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"
48 #include "x86/common_x86_asm.h"
50 #define FILE_DEBUG_FLAG DEBUG_MIPTREE
52 static void *intel_miptree_map_raw(struct brw_context
*brw
,
53 struct intel_mipmap_tree
*mt
,
56 static void intel_miptree_unmap_raw(struct intel_mipmap_tree
*mt
);
59 intel_miptree_alloc_aux(struct brw_context
*brw
,
60 struct intel_mipmap_tree
*mt
);
63 intel_miptree_supports_mcs(struct brw_context
*brw
,
64 const struct intel_mipmap_tree
*mt
)
66 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
68 /* MCS compression only applies to multisampled miptrees */
69 if (mt
->surf
.samples
<= 1)
72 /* Prior to Gen7, all MSAA surfaces used IMS layout. */
76 /* In Gen7, IMS layout is only used for depth and stencil buffers. */
77 switch (_mesa_get_format_base_format(mt
->format
)) {
78 case GL_DEPTH_COMPONENT
:
79 case GL_STENCIL_INDEX
:
80 case GL_DEPTH_STENCIL
:
83 /* From the Ivy Bridge PRM, Vol4 Part1 p77 ("MCS Enable"):
85 * This field must be set to 0 for all SINT MSRTs when all RT channels
88 * In practice this means that we have to disable MCS for all signed
89 * integer MSAA buffers. The alternative, to disable MCS only when one
90 * of the render target channels is disabled, is impractical because it
91 * would require converting between CMS and UMS MSAA layouts on the fly,
94 if (devinfo
->gen
== 7 && _mesa_get_format_datatype(mt
->format
) == GL_INT
) {
103 intel_tiling_supports_ccs(const struct brw_context
*brw
,
104 enum isl_tiling tiling
)
106 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
108 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
109 * Target(s)", beneath the "Fast Color Clear" bullet (p326):
111 * - Support is limited to tiled render targets.
113 * Gen9 changes the restriction to Y-tile only.
115 if (devinfo
->gen
>= 9)
116 return tiling
== ISL_TILING_Y0
;
117 else if (devinfo
->gen
>= 7)
118 return tiling
!= ISL_TILING_LINEAR
;
124 * For a single-sampled render target ("non-MSRT"), determine if an MCS buffer
125 * can be used. This doesn't (and should not) inspect any of the properties of
128 * From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render Target(s)",
129 * beneath the "Fast Color Clear" bullet (p326):
131 * - Support is for non-mip-mapped and non-array surface types only.
133 * And then later, on p327:
135 * - MCS buffer for non-MSRT is supported only for RT formats 32bpp,
138 * From the Skylake documentation, it is made clear that X-tiling is no longer
141 * - MCS and Lossless compression is supported for TiledY/TileYs/TileYf
145 intel_miptree_supports_ccs(struct brw_context
*brw
,
146 const struct intel_mipmap_tree
*mt
)
148 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
150 /* MCS support does not exist prior to Gen7 */
151 if (devinfo
->gen
< 7)
154 /* This function applies only to non-multisampled render targets. */
155 if (mt
->surf
.samples
> 1)
158 /* MCS is only supported for color buffers */
159 switch (_mesa_get_format_base_format(mt
->format
)) {
160 case GL_DEPTH_COMPONENT
:
161 case GL_DEPTH_STENCIL
:
162 case GL_STENCIL_INDEX
:
166 if (mt
->cpp
!= 4 && mt
->cpp
!= 8 && mt
->cpp
!= 16)
169 const bool mip_mapped
= mt
->first_level
!= 0 || mt
->last_level
!= 0;
170 const bool arrayed
= mt
->surf
.logical_level0_px
.array_len
> 1 ||
171 mt
->surf
.logical_level0_px
.depth
> 1;
174 /* Multisample surfaces with the CMS layout are not layered surfaces,
175 * yet still have physical_depth0 > 1. Assert that we don't
176 * accidentally reject a multisampled surface here. We should have
177 * rejected it earlier by explicitly checking the sample count.
179 assert(mt
->surf
.samples
== 1);
182 /* Handle the hardware restrictions...
184 * All GENs have the following restriction: "MCS buffer for non-MSRT is
185 * supported only for RT formats 32bpp, 64bpp, and 128bpp."
187 * From the HSW PRM Volume 7: 3D-Media-GPGPU, page 652: (Color Clear of
188 * Non-MultiSampler Render Target Restrictions) Support is for
189 * non-mip-mapped and non-array surface types only.
191 * From the BDW PRM Volume 7: 3D-Media-GPGPU, page 649: (Color Clear of
192 * Non-MultiSampler Render Target Restriction). Mip-mapped and arrayed
193 * surfaces are supported with MCS buffer layout with these alignments in
194 * the RT space: Horizontal Alignment = 256 and Vertical Alignment = 128.
196 * From the SKL PRM Volume 7: 3D-Media-GPGPU, page 632: (Color Clear of
197 * Non-MultiSampler Render Target Restriction). Mip-mapped and arrayed
198 * surfaces are supported with MCS buffer layout with these alignments in
199 * the RT space: Horizontal Alignment = 128 and Vertical Alignment = 64.
201 if (devinfo
->gen
< 8 && (mip_mapped
|| arrayed
))
204 /* There's no point in using an MCS buffer if the surface isn't in a
207 if (!brw
->mesa_format_supports_render
[mt
->format
])
214 intel_tiling_supports_hiz(const struct brw_context
*brw
,
215 enum isl_tiling tiling
)
217 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
219 if (devinfo
->gen
< 6)
222 return tiling
== ISL_TILING_Y0
;
226 intel_miptree_supports_hiz(const struct brw_context
*brw
,
227 const struct intel_mipmap_tree
*mt
)
232 switch (mt
->format
) {
233 case MESA_FORMAT_Z_FLOAT32
:
234 case MESA_FORMAT_Z32_FLOAT_S8X24_UINT
:
235 case MESA_FORMAT_Z24_UNORM_X8_UINT
:
236 case MESA_FORMAT_Z24_UNORM_S8_UINT
:
237 case MESA_FORMAT_Z_UNORM16
:
245 intel_miptree_supports_ccs_e(struct brw_context
*brw
,
246 const struct intel_mipmap_tree
*mt
)
248 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
250 if (devinfo
->gen
< 9)
253 /* For now compression is only enabled for integer formats even though
254 * there exist supported floating point formats also. This is a heuristic
255 * decision based on current public benchmarks. In none of the cases these
256 * formats provided any improvement but a few cases were seen to regress.
257 * Hence these are left to to be enabled in the future when they are known
260 if (_mesa_get_format_datatype(mt
->format
) == GL_FLOAT
)
263 if (!intel_miptree_supports_ccs(brw
, mt
))
266 /* Many window system buffers are sRGB even if they are never rendered as
267 * sRGB. For those, we want CCS_E for when sRGBEncode is false. When the
268 * surface is used as sRGB, we fall back to CCS_D.
270 mesa_format linear_format
= _mesa_get_srgb_format_linear(mt
->format
);
271 enum isl_format isl_format
= brw_isl_format_for_mesa_format(linear_format
);
272 return isl_format_supports_ccs_e(&brw
->screen
->devinfo
, isl_format
);
276 * Determine depth format corresponding to a depth+stencil format,
277 * for separate stencil.
280 intel_depth_format_for_depthstencil_format(mesa_format format
) {
282 case MESA_FORMAT_Z24_UNORM_S8_UINT
:
283 return MESA_FORMAT_Z24_UNORM_X8_UINT
;
284 case MESA_FORMAT_Z32_FLOAT_S8X24_UINT
:
285 return MESA_FORMAT_Z_FLOAT32
;
292 create_mapping_table(GLenum target
, unsigned first_level
, unsigned last_level
,
293 unsigned depth0
, struct intel_mipmap_level
*table
)
295 for (unsigned level
= first_level
; level
<= last_level
; level
++) {
297 target
== GL_TEXTURE_3D
? minify(depth0
, level
) : depth0
;
299 table
[level
].slice
= calloc(d
, sizeof(*table
[0].slice
));
300 if (!table
[level
].slice
)
307 for (unsigned level
= first_level
; level
<= last_level
; level
++)
308 free(table
[level
].slice
);
314 needs_separate_stencil(const struct brw_context
*brw
,
315 struct intel_mipmap_tree
*mt
,
318 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
320 if (_mesa_get_format_base_format(format
) != GL_DEPTH_STENCIL
)
323 if (devinfo
->must_use_separate_stencil
)
326 return brw
->has_separate_stencil
&&
327 intel_miptree_supports_hiz(brw
, mt
);
331 * Choose the aux usage for this miptree. This function must be called fairly
332 * late in the miptree create process after we have a tiling.
335 intel_miptree_choose_aux_usage(struct brw_context
*brw
,
336 struct intel_mipmap_tree
*mt
)
338 assert(mt
->aux_usage
== ISL_AUX_USAGE_NONE
);
340 if (intel_miptree_supports_mcs(brw
, mt
)) {
341 assert(mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
342 mt
->aux_usage
= ISL_AUX_USAGE_MCS
;
343 } else if (intel_tiling_supports_ccs(brw
, mt
->surf
.tiling
) &&
344 intel_miptree_supports_ccs(brw
, mt
)) {
345 if (!unlikely(INTEL_DEBUG
& DEBUG_NO_RBC
) &&
346 intel_miptree_supports_ccs_e(brw
, mt
)) {
347 mt
->aux_usage
= ISL_AUX_USAGE_CCS_E
;
349 mt
->aux_usage
= ISL_AUX_USAGE_CCS_D
;
351 } else if (intel_tiling_supports_hiz(brw
, mt
->surf
.tiling
) &&
352 intel_miptree_supports_hiz(brw
, mt
)) {
353 mt
->aux_usage
= ISL_AUX_USAGE_HIZ
;
356 /* We can do fast-clear on all auxiliary surface types that are
357 * allocated through the normal texture creation paths.
359 if (mt
->aux_usage
!= ISL_AUX_USAGE_NONE
)
360 mt
->supports_fast_clear
= true;
365 * Choose an appropriate uncompressed format for a requested
366 * compressed format, if unsupported.
369 intel_lower_compressed_format(struct brw_context
*brw
, mesa_format format
)
371 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
373 /* No need to lower ETC formats on these platforms,
374 * they are supported natively.
376 if (devinfo
->gen
>= 8 || devinfo
->is_baytrail
)
380 case MESA_FORMAT_ETC1_RGB8
:
381 return MESA_FORMAT_R8G8B8X8_UNORM
;
382 case MESA_FORMAT_ETC2_RGB8
:
383 return MESA_FORMAT_R8G8B8X8_UNORM
;
384 case MESA_FORMAT_ETC2_SRGB8
:
385 case MESA_FORMAT_ETC2_SRGB8_ALPHA8_EAC
:
386 case MESA_FORMAT_ETC2_SRGB8_PUNCHTHROUGH_ALPHA1
:
387 return MESA_FORMAT_B8G8R8A8_SRGB
;
388 case MESA_FORMAT_ETC2_RGBA8_EAC
:
389 case MESA_FORMAT_ETC2_RGB8_PUNCHTHROUGH_ALPHA1
:
390 return MESA_FORMAT_R8G8B8A8_UNORM
;
391 case MESA_FORMAT_ETC2_R11_EAC
:
392 return MESA_FORMAT_R_UNORM16
;
393 case MESA_FORMAT_ETC2_SIGNED_R11_EAC
:
394 return MESA_FORMAT_R_SNORM16
;
395 case MESA_FORMAT_ETC2_RG11_EAC
:
396 return MESA_FORMAT_R16G16_UNORM
;
397 case MESA_FORMAT_ETC2_SIGNED_RG11_EAC
:
398 return MESA_FORMAT_R16G16_SNORM
;
400 /* Non ETC1 / ETC2 format */
406 brw_get_num_logical_layers(const struct intel_mipmap_tree
*mt
, unsigned level
)
408 if (mt
->surf
.dim
== ISL_SURF_DIM_3D
)
409 return minify(mt
->surf
.logical_level0_px
.depth
, level
);
411 return mt
->surf
.logical_level0_px
.array_len
;
414 UNUSED
static unsigned
415 get_num_phys_layers(const struct isl_surf
*surf
, unsigned level
)
417 /* In case of physical dimensions one needs to consider also the layout.
418 * See isl_calc_phys_level0_extent_sa().
420 if (surf
->dim
!= ISL_SURF_DIM_3D
)
421 return surf
->phys_level0_sa
.array_len
;
423 if (surf
->dim_layout
== ISL_DIM_LAYOUT_GEN4_2D
)
424 return minify(surf
->phys_level0_sa
.array_len
, level
);
426 return minify(surf
->phys_level0_sa
.depth
, level
);
429 /** \brief Assert that the level and layer are valid for the miptree. */
431 intel_miptree_check_level_layer(const struct intel_mipmap_tree
*mt
,
439 assert(level
>= mt
->first_level
);
440 assert(level
<= mt
->last_level
);
441 assert(layer
< get_num_phys_layers(&mt
->surf
, level
));
444 static enum isl_aux_state
**
445 create_aux_state_map(struct intel_mipmap_tree
*mt
,
446 enum isl_aux_state initial
)
448 const uint32_t levels
= mt
->last_level
+ 1;
450 uint32_t total_slices
= 0;
451 for (uint32_t level
= 0; level
< levels
; level
++)
452 total_slices
+= brw_get_num_logical_layers(mt
, level
);
454 const size_t per_level_array_size
= levels
* sizeof(enum isl_aux_state
*);
456 /* We're going to allocate a single chunk of data for both the per-level
457 * reference array and the arrays of aux_state. This makes cleanup
458 * significantly easier.
460 const size_t total_size
= per_level_array_size
+
461 total_slices
* sizeof(enum isl_aux_state
);
462 void *data
= malloc(total_size
);
466 enum isl_aux_state
**per_level_arr
= data
;
467 enum isl_aux_state
*s
= data
+ per_level_array_size
;
468 for (uint32_t level
= 0; level
< levels
; level
++) {
469 per_level_arr
[level
] = s
;
470 const unsigned level_layers
= brw_get_num_logical_layers(mt
, level
);
471 for (uint32_t a
= 0; a
< level_layers
; a
++)
474 assert((void *)s
== data
+ total_size
);
476 return per_level_arr
;
480 free_aux_state_map(enum isl_aux_state
**state
)
486 need_to_retile_as_linear(struct brw_context
*brw
, unsigned row_pitch
,
487 enum isl_tiling tiling
, unsigned samples
)
492 if (tiling
== ISL_TILING_LINEAR
)
495 /* If the width is much smaller than a tile, don't bother tiling. */
499 if (ALIGN(row_pitch
, 512) >= 32768) {
500 perf_debug("row pitch %u too large to blit, falling back to untiled",
509 need_to_retile_as_x(const struct brw_context
*brw
, uint64_t size
,
510 enum isl_tiling tiling
)
512 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
514 /* If the BO is too large to fit in the aperture, we need to use the
515 * BLT engine to support it. Prior to Sandybridge, the BLT paths can't
516 * handle Y-tiling, so we need to fall back to X.
518 if (devinfo
->gen
< 6 && size
>= brw
->max_gtt_map_object_size
&&
519 tiling
== ISL_TILING_Y0
)
525 static struct intel_mipmap_tree
*
526 make_surface(struct brw_context
*brw
, GLenum target
, mesa_format format
,
527 unsigned first_level
, unsigned last_level
,
528 unsigned width0
, unsigned height0
, unsigned depth0
,
529 unsigned num_samples
, isl_tiling_flags_t tiling_flags
,
530 isl_surf_usage_flags_t isl_usage_flags
, uint32_t alloc_flags
,
531 unsigned row_pitch
, struct brw_bo
*bo
)
533 struct intel_mipmap_tree
*mt
= calloc(sizeof(*mt
), 1);
537 if (!create_mapping_table(target
, first_level
, last_level
, depth0
,
545 if (target
== GL_TEXTURE_CUBE_MAP
||
546 target
== GL_TEXTURE_CUBE_MAP_ARRAY
)
547 isl_usage_flags
|= ISL_SURF_USAGE_CUBE_BIT
;
549 DBG("%s: %s %s %ux %u:%u:%u %d..%d <-- %p\n",
551 _mesa_enum_to_string(target
),
552 _mesa_get_format_name(format
),
553 num_samples
, width0
, height0
, depth0
,
554 first_level
, last_level
, mt
);
556 struct isl_surf_init_info init_info
= {
557 .dim
= get_isl_surf_dim(target
),
558 .format
= translate_tex_format(brw
, format
, false),
561 .depth
= target
== GL_TEXTURE_3D
? depth0
: 1,
562 .levels
= last_level
- first_level
+ 1,
563 .array_len
= target
== GL_TEXTURE_3D
? 1 : depth0
,
564 .samples
= num_samples
,
565 .row_pitch
= row_pitch
,
566 .usage
= isl_usage_flags
,
567 .tiling_flags
= tiling_flags
,
570 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
573 /* Depth surfaces are always Y-tiled and stencil is always W-tiled, although
574 * on gen7 platforms we also need to create Y-tiled copies of stencil for
575 * texturing since the hardware can't sample from W-tiled surfaces. For
576 * everything else, check for corner cases needing special treatment.
578 bool is_depth_stencil
=
579 mt
->surf
.usage
& (ISL_SURF_USAGE_STENCIL_BIT
| ISL_SURF_USAGE_DEPTH_BIT
);
580 if (!is_depth_stencil
) {
581 if (need_to_retile_as_linear(brw
, mt
->surf
.row_pitch
,
582 mt
->surf
.tiling
, mt
->surf
.samples
)) {
583 init_info
.tiling_flags
= 1u << ISL_TILING_LINEAR
;
584 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
586 } else if (need_to_retile_as_x(brw
, mt
->surf
.size
, mt
->surf
.tiling
)) {
587 init_info
.tiling_flags
= 1u << ISL_TILING_X
;
588 if (!isl_surf_init_s(&brw
->isl_dev
, &mt
->surf
, &init_info
))
593 /* In case of linear the buffer gets padded by fixed 64 bytes and therefore
594 * the size may not be multiple of row_pitch.
595 * See isl_apply_surface_padding().
597 if (mt
->surf
.tiling
!= ISL_TILING_LINEAR
)
598 assert(mt
->surf
.size
% mt
->surf
.row_pitch
== 0);
601 mt
->bo
= brw_bo_alloc_tiled(brw
->bufmgr
, "isl-miptree",
603 isl_tiling_to_i915_tiling(
605 mt
->surf
.row_pitch
, alloc_flags
);
612 mt
->first_level
= first_level
;
613 mt
->last_level
= last_level
;
616 mt
->aux_state
= NULL
;
617 mt
->cpp
= isl_format_get_layout(mt
->surf
.format
)->bpb
/ 8;
618 mt
->compressed
= _mesa_is_format_compressed(format
);
619 mt
->drm_modifier
= DRM_FORMAT_MOD_INVALID
;
624 intel_miptree_release(&mt
);
629 make_separate_stencil_surface(struct brw_context
*brw
,
630 struct intel_mipmap_tree
*mt
)
632 mt
->stencil_mt
= make_surface(brw
, mt
->target
, MESA_FORMAT_S_UINT8
,
633 0, mt
->surf
.levels
- 1,
634 mt
->surf
.logical_level0_px
.width
,
635 mt
->surf
.logical_level0_px
.height
,
636 mt
->surf
.dim
== ISL_SURF_DIM_3D
?
637 mt
->surf
.logical_level0_px
.depth
:
638 mt
->surf
.logical_level0_px
.array_len
,
639 mt
->surf
.samples
, ISL_TILING_W_BIT
,
640 ISL_SURF_USAGE_STENCIL_BIT
|
641 ISL_SURF_USAGE_TEXTURE_BIT
,
642 BO_ALLOC_BUSY
, 0, NULL
);
647 mt
->stencil_mt
->r8stencil_needs_update
= true;
652 static struct intel_mipmap_tree
*
653 miptree_create(struct brw_context
*brw
,
662 enum intel_miptree_create_flags flags
)
664 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
666 if (format
== MESA_FORMAT_S_UINT8
)
667 return make_surface(brw
, target
, format
, first_level
, last_level
,
668 width0
, height0
, depth0
, num_samples
,
670 ISL_SURF_USAGE_STENCIL_BIT
|
671 ISL_SURF_USAGE_TEXTURE_BIT
,
676 const GLenum base_format
= _mesa_get_format_base_format(format
);
677 if ((base_format
== GL_DEPTH_COMPONENT
||
678 base_format
== GL_DEPTH_STENCIL
) &&
679 !(flags
& MIPTREE_CREATE_LINEAR
)) {
680 /* Fix up the Z miptree format for how we're splitting out separate
681 * stencil. Gen7 expects there to be no stencil bits in its depth buffer.
683 const mesa_format depth_only_format
=
684 intel_depth_format_for_depthstencil_format(format
);
685 struct intel_mipmap_tree
*mt
= make_surface(
686 brw
, target
, devinfo
->gen
>= 6 ? depth_only_format
: format
,
687 first_level
, last_level
,
688 width0
, height0
, depth0
, num_samples
, ISL_TILING_Y0_BIT
,
689 ISL_SURF_USAGE_DEPTH_BIT
| ISL_SURF_USAGE_TEXTURE_BIT
,
690 BO_ALLOC_BUSY
, 0, NULL
);
692 if (needs_separate_stencil(brw
, mt
, format
) &&
693 !make_separate_stencil_surface(brw
, mt
)) {
694 intel_miptree_release(&mt
);
698 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
699 intel_miptree_choose_aux_usage(brw
, mt
);
704 mesa_format tex_format
= format
;
705 mesa_format etc_format
= MESA_FORMAT_NONE
;
706 uint32_t alloc_flags
= 0;
708 format
= intel_lower_compressed_format(brw
, format
);
710 etc_format
= (format
!= tex_format
) ? tex_format
: MESA_FORMAT_NONE
;
712 if (flags
& MIPTREE_CREATE_BUSY
)
713 alloc_flags
|= BO_ALLOC_BUSY
;
715 isl_tiling_flags_t tiling_flags
= (flags
& MIPTREE_CREATE_LINEAR
) ?
716 ISL_TILING_LINEAR_BIT
: ISL_TILING_ANY_MASK
;
718 /* TODO: This used to be because there wasn't BLORP to handle Y-tiling. */
719 if (devinfo
->gen
< 6)
720 tiling_flags
&= ~ISL_TILING_Y0_BIT
;
722 struct intel_mipmap_tree
*mt
= make_surface(
724 first_level
, last_level
,
725 width0
, height0
, depth0
,
726 num_samples
, tiling_flags
,
727 ISL_SURF_USAGE_RENDER_TARGET_BIT
|
728 ISL_SURF_USAGE_TEXTURE_BIT
,
729 alloc_flags
, 0, NULL
);
733 mt
->etc_format
= etc_format
;
735 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
736 intel_miptree_choose_aux_usage(brw
, mt
);
741 struct intel_mipmap_tree
*
742 intel_miptree_create(struct brw_context
*brw
,
751 enum intel_miptree_create_flags flags
)
753 assert(num_samples
> 0);
755 struct intel_mipmap_tree
*mt
= miptree_create(
757 first_level
, last_level
,
758 width0
, height0
, depth0
, num_samples
,
765 if (!intel_miptree_alloc_aux(brw
, mt
)) {
766 intel_miptree_release(&mt
);
773 struct intel_mipmap_tree
*
774 intel_miptree_create_for_bo(struct brw_context
*brw
,
782 enum intel_miptree_create_flags flags
)
784 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
785 struct intel_mipmap_tree
*mt
;
786 uint32_t tiling
, swizzle
;
787 const GLenum target
= depth
> 1 ? GL_TEXTURE_2D_ARRAY
: GL_TEXTURE_2D
;
788 const GLenum base_format
= _mesa_get_format_base_format(format
);
790 if ((base_format
== GL_DEPTH_COMPONENT
||
791 base_format
== GL_DEPTH_STENCIL
)) {
792 const mesa_format depth_only_format
=
793 intel_depth_format_for_depthstencil_format(format
);
794 mt
= make_surface(brw
, target
,
795 devinfo
->gen
>= 6 ? depth_only_format
: format
,
796 0, 0, width
, height
, depth
, 1, ISL_TILING_Y0_BIT
,
797 ISL_SURF_USAGE_DEPTH_BIT
| ISL_SURF_USAGE_TEXTURE_BIT
,
798 BO_ALLOC_BUSY
, pitch
, bo
);
802 brw_bo_reference(bo
);
804 if (!(flags
& MIPTREE_CREATE_NO_AUX
))
805 intel_miptree_choose_aux_usage(brw
, mt
);
808 } else if (format
== MESA_FORMAT_S_UINT8
) {
809 mt
= make_surface(brw
, target
, MESA_FORMAT_S_UINT8
,
810 0, 0, width
, height
, depth
, 1,
812 ISL_SURF_USAGE_STENCIL_BIT
|
813 ISL_SURF_USAGE_TEXTURE_BIT
,
814 BO_ALLOC_BUSY
, pitch
, bo
);
818 assert(bo
->size
>= mt
->surf
.size
);
820 brw_bo_reference(bo
);
824 brw_bo_get_tiling(bo
, &tiling
, &swizzle
);
826 /* Nothing will be able to use this miptree with the BO if the offset isn't
829 if (tiling
!= I915_TILING_NONE
)
830 assert(offset
% 4096 == 0);
832 /* miptrees can't handle negative pitch. If you need flipping of images,
833 * that's outside of the scope of the mt.
837 /* The BO already has a tiling format and we shouldn't confuse the lower
838 * layers by making it try to find a tiling format again.
840 assert((flags
& MIPTREE_CREATE_LINEAR
) == 0);
842 mt
= make_surface(brw
, target
, format
,
843 0, 0, width
, height
, depth
, 1,
844 1lu << isl_tiling_from_i915_tiling(tiling
),
845 ISL_SURF_USAGE_RENDER_TARGET_BIT
|
846 ISL_SURF_USAGE_TEXTURE_BIT
,
851 brw_bo_reference(bo
);
855 if (!(flags
& MIPTREE_CREATE_NO_AUX
)) {
856 intel_miptree_choose_aux_usage(brw
, mt
);
858 if (!intel_miptree_alloc_aux(brw
, mt
)) {
859 intel_miptree_release(&mt
);
867 static struct intel_mipmap_tree
*
868 miptree_create_for_planar_image(struct brw_context
*brw
,
869 __DRIimage
*image
, GLenum target
)
871 const struct intel_image_format
*f
= image
->planar_format
;
872 struct intel_mipmap_tree
*planar_mt
= NULL
;
874 for (int i
= 0; i
< f
->nplanes
; i
++) {
875 const int index
= f
->planes
[i
].buffer_index
;
876 const uint32_t dri_format
= f
->planes
[i
].dri_format
;
877 const mesa_format format
= driImageFormatToGLFormat(dri_format
);
878 const uint32_t width
= image
->width
>> f
->planes
[i
].width_shift
;
879 const uint32_t height
= image
->height
>> f
->planes
[i
].height_shift
;
881 /* Disable creation of the texture's aux buffers because the driver
882 * exposes no EGL API to manage them. That is, there is no API for
883 * resolving the aux buffer's content to the main buffer nor for
884 * invalidating the aux buffer's content.
886 struct intel_mipmap_tree
*mt
=
887 intel_miptree_create_for_bo(brw
, image
->bo
, format
,
888 image
->offsets
[index
],
890 image
->strides
[index
],
891 MIPTREE_CREATE_NO_AUX
);
900 planar_mt
->plane
[i
- 1] = mt
;
903 planar_mt
->drm_modifier
= image
->modifier
;
909 create_ccs_buf_for_image(struct brw_context
*brw
,
911 struct intel_mipmap_tree
*mt
,
912 enum isl_aux_state initial_state
)
914 struct isl_surf temp_ccs_surf
;
916 /* CCS is only supported for very simple miptrees */
917 assert(image
->aux_offset
!= 0 && image
->aux_pitch
!= 0);
918 assert(image
->tile_x
== 0 && image
->tile_y
== 0);
919 assert(mt
->surf
.samples
== 1);
920 assert(mt
->surf
.levels
== 1);
921 assert(mt
->surf
.logical_level0_px
.depth
== 1);
922 assert(mt
->surf
.logical_level0_px
.array_len
== 1);
923 assert(mt
->first_level
== 0);
924 assert(mt
->last_level
== 0);
926 /* We shouldn't already have a CCS */
927 assert(!mt
->mcs_buf
);
929 if (!isl_surf_get_ccs_surf(&brw
->isl_dev
, &mt
->surf
, &temp_ccs_surf
,
933 assert(image
->aux_offset
< image
->bo
->size
);
934 assert(temp_ccs_surf
.size
<= image
->bo
->size
- image
->aux_offset
);
936 mt
->mcs_buf
= calloc(sizeof(*mt
->mcs_buf
), 1);
937 if (mt
->mcs_buf
== NULL
)
940 mt
->aux_state
= create_aux_state_map(mt
, initial_state
);
941 if (!mt
->aux_state
) {
947 mt
->mcs_buf
->bo
= image
->bo
;
948 brw_bo_reference(image
->bo
);
950 mt
->mcs_buf
->offset
= image
->aux_offset
;
951 mt
->mcs_buf
->size
= image
->bo
->size
- image
->aux_offset
;
952 mt
->mcs_buf
->pitch
= image
->aux_pitch
;
953 mt
->mcs_buf
->qpitch
= 0;
954 mt
->mcs_buf
->surf
= temp_ccs_surf
;
959 struct intel_mipmap_tree
*
960 intel_miptree_create_for_dri_image(struct brw_context
*brw
,
961 __DRIimage
*image
, GLenum target
,
963 bool is_winsys_image
)
965 if (image
->planar_format
&& image
->planar_format
->nplanes
> 1)
966 return miptree_create_for_planar_image(brw
, image
, target
);
968 if (image
->planar_format
)
969 assert(image
->planar_format
->planes
[0].dri_format
== image
->dri_format
);
971 if (!brw
->ctx
.TextureFormatSupported
[format
]) {
972 /* The texture storage paths in core Mesa detect if the driver does not
973 * support the user-requested format, and then searches for a
974 * fallback format. The DRIimage code bypasses core Mesa, though. So we
975 * do the fallbacks here for important formats.
977 * We must support DRM_FOURCC_XBGR8888 textures because the Android
978 * framework produces HAL_PIXEL_FORMAT_RGBX8888 winsys surfaces, which
979 * the Chrome OS compositor consumes as dma_buf EGLImages.
981 format
= _mesa_format_fallback_rgbx_to_rgba(format
);
984 if (!brw
->ctx
.TextureFormatSupported
[format
])
987 const struct isl_drm_modifier_info
*mod_info
=
988 isl_drm_modifier_get_info(image
->modifier
);
990 enum intel_miptree_create_flags mt_create_flags
= 0;
992 /* If this image comes in from a window system, we have different
993 * requirements than if it comes in via an EGL import operation. Window
994 * system images can use any form of auxiliary compression we wish because
995 * they get "flushed" before being handed off to the window system and we
996 * have the opportunity to do resolves. Non window-system images, on the
997 * other hand, have no resolve point so we can't have aux without a
1000 if (!is_winsys_image
)
1001 mt_create_flags
|= MIPTREE_CREATE_NO_AUX
;
1003 /* If we have a modifier which specifies aux, don't create one yet */
1004 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
)
1005 mt_create_flags
|= MIPTREE_CREATE_NO_AUX
;
1007 /* Disable creation of the texture's aux buffers because the driver exposes
1008 * no EGL API to manage them. That is, there is no API for resolving the aux
1009 * buffer's content to the main buffer nor for invalidating the aux buffer's
1012 struct intel_mipmap_tree
*mt
=
1013 intel_miptree_create_for_bo(brw
, image
->bo
, format
,
1014 image
->offset
, image
->width
, image
->height
, 1,
1015 image
->pitch
, mt_create_flags
);
1019 mt
->target
= target
;
1020 mt
->level
[0].level_x
= image
->tile_x
;
1021 mt
->level
[0].level_y
= image
->tile_y
;
1022 mt
->drm_modifier
= image
->modifier
;
1024 /* From "OES_EGL_image" error reporting. We report GL_INVALID_OPERATION
1025 * for EGL images from non-tile aligned sufaces in gen4 hw and earlier which has
1026 * trouble resolving back to destination image due to alignment issues.
1028 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1029 if (!devinfo
->has_surface_tile_offset
) {
1030 uint32_t draw_x
, draw_y
;
1031 intel_miptree_get_tile_offsets(mt
, 0, 0, &draw_x
, &draw_y
);
1033 if (draw_x
!= 0 || draw_y
!= 0) {
1034 _mesa_error(&brw
->ctx
, GL_INVALID_OPERATION
, __func__
);
1035 intel_miptree_release(&mt
);
1040 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
1041 assert(mod_info
->aux_usage
== ISL_AUX_USAGE_CCS_E
);
1043 mt
->aux_usage
= mod_info
->aux_usage
;
1044 /* If we are a window system buffer, then we can support fast-clears
1045 * even if the modifier doesn't support them by doing a partial resolve
1046 * as part of the flush operation.
1048 mt
->supports_fast_clear
=
1049 is_winsys_image
|| mod_info
->supports_clear_color
;
1051 /* We don't know the actual state of the surface when we get it but we
1052 * can make a pretty good guess based on the modifier. What we do know
1053 * for sure is that it isn't in the AUX_INVALID state, so we just assume
1054 * a worst case of compression.
1056 enum isl_aux_state initial_state
=
1057 isl_drm_modifier_get_default_aux_state(image
->modifier
);
1059 if (!create_ccs_buf_for_image(brw
, image
, mt
, initial_state
)) {
1060 intel_miptree_release(&mt
);
1065 /* Don't assume coherency for imported EGLimages. We don't know what
1066 * external clients are going to do with it. They may scan it out.
1068 image
->bo
->cache_coherent
= false;
1074 * For a singlesample renderbuffer, this simply wraps the given BO with a
1077 * For a multisample renderbuffer, this wraps the window system's
1078 * (singlesample) BO with a singlesample miptree attached to the
1079 * intel_renderbuffer, then creates a multisample miptree attached to irb->mt
1080 * that will contain the actual rendering (which is lazily resolved to
1081 * irb->singlesample_mt).
1084 intel_update_winsys_renderbuffer_miptree(struct brw_context
*intel
,
1085 struct intel_renderbuffer
*irb
,
1086 struct intel_mipmap_tree
*singlesample_mt
,
1087 uint32_t width
, uint32_t height
,
1090 struct intel_mipmap_tree
*multisample_mt
= NULL
;
1091 struct gl_renderbuffer
*rb
= &irb
->Base
.Base
;
1092 mesa_format format
= rb
->Format
;
1093 const unsigned num_samples
= MAX2(rb
->NumSamples
, 1);
1095 /* Only the front and back buffers, which are color buffers, are allocated
1096 * through the image loader.
1098 assert(_mesa_get_format_base_format(format
) == GL_RGB
||
1099 _mesa_get_format_base_format(format
) == GL_RGBA
);
1101 assert(singlesample_mt
);
1103 if (num_samples
== 1) {
1104 intel_miptree_release(&irb
->mt
);
1105 irb
->mt
= singlesample_mt
;
1107 assert(!irb
->singlesample_mt
);
1109 intel_miptree_release(&irb
->singlesample_mt
);
1110 irb
->singlesample_mt
= singlesample_mt
;
1113 irb
->mt
->surf
.logical_level0_px
.width
!= width
||
1114 irb
->mt
->surf
.logical_level0_px
.height
!= height
) {
1115 multisample_mt
= intel_miptree_create_for_renderbuffer(intel
,
1120 if (!multisample_mt
)
1123 irb
->need_downsample
= false;
1124 intel_miptree_release(&irb
->mt
);
1125 irb
->mt
= multisample_mt
;
1131 intel_miptree_release(&irb
->mt
);
1135 struct intel_mipmap_tree
*
1136 intel_miptree_create_for_renderbuffer(struct brw_context
*brw
,
1140 uint32_t num_samples
)
1142 struct intel_mipmap_tree
*mt
;
1144 GLenum target
= num_samples
> 1 ? GL_TEXTURE_2D_MULTISAMPLE
: GL_TEXTURE_2D
;
1146 mt
= intel_miptree_create(brw
, target
, format
, 0, 0,
1147 width
, height
, depth
, num_samples
,
1148 MIPTREE_CREATE_BUSY
);
1155 intel_miptree_release(&mt
);
1160 intel_miptree_reference(struct intel_mipmap_tree
**dst
,
1161 struct intel_mipmap_tree
*src
)
1166 intel_miptree_release(dst
);
1170 DBG("%s %p refcount now %d\n", __func__
, src
, src
->refcount
);
1177 intel_miptree_aux_buffer_free(struct intel_miptree_aux_buffer
*aux_buf
)
1179 if (aux_buf
== NULL
)
1182 brw_bo_unreference(aux_buf
->bo
);
1188 intel_miptree_release(struct intel_mipmap_tree
**mt
)
1193 DBG("%s %p refcount will be %d\n", __func__
, *mt
, (*mt
)->refcount
- 1);
1194 if (--(*mt
)->refcount
<= 0) {
1197 DBG("%s deleting %p\n", __func__
, *mt
);
1199 brw_bo_unreference((*mt
)->bo
);
1200 intel_miptree_release(&(*mt
)->stencil_mt
);
1201 intel_miptree_release(&(*mt
)->r8stencil_mt
);
1202 intel_miptree_aux_buffer_free((*mt
)->hiz_buf
);
1203 intel_miptree_aux_buffer_free((*mt
)->mcs_buf
);
1204 free_aux_state_map((*mt
)->aux_state
);
1206 intel_miptree_release(&(*mt
)->plane
[0]);
1207 intel_miptree_release(&(*mt
)->plane
[1]);
1209 for (i
= 0; i
< MAX_TEXTURE_LEVELS
; i
++) {
1210 free((*mt
)->level
[i
].slice
);
1220 intel_get_image_dims(struct gl_texture_image
*image
,
1221 int *width
, int *height
, int *depth
)
1223 switch (image
->TexObject
->Target
) {
1224 case GL_TEXTURE_1D_ARRAY
:
1225 /* For a 1D Array texture the OpenGL API will treat the image height as
1226 * the number of array slices. For Intel hardware, we treat the 1D array
1227 * as a 2D Array with a height of 1. So, here we want to swap image
1230 assert(image
->Depth
== 1);
1231 *width
= image
->Width
;
1233 *depth
= image
->Height
;
1235 case GL_TEXTURE_CUBE_MAP
:
1236 /* For Cube maps, the mesa/main api layer gives us a depth of 1 even
1237 * though we really have 6 slices.
1239 assert(image
->Depth
== 1);
1240 *width
= image
->Width
;
1241 *height
= image
->Height
;
1245 *width
= image
->Width
;
1246 *height
= image
->Height
;
1247 *depth
= image
->Depth
;
1253 * Can the image be pulled into a unified mipmap tree? This mirrors
1254 * the completeness test in a lot of ways.
1256 * Not sure whether I want to pass gl_texture_image here.
1259 intel_miptree_match_image(struct intel_mipmap_tree
*mt
,
1260 struct gl_texture_image
*image
)
1262 struct intel_texture_image
*intelImage
= intel_texture_image(image
);
1263 GLuint level
= intelImage
->base
.Base
.Level
;
1264 int width
, height
, depth
;
1266 /* glTexImage* choose the texture object based on the target passed in, and
1267 * objects can't change targets over their lifetimes, so this should be
1270 assert(image
->TexObject
->Target
== mt
->target
);
1272 mesa_format mt_format
= mt
->format
;
1273 if (mt
->format
== MESA_FORMAT_Z24_UNORM_X8_UINT
&& mt
->stencil_mt
)
1274 mt_format
= MESA_FORMAT_Z24_UNORM_S8_UINT
;
1275 if (mt
->format
== MESA_FORMAT_Z_FLOAT32
&& mt
->stencil_mt
)
1276 mt_format
= MESA_FORMAT_Z32_FLOAT_S8X24_UINT
;
1277 if (mt
->etc_format
!= MESA_FORMAT_NONE
)
1278 mt_format
= mt
->etc_format
;
1280 if (image
->TexFormat
!= mt_format
)
1283 intel_get_image_dims(image
, &width
, &height
, &depth
);
1285 if (mt
->target
== GL_TEXTURE_CUBE_MAP
)
1288 if (level
>= mt
->surf
.levels
)
1291 const unsigned level_depth
=
1292 mt
->surf
.dim
== ISL_SURF_DIM_3D
?
1293 minify(mt
->surf
.logical_level0_px
.depth
, level
) :
1294 mt
->surf
.logical_level0_px
.array_len
;
1296 return width
== minify(mt
->surf
.logical_level0_px
.width
, level
) &&
1297 height
== minify(mt
->surf
.logical_level0_px
.height
, level
) &&
1298 depth
== level_depth
&&
1299 MAX2(image
->NumSamples
, 1) == mt
->surf
.samples
;
1303 intel_miptree_get_image_offset(const struct intel_mipmap_tree
*mt
,
1304 GLuint level
, GLuint slice
,
1305 GLuint
*x
, GLuint
*y
)
1307 if (level
== 0 && slice
== 0) {
1308 *x
= mt
->level
[0].level_x
;
1309 *y
= mt
->level
[0].level_y
;
1313 uint32_t x_offset_sa
, y_offset_sa
;
1315 /* Miptree itself can have an offset only if it represents a single
1316 * slice in an imported buffer object.
1317 * See intel_miptree_create_for_dri_image().
1319 assert(mt
->level
[0].level_x
== 0);
1320 assert(mt
->level
[0].level_y
== 0);
1322 /* Given level is relative to level zero while the miptree may be
1323 * represent just a subset of all levels starting from 'first_level'.
1325 assert(level
>= mt
->first_level
);
1326 level
-= mt
->first_level
;
1328 const unsigned z
= mt
->surf
.dim
== ISL_SURF_DIM_3D
? slice
: 0;
1329 slice
= mt
->surf
.dim
== ISL_SURF_DIM_3D
? 0 : slice
;
1330 isl_surf_get_image_offset_el(&mt
->surf
, level
, slice
, z
,
1331 &x_offset_sa
, &y_offset_sa
);
1339 * This function computes the tile_w (in bytes) and tile_h (in rows) of
1340 * different tiling patterns. If the BO is untiled, tile_w is set to cpp
1341 * and tile_h is set to 1.
1344 intel_get_tile_dims(enum isl_tiling tiling
, uint32_t cpp
,
1345 uint32_t *tile_w
, uint32_t *tile_h
)
1356 case ISL_TILING_LINEAR
:
1361 unreachable("not reached");
1367 * This function computes masks that may be used to select the bits of the X
1368 * and Y coordinates that indicate the offset within a tile. If the BO is
1369 * untiled, the masks are set to 0.
1372 intel_get_tile_masks(enum isl_tiling tiling
, uint32_t cpp
,
1373 uint32_t *mask_x
, uint32_t *mask_y
)
1375 uint32_t tile_w_bytes
, tile_h
;
1377 intel_get_tile_dims(tiling
, cpp
, &tile_w_bytes
, &tile_h
);
1379 *mask_x
= tile_w_bytes
/ cpp
- 1;
1380 *mask_y
= tile_h
- 1;
1384 * Compute the offset (in bytes) from the start of the BO to the given x
1385 * and y coordinate. For tiled BOs, caller must ensure that x and y are
1386 * multiples of the tile size.
1389 intel_miptree_get_aligned_offset(const struct intel_mipmap_tree
*mt
,
1390 uint32_t x
, uint32_t y
)
1393 uint32_t pitch
= mt
->surf
.row_pitch
;
1395 switch (mt
->surf
.tiling
) {
1397 unreachable("not reached");
1398 case ISL_TILING_LINEAR
:
1399 return y
* pitch
+ x
* cpp
;
1401 assert((x
% (512 / cpp
)) == 0);
1402 assert((y
% 8) == 0);
1403 return y
* pitch
+ x
/ (512 / cpp
) * 4096;
1405 assert((x
% (128 / cpp
)) == 0);
1406 assert((y
% 32) == 0);
1407 return y
* pitch
+ x
/ (128 / cpp
) * 4096;
1412 * Rendering with tiled buffers requires that the base address of the buffer
1413 * be aligned to a page boundary. For renderbuffers, and sometimes with
1414 * textures, we may want the surface to point at a texture image level that
1415 * isn't at a page boundary.
1417 * This function returns an appropriately-aligned base offset
1418 * according to the tiling restrictions, plus any required x/y offset
1422 intel_miptree_get_tile_offsets(const struct intel_mipmap_tree
*mt
,
1423 GLuint level
, GLuint slice
,
1428 uint32_t mask_x
, mask_y
;
1430 intel_get_tile_masks(mt
->surf
.tiling
, mt
->cpp
, &mask_x
, &mask_y
);
1431 intel_miptree_get_image_offset(mt
, level
, slice
, &x
, &y
);
1433 *tile_x
= x
& mask_x
;
1434 *tile_y
= y
& mask_y
;
1436 return intel_miptree_get_aligned_offset(mt
, x
& ~mask_x
, y
& ~mask_y
);
1440 intel_miptree_copy_slice_sw(struct brw_context
*brw
,
1441 struct intel_mipmap_tree
*src_mt
,
1442 unsigned src_level
, unsigned src_layer
,
1443 struct intel_mipmap_tree
*dst_mt
,
1444 unsigned dst_level
, unsigned dst_layer
,
1445 unsigned width
, unsigned height
)
1448 ptrdiff_t src_stride
, dst_stride
;
1449 const unsigned cpp
= (isl_format_get_layout(dst_mt
->surf
.format
)->bpb
/ 8);
1451 intel_miptree_map(brw
, src_mt
,
1452 src_level
, src_layer
,
1455 GL_MAP_READ_BIT
| BRW_MAP_DIRECT_BIT
,
1458 intel_miptree_map(brw
, dst_mt
,
1459 dst_level
, dst_layer
,
1462 GL_MAP_WRITE_BIT
| GL_MAP_INVALIDATE_RANGE_BIT
|
1466 DBG("sw blit %s mt %p %p/%"PRIdPTR
" -> %s mt %p %p/%"PRIdPTR
" (%dx%d)\n",
1467 _mesa_get_format_name(src_mt
->format
),
1468 src_mt
, src
, src_stride
,
1469 _mesa_get_format_name(dst_mt
->format
),
1470 dst_mt
, dst
, dst_stride
,
1473 int row_size
= cpp
* width
;
1474 if (src_stride
== row_size
&&
1475 dst_stride
== row_size
) {
1476 memcpy(dst
, src
, row_size
* height
);
1478 for (int i
= 0; i
< height
; i
++) {
1479 memcpy(dst
, src
, row_size
);
1485 intel_miptree_unmap(brw
, dst_mt
, dst_level
, dst_layer
);
1486 intel_miptree_unmap(brw
, src_mt
, src_level
, src_layer
);
1488 /* Don't forget to copy the stencil data over, too. We could have skipped
1489 * passing BRW_MAP_DIRECT_BIT, but that would have meant intel_miptree_map
1490 * shuffling the two data sources in/out of temporary storage instead of
1491 * the direct mapping we get this way.
1493 if (dst_mt
->stencil_mt
) {
1494 assert(src_mt
->stencil_mt
);
1495 intel_miptree_copy_slice_sw(brw
,
1496 src_mt
->stencil_mt
, src_level
, src_layer
,
1497 dst_mt
->stencil_mt
, dst_level
, dst_layer
,
1503 intel_miptree_copy_slice(struct brw_context
*brw
,
1504 struct intel_mipmap_tree
*src_mt
,
1505 unsigned src_level
, unsigned src_layer
,
1506 struct intel_mipmap_tree
*dst_mt
,
1507 unsigned dst_level
, unsigned dst_layer
)
1510 mesa_format format
= src_mt
->format
;
1511 unsigned width
= minify(src_mt
->surf
.phys_level0_sa
.width
,
1512 src_level
- src_mt
->first_level
);
1513 unsigned height
= minify(src_mt
->surf
.phys_level0_sa
.height
,
1514 src_level
- src_mt
->first_level
);
1516 assert(src_layer
< get_num_phys_layers(&src_mt
->surf
,
1517 src_level
- src_mt
->first_level
));
1519 assert(src_mt
->format
== dst_mt
->format
);
1521 if (dst_mt
->compressed
) {
1523 _mesa_get_format_block_size(dst_mt
->format
, &i
, &j
);
1524 height
= ALIGN_NPOT(height
, j
) / j
;
1525 width
= ALIGN_NPOT(width
, i
) / i
;
1528 /* If it's a packed depth/stencil buffer with separate stencil, the blit
1529 * below won't apply since we can't do the depth's Y tiling or the
1530 * stencil's W tiling in the blitter.
1532 if (src_mt
->stencil_mt
) {
1533 intel_miptree_copy_slice_sw(brw
,
1534 src_mt
, src_level
, src_layer
,
1535 dst_mt
, dst_level
, dst_layer
,
1540 uint32_t dst_x
, dst_y
, src_x
, src_y
;
1541 intel_miptree_get_image_offset(dst_mt
, dst_level
, dst_layer
,
1543 intel_miptree_get_image_offset(src_mt
, src_level
, src_layer
,
1546 DBG("validate blit mt %s %p %d,%d/%d -> mt %s %p %d,%d/%d (%dx%d)\n",
1547 _mesa_get_format_name(src_mt
->format
),
1548 src_mt
, src_x
, src_y
, src_mt
->surf
.row_pitch
,
1549 _mesa_get_format_name(dst_mt
->format
),
1550 dst_mt
, dst_x
, dst_y
, dst_mt
->surf
.row_pitch
,
1553 if (!intel_miptree_blit(brw
,
1554 src_mt
, src_level
, src_layer
, 0, 0, false,
1555 dst_mt
, dst_level
, dst_layer
, 0, 0, false,
1556 width
, height
, GL_COPY
)) {
1557 perf_debug("miptree validate blit for %s failed\n",
1558 _mesa_get_format_name(format
));
1560 intel_miptree_copy_slice_sw(brw
,
1561 src_mt
, src_level
, src_layer
,
1562 dst_mt
, dst_level
, dst_layer
,
1568 * Copies the image's current data to the given miptree, and associates that
1569 * miptree with the image.
1572 intel_miptree_copy_teximage(struct brw_context
*brw
,
1573 struct intel_texture_image
*intelImage
,
1574 struct intel_mipmap_tree
*dst_mt
)
1576 struct intel_mipmap_tree
*src_mt
= intelImage
->mt
;
1577 struct intel_texture_object
*intel_obj
=
1578 intel_texture_object(intelImage
->base
.Base
.TexObject
);
1579 int level
= intelImage
->base
.Base
.Level
;
1580 const unsigned face
= intelImage
->base
.Base
.Face
;
1581 unsigned start_layer
, end_layer
;
1583 if (intel_obj
->base
.Target
== GL_TEXTURE_1D_ARRAY
) {
1585 assert(intelImage
->base
.Base
.Height
);
1587 end_layer
= intelImage
->base
.Base
.Height
- 1;
1588 } else if (face
> 0) {
1592 assert(intelImage
->base
.Base
.Depth
);
1594 end_layer
= intelImage
->base
.Base
.Depth
- 1;
1597 for (unsigned i
= start_layer
; i
<= end_layer
; i
++) {
1598 intel_miptree_copy_slice(brw
,
1603 intel_miptree_reference(&intelImage
->mt
, dst_mt
);
1604 intel_obj
->needs_validate
= true;
1608 intel_miptree_init_mcs(struct brw_context
*brw
,
1609 struct intel_mipmap_tree
*mt
,
1612 assert(mt
->mcs_buf
!= NULL
);
1614 /* From the Ivy Bridge PRM, Vol 2 Part 1 p326:
1616 * When MCS buffer is enabled and bound to MSRT, it is required that it
1617 * is cleared prior to any rendering.
1619 * Since we don't use the MCS buffer for any purpose other than rendering,
1620 * it makes sense to just clear it immediately upon allocation.
1622 * Note: the clear value for MCS buffers is all 1's, so we memset to 0xff.
1624 void *map
= brw_bo_map(brw
, mt
->mcs_buf
->bo
, MAP_WRITE
);
1625 if (unlikely(map
== NULL
)) {
1626 fprintf(stderr
, "Failed to map mcs buffer into GTT\n");
1627 brw_bo_unreference(mt
->mcs_buf
->bo
);
1632 memset(data
, init_value
, mt
->mcs_buf
->size
);
1633 brw_bo_unmap(mt
->mcs_buf
->bo
);
1636 static struct intel_miptree_aux_buffer
*
1637 intel_alloc_aux_buffer(struct brw_context
*brw
,
1639 const struct isl_surf
*aux_surf
,
1640 uint32_t alloc_flags
,
1641 struct intel_mipmap_tree
*mt
)
1643 struct intel_miptree_aux_buffer
*buf
= calloc(sizeof(*buf
), 1);
1647 buf
->size
= aux_surf
->size
;
1648 buf
->pitch
= aux_surf
->row_pitch
;
1649 buf
->qpitch
= isl_surf_get_array_pitch_sa_rows(aux_surf
);
1651 /* ISL has stricter set of alignment rules then the drm allocator.
1652 * Therefore one can pass the ISL dimensions in terms of bytes instead of
1653 * trying to recalculate based on different format block sizes.
1655 buf
->bo
= brw_bo_alloc_tiled(brw
->bufmgr
, name
, buf
->size
,
1656 I915_TILING_Y
, buf
->pitch
, alloc_flags
);
1662 buf
->surf
= *aux_surf
;
1668 intel_miptree_alloc_mcs(struct brw_context
*brw
,
1669 struct intel_mipmap_tree
*mt
,
1672 assert(brw
->screen
->devinfo
.gen
>= 7); /* MCS only used on Gen7+ */
1673 assert(mt
->mcs_buf
== NULL
);
1674 assert(mt
->aux_usage
== ISL_AUX_USAGE_MCS
);
1676 /* Multisampled miptrees are only supported for single level. */
1677 assert(mt
->first_level
== 0);
1678 enum isl_aux_state
**aux_state
=
1679 create_aux_state_map(mt
, ISL_AUX_STATE_CLEAR
);
1683 struct isl_surf temp_mcs_surf
;
1685 MAYBE_UNUSED
bool ok
=
1686 isl_surf_get_mcs_surf(&brw
->isl_dev
, &mt
->surf
, &temp_mcs_surf
);
1689 /* Buffer needs to be initialised requiring the buffer to be immediately
1690 * mapped to cpu space for writing. Therefore do not use the gpu access
1691 * flag which can cause an unnecessary delay if the backing pages happened
1692 * to be just used by the GPU.
1694 const uint32_t alloc_flags
= 0;
1695 mt
->mcs_buf
= intel_alloc_aux_buffer(brw
, "mcs-miptree",
1696 &temp_mcs_surf
, alloc_flags
, mt
);
1702 mt
->aux_state
= aux_state
;
1704 intel_miptree_init_mcs(brw
, mt
, 0xFF);
1710 intel_miptree_alloc_ccs(struct brw_context
*brw
,
1711 struct intel_mipmap_tree
*mt
)
1713 assert(mt
->mcs_buf
== NULL
);
1714 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
||
1715 mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
1717 struct isl_surf temp_ccs_surf
;
1719 if (!isl_surf_get_ccs_surf(&brw
->isl_dev
, &mt
->surf
, &temp_ccs_surf
, 0))
1722 assert(temp_ccs_surf
.size
&&
1723 (temp_ccs_surf
.size
% temp_ccs_surf
.row_pitch
== 0));
1725 enum isl_aux_state
**aux_state
=
1726 create_aux_state_map(mt
, ISL_AUX_STATE_PASS_THROUGH
);
1730 /* When CCS_E is used, we need to ensure that the CCS starts off in a valid
1731 * state. From the Sky Lake PRM, "MCS Buffer for Render Target(s)":
1733 * "If Software wants to enable Color Compression without Fast clear,
1734 * Software needs to initialize MCS with zeros."
1736 * A CCS value of 0 indicates that the corresponding block is in the
1737 * pass-through state which is what we want.
1739 * For CCS_D, on the other hand, we don't care as we're about to perform a
1740 * fast-clear operation. In that case, being hot in caches more useful.
1742 const uint32_t alloc_flags
= mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
?
1743 BO_ALLOC_ZEROED
: BO_ALLOC_BUSY
;
1744 mt
->mcs_buf
= intel_alloc_aux_buffer(brw
, "ccs-miptree",
1745 &temp_ccs_surf
, alloc_flags
, mt
);
1751 mt
->aux_state
= aux_state
;
1757 * Helper for intel_miptree_alloc_hiz() that sets
1758 * \c mt->level[level].has_hiz. Return true if and only if
1759 * \c has_hiz was set.
1762 intel_miptree_level_enable_hiz(struct brw_context
*brw
,
1763 struct intel_mipmap_tree
*mt
,
1766 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1768 assert(mt
->hiz_buf
);
1769 assert(mt
->surf
.size
> 0);
1771 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
1772 uint32_t width
= minify(mt
->surf
.phys_level0_sa
.width
, level
);
1773 uint32_t height
= minify(mt
->surf
.phys_level0_sa
.height
, level
);
1775 /* Disable HiZ for LOD > 0 unless the width is 8 aligned
1776 * and the height is 4 aligned. This allows our HiZ support
1777 * to fulfill Haswell restrictions for HiZ ops. For LOD == 0,
1778 * we can grow the width & height to allow the HiZ op to
1779 * force the proper size alignments.
1781 if (level
> 0 && ((width
& 7) || (height
& 3))) {
1782 DBG("mt %p level %d: HiZ DISABLED\n", mt
, level
);
1787 DBG("mt %p level %d: HiZ enabled\n", mt
, level
);
1788 mt
->level
[level
].has_hiz
= true;
1793 intel_miptree_alloc_hiz(struct brw_context
*brw
,
1794 struct intel_mipmap_tree
*mt
)
1796 assert(mt
->hiz_buf
== NULL
);
1797 assert(mt
->aux_usage
== ISL_AUX_USAGE_HIZ
);
1799 enum isl_aux_state
**aux_state
=
1800 create_aux_state_map(mt
, ISL_AUX_STATE_AUX_INVALID
);
1804 struct isl_surf temp_hiz_surf
;
1806 MAYBE_UNUSED
bool ok
=
1807 isl_surf_get_hiz_surf(&brw
->isl_dev
, &mt
->surf
, &temp_hiz_surf
);
1810 const uint32_t alloc_flags
= BO_ALLOC_BUSY
;
1811 mt
->hiz_buf
= intel_alloc_aux_buffer(brw
, "hiz-miptree",
1812 &temp_hiz_surf
, alloc_flags
, mt
);
1819 for (unsigned level
= mt
->first_level
; level
<= mt
->last_level
; ++level
)
1820 intel_miptree_level_enable_hiz(brw
, mt
, level
);
1822 mt
->aux_state
= aux_state
;
1829 * Allocate the initial aux surface for a miptree based on mt->aux_usage
1831 * Since MCS, HiZ, and CCS_E can compress more than just clear color, we
1832 * create the auxiliary surfaces up-front. CCS_D, on the other hand, can only
1833 * compress clear color so we wait until an actual fast-clear to allocate it.
1836 intel_miptree_alloc_aux(struct brw_context
*brw
,
1837 struct intel_mipmap_tree
*mt
)
1839 switch (mt
->aux_usage
) {
1840 case ISL_AUX_USAGE_NONE
:
1843 case ISL_AUX_USAGE_HIZ
:
1844 assert(!_mesa_is_format_color_format(mt
->format
));
1845 if (!intel_miptree_alloc_hiz(brw
, mt
))
1849 case ISL_AUX_USAGE_MCS
:
1850 assert(_mesa_is_format_color_format(mt
->format
));
1851 assert(mt
->surf
.samples
> 1);
1852 if (!intel_miptree_alloc_mcs(brw
, mt
, mt
->surf
.samples
))
1856 case ISL_AUX_USAGE_CCS_D
:
1857 /* Since CCS_D can only compress clear color so we wait until an actual
1858 * fast-clear to allocate it.
1862 case ISL_AUX_USAGE_CCS_E
:
1863 assert(_mesa_is_format_color_format(mt
->format
));
1864 assert(mt
->surf
.samples
== 1);
1865 if (!intel_miptree_alloc_ccs(brw
, mt
))
1870 unreachable("Invalid aux usage");
1875 * Can the miptree sample using the hiz buffer?
1878 intel_miptree_sample_with_hiz(struct brw_context
*brw
,
1879 struct intel_mipmap_tree
*mt
)
1881 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1883 /* It's unclear how well supported sampling from the hiz buffer is on GEN8,
1884 * so keep things conservative for now and never enable it unless we're SKL+.
1886 if (devinfo
->gen
< 9) {
1894 /* It seems the hardware won't fallback to the depth buffer if some of the
1895 * mipmap levels aren't available in the HiZ buffer. So we need all levels
1896 * of the texture to be HiZ enabled.
1898 for (unsigned level
= 0; level
< mt
->surf
.levels
; ++level
) {
1899 if (!intel_miptree_level_has_hiz(mt
, level
))
1903 /* If compressed multisampling is enabled, then we use it for the auxiliary
1906 * From the BDW PRM (Volume 2d: Command Reference: Structures
1907 * RENDER_SURFACE_STATE.AuxiliarySurfaceMode):
1909 * "If this field is set to AUX_HIZ, Number of Multisamples must be
1910 * MULTISAMPLECOUNT_1, and Surface Type cannot be SURFTYPE_3D.
1912 * There is no such blurb for 1D textures, but there is sufficient evidence
1913 * that this is broken on SKL+.
1915 return (mt
->surf
.samples
== 1 &&
1916 mt
->target
!= GL_TEXTURE_3D
&&
1917 mt
->target
!= GL_TEXTURE_1D
/* gen9+ restriction */);
1921 * Does the miptree slice have hiz enabled?
1924 intel_miptree_level_has_hiz(const struct intel_mipmap_tree
*mt
, uint32_t level
)
1926 intel_miptree_check_level_layer(mt
, level
, 0);
1927 return mt
->level
[level
].has_hiz
;
1930 static inline uint32_t
1931 miptree_level_range_length(const struct intel_mipmap_tree
*mt
,
1932 uint32_t start_level
, uint32_t num_levels
)
1934 assert(start_level
>= mt
->first_level
);
1935 assert(start_level
<= mt
->last_level
);
1937 if (num_levels
== INTEL_REMAINING_LAYERS
)
1938 num_levels
= mt
->last_level
- start_level
+ 1;
1939 /* Check for overflow */
1940 assert(start_level
+ num_levels
>= start_level
);
1941 assert(start_level
+ num_levels
<= mt
->last_level
+ 1);
1946 static inline uint32_t
1947 miptree_layer_range_length(const struct intel_mipmap_tree
*mt
, uint32_t level
,
1948 uint32_t start_layer
, uint32_t num_layers
)
1950 assert(level
<= mt
->last_level
);
1952 const uint32_t total_num_layers
= brw_get_num_logical_layers(mt
, level
);
1953 assert(start_layer
< total_num_layers
);
1954 if (num_layers
== INTEL_REMAINING_LAYERS
)
1955 num_layers
= total_num_layers
- start_layer
;
1956 /* Check for overflow */
1957 assert(start_layer
+ num_layers
>= start_layer
);
1958 assert(start_layer
+ num_layers
<= total_num_layers
);
1964 intel_miptree_has_color_unresolved(const struct intel_mipmap_tree
*mt
,
1965 unsigned start_level
, unsigned num_levels
,
1966 unsigned start_layer
, unsigned num_layers
)
1968 assert(_mesa_is_format_color_format(mt
->format
));
1973 /* Clamp the level range to fit the miptree */
1974 num_levels
= miptree_level_range_length(mt
, start_level
, num_levels
);
1976 for (uint32_t l
= 0; l
< num_levels
; l
++) {
1977 const uint32_t level
= start_level
+ l
;
1978 const uint32_t level_layers
=
1979 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
1980 for (unsigned a
= 0; a
< level_layers
; a
++) {
1981 enum isl_aux_state aux_state
=
1982 intel_miptree_get_aux_state(mt
, level
, start_layer
+ a
);
1983 assert(aux_state
!= ISL_AUX_STATE_AUX_INVALID
);
1984 if (aux_state
!= ISL_AUX_STATE_PASS_THROUGH
)
1993 intel_miptree_check_color_resolve(const struct brw_context
*brw
,
1994 const struct intel_mipmap_tree
*mt
,
1995 unsigned level
, unsigned layer
)
2000 /* Fast color clear is supported for mipmapped surfaces only on Gen8+. */
2001 assert(brw
->screen
->devinfo
.gen
>= 8 ||
2002 (level
== 0 && mt
->first_level
== 0 && mt
->last_level
== 0));
2004 /* Compression of arrayed msaa surfaces is supported. */
2005 if (mt
->surf
.samples
> 1)
2008 /* Fast color clear is supported for non-msaa arrays only on Gen8+. */
2009 assert(brw
->screen
->devinfo
.gen
>= 8 ||
2011 mt
->surf
.logical_level0_px
.depth
== 1 &&
2012 mt
->surf
.logical_level0_px
.array_len
== 1));
2018 static enum blorp_fast_clear_op
2019 get_ccs_d_resolve_op(enum isl_aux_state aux_state
,
2020 enum isl_aux_usage aux_usage
,
2021 bool fast_clear_supported
)
2023 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_CCS_D
);
2025 const bool ccs_supported
= aux_usage
== ISL_AUX_USAGE_CCS_D
;
2027 assert(ccs_supported
== fast_clear_supported
);
2029 switch (aux_state
) {
2030 case ISL_AUX_STATE_CLEAR
:
2031 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2033 return BLORP_FAST_CLEAR_OP_RESOLVE_FULL
;
2035 return BLORP_FAST_CLEAR_OP_NONE
;
2037 case ISL_AUX_STATE_PASS_THROUGH
:
2038 return BLORP_FAST_CLEAR_OP_NONE
;
2040 case ISL_AUX_STATE_RESOLVED
:
2041 case ISL_AUX_STATE_AUX_INVALID
:
2042 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2043 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2047 unreachable("Invalid aux state for CCS_D");
2050 static enum blorp_fast_clear_op
2051 get_ccs_e_resolve_op(enum isl_aux_state aux_state
,
2052 enum isl_aux_usage aux_usage
,
2053 bool fast_clear_supported
)
2055 /* CCS_E surfaces can be accessed as CCS_D if we're careful. */
2056 assert(aux_usage
== ISL_AUX_USAGE_NONE
||
2057 aux_usage
== ISL_AUX_USAGE_CCS_D
||
2058 aux_usage
== ISL_AUX_USAGE_CCS_E
);
2060 if (aux_usage
== ISL_AUX_USAGE_CCS_D
)
2061 assert(fast_clear_supported
);
2063 switch (aux_state
) {
2064 case ISL_AUX_STATE_CLEAR
:
2065 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2066 if (fast_clear_supported
)
2067 return BLORP_FAST_CLEAR_OP_NONE
;
2068 else if (aux_usage
== ISL_AUX_USAGE_CCS_E
)
2069 return BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL
;
2071 return BLORP_FAST_CLEAR_OP_RESOLVE_FULL
;
2073 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2074 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2075 return BLORP_FAST_CLEAR_OP_RESOLVE_FULL
;
2076 else if (!fast_clear_supported
)
2077 return BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL
;
2079 return BLORP_FAST_CLEAR_OP_NONE
;
2081 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2082 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2083 return BLORP_FAST_CLEAR_OP_RESOLVE_FULL
;
2085 return BLORP_FAST_CLEAR_OP_NONE
;
2087 case ISL_AUX_STATE_PASS_THROUGH
:
2088 return BLORP_FAST_CLEAR_OP_NONE
;
2090 case ISL_AUX_STATE_RESOLVED
:
2091 case ISL_AUX_STATE_AUX_INVALID
:
2095 unreachable("Invalid aux state for CCS_E");
2099 intel_miptree_prepare_ccs_access(struct brw_context
*brw
,
2100 struct intel_mipmap_tree
*mt
,
2101 uint32_t level
, uint32_t layer
,
2102 enum isl_aux_usage aux_usage
,
2103 bool fast_clear_supported
)
2105 enum isl_aux_state aux_state
= intel_miptree_get_aux_state(mt
, level
, layer
);
2107 enum blorp_fast_clear_op resolve_op
;
2108 if (mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2109 resolve_op
= get_ccs_e_resolve_op(aux_state
, aux_usage
,
2110 fast_clear_supported
);
2112 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2113 resolve_op
= get_ccs_d_resolve_op(aux_state
, aux_usage
,
2114 fast_clear_supported
);
2117 if (resolve_op
!= BLORP_FAST_CLEAR_OP_NONE
) {
2118 intel_miptree_check_color_resolve(brw
, mt
, level
, layer
);
2119 brw_blorp_resolve_color(brw
, mt
, level
, layer
, resolve_op
);
2121 switch (resolve_op
) {
2122 case BLORP_FAST_CLEAR_OP_RESOLVE_FULL
:
2123 /* The CCS full resolve operation destroys the CCS and sets it to the
2124 * pass-through state. (You can also think of this as being both a
2125 * resolve and an ambiguate in one operation.)
2127 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2128 ISL_AUX_STATE_PASS_THROUGH
);
2131 case BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL
:
2132 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2133 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2137 unreachable("Invalid resolve op");
2143 intel_miptree_finish_ccs_write(struct brw_context
*brw
,
2144 struct intel_mipmap_tree
*mt
,
2145 uint32_t level
, uint32_t layer
,
2146 enum isl_aux_usage aux_usage
)
2148 assert(aux_usage
== ISL_AUX_USAGE_NONE
||
2149 aux_usage
== ISL_AUX_USAGE_CCS_D
||
2150 aux_usage
== ISL_AUX_USAGE_CCS_E
);
2152 enum isl_aux_state aux_state
= intel_miptree_get_aux_state(mt
, level
, layer
);
2154 if (mt
->aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2155 switch (aux_state
) {
2156 case ISL_AUX_STATE_CLEAR
:
2157 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2158 assert(aux_usage
== ISL_AUX_USAGE_CCS_E
||
2159 aux_usage
== ISL_AUX_USAGE_CCS_D
);
2161 if (aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2162 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2163 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2164 } else if (aux_state
!= ISL_AUX_STATE_PARTIAL_CLEAR
) {
2165 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2166 ISL_AUX_STATE_PARTIAL_CLEAR
);
2170 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2171 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2172 assert(aux_usage
== ISL_AUX_USAGE_CCS_E
);
2173 break; /* Nothing to do */
2175 case ISL_AUX_STATE_PASS_THROUGH
:
2176 if (aux_usage
== ISL_AUX_USAGE_CCS_E
) {
2177 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2178 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2184 case ISL_AUX_STATE_RESOLVED
:
2185 case ISL_AUX_STATE_AUX_INVALID
:
2186 unreachable("Invalid aux state for CCS_E");
2189 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2190 /* CCS_D is a bit simpler */
2191 switch (aux_state
) {
2192 case ISL_AUX_STATE_CLEAR
:
2193 assert(aux_usage
== ISL_AUX_USAGE_CCS_D
);
2194 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2195 ISL_AUX_STATE_PARTIAL_CLEAR
);
2198 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2199 assert(aux_usage
== ISL_AUX_USAGE_CCS_D
);
2200 break; /* Nothing to do */
2202 case ISL_AUX_STATE_PASS_THROUGH
:
2206 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2207 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2208 case ISL_AUX_STATE_RESOLVED
:
2209 case ISL_AUX_STATE_AUX_INVALID
:
2210 unreachable("Invalid aux state for CCS_D");
2216 intel_miptree_prepare_mcs_access(struct brw_context
*brw
,
2217 struct intel_mipmap_tree
*mt
,
2219 enum isl_aux_usage aux_usage
,
2220 bool fast_clear_supported
)
2222 assert(aux_usage
== ISL_AUX_USAGE_MCS
);
2224 switch (intel_miptree_get_aux_state(mt
, 0, layer
)) {
2225 case ISL_AUX_STATE_CLEAR
:
2226 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2227 if (!fast_clear_supported
) {
2228 brw_blorp_mcs_partial_resolve(brw
, mt
, layer
, 1);
2229 intel_miptree_set_aux_state(brw
, mt
, 0, layer
, 1,
2230 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2234 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2235 break; /* Nothing to do */
2237 case ISL_AUX_STATE_RESOLVED
:
2238 case ISL_AUX_STATE_PASS_THROUGH
:
2239 case ISL_AUX_STATE_AUX_INVALID
:
2240 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2241 unreachable("Invalid aux state for MCS");
2246 intel_miptree_finish_mcs_write(struct brw_context
*brw
,
2247 struct intel_mipmap_tree
*mt
,
2249 enum isl_aux_usage aux_usage
)
2251 assert(aux_usage
== ISL_AUX_USAGE_MCS
);
2253 switch (intel_miptree_get_aux_state(mt
, 0, layer
)) {
2254 case ISL_AUX_STATE_CLEAR
:
2255 intel_miptree_set_aux_state(brw
, mt
, 0, layer
, 1,
2256 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2259 case ISL_AUX_STATE_COMPRESSED_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_prepare_hiz_access(struct brw_context
*brw
,
2273 struct intel_mipmap_tree
*mt
,
2274 uint32_t level
, uint32_t layer
,
2275 enum isl_aux_usage aux_usage
,
2276 bool fast_clear_supported
)
2278 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_HIZ
);
2280 enum blorp_hiz_op hiz_op
= BLORP_HIZ_OP_NONE
;
2281 switch (intel_miptree_get_aux_state(mt
, level
, layer
)) {
2282 case ISL_AUX_STATE_CLEAR
:
2283 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2284 if (aux_usage
!= ISL_AUX_USAGE_HIZ
|| !fast_clear_supported
)
2285 hiz_op
= BLORP_HIZ_OP_DEPTH_RESOLVE
;
2288 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2289 if (aux_usage
!= ISL_AUX_USAGE_HIZ
)
2290 hiz_op
= BLORP_HIZ_OP_DEPTH_RESOLVE
;
2293 case ISL_AUX_STATE_PASS_THROUGH
:
2294 case ISL_AUX_STATE_RESOLVED
:
2297 case ISL_AUX_STATE_AUX_INVALID
:
2298 if (aux_usage
== ISL_AUX_USAGE_HIZ
)
2299 hiz_op
= BLORP_HIZ_OP_HIZ_RESOLVE
;
2302 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2303 unreachable("Invalid HiZ state");
2306 if (hiz_op
!= BLORP_HIZ_OP_NONE
) {
2307 intel_hiz_exec(brw
, mt
, level
, layer
, 1, hiz_op
);
2310 case BLORP_HIZ_OP_DEPTH_RESOLVE
:
2311 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2312 ISL_AUX_STATE_RESOLVED
);
2315 case BLORP_HIZ_OP_HIZ_RESOLVE
:
2316 /* The HiZ resolve operation is actually an ambiguate */
2317 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2318 ISL_AUX_STATE_PASS_THROUGH
);
2322 unreachable("Invalid HiZ op");
2328 intel_miptree_finish_hiz_write(struct brw_context
*brw
,
2329 struct intel_mipmap_tree
*mt
,
2330 uint32_t level
, uint32_t layer
,
2331 enum isl_aux_usage aux_usage
)
2333 assert(aux_usage
== ISL_AUX_USAGE_NONE
|| aux_usage
== ISL_AUX_USAGE_HIZ
);
2335 switch (intel_miptree_get_aux_state(mt
, level
, layer
)) {
2336 case ISL_AUX_STATE_CLEAR
:
2337 assert(aux_usage
== ISL_AUX_USAGE_HIZ
);
2338 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2339 ISL_AUX_STATE_COMPRESSED_CLEAR
);
2342 case ISL_AUX_STATE_COMPRESSED_NO_CLEAR
:
2343 case ISL_AUX_STATE_COMPRESSED_CLEAR
:
2344 assert(aux_usage
== ISL_AUX_USAGE_HIZ
);
2345 break; /* Nothing to do */
2347 case ISL_AUX_STATE_RESOLVED
:
2348 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
2349 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2350 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2352 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2353 ISL_AUX_STATE_AUX_INVALID
);
2357 case ISL_AUX_STATE_PASS_THROUGH
:
2358 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
2359 intel_miptree_set_aux_state(brw
, mt
, level
, layer
, 1,
2360 ISL_AUX_STATE_COMPRESSED_NO_CLEAR
);
2364 case ISL_AUX_STATE_AUX_INVALID
:
2365 assert(aux_usage
!= ISL_AUX_USAGE_HIZ
);
2368 case ISL_AUX_STATE_PARTIAL_CLEAR
:
2369 unreachable("Invalid HiZ state");
2374 intel_miptree_prepare_access(struct brw_context
*brw
,
2375 struct intel_mipmap_tree
*mt
,
2376 uint32_t start_level
, uint32_t num_levels
,
2377 uint32_t start_layer
, uint32_t num_layers
,
2378 enum isl_aux_usage aux_usage
,
2379 bool fast_clear_supported
)
2381 num_levels
= miptree_level_range_length(mt
, start_level
, num_levels
);
2383 switch (mt
->aux_usage
) {
2384 case ISL_AUX_USAGE_NONE
:
2388 case ISL_AUX_USAGE_MCS
:
2389 assert(mt
->mcs_buf
);
2390 assert(start_level
== 0 && num_levels
== 1);
2391 const uint32_t level_layers
=
2392 miptree_layer_range_length(mt
, 0, start_layer
, num_layers
);
2393 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2394 intel_miptree_prepare_mcs_access(brw
, mt
, start_layer
+ a
,
2395 aux_usage
, fast_clear_supported
);
2399 case ISL_AUX_USAGE_CCS_D
:
2400 case ISL_AUX_USAGE_CCS_E
:
2404 for (uint32_t l
= 0; l
< num_levels
; l
++) {
2405 const uint32_t level
= start_level
+ l
;
2406 const uint32_t level_layers
=
2407 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2408 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2409 intel_miptree_prepare_ccs_access(brw
, mt
, level
,
2411 aux_usage
, fast_clear_supported
);
2416 case ISL_AUX_USAGE_HIZ
:
2417 assert(mt
->hiz_buf
);
2418 for (uint32_t l
= 0; l
< num_levels
; l
++) {
2419 const uint32_t level
= start_level
+ l
;
2420 if (!intel_miptree_level_has_hiz(mt
, level
))
2423 const uint32_t level_layers
=
2424 miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2425 for (uint32_t a
= 0; a
< level_layers
; a
++) {
2426 intel_miptree_prepare_hiz_access(brw
, mt
, level
, start_layer
+ a
,
2427 aux_usage
, fast_clear_supported
);
2433 unreachable("Invalid aux usage");
2438 intel_miptree_finish_write(struct brw_context
*brw
,
2439 struct intel_mipmap_tree
*mt
, uint32_t level
,
2440 uint32_t start_layer
, uint32_t num_layers
,
2441 enum isl_aux_usage aux_usage
)
2443 num_layers
= miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2445 switch (mt
->aux_usage
) {
2446 case ISL_AUX_USAGE_NONE
:
2450 case ISL_AUX_USAGE_MCS
:
2451 assert(mt
->mcs_buf
);
2452 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2453 intel_miptree_finish_mcs_write(brw
, mt
, start_layer
+ a
,
2458 case ISL_AUX_USAGE_CCS_D
:
2459 case ISL_AUX_USAGE_CCS_E
:
2463 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2464 intel_miptree_finish_ccs_write(brw
, mt
, level
, start_layer
+ a
,
2469 case ISL_AUX_USAGE_HIZ
:
2470 if (!intel_miptree_level_has_hiz(mt
, level
))
2473 for (uint32_t a
= 0; a
< num_layers
; a
++) {
2474 intel_miptree_finish_hiz_write(brw
, mt
, level
, start_layer
+ a
,
2480 unreachable("Invavlid aux usage");
2485 intel_miptree_get_aux_state(const struct intel_mipmap_tree
*mt
,
2486 uint32_t level
, uint32_t layer
)
2488 intel_miptree_check_level_layer(mt
, level
, layer
);
2490 if (_mesa_is_format_color_format(mt
->format
)) {
2491 assert(mt
->mcs_buf
!= NULL
);
2492 assert(mt
->surf
.samples
== 1 ||
2493 mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
2494 } else if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2495 unreachable("Cannot get aux state for stencil");
2497 assert(intel_miptree_level_has_hiz(mt
, level
));
2500 return mt
->aux_state
[level
][layer
];
2504 intel_miptree_set_aux_state(struct brw_context
*brw
,
2505 struct intel_mipmap_tree
*mt
, uint32_t level
,
2506 uint32_t start_layer
, uint32_t num_layers
,
2507 enum isl_aux_state aux_state
)
2509 num_layers
= miptree_layer_range_length(mt
, level
, start_layer
, num_layers
);
2511 if (_mesa_is_format_color_format(mt
->format
)) {
2512 assert(mt
->mcs_buf
!= NULL
);
2513 assert(mt
->surf
.samples
== 1 ||
2514 mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
);
2515 } else if (mt
->format
== MESA_FORMAT_S_UINT8
) {
2516 unreachable("Cannot get aux state for stencil");
2518 assert(intel_miptree_level_has_hiz(mt
, level
));
2521 for (unsigned a
= 0; a
< num_layers
; a
++) {
2522 if (mt
->aux_state
[level
][start_layer
+ a
] != aux_state
) {
2523 mt
->aux_state
[level
][start_layer
+ a
] = aux_state
;
2524 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
2529 /* On Gen9 color buffers may be compressed by the hardware (lossless
2530 * compression). There are, however, format restrictions and care needs to be
2531 * taken that the sampler engine is capable for re-interpreting a buffer with
2532 * format different the buffer was originally written with.
2534 * For example, SRGB formats are not compressible and the sampler engine isn't
2535 * capable of treating RGBA_UNORM as SRGB_ALPHA. In such a case the underlying
2536 * color buffer needs to be resolved so that the sampling surface can be
2537 * sampled as non-compressed (i.e., without the auxiliary MCS buffer being
2541 can_texture_with_ccs(struct brw_context
*brw
,
2542 struct intel_mipmap_tree
*mt
,
2543 enum isl_format view_format
)
2545 if (mt
->aux_usage
!= ISL_AUX_USAGE_CCS_E
)
2548 if (!isl_formats_are_ccs_e_compatible(&brw
->screen
->devinfo
,
2549 mt
->surf
.format
, view_format
)) {
2550 perf_debug("Incompatible sampling format (%s) for rbc (%s)\n",
2551 isl_format_get_layout(view_format
)->name
,
2552 _mesa_get_format_name(mt
->format
));
2560 intel_miptree_texture_aux_usage(struct brw_context
*brw
,
2561 struct intel_mipmap_tree
*mt
,
2562 enum isl_format view_format
)
2564 switch (mt
->aux_usage
) {
2565 case ISL_AUX_USAGE_HIZ
:
2566 if (intel_miptree_sample_with_hiz(brw
, mt
))
2567 return ISL_AUX_USAGE_HIZ
;
2570 case ISL_AUX_USAGE_MCS
:
2571 return ISL_AUX_USAGE_MCS
;
2573 case ISL_AUX_USAGE_CCS_D
:
2574 case ISL_AUX_USAGE_CCS_E
:
2576 assert(mt
->aux_usage
== ISL_AUX_USAGE_CCS_D
);
2577 return ISL_AUX_USAGE_NONE
;
2580 /* If we don't have any unresolved color, report an aux usage of
2581 * ISL_AUX_USAGE_NONE. This way, texturing won't even look at the
2582 * aux surface and we can save some bandwidth.
2584 if (!intel_miptree_has_color_unresolved(mt
, 0, INTEL_REMAINING_LEVELS
,
2585 0, INTEL_REMAINING_LAYERS
))
2586 return ISL_AUX_USAGE_NONE
;
2588 if (can_texture_with_ccs(brw
, mt
, view_format
))
2589 return ISL_AUX_USAGE_CCS_E
;
2596 return ISL_AUX_USAGE_NONE
;
2600 isl_formats_are_fast_clear_compatible(enum isl_format a
, enum isl_format b
)
2602 /* On gen8 and earlier, the hardware was only capable of handling 0/1 clear
2603 * values so sRGB curve application was a no-op for all fast-clearable
2606 * On gen9+, the hardware supports arbitrary clear values. For sRGB clear
2607 * values, the hardware interprets the floats, not as what would be
2608 * returned from the sampler (or written by the shader), but as being
2609 * between format conversion and sRGB curve application. This means that
2610 * we can switch between sRGB and UNORM without having to whack the clear
2613 return isl_format_srgb_to_linear(a
) == isl_format_srgb_to_linear(b
);
2617 intel_miptree_prepare_texture(struct brw_context
*brw
,
2618 struct intel_mipmap_tree
*mt
,
2619 enum isl_format view_format
,
2620 uint32_t start_level
, uint32_t num_levels
,
2621 uint32_t start_layer
, uint32_t num_layers
,
2624 enum isl_aux_usage aux_usage
= disable_aux
? ISL_AUX_USAGE_NONE
:
2625 intel_miptree_texture_aux_usage(brw
, mt
, view_format
);
2626 bool clear_supported
= aux_usage
!= ISL_AUX_USAGE_NONE
;
2628 /* Clear color is specified as ints or floats and the conversion is done by
2629 * the sampler. If we have a texture view, we would have to perform the
2630 * clear color conversion manually. Just disable clear color.
2632 if (!isl_formats_are_fast_clear_compatible(mt
->surf
.format
, view_format
))
2633 clear_supported
= false;
2635 intel_miptree_prepare_access(brw
, mt
, start_level
, num_levels
,
2636 start_layer
, num_layers
,
2637 aux_usage
, clear_supported
);
2641 intel_miptree_prepare_image(struct brw_context
*brw
,
2642 struct intel_mipmap_tree
*mt
)
2644 /* The data port doesn't understand any compression */
2645 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2646 0, INTEL_REMAINING_LAYERS
,
2647 ISL_AUX_USAGE_NONE
, false);
2651 intel_miptree_render_aux_usage(struct brw_context
*brw
,
2652 struct intel_mipmap_tree
*mt
,
2653 enum isl_format render_format
,
2656 switch (mt
->aux_usage
) {
2657 case ISL_AUX_USAGE_MCS
:
2658 assert(mt
->mcs_buf
);
2659 return ISL_AUX_USAGE_MCS
;
2661 case ISL_AUX_USAGE_CCS_D
:
2662 return mt
->mcs_buf
? ISL_AUX_USAGE_CCS_D
: ISL_AUX_USAGE_NONE
;
2664 case ISL_AUX_USAGE_CCS_E
: {
2665 /* If the format supports CCS_E, then we can just use it */
2666 if (isl_format_supports_ccs_e(&brw
->screen
->devinfo
, render_format
))
2667 return ISL_AUX_USAGE_CCS_E
;
2669 /* Otherwise, we have to fall back to CCS_D */
2671 /* gen9 hardware technically supports non-0/1 clear colors with sRGB
2672 * formats. However, there are issues with blending where it doesn't
2673 * properly apply the sRGB curve to the clear color when blending.
2675 if (blend_enabled
&& isl_format_is_srgb(render_format
) &&
2676 !isl_color_value_is_zero_one(mt
->fast_clear_color
, render_format
))
2677 return ISL_AUX_USAGE_NONE
;
2679 return ISL_AUX_USAGE_CCS_D
;
2683 return ISL_AUX_USAGE_NONE
;
2688 intel_miptree_prepare_render(struct brw_context
*brw
,
2689 struct intel_mipmap_tree
*mt
, uint32_t level
,
2690 uint32_t start_layer
, uint32_t layer_count
,
2691 enum isl_format render_format
,
2694 enum isl_aux_usage aux_usage
=
2695 intel_miptree_render_aux_usage(brw
, mt
, render_format
, blend_enabled
);
2696 intel_miptree_prepare_access(brw
, mt
, level
, 1, start_layer
, layer_count
,
2697 aux_usage
, aux_usage
!= ISL_AUX_USAGE_NONE
);
2701 intel_miptree_finish_render(struct brw_context
*brw
,
2702 struct intel_mipmap_tree
*mt
, uint32_t level
,
2703 uint32_t start_layer
, uint32_t layer_count
,
2704 enum isl_format render_format
,
2707 assert(_mesa_is_format_color_format(mt
->format
));
2709 enum isl_aux_usage aux_usage
=
2710 intel_miptree_render_aux_usage(brw
, mt
, render_format
, blend_enabled
);
2711 intel_miptree_finish_write(brw
, mt
, level
, start_layer
, layer_count
,
2716 intel_miptree_prepare_depth(struct brw_context
*brw
,
2717 struct intel_mipmap_tree
*mt
, uint32_t level
,
2718 uint32_t start_layer
, uint32_t layer_count
)
2720 intel_miptree_prepare_access(brw
, mt
, level
, 1, start_layer
, layer_count
,
2721 mt
->aux_usage
, mt
->hiz_buf
!= NULL
);
2725 intel_miptree_finish_depth(struct brw_context
*brw
,
2726 struct intel_mipmap_tree
*mt
, uint32_t level
,
2727 uint32_t start_layer
, uint32_t layer_count
,
2730 if (depth_written
) {
2731 intel_miptree_finish_write(brw
, mt
, level
, start_layer
, layer_count
,
2732 mt
->hiz_buf
!= NULL
);
2737 intel_miptree_prepare_external(struct brw_context
*brw
,
2738 struct intel_mipmap_tree
*mt
)
2740 enum isl_aux_usage aux_usage
= ISL_AUX_USAGE_NONE
;
2741 bool supports_fast_clear
= false;
2743 const struct isl_drm_modifier_info
*mod_info
=
2744 isl_drm_modifier_get_info(mt
->drm_modifier
);
2746 if (mod_info
&& mod_info
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
2747 /* CCS_E is the only supported aux for external images and it's only
2748 * supported on very simple images.
2750 assert(mod_info
->aux_usage
== ISL_AUX_USAGE_CCS_E
);
2751 assert(_mesa_is_format_color_format(mt
->format
));
2752 assert(mt
->first_level
== 0 && mt
->last_level
== 0);
2753 assert(mt
->surf
.logical_level0_px
.depth
== 1);
2754 assert(mt
->surf
.logical_level0_px
.array_len
== 1);
2755 assert(mt
->surf
.samples
== 1);
2756 assert(mt
->mcs_buf
!= NULL
);
2758 aux_usage
= mod_info
->aux_usage
;
2759 supports_fast_clear
= mod_info
->supports_clear_color
;
2762 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2763 0, INTEL_REMAINING_LAYERS
,
2764 aux_usage
, supports_fast_clear
);
2768 * Make it possible to share the BO backing the given miptree with another
2769 * process or another miptree.
2771 * Fast color clears are unsafe with shared buffers, so we need to resolve and
2772 * then discard the MCS buffer, if present. We also set the no_ccs flag to
2773 * ensure that no MCS buffer gets allocated in the future.
2775 * HiZ is similarly unsafe with shared buffers.
2778 intel_miptree_make_shareable(struct brw_context
*brw
,
2779 struct intel_mipmap_tree
*mt
)
2781 /* MCS buffers are also used for multisample buffers, but we can't resolve
2782 * away a multisample MCS buffer because it's an integral part of how the
2783 * pixel data is stored. Fortunately this code path should never be
2784 * reached for multisample buffers.
2786 assert(mt
->surf
.msaa_layout
== ISL_MSAA_LAYOUT_NONE
||
2787 mt
->surf
.samples
== 1);
2789 intel_miptree_prepare_access(brw
, mt
, 0, INTEL_REMAINING_LEVELS
,
2790 0, INTEL_REMAINING_LAYERS
,
2791 ISL_AUX_USAGE_NONE
, false);
2794 brw_bo_unreference(mt
->mcs_buf
->bo
);
2798 /* Any pending MCS/CCS operations are no longer needed. Trying to
2799 * execute any will likely crash due to the missing aux buffer. So let's
2800 * delete all pending ops.
2802 free(mt
->aux_state
);
2803 mt
->aux_state
= NULL
;
2804 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
2808 intel_miptree_aux_buffer_free(mt
->hiz_buf
);
2811 for (uint32_t l
= mt
->first_level
; l
<= mt
->last_level
; ++l
) {
2812 mt
->level
[l
].has_hiz
= false;
2815 /* Any pending HiZ operations are no longer needed. Trying to execute
2816 * any will likely crash due to the missing aux buffer. So let's delete
2819 free(mt
->aux_state
);
2820 mt
->aux_state
= NULL
;
2821 brw
->ctx
.NewDriverState
|= BRW_NEW_AUX_STATE
;
2824 mt
->aux_usage
= ISL_AUX_USAGE_NONE
;
2825 mt
->supports_fast_clear
= false;
2830 * \brief Get pointer offset into stencil buffer.
2832 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
2833 * must decode the tile's layout in software.
2836 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
2838 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
2840 * Even though the returned offset is always positive, the return type is
2842 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
2843 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
2846 intel_offset_S8(uint32_t stride
, uint32_t x
, uint32_t y
, bool swizzled
)
2848 uint32_t tile_size
= 4096;
2849 uint32_t tile_width
= 64;
2850 uint32_t tile_height
= 64;
2851 uint32_t row_size
= 64 * stride
/ 2; /* Two rows are interleaved. */
2853 uint32_t tile_x
= x
/ tile_width
;
2854 uint32_t tile_y
= y
/ tile_height
;
2856 /* The byte's address relative to the tile's base addres. */
2857 uint32_t byte_x
= x
% tile_width
;
2858 uint32_t byte_y
= y
% tile_height
;
2860 uintptr_t u
= tile_y
* row_size
2861 + tile_x
* tile_size
2862 + 512 * (byte_x
/ 8)
2864 + 32 * ((byte_y
/ 4) % 2)
2865 + 16 * ((byte_x
/ 4) % 2)
2866 + 8 * ((byte_y
/ 2) % 2)
2867 + 4 * ((byte_x
/ 2) % 2)
2872 /* adjust for bit6 swizzling */
2873 if (((byte_x
/ 8) % 2) == 1) {
2874 if (((byte_y
/ 8) % 2) == 0) {
2886 intel_miptree_updownsample(struct brw_context
*brw
,
2887 struct intel_mipmap_tree
*src
,
2888 struct intel_mipmap_tree
*dst
)
2890 unsigned src_w
= src
->surf
.logical_level0_px
.width
;
2891 unsigned src_h
= src
->surf
.logical_level0_px
.height
;
2892 unsigned dst_w
= dst
->surf
.logical_level0_px
.width
;
2893 unsigned dst_h
= dst
->surf
.logical_level0_px
.height
;
2895 brw_blorp_blit_miptrees(brw
,
2896 src
, 0 /* level */, 0 /* layer */,
2897 src
->format
, SWIZZLE_XYZW
,
2898 dst
, 0 /* level */, 0 /* layer */, dst
->format
,
2901 GL_NEAREST
, false, false /*mirror x, y*/,
2904 if (src
->stencil_mt
) {
2905 src_w
= src
->stencil_mt
->surf
.logical_level0_px
.width
;
2906 src_h
= src
->stencil_mt
->surf
.logical_level0_px
.height
;
2907 dst_w
= dst
->stencil_mt
->surf
.logical_level0_px
.width
;
2908 dst_h
= dst
->stencil_mt
->surf
.logical_level0_px
.height
;
2910 brw_blorp_blit_miptrees(brw
,
2911 src
->stencil_mt
, 0 /* level */, 0 /* layer */,
2912 src
->stencil_mt
->format
, SWIZZLE_XYZW
,
2913 dst
->stencil_mt
, 0 /* level */, 0 /* layer */,
2914 dst
->stencil_mt
->format
,
2917 GL_NEAREST
, false, false /*mirror x, y*/,
2918 false, false /* decode/encode srgb */);
2923 intel_update_r8stencil(struct brw_context
*brw
,
2924 struct intel_mipmap_tree
*mt
)
2926 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
2928 assert(devinfo
->gen
>= 7);
2929 struct intel_mipmap_tree
*src
=
2930 mt
->format
== MESA_FORMAT_S_UINT8
? mt
: mt
->stencil_mt
;
2931 if (!src
|| devinfo
->gen
>= 8 || !src
->r8stencil_needs_update
)
2934 assert(src
->surf
.size
> 0);
2936 if (!mt
->r8stencil_mt
) {
2937 assert(devinfo
->gen
> 6); /* Handle MIPTREE_LAYOUT_GEN6_HIZ_STENCIL */
2938 mt
->r8stencil_mt
= make_surface(
2941 MESA_FORMAT_R_UINT8
,
2942 src
->first_level
, src
->last_level
,
2943 src
->surf
.logical_level0_px
.width
,
2944 src
->surf
.logical_level0_px
.height
,
2945 src
->surf
.dim
== ISL_SURF_DIM_3D
?
2946 src
->surf
.logical_level0_px
.depth
:
2947 src
->surf
.logical_level0_px
.array_len
,
2950 ISL_SURF_USAGE_TEXTURE_BIT
,
2951 BO_ALLOC_BUSY
, 0, NULL
);
2952 assert(mt
->r8stencil_mt
);
2955 struct intel_mipmap_tree
*dst
= mt
->r8stencil_mt
;
2957 for (int level
= src
->first_level
; level
<= src
->last_level
; level
++) {
2958 const unsigned depth
= src
->surf
.dim
== ISL_SURF_DIM_3D
?
2959 minify(src
->surf
.phys_level0_sa
.depth
, level
) :
2960 src
->surf
.phys_level0_sa
.array_len
;
2962 for (unsigned layer
= 0; layer
< depth
; layer
++) {
2963 brw_blorp_copy_miptrees(brw
,
2967 minify(src
->surf
.logical_level0_px
.width
,
2969 minify(src
->surf
.logical_level0_px
.height
,
2974 brw_render_cache_set_check_flush(brw
, dst
->bo
);
2975 src
->r8stencil_needs_update
= false;
2979 intel_miptree_map_raw(struct brw_context
*brw
,
2980 struct intel_mipmap_tree
*mt
,
2983 struct brw_bo
*bo
= mt
->bo
;
2985 if (brw_batch_references(&brw
->batch
, bo
))
2986 intel_batchbuffer_flush(brw
);
2988 return brw_bo_map(brw
, bo
, mode
);
2992 intel_miptree_unmap_raw(struct intel_mipmap_tree
*mt
)
2994 brw_bo_unmap(mt
->bo
);
2998 intel_miptree_map_gtt(struct brw_context
*brw
,
2999 struct intel_mipmap_tree
*mt
,
3000 struct intel_miptree_map
*map
,
3001 unsigned int level
, unsigned int slice
)
3003 unsigned int bw
, bh
;
3005 unsigned int image_x
, image_y
;
3006 intptr_t x
= map
->x
;
3007 intptr_t y
= map
->y
;
3009 /* For compressed formats, the stride is the number of bytes per
3010 * row of blocks. intel_miptree_get_image_offset() already does
3013 _mesa_get_format_block_size(mt
->format
, &bw
, &bh
);
3014 assert(y
% bh
== 0);
3015 assert(x
% bw
== 0);
3019 base
= intel_miptree_map_raw(brw
, mt
, map
->mode
);
3026 /* Note that in the case of cube maps, the caller must have passed the
3027 * slice number referencing the face.
3029 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3033 map
->stride
= mt
->surf
.row_pitch
;
3034 map
->ptr
= base
+ y
* map
->stride
+ x
* mt
->cpp
;
3037 DBG("%s: %d,%d %dx%d from mt %p (%s) "
3038 "%"PRIiPTR
",%"PRIiPTR
" = %p/%d\n", __func__
,
3039 map
->x
, map
->y
, map
->w
, map
->h
,
3040 mt
, _mesa_get_format_name(mt
->format
),
3041 x
, y
, map
->ptr
, map
->stride
);
3045 intel_miptree_unmap_gtt(struct intel_mipmap_tree
*mt
)
3047 intel_miptree_unmap_raw(mt
);
3051 intel_miptree_map_blit(struct brw_context
*brw
,
3052 struct intel_mipmap_tree
*mt
,
3053 struct intel_miptree_map
*map
,
3054 unsigned int level
, unsigned int slice
)
3056 map
->linear_mt
= intel_miptree_create(brw
, GL_TEXTURE_2D
, mt
->format
,
3057 /* first_level */ 0,
3061 MIPTREE_CREATE_LINEAR
);
3063 if (!map
->linear_mt
) {
3064 fprintf(stderr
, "Failed to allocate blit temporary\n");
3067 map
->stride
= map
->linear_mt
->surf
.row_pitch
;
3069 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3070 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3071 * invalidate is set, since we'll be writing the whole rectangle from our
3072 * temporary buffer back out.
3074 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3075 if (!intel_miptree_copy(brw
,
3076 mt
, level
, slice
, map
->x
, map
->y
,
3077 map
->linear_mt
, 0, 0, 0, 0,
3079 fprintf(stderr
, "Failed to blit\n");
3084 map
->ptr
= intel_miptree_map_raw(brw
, map
->linear_mt
, map
->mode
);
3086 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__
,
3087 map
->x
, map
->y
, map
->w
, map
->h
,
3088 mt
, _mesa_get_format_name(mt
->format
),
3089 level
, slice
, map
->ptr
, map
->stride
);
3094 intel_miptree_release(&map
->linear_mt
);
3100 intel_miptree_unmap_blit(struct brw_context
*brw
,
3101 struct intel_mipmap_tree
*mt
,
3102 struct intel_miptree_map
*map
,
3106 struct gl_context
*ctx
= &brw
->ctx
;
3108 intel_miptree_unmap_raw(map
->linear_mt
);
3110 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3111 bool ok
= intel_miptree_copy(brw
,
3112 map
->linear_mt
, 0, 0, 0, 0,
3113 mt
, level
, slice
, map
->x
, map
->y
,
3115 WARN_ONCE(!ok
, "Failed to blit from linear temporary mapping");
3118 intel_miptree_release(&map
->linear_mt
);
3122 * "Map" a buffer by copying it to an untiled temporary using MOVNTDQA.
3124 #if defined(USE_SSE41)
3126 intel_miptree_map_movntdqa(struct brw_context
*brw
,
3127 struct intel_mipmap_tree
*mt
,
3128 struct intel_miptree_map
*map
,
3129 unsigned int level
, unsigned int slice
)
3131 assert(map
->mode
& GL_MAP_READ_BIT
);
3132 assert(!(map
->mode
& GL_MAP_WRITE_BIT
));
3134 DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__
,
3135 map
->x
, map
->y
, map
->w
, map
->h
,
3136 mt
, _mesa_get_format_name(mt
->format
),
3137 level
, slice
, map
->ptr
, map
->stride
);
3139 /* Map the original image */
3142 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3146 void *src
= intel_miptree_map_raw(brw
, mt
, map
->mode
);
3152 src
+= image_y
* mt
->surf
.row_pitch
;
3153 src
+= image_x
* mt
->cpp
;
3155 /* Due to the pixel offsets for the particular image being mapped, our
3156 * src pointer may not be 16-byte aligned. However, if the pitch is
3157 * divisible by 16, then the amount by which it's misaligned will remain
3158 * consistent from row to row.
3160 assert((mt
->surf
.row_pitch
% 16) == 0);
3161 const int misalignment
= ((uintptr_t) src
) & 15;
3163 /* Create an untiled temporary buffer for the mapping. */
3164 const unsigned width_bytes
= _mesa_format_row_stride(mt
->format
, map
->w
);
3166 map
->stride
= ALIGN(misalignment
+ width_bytes
, 16);
3168 map
->buffer
= _mesa_align_malloc(map
->stride
* map
->h
, 16);
3169 /* Offset the destination so it has the same misalignment as src. */
3170 map
->ptr
= map
->buffer
+ misalignment
;
3172 assert((((uintptr_t) map
->ptr
) & 15) == misalignment
);
3174 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3175 void *dst_ptr
= map
->ptr
+ y
* map
->stride
;
3176 void *src_ptr
= src
+ y
* mt
->surf
.row_pitch
;
3178 _mesa_streaming_load_memcpy(dst_ptr
, src_ptr
, width_bytes
);
3181 intel_miptree_unmap_raw(mt
);
3185 intel_miptree_unmap_movntdqa(struct brw_context
*brw
,
3186 struct intel_mipmap_tree
*mt
,
3187 struct intel_miptree_map
*map
,
3191 _mesa_align_free(map
->buffer
);
3198 intel_miptree_map_s8(struct brw_context
*brw
,
3199 struct intel_mipmap_tree
*mt
,
3200 struct intel_miptree_map
*map
,
3201 unsigned int level
, unsigned int slice
)
3203 map
->stride
= map
->w
;
3204 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
3208 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3209 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3210 * invalidate is set, since we'll be writing the whole rectangle from our
3211 * temporary buffer back out.
3213 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3214 uint8_t *untiled_s8_map
= map
->ptr
;
3215 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
, GL_MAP_READ_BIT
);
3216 unsigned int image_x
, image_y
;
3218 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3220 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3221 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3222 ptrdiff_t offset
= intel_offset_S8(mt
->surf
.row_pitch
,
3223 x
+ image_x
+ map
->x
,
3224 y
+ image_y
+ map
->y
,
3225 brw
->has_swizzling
);
3226 untiled_s8_map
[y
* map
->w
+ x
] = tiled_s8_map
[offset
];
3230 intel_miptree_unmap_raw(mt
);
3232 DBG("%s: %d,%d %dx%d from mt %p %d,%d = %p/%d\n", __func__
,
3233 map
->x
, map
->y
, map
->w
, map
->h
,
3234 mt
, map
->x
+ image_x
, map
->y
+ image_y
, map
->ptr
, map
->stride
);
3236 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__
,
3237 map
->x
, map
->y
, map
->w
, map
->h
,
3238 mt
, map
->ptr
, map
->stride
);
3243 intel_miptree_unmap_s8(struct brw_context
*brw
,
3244 struct intel_mipmap_tree
*mt
,
3245 struct intel_miptree_map
*map
,
3249 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3250 unsigned int image_x
, image_y
;
3251 uint8_t *untiled_s8_map
= map
->ptr
;
3252 uint8_t *tiled_s8_map
= intel_miptree_map_raw(brw
, mt
, GL_MAP_WRITE_BIT
);
3254 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3256 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3257 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3258 ptrdiff_t offset
= intel_offset_S8(mt
->surf
.row_pitch
,
3259 image_x
+ x
+ map
->x
,
3260 image_y
+ y
+ map
->y
,
3261 brw
->has_swizzling
);
3262 tiled_s8_map
[offset
] = untiled_s8_map
[y
* map
->w
+ x
];
3266 intel_miptree_unmap_raw(mt
);
3273 intel_miptree_map_etc(struct brw_context
*brw
,
3274 struct intel_mipmap_tree
*mt
,
3275 struct intel_miptree_map
*map
,
3279 assert(mt
->etc_format
!= MESA_FORMAT_NONE
);
3280 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
) {
3281 assert(mt
->format
== MESA_FORMAT_R8G8B8X8_UNORM
);
3284 assert(map
->mode
& GL_MAP_WRITE_BIT
);
3285 assert(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
);
3287 map
->stride
= _mesa_format_row_stride(mt
->etc_format
, map
->w
);
3288 map
->buffer
= malloc(_mesa_format_image_size(mt
->etc_format
,
3289 map
->w
, map
->h
, 1));
3290 map
->ptr
= map
->buffer
;
3294 intel_miptree_unmap_etc(struct brw_context
*brw
,
3295 struct intel_mipmap_tree
*mt
,
3296 struct intel_miptree_map
*map
,
3302 intel_miptree_get_image_offset(mt
, level
, slice
, &image_x
, &image_y
);
3307 uint8_t *dst
= intel_miptree_map_raw(brw
, mt
, GL_MAP_WRITE_BIT
)
3308 + image_y
* mt
->surf
.row_pitch
3309 + image_x
* mt
->cpp
;
3311 if (mt
->etc_format
== MESA_FORMAT_ETC1_RGB8
)
3312 _mesa_etc1_unpack_rgba8888(dst
, mt
->surf
.row_pitch
,
3313 map
->ptr
, map
->stride
,
3316 _mesa_unpack_etc2_format(dst
, mt
->surf
.row_pitch
,
3317 map
->ptr
, map
->stride
,
3318 map
->w
, map
->h
, mt
->etc_format
);
3320 intel_miptree_unmap_raw(mt
);
3325 * Mapping function for packed depth/stencil miptrees backed by real separate
3326 * miptrees for depth and stencil.
3328 * On gen7, and to support HiZ pre-gen7, we have to have the stencil buffer
3329 * separate from the depth buffer. Yet at the GL API level, we have to expose
3330 * packed depth/stencil textures and FBO attachments, and Mesa core expects to
3331 * be able to map that memory for texture storage and glReadPixels-type
3332 * operations. We give Mesa core that access by mallocing a temporary and
3333 * copying the data between the actual backing store and the temporary.
3336 intel_miptree_map_depthstencil(struct brw_context
*brw
,
3337 struct intel_mipmap_tree
*mt
,
3338 struct intel_miptree_map
*map
,
3339 unsigned int level
, unsigned int slice
)
3341 struct intel_mipmap_tree
*z_mt
= mt
;
3342 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
3343 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
3344 int packed_bpp
= map_z32f_x24s8
? 8 : 4;
3346 map
->stride
= map
->w
* packed_bpp
;
3347 map
->buffer
= map
->ptr
= malloc(map
->stride
* map
->h
);
3351 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
3352 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
3353 * invalidate is set, since we'll be writing the whole rectangle from our
3354 * temporary buffer back out.
3356 if (!(map
->mode
& GL_MAP_INVALIDATE_RANGE_BIT
)) {
3357 uint32_t *packed_map
= map
->ptr
;
3358 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
, GL_MAP_READ_BIT
);
3359 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
, GL_MAP_READ_BIT
);
3360 unsigned int s_image_x
, s_image_y
;
3361 unsigned int z_image_x
, z_image_y
;
3363 intel_miptree_get_image_offset(s_mt
, level
, slice
,
3364 &s_image_x
, &s_image_y
);
3365 intel_miptree_get_image_offset(z_mt
, level
, slice
,
3366 &z_image_x
, &z_image_y
);
3368 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3369 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3370 int map_x
= map
->x
+ x
, map_y
= map
->y
+ y
;
3371 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->surf
.row_pitch
,
3374 brw
->has_swizzling
);
3375 ptrdiff_t z_offset
= ((map_y
+ z_image_y
) *
3376 (z_mt
->surf
.row_pitch
/ 4) +
3377 (map_x
+ z_image_x
));
3378 uint8_t s
= s_map
[s_offset
];
3379 uint32_t z
= z_map
[z_offset
];
3381 if (map_z32f_x24s8
) {
3382 packed_map
[(y
* map
->w
+ x
) * 2 + 0] = z
;
3383 packed_map
[(y
* map
->w
+ x
) * 2 + 1] = s
;
3385 packed_map
[y
* map
->w
+ x
] = (s
<< 24) | (z
& 0x00ffffff);
3390 intel_miptree_unmap_raw(s_mt
);
3391 intel_miptree_unmap_raw(z_mt
);
3393 DBG("%s: %d,%d %dx%d from z mt %p %d,%d, s mt %p %d,%d = %p/%d\n",
3395 map
->x
, map
->y
, map
->w
, map
->h
,
3396 z_mt
, map
->x
+ z_image_x
, map
->y
+ z_image_y
,
3397 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
3398 map
->ptr
, map
->stride
);
3400 DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__
,
3401 map
->x
, map
->y
, map
->w
, map
->h
,
3402 mt
, map
->ptr
, map
->stride
);
3407 intel_miptree_unmap_depthstencil(struct brw_context
*brw
,
3408 struct intel_mipmap_tree
*mt
,
3409 struct intel_miptree_map
*map
,
3413 struct intel_mipmap_tree
*z_mt
= mt
;
3414 struct intel_mipmap_tree
*s_mt
= mt
->stencil_mt
;
3415 bool map_z32f_x24s8
= mt
->format
== MESA_FORMAT_Z_FLOAT32
;
3417 if (map
->mode
& GL_MAP_WRITE_BIT
) {
3418 uint32_t *packed_map
= map
->ptr
;
3419 uint8_t *s_map
= intel_miptree_map_raw(brw
, s_mt
, GL_MAP_WRITE_BIT
);
3420 uint32_t *z_map
= intel_miptree_map_raw(brw
, z_mt
, GL_MAP_WRITE_BIT
);
3421 unsigned int s_image_x
, s_image_y
;
3422 unsigned int z_image_x
, z_image_y
;
3424 intel_miptree_get_image_offset(s_mt
, level
, slice
,
3425 &s_image_x
, &s_image_y
);
3426 intel_miptree_get_image_offset(z_mt
, level
, slice
,
3427 &z_image_x
, &z_image_y
);
3429 for (uint32_t y
= 0; y
< map
->h
; y
++) {
3430 for (uint32_t x
= 0; x
< map
->w
; x
++) {
3431 ptrdiff_t s_offset
= intel_offset_S8(s_mt
->surf
.row_pitch
,
3432 x
+ s_image_x
+ map
->x
,
3433 y
+ s_image_y
+ map
->y
,
3434 brw
->has_swizzling
);
3435 ptrdiff_t z_offset
= ((y
+ z_image_y
+ map
->y
) *
3436 (z_mt
->surf
.row_pitch
/ 4) +
3437 (x
+ z_image_x
+ map
->x
));
3439 if (map_z32f_x24s8
) {
3440 z_map
[z_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 0];
3441 s_map
[s_offset
] = packed_map
[(y
* map
->w
+ x
) * 2 + 1];
3443 uint32_t packed
= packed_map
[y
* map
->w
+ x
];
3444 s_map
[s_offset
] = packed
>> 24;
3445 z_map
[z_offset
] = packed
;
3450 intel_miptree_unmap_raw(s_mt
);
3451 intel_miptree_unmap_raw(z_mt
);
3453 DBG("%s: %d,%d %dx%d from z mt %p (%s) %d,%d, s mt %p %d,%d = %p/%d\n",
3455 map
->x
, map
->y
, map
->w
, map
->h
,
3456 z_mt
, _mesa_get_format_name(z_mt
->format
),
3457 map
->x
+ z_image_x
, map
->y
+ z_image_y
,
3458 s_mt
, map
->x
+ s_image_x
, map
->y
+ s_image_y
,
3459 map
->ptr
, map
->stride
);
3466 * Create and attach a map to the miptree at (level, slice). Return the
3469 static struct intel_miptree_map
*
3470 intel_miptree_attach_map(struct intel_mipmap_tree
*mt
,
3479 struct intel_miptree_map
*map
= calloc(1, sizeof(*map
));
3484 assert(mt
->level
[level
].slice
[slice
].map
== NULL
);
3485 mt
->level
[level
].slice
[slice
].map
= map
;
3497 * Release the map at (level, slice).
3500 intel_miptree_release_map(struct intel_mipmap_tree
*mt
,
3504 struct intel_miptree_map
**map
;
3506 map
= &mt
->level
[level
].slice
[slice
].map
;
3512 can_blit_slice(struct intel_mipmap_tree
*mt
,
3513 unsigned int level
, unsigned int slice
)
3515 /* See intel_miptree_blit() for details on the 32k pitch limit. */
3516 if (mt
->surf
.row_pitch
>= 32768)
3523 use_intel_mipree_map_blit(struct brw_context
*brw
,
3524 struct intel_mipmap_tree
*mt
,
3529 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
3531 if (devinfo
->has_llc
&&
3532 /* It's probably not worth swapping to the blit ring because of
3533 * all the overhead involved.
3535 !(mode
& GL_MAP_WRITE_BIT
) &&
3537 (mt
->surf
.tiling
== ISL_TILING_X
||
3538 /* Prior to Sandybridge, the blitter can't handle Y tiling */
3539 (devinfo
->gen
>= 6 && mt
->surf
.tiling
== ISL_TILING_Y0
) ||
3540 /* Fast copy blit on skl+ supports all tiling formats. */
3541 devinfo
->gen
>= 9) &&
3542 can_blit_slice(mt
, level
, slice
))
3545 if (mt
->surf
.tiling
!= ISL_TILING_LINEAR
&&
3546 mt
->bo
->size
>= brw
->max_gtt_map_object_size
) {
3547 assert(can_blit_slice(mt
, level
, slice
));
3555 * Parameter \a out_stride has type ptrdiff_t not because the buffer stride may
3556 * exceed 32 bits but to diminish the likelihood subtle bugs in pointer
3557 * arithmetic overflow.
3559 * If you call this function and use \a out_stride, then you're doing pointer
3560 * arithmetic on \a out_ptr. The type of \a out_stride doesn't prevent all
3561 * bugs. The caller must still take care to avoid 32-bit overflow errors in
3562 * all arithmetic expressions that contain buffer offsets and pixel sizes,
3563 * which usually have type uint32_t or GLuint.
3566 intel_miptree_map(struct brw_context
*brw
,
3567 struct intel_mipmap_tree
*mt
,
3576 ptrdiff_t *out_stride
)
3578 struct intel_miptree_map
*map
;
3580 assert(mt
->surf
.samples
== 1);
3582 map
= intel_miptree_attach_map(mt
, level
, slice
, x
, y
, w
, h
, mode
);
3589 intel_miptree_access_raw(brw
, mt
, level
, slice
,
3590 map
->mode
& GL_MAP_WRITE_BIT
);
3592 if (mt
->format
== MESA_FORMAT_S_UINT8
) {
3593 intel_miptree_map_s8(brw
, mt
, map
, level
, slice
);
3594 } else if (mt
->etc_format
!= MESA_FORMAT_NONE
&&
3595 !(mode
& BRW_MAP_DIRECT_BIT
)) {
3596 intel_miptree_map_etc(brw
, mt
, map
, level
, slice
);
3597 } else if (mt
->stencil_mt
&& !(mode
& BRW_MAP_DIRECT_BIT
)) {
3598 intel_miptree_map_depthstencil(brw
, mt
, map
, level
, slice
);
3599 } else if (use_intel_mipree_map_blit(brw
, mt
, mode
, level
, slice
)) {
3600 intel_miptree_map_blit(brw
, mt
, map
, level
, slice
);
3601 #if defined(USE_SSE41)
3602 } else if (!(mode
& GL_MAP_WRITE_BIT
) &&
3603 !mt
->compressed
&& cpu_has_sse4_1
&&
3604 (mt
->surf
.row_pitch
% 16 == 0)) {
3605 intel_miptree_map_movntdqa(brw
, mt
, map
, level
, slice
);
3608 intel_miptree_map_gtt(brw
, mt
, map
, level
, slice
);
3611 *out_ptr
= map
->ptr
;
3612 *out_stride
= map
->stride
;
3614 if (map
->ptr
== NULL
)
3615 intel_miptree_release_map(mt
, level
, slice
);
3619 intel_miptree_unmap(struct brw_context
*brw
,
3620 struct intel_mipmap_tree
*mt
,
3624 struct intel_miptree_map
*map
= mt
->level
[level
].slice
[slice
].map
;
3626 assert(mt
->surf
.samples
== 1);
3631 DBG("%s: mt %p (%s) level %d slice %d\n", __func__
,
3632 mt
, _mesa_get_format_name(mt
->format
), level
, slice
);
3634 if (mt
->format
== MESA_FORMAT_S_UINT8
) {
3635 intel_miptree_unmap_s8(brw
, mt
, map
, level
, slice
);
3636 } else if (mt
->etc_format
!= MESA_FORMAT_NONE
&&
3637 !(map
->mode
& BRW_MAP_DIRECT_BIT
)) {
3638 intel_miptree_unmap_etc(brw
, mt
, map
, level
, slice
);
3639 } else if (mt
->stencil_mt
&& !(map
->mode
& BRW_MAP_DIRECT_BIT
)) {
3640 intel_miptree_unmap_depthstencil(brw
, mt
, map
, level
, slice
);
3641 } else if (map
->linear_mt
) {
3642 intel_miptree_unmap_blit(brw
, mt
, map
, level
, slice
);
3643 #if defined(USE_SSE41)
3644 } else if (map
->buffer
&& cpu_has_sse4_1
) {
3645 intel_miptree_unmap_movntdqa(brw
, mt
, map
, level
, slice
);
3648 intel_miptree_unmap_gtt(mt
);
3651 intel_miptree_release_map(mt
, level
, slice
);
3655 get_isl_surf_dim(GLenum target
)
3659 case GL_TEXTURE_1D_ARRAY
:
3660 return ISL_SURF_DIM_1D
;
3663 case GL_TEXTURE_2D_ARRAY
:
3664 case GL_TEXTURE_RECTANGLE
:
3665 case GL_TEXTURE_CUBE_MAP
:
3666 case GL_TEXTURE_CUBE_MAP_ARRAY
:
3667 case GL_TEXTURE_2D_MULTISAMPLE
:
3668 case GL_TEXTURE_2D_MULTISAMPLE_ARRAY
:
3669 case GL_TEXTURE_EXTERNAL_OES
:
3670 return ISL_SURF_DIM_2D
;
3673 return ISL_SURF_DIM_3D
;
3676 unreachable("Invalid texture target");
3680 get_isl_dim_layout(const struct gen_device_info
*devinfo
,
3681 enum isl_tiling tiling
, GLenum target
)
3685 case GL_TEXTURE_1D_ARRAY
:
3686 return (devinfo
->gen
>= 9 && tiling
== ISL_TILING_LINEAR
?
3687 ISL_DIM_LAYOUT_GEN9_1D
: ISL_DIM_LAYOUT_GEN4_2D
);
3690 case GL_TEXTURE_2D_ARRAY
:
3691 case GL_TEXTURE_RECTANGLE
:
3692 case GL_TEXTURE_2D_MULTISAMPLE
:
3693 case GL_TEXTURE_2D_MULTISAMPLE_ARRAY
:
3694 case GL_TEXTURE_EXTERNAL_OES
:
3695 return ISL_DIM_LAYOUT_GEN4_2D
;
3697 case GL_TEXTURE_CUBE_MAP
:
3698 case GL_TEXTURE_CUBE_MAP_ARRAY
:
3699 return (devinfo
->gen
== 4 ? ISL_DIM_LAYOUT_GEN4_3D
:
3700 ISL_DIM_LAYOUT_GEN4_2D
);
3703 return (devinfo
->gen
>= 9 ?
3704 ISL_DIM_LAYOUT_GEN4_2D
: ISL_DIM_LAYOUT_GEN4_3D
);
3707 unreachable("Invalid texture target");
3711 intel_miptree_get_aux_isl_usage(const struct brw_context
*brw
,
3712 const struct intel_mipmap_tree
*mt
)
3715 return ISL_AUX_USAGE_HIZ
;
3718 return ISL_AUX_USAGE_NONE
;
3720 return mt
->aux_usage
;