3083 x
, y
, map
->ptr
, map
->stride
);
3085 map
->unmap
= intel_miptree_unmap_gtt
;
3089 intel_miptree_unmap_blit(struct brw_context
*brw
,
3090 struct intel_mipmap_tree
*mt
,
3091 struct intel_miptree_map
*map
,
3095 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
3096 struct gl_context
*ctx
= &brw
->ctx
;
3098 intel_miptree_unmap_raw(map
->linear_mt
);
3100 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3101 if (devinfo
->gen
>= 6) {
3102 brw_blorp_copy_miptrees(brw
, map
->linear_mt
, 0, 0,
3104 0, 0, map
->x
, map
->y
, map
->w
, map
->h
);
3106 bool ok
= intel_miptree_copy(brw
,
3107 map
->linear_mt
, 0, 0, 0, 0,
3108 mt
, level
, slice
, map
->x
, map
->y
,
3110 WARN_ONCE(!ok
, "Failed to blit from linear temporary mapping");
3114 intel_miptree_release(&map
->linear_mt
);
3118 intel_miptree_map_blit(struct brw_context
*brw
,
3119 struct intel_mipmap_tree
*mt
,
3120 struct intel_miptree_map
*map
,
3121 unsigned int level
, unsigned int slice
)
3123 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
3124 map
->linear_mt
= intel_miptree_create(brw
, GL_TEXTURE_2D
, mt
->format
,
3125 /* first_level */ 0,
3129 MIPTREE_CREATE_LINEAR
);
3131 if (!map
->linear_mt
) {
3132 fprintf(stderr
, "Failed to allocate blit temporary\n");
3135 map
->stride
= map
->linear_mt
->surf
.row_pitch
;
3137 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3138 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3139 * invalidate is set, since we'll be writing the whole rectangle from our
3140 * temporary buffer back out.
3142 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3143 if (devinfo
->gen
>= 6) {
3144 brw_blorp_copy_miptrees(brw
, mt
, level
, slice
,
3145 map
->linear_mt
, 0, 0,
3146 map
->x
, map
->y
, 0, 0, map
->w
, map
->h
);
3148 if (!intel_miptree_copy(brw
,
3149 mt
, level
, slice
, map
->x
, map
->y
,
3150 map
->linear_mt
, 0, 0, 0, 0,
3152 fprintf(stderr
, "Failed to blit\n");
3158 map
->ptr
= intel_miptree_map_raw(brw
, map
->linear_mt
, map
->mode
);
3160 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__
,
3161 map
->x
, map
->y
, map
->w
, map
->h
,
3162 mt
, _mesa_get_format_name(mt
->format
),
3163 level
, slice
, map
->ptr
, map
->stride
);
3165 map
->unmap
= intel_miptree_unmap_blit
;
3169 intel_miptree_release(&map
->linear_mt
);
3175 * "Map" a buffer by copying it to an untiled temporary using MOVNTDQA.
3177 #if defined(USE_SSE41)
3179 intel_miptree_unmap_movntdqa(struct brw_context
*brw
,
3180 struct intel_mipmap_tree
*mt
,
3181 struct intel_miptree_map
*map
,
3185 _mesa_align_free(map
->buffer
);
3191 intel_miptree_map_movntdqa(struct brw_context
*brw
,
3192 struct intel_mipmap_tree
*mt
,
3193 struct intel_miptree_map
*map
,
3194 unsigned int level
, unsigned int slice
)
3196 assert(map
->mode
& GL_MAP_READ_BIT
);
3197 assert(!(map
->mode
& GL_MAP_WRITE_BIT
));
3199 intel_miptree_access_raw(brw
, mt
, level
, slice
, false);
3201 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__
,
3202 map
->x
, map
->y
, map
->w
, map
->h
,
3203 mt
, _mesa_get_format_name(mt
->format
),
3204 level
, slice
, map
->ptr
, map
->stride
);
3206 /* Map the original image */
3209 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3213 void *src
= intel_miptree_map_raw(brw
, mt
, map
->mode
);
3219 src
+= image_y
* mt
->surf
.row_pitch
;
3220 src
+= image_x
* mt
->cpp
;
3222 /* Due to the pixel offsets for the particular image being mapped, our
3223 * src pointer may not be 16-byte aligned. However, if the pitch is
3224 * divisible by 16, then the amount by which it's misaligned will remain
3225 * consistent from row to row.
3227 assert((mt
->surf
.row_pitch
% 16) == 0);
3228 const int misalignment
= ((uintptr_t) src
) & 15;
3230 /* Create an untiled temporary buffer for the mapping. */
3231 const unsigned width_bytes
= _mesa_format_row_stride(mt
->format
, map
->w
);
3233 map
->stride
= ALIGN(misalignment
+ width_bytes
, 16);
3235 map
->buffer
= _mesa_align_malloc(map
->stride
* map
->h
, 16);
3236 /* Offset the destination so it has the same misalignment as src. */
3237 map
->ptr
= map
->buffer
+ misalignment
;
3239 assert((((uintptr_t) map
->ptr
) & 15) == misalignment
);
3241 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3242 void *dst_ptr
= map
->ptr
+ y
* map
->stride
;
3243 void *src_ptr
= src
+ y
* mt
->surf
.row_pitch
;
3245 _mesa_streaming_load_memcpy(dst_ptr
, src_ptr
, width_bytes
);
3248 intel_miptree_unmap_raw(mt
);
3250 map
->unmap
= intel_miptree_unmap_movntdqa
;
3255 intel_miptree_unmap_s8(struct brw_context
*brw
,
3256 struct intel_mipmap_tree
*mt
,
3257 struct intel_miptree_map
*map
,
3261 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3262 unsigned int image_x
, image_y
;
3263 uint8_t *untiled_s8_map
= map
->ptr
;
3264 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
, GL_MAP_WRITE_BIT
);
3266 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3268 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3269 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3270 ptrdiff_t offset
= intel_offset_S8(mt
->surf
.row_pitch
,
3271 image_x
+ x
+ map
->x
,
3272 image_y
+ y
+ map
->y
,
3273 brw
->has_swizzling
);
3274 tiled_s8_map
[offset
] = untiled_s8_map
[y
* map
->w
+ x
];
3278 intel_miptree_unmap_raw(mt
);
3285 intel_miptree_map_s8(struct brw_context
*brw
,
3286 struct intel_mipmap_tree
*mt
,
3287 struct intel_miptree_map
*map
,
3288 unsigned int level
, unsigned int slice
)
3290 map
->stride
= map
->w
;
3291 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
3295 intel_miptree_access_raw(brw
, mt
, level
, slice
,
3296 map
->mode
& GL_MAP_WRITE_BIT
);
3298 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3299 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3300 * invalidate is set, since we'll be writing the whole rectangle from our
3301 * temporary buffer back out.
3303 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3304 uint8_t *untiled_s8_map
= map
->ptr
;
3305 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
, GL_MAP_READ_BIT
);
3306 unsigned int image_x
, image_y
;
3308 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3310 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3311 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3312 ptrdiff_t offset
= intel_offset_S8(mt
->surf
.row_pitch
,
3313 x
+ image_x
+ map
->x
,
3314 y
+ image_y
+ map
->y
,
3315 brw
->has_swizzling
);
3316 untiled_s8_map
[y
* map
->w
+ x
] = tiled_s8_map
[offset
];
3320 intel_miptree_unmap_raw(mt
);
3322 DBG("%s: %d,%d %dx%d from mt %p %d,%d = %p/%d\n", __func__
,
3323 map
->x
, map
->y
, map
->w
, map
->h
,
3324 mt
, map
->x
+ image_x
, map
->y
+ image_y
, map
->ptr
, map
->stride
);
3326 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__
,
3327 map
->x
, map
->y
, map
->w
, map
->h
,
3328 mt
, map
->ptr
, map
->stride
);
3331 map
->unmap
= intel_miptree_unmap_s8
;
3335 intel_miptree_unmap_etc(struct brw_context
*brw
,
3336 struct intel_mipmap_tree
*mt
,
3337 struct intel_miptree_map
*map
,
3343 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3348 uint8_t *dst
= intel_miptree_map_raw(brw
, mt
, GL_MAP_WRITE_BIT
)
3349 + image_y
* mt
->surf
.row_pitch
3350 + image_x
* mt
->cpp
;
3352 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
)
3353 _mesa_etc1_unpack_rgba8888(dst
, mt
->surf
.row_pitch
,
3354 map
->ptr
, map
->stride
,
3357 _mesa_unpack_etc2_format(dst
, mt
->surf
.row_pitch
,
3358 map
->ptr
, map
->stride
,
3359 map
->w
, map
->h
, mt
->etc_format
);
3361 intel_miptree_unmap_raw(mt
);
3366 intel_miptree_map_etc(struct brw_context
*brw
,
3367 struct intel_mipmap_tree
*mt
,
3368 struct intel_miptree_map
*map
,
3372 assert(mt
->etc_format
!= MESA_FORMAT_NONE
);
3373 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
) {
3374 assert(mt
->format
== MESA_FORMAT_R8G8B8X8_UNORM
);
3377 assert(map
->mode
& GL_MAP_WRITE_BIT
);
3378 assert(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
);
3380 intel_miptree_access_raw(brw
, mt
, level
, slice
, true);
3382 map
->stride
= _mesa_format_row_stride(mt
->etc_format
, map
->w
);
3383 map
->buffer
= malloc(_mesa_format_image_size(mt
->etc_format
,
3384 map
->w
, map
->h
, 1));
3385 map
->ptr
= map
->buffer
;
3386 map
->unmap
= intel_miptree_unmap_etc
;
3390 * Mapping functions for packed depth/stencil miptrees backed by real separate
3391 * miptrees for depth and stencil.
3393 * On gen7, and to support HiZ pre-gen7, we have to have the stencil buffer
3394 * separate from the depth buffer. Yet at the GL API level, we have to expose
3395 * packed depth/stencil textures and FBO attachments, and Mesa core expects to
3396 * be able to map that memory for texture storage and glReadPixels-type
3397 * operations. We give Mesa core that access by mallocing a temporary and
3398 * copying the data between the actual backing store and the temporary.
3401 intel_miptree_unmap_depthstencil(struct brw_context
*brw
,
3402 struct intel_mipmap_tree
*mt
,
3403 struct intel_miptree_map
*map
,
3407 struct intel_mipmap_tree
*z_mt
= mt
;
3408 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
3409 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
3411 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3412 uint32_t *packed_map
= map
->ptr
;
3413 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
, GL_MAP_WRITE_BIT
);
3414 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
, GL_MAP_WRITE_BIT
);
3415 unsigned int s_image_x
, s_image_y
;
3416 unsigned int z_image_x
, z_image_y
;
3418 intel_miptree_get_image_offset(s_mt
, level
, slice
,
3419 &s_image_x
, &s_image_y
);
3420 intel_miptree_get_image_offset(z_mt
, level
, slice
,
3421 &z_image_x
, &z_image_y
);
3423 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3424 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3425 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->surf
.row_pitch
,
3426 x
+ s_image_x
+ map
->x
,
3427 y
+ s_image_y
+ map
->y
,
3428 brw
->has_swizzling
);
3429 ptrdiff_t z_offset
= ((y
+ z_image_y
+ map
->y
) *
3430 (z_mt
->surf
.row_pitch
/ 4) +
3431 (x
+ z_image_x
+ map
->x
));
3433 if (map_z32f_x24s8
) {
3434 z_map
[z_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 0];
3435 s_map
[s_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 1];
3437 uint32_t packed
= packed_map
[y
* map
->w
+ x
];
3438 s_map
[s_offset
] = packed
>> 24;
3439 z_map
[z_offset
] = packed
;
3444 intel_miptree_unmap_raw(s_mt
);
3445 intel_miptree_unmap_raw(z_mt
);
3447 DBG("%s: %d,%d %dx%d from z mt %p (%s) %d,%d, s mt %p %d,%d = %p/%d\n",
3449 map
->x
, map
->y
, map
->w
, map
->h
,
3450 z_mt
, _mesa_get_format_name(z_mt
->format
),
3451 map
->x
+ z_image_x
, map
->y
+ z_image_y
,
3452 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
3453 map
->ptr
, map
->stride
);
3460 intel_miptree_map_depthstencil(struct brw_context
*brw
,
3461 struct intel_mipmap_tree
*mt
,
3462 struct intel_miptree_map
*map
,
3463 unsigned int level
, unsigned int slice
)
3465 struct intel_mipmap_tree
*z_mt
= mt
;
3466 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
3467 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
3468 int packed_bpp
= map_z32f_x24s8
? 8 : 4;
3470 map
->stride
= map
->w
* packed_bpp
;
3471 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
3475 intel_miptree_access_raw(brw
, mt
, level
, slice
,
3476 map
->mode
& GL_MAP_WRITE_BIT
);
3478 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3479 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3480 * invalidate is set, since we'll be writing the whole rectangle from our
3481 * temporary buffer back out.
3483 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3484 uint32_t *packed_map
= map
->ptr
;
3485 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
, GL_MAP_READ_BIT
);
3486 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
, GL_MAP_READ_BIT
);
3487 unsigned int s_image_x
, s_image_y
;
3488 unsigned int z_image_x
, z_image_y
;
3490 intel_miptree_get_image_offset(s_mt
, level
, slice
,
3491 &s_image_x
, &s_image_y
);
3492 intel_miptree_get_image_offset(z_mt
, level
, slice
,
3493 &z_image_x
, &z_image_y
);
3495 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3496 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3497 int map_x
= map
->x
+ x
, map_y
= map
->y
+ y
;
3498 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->surf
.row_pitch
,
3501 brw
->has_swizzling
);
3502 ptrdiff_t z_offset
= ((map_y
+ z_image_y
) *
3503 (z_mt
->surf
.row_pitch
/ 4) +
3504 (map_x
+ z_image_x
));
3505 uint8_t s
= s_map
[s_offset
];
3506 uint32_t z
= z_map
[z_offset
];
3508 if (map_z32f_x24s8
) {
3509 packed_map
[(y
* map
->w
+ x
) * 2 + 0] = z
;
3510 packed_map
[(y
* map
->w
+ x
) * 2 + 1] = s
;
3512 packed_map
[y
* map
->w
+ x
] = (s
<< 24) | (z
& 0x00ffffff);
3517 intel_miptree_unmap_raw(s_mt
);
3518 intel_miptree_unmap_raw(z_mt
);
3520 DBG("%s: %d,%d %dx%d from z mt %p %d,%d, s mt %p %d,%d = %p/%d\n",
3522 map
->x
, map
->y
, map
->w
, map
->h
,
3523 z_mt
, map
->x
+ z_image_x
, map
->y
+ z_image_y
,
3524 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
3525 map
->ptr
, map
->stride
);
3527 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__
,
3528 map
->x
, map
->y
, map
->w
, map
->h
,
3529 mt
, map
->ptr
, map
->stride
);
3532 map
->unmap
= intel_miptree_unmap_depthstencil
;
3536 * Create and attach a map to the miptree at (level, slice). Return the
3539 static struct intel_miptree_map
*
3540 intel_miptree_attach_map(struct intel_mipmap_tree
*mt
,
3549 struct intel_miptree_map
*map
= calloc(1, sizeof(*map
));
3554 assert(mt
->level
[level
].slice
[slice
].map
== NULL
);
3555 mt
->level
[level
].slice
[slice
].map
= map
;
3567 * Release the map at (level, slice).
3570 intel_miptree_release_map(struct intel_mipmap_tree
*mt
,
3574 struct intel_miptree_map
**map
;
3576 map
= &mt
->level
[level
].slice
[slice
].map
;
3582 can_blit_slice(struct intel_mipmap_tree
*mt
,
3583 unsigned int level
, unsigned int slice
)
3585 /* See intel_miptree_blit() for details on the 32k pitch limit. */
3586 if (mt
->surf
.row_pitch
>= 32768)
3593 use_intel_mipree_map_blit(struct brw_context
*brw
,
3594 struct intel_mipmap_tree
*mt
,
3599 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
3601 if (devinfo
->has_llc
&&
3602 /* It's probably not worth swapping to the blit ring because of
3603 * all the overhead involved.
3605 !(mode
& GL_MAP_WRITE_BIT
) &&
3607 (mt
->surf
.tiling
== ISL_TILING_X
||
3608 /* Prior to Sandybridge, the blitter can't handle Y tiling */
3609 (devinfo
->gen
>= 6 && mt
->surf
.tiling
== ISL_TILING_Y0
) ||
3610 /* Fast copy blit on skl+ supports all tiling formats. */
3611 devinfo
->gen
>= 9) &&
3612 can_blit_slice(mt
, level
, slice
))
3615 if (mt
->surf
.tiling
!= ISL_TILING_LINEAR
&&
3616 mt
->bo
->size
>= brw
->max_gtt_map_object_size
) {
3617 assert(can_blit_slice(mt
, level
, slice
));
3625 * Parameter \a out_stride has type ptrdiff_t not because the buffer stride may
3626 * exceed 32 bits but to diminish the likelihood subtle bugs in pointer
3627 * arithmetic overflow.
3629 * If you call this function and use \a out_stride, then you're doing pointer
3630 * arithmetic on \a out_ptr. The type of \a out_stride doesn't prevent all
3631 * bugs. The caller must still take care to avoid 32-bit overflow errors in
3632 * all arithmetic expressions that contain buffer offsets and pixel sizes,
3633 * which usually have type uint32_t or GLuint.
3636 intel_miptree_map(struct brw_context
*brw
,
3637 struct intel_mipmap_tree
*mt
,
3646 ptrdiff_t *out_stride
)
3648 struct intel_miptree_map
*map
;
3650 assert(mt
->surf
.samples
== 1);
3652 map
= intel_miptree_attach_map(mt
, level
, slice
, x
, y
, w
, h
, mode
);
3659 if (mt
->format
== MESA_FORMAT_S_UINT8
) {
3660 intel_miptree_map_s8(brw
, mt
, map
, level
, slice
);
3661 } else if (mt
->etc_format
!= MESA_FORMAT_NONE
&&
3662 !(mode
& BRW_MAP_DIRECT_BIT
)) {
3663 intel_miptree_map_etc(brw
, mt
, map
, level
, slice
);
3664 } else if (mt
->stencil_mt
&& !(mode
& BRW_MAP_DIRECT_BIT
)) {
3665 intel_miptree_map_depthstencil(brw
, mt
, map
, level
, slice
);
3666 } else if (use_intel_mipree_map_blit(brw
, mt
, mode
, level
, slice
)) {
3667 intel_miptree_map_blit(brw
, mt
, map
, level
, slice
);
3668 #if defined(USE_SSE41)
3669 } else if (!(mode
& GL_MAP_WRITE_BIT
) &&
3670 !mt
->compressed
&& cpu_has_sse4_1
&&
3671 (mt
->surf
.row_pitch
% 16 == 0)) {
3672 intel_miptree_map_movntdqa(brw
, mt
, map
, level
, slice
);
3675 intel_miptree_map_gtt(brw
, mt
, map
, level
, slice
);
3678 *out_ptr
= map
->ptr
;
3679 *out_stride
= map
->stride
;
3681 if (map
->ptr
== NULL
)
3682 intel_miptree_release_map(mt
, level
, slice
);
3686 intel_miptree_unmap(struct brw_context
*brw
,
3687 struct intel_mipmap_tree
*mt
,
3691 struct intel_miptree_map
*map
= mt
->level
[level
].slice
[slice
].map
;
3693 assert(mt
->surf
.samples
== 1);
3698 DBG("%s: mt %p (%s) level %d slice %d\n", __func__
,
3699 mt
, _mesa_get_format_name(mt
->format
), level
, slice
);
3702 map
->unmap(brw
, mt
, map
, level
, slice
);
3704 intel_miptree_release_map(mt
, level
, slice
);
3708 get_isl_surf_dim(GLenum target
)
3712 case GL_TEXTURE_1D_ARRAY
:
3713 return ISL_SURF_DIM_1D
;
3716 case GL_TEXTURE_2D_ARRAY
:
3717 case GL_TEXTURE_RECTANGLE
:
3718 case GL_TEXTURE_CUBE_MAP
:
3719 case GL_TEXTURE_CUBE_MAP_ARRAY
:
3720 case GL_TEXTURE_2D_MULTISAMPLE
:
3721 case GL_TEXTURE_2D_MULTISAMPLE_ARRAY
:
3722 case GL_TEXTURE_EXTERNAL_OES
:
3723 return ISL_SURF_DIM_2D
;
3726 return ISL_SURF_DIM_3D
;
3729 unreachable("Invalid texture target");
3733 get_isl_dim_layout(const struct gen_device_info
*devinfo
,
3734 enum isl_tiling tiling
, GLenum target
)
3738 case GL_TEXTURE_1D_ARRAY
:
3739 return (devinfo
->gen
>= 9 && tiling
== ISL_TILING_LINEAR
?
3740 ISL_DIM_LAYOUT_GEN9_1D
: ISL_DIM_LAYOUT_GEN4_2D
);
3743 case GL_TEXTURE_2D_ARRAY
:
3744 case GL_TEXTURE_RECTANGLE
:
3745 case GL_TEXTURE_2D_MULTISAMPLE
:
3746 case GL_TEXTURE_2D_MULTISAMPLE_ARRAY
:
3747 case GL_TEXTURE_EXTERNAL_OES
:
3748 return ISL_DIM_LAYOUT_GEN4_2D
;
3750 case GL_TEXTURE_CUBE_MAP
:
3751 case GL_TEXTURE_CUBE_MAP_ARRAY
:
3752 return (devinfo
->gen
== 4 ? ISL_DIM_LAYOUT_GEN4_3D
:
3753 ISL_DIM_LAYOUT_GEN4_2D
);
3756 return (devinfo
->gen
>= 9 ?
3757 ISL_DIM_LAYOUT_GEN4_2D
: ISL_DIM_LAYOUT_GEN4_3D
);
3760 unreachable("Invalid texture target");
3764 intel_miptree_set_clear_color(struct brw_context
*brw
,
3765 struct intel_mipmap_tree
*mt
,
3766 union isl_color_value clear_color
)
3768 if (memcmp(&mt
->fast_clear_color
, &clear_color
, sizeof(clear_color
)) != 0) {
3769 mt
->fast_clear_color
= clear_color
;
3770 if (mt
->aux_buf
->clear_color_bo
) {
3771 /* We can't update the clear color while the hardware is still using
3772 * the previous one for a resolve or sampling from it. Make sure that
3773 * there are no pending commands at this point.
3775 brw_emit_pipe_control_flush(brw
, PIPE_CONTROL_CS_STALL
);
3776 for (int i
= 0; i
< 4; i
++) {
3777 brw_store_data_imm32(brw
, mt
->aux_buf
->clear_color_bo
,
3778 mt
->aux_buf
->clear_color_offset
+ i
* 4,
3779 mt
->fast_clear_color
.u32
[i
]);
3781 brw_emit_pipe_control_flush(brw
, PIPE_CONTROL_STATE_CACHE_INVALIDATE
);
3783 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
3789 union isl_color_value
3790 intel_miptree_get_clear_color(const struct gen_device_info
*devinfo
,
3791 const struct intel_mipmap_tree
*mt
,
3792 enum isl_format view_format
, bool sampling
,
3793 struct brw_bo
**clear_color_bo
,
3794 uint32_t *clear_color_offset
)
3796 assert(mt
->aux_buf
);
3798 if (devinfo
->gen
== 10 && isl_format_is_srgb(view_format
) && sampling
) {
3799 /* The gen10 sampler doesn't gamma-correct the clear color. In this case,
3800 * we switch to using the inline clear color and do the sRGB color
3801 * conversion process defined in the OpenGL spec. The red, green, and
3802 * blue channels take part in gamma correction, while the alpha channel
3805 union isl_color_value srgb_decoded_value
= mt
->fast_clear_color
;
3806 for (unsigned i
= 0; i
< 3; i
++) {
3807 srgb_decoded_value
.f32
[i
] =
3808 util_format_srgb_to_linear_float(mt
->fast_clear_color
.f32
[i
]);
3810 *clear_color_bo
= 0;
3811 *clear_color_offset
= 0;
3812 return srgb_decoded_value
;
3814 *clear_color_bo
= mt
->aux_buf
->clear_color_bo
;
3815 *clear_color_offset
= mt
->aux_buf
->clear_color_offset
;
3816 return mt
->fast_clear_color
;