2 * Copyright 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "genxml/genX_bits.h"
36 #include "isl_gen12.h"
40 isl_memcpy_linear_to_tiled(uint32_t xt1
, uint32_t xt2
,
41 uint32_t yt1
, uint32_t yt2
,
42 char *dst
, const char *src
,
43 uint32_t dst_pitch
, int32_t src_pitch
,
45 enum isl_tiling tiling
,
46 isl_memcpy_type copy_type
)
49 if (copy_type
== ISL_MEMCPY_STREAMING_LOAD
) {
50 _isl_memcpy_linear_to_tiled_sse41(
51 xt1
, xt2
, yt1
, yt2
, dst
, src
, dst_pitch
, src_pitch
, has_swizzling
,
57 _isl_memcpy_linear_to_tiled(
58 xt1
, xt2
, yt1
, yt2
, dst
, src
, dst_pitch
, src_pitch
, has_swizzling
,
63 isl_memcpy_tiled_to_linear(uint32_t xt1
, uint32_t xt2
,
64 uint32_t yt1
, uint32_t yt2
,
65 char *dst
, const char *src
,
66 int32_t dst_pitch
, uint32_t src_pitch
,
68 enum isl_tiling tiling
,
69 isl_memcpy_type copy_type
)
72 if (copy_type
== ISL_MEMCPY_STREAMING_LOAD
) {
73 _isl_memcpy_tiled_to_linear_sse41(
74 xt1
, xt2
, yt1
, yt2
, dst
, src
, dst_pitch
, src_pitch
, has_swizzling
,
80 _isl_memcpy_tiled_to_linear(
81 xt1
, xt2
, yt1
, yt2
, dst
, src
, dst_pitch
, src_pitch
, has_swizzling
,
85 void PRINTFLIKE(3, 4) UNUSED
86 __isl_finishme(const char *file
, int line
, const char *fmt
, ...)
92 vsnprintf(buf
, sizeof(buf
), fmt
, ap
);
95 fprintf(stderr
, "%s:%d: FINISHME: %s\n", file
, line
, buf
);
99 isl_device_setup_mocs(struct isl_device
*dev
)
101 if (dev
->info
->gen
>= 12) {
102 /* TODO: Set PTE to MOCS 61 when the kernel is ready */
103 /* TC=1/LLC Only, LeCC=1/Uncacheable, LRUM=0, L3CC=1/Uncacheable */
104 dev
->mocs
.external
= 3 << 1;
105 /* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */
106 dev
->mocs
.internal
= 2 << 1;
107 } else if (dev
->info
->gen
>= 9) {
108 /* TC=LLC/eLLC, LeCC=PTE, LRUM=3, L3CC=WB */
109 dev
->mocs
.external
= 1 << 1;
110 /* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */
111 dev
->mocs
.internal
= 2 << 1;
112 } else if (dev
->info
->gen
>= 8) {
113 /* MEMORY_OBJECT_CONTROL_STATE:
114 * .MemoryTypeLLCeLLCCacheabilityControl = UCwithFenceifcoherentcycle,
115 * .TargetCache = L3DefertoPATforLLCeLLCselection,
118 dev
->mocs
.external
= 0x18;
119 /* MEMORY_OBJECT_CONTROL_STATE:
120 * .MemoryTypeLLCeLLCCacheabilityControl = WB,
121 * .TargetCache = L3DefertoPATforLLCeLLCselection,
124 dev
->mocs
.internal
= 0x78;
125 } else if (dev
->info
->gen
>= 7) {
126 if (dev
->info
->is_haswell
) {
127 /* MEMORY_OBJECT_CONTROL_STATE:
128 * .LLCeLLCCacheabilityControlLLCCC = 0,
129 * .L3CacheabilityControlL3CC = 1,
131 dev
->mocs
.internal
= 1;
132 dev
->mocs
.external
= 1;
134 /* MEMORY_OBJECT_CONTROL_STATE:
135 * .GraphicsDataTypeGFDT = 0,
136 * .LLCCacheabilityControlLLCCC = 0,
137 * .L3CacheabilityControlL3CC = 1,
139 dev
->mocs
.internal
= 1;
140 dev
->mocs
.external
= 1;
143 dev
->mocs
.internal
= 0;
144 dev
->mocs
.external
= 0;
149 isl_device_init(struct isl_device
*dev
,
150 const struct gen_device_info
*info
,
151 bool has_bit6_swizzling
)
153 /* Gen8+ don't have bit6 swizzling, ensure callsite is not confused. */
154 assert(!(has_bit6_swizzling
&& info
->gen
>= 8));
157 dev
->use_separate_stencil
= ISL_DEV_GEN(dev
) >= 6;
158 dev
->has_bit6_swizzling
= has_bit6_swizzling
;
160 /* The ISL_DEV macros may be defined in the CFLAGS, thus hardcoding some
161 * device properties at buildtime. Verify that the macros with the device
162 * properties chosen during runtime.
164 ISL_DEV_GEN_SANITIZE(dev
);
165 ISL_DEV_USE_SEPARATE_STENCIL_SANITIZE(dev
);
167 /* Did we break hiz or stencil? */
168 if (ISL_DEV_USE_SEPARATE_STENCIL(dev
))
169 assert(info
->has_hiz_and_separate_stencil
);
170 if (info
->must_use_separate_stencil
)
171 assert(ISL_DEV_USE_SEPARATE_STENCIL(dev
));
173 dev
->ss
.size
= RENDER_SURFACE_STATE_length(info
) * 4;
174 dev
->ss
.align
= isl_align(dev
->ss
.size
, 32);
176 dev
->ss
.clear_color_state_size
=
177 isl_align(CLEAR_COLOR_length(info
) * 4, 64);
178 dev
->ss
.clear_color_state_offset
=
179 RENDER_SURFACE_STATE_ClearValueAddress_start(info
) / 32 * 4;
181 dev
->ss
.clear_value_size
=
182 isl_align(RENDER_SURFACE_STATE_RedClearColor_bits(info
) +
183 RENDER_SURFACE_STATE_GreenClearColor_bits(info
) +
184 RENDER_SURFACE_STATE_BlueClearColor_bits(info
) +
185 RENDER_SURFACE_STATE_AlphaClearColor_bits(info
), 32) / 8;
187 dev
->ss
.clear_value_offset
=
188 RENDER_SURFACE_STATE_RedClearColor_start(info
) / 32 * 4;
190 assert(RENDER_SURFACE_STATE_SurfaceBaseAddress_start(info
) % 8 == 0);
191 dev
->ss
.addr_offset
=
192 RENDER_SURFACE_STATE_SurfaceBaseAddress_start(info
) / 8;
194 /* The "Auxiliary Surface Base Address" field starts a bit higher up
195 * because the bottom 12 bits are used for other things. Round down to
196 * the nearest dword before.
198 dev
->ss
.aux_addr_offset
=
199 (RENDER_SURFACE_STATE_AuxiliarySurfaceBaseAddress_start(info
) & ~31) / 8;
201 dev
->ds
.size
= _3DSTATE_DEPTH_BUFFER_length(info
) * 4;
202 assert(_3DSTATE_DEPTH_BUFFER_SurfaceBaseAddress_start(info
) % 8 == 0);
203 dev
->ds
.depth_offset
=
204 _3DSTATE_DEPTH_BUFFER_SurfaceBaseAddress_start(info
) / 8;
206 if (dev
->use_separate_stencil
) {
207 dev
->ds
.size
+= _3DSTATE_STENCIL_BUFFER_length(info
) * 4 +
208 _3DSTATE_HIER_DEPTH_BUFFER_length(info
) * 4 +
209 _3DSTATE_CLEAR_PARAMS_length(info
) * 4;
211 assert(_3DSTATE_STENCIL_BUFFER_SurfaceBaseAddress_start(info
) % 8 == 0);
212 dev
->ds
.stencil_offset
=
213 _3DSTATE_DEPTH_BUFFER_length(info
) * 4 +
214 _3DSTATE_STENCIL_BUFFER_SurfaceBaseAddress_start(info
) / 8;
216 assert(_3DSTATE_HIER_DEPTH_BUFFER_SurfaceBaseAddress_start(info
) % 8 == 0);
218 _3DSTATE_DEPTH_BUFFER_length(info
) * 4 +
219 _3DSTATE_STENCIL_BUFFER_length(info
) * 4 +
220 _3DSTATE_HIER_DEPTH_BUFFER_SurfaceBaseAddress_start(info
) / 8;
222 dev
->ds
.stencil_offset
= 0;
223 dev
->ds
.hiz_offset
= 0;
226 if (ISL_DEV_GEN(dev
) >= 12) {
227 dev
->ds
.size
+= GEN12_MI_LOAD_REGISTER_IMM_length
* 4 * 2;
230 isl_device_setup_mocs(dev
);
234 * @brief Query the set of multisamples supported by the device.
236 * This function always returns non-zero, as ISL_SAMPLE_COUNT_1_BIT is always
239 isl_sample_count_mask_t ATTRIBUTE_CONST
240 isl_device_get_sample_counts(struct isl_device
*dev
)
242 if (ISL_DEV_GEN(dev
) >= 9) {
243 return ISL_SAMPLE_COUNT_1_BIT
|
244 ISL_SAMPLE_COUNT_2_BIT
|
245 ISL_SAMPLE_COUNT_4_BIT
|
246 ISL_SAMPLE_COUNT_8_BIT
|
247 ISL_SAMPLE_COUNT_16_BIT
;
248 } else if (ISL_DEV_GEN(dev
) >= 8) {
249 return ISL_SAMPLE_COUNT_1_BIT
|
250 ISL_SAMPLE_COUNT_2_BIT
|
251 ISL_SAMPLE_COUNT_4_BIT
|
252 ISL_SAMPLE_COUNT_8_BIT
;
253 } else if (ISL_DEV_GEN(dev
) >= 7) {
254 return ISL_SAMPLE_COUNT_1_BIT
|
255 ISL_SAMPLE_COUNT_4_BIT
|
256 ISL_SAMPLE_COUNT_8_BIT
;
257 } else if (ISL_DEV_GEN(dev
) >= 6) {
258 return ISL_SAMPLE_COUNT_1_BIT
|
259 ISL_SAMPLE_COUNT_4_BIT
;
261 return ISL_SAMPLE_COUNT_1_BIT
;
266 * @param[out] info is written only on success
269 isl_tiling_get_info(enum isl_tiling tiling
,
271 struct isl_tile_info
*tile_info
)
273 const uint32_t bs
= format_bpb
/ 8;
274 struct isl_extent2d logical_el
, phys_B
;
276 if (tiling
!= ISL_TILING_LINEAR
&& !isl_is_pow2(format_bpb
)) {
277 /* It is possible to have non-power-of-two formats in a tiled buffer.
278 * The easiest way to handle this is to treat the tile as if it is three
279 * times as wide. This way no pixel will ever cross a tile boundary.
280 * This really only works on legacy X and Y tiling formats.
282 assert(tiling
== ISL_TILING_X
|| tiling
== ISL_TILING_Y0
);
283 assert(bs
% 3 == 0 && isl_is_pow2(format_bpb
/ 3));
284 isl_tiling_get_info(tiling
, format_bpb
/ 3, tile_info
);
289 case ISL_TILING_LINEAR
:
291 logical_el
= isl_extent2d(1, 1);
292 phys_B
= isl_extent2d(bs
, 1);
297 logical_el
= isl_extent2d(512 / bs
, 8);
298 phys_B
= isl_extent2d(512, 8);
303 logical_el
= isl_extent2d(128 / bs
, 32);
304 phys_B
= isl_extent2d(128, 32);
309 logical_el
= isl_extent2d(64, 64);
310 /* From the Broadwell PRM Vol 2d, RENDER_SURFACE_STATE::SurfacePitch:
312 * "If the surface is a stencil buffer (and thus has Tile Mode set
313 * to TILEMODE_WMAJOR), the pitch must be set to 2x the value
314 * computed based on width, as the stencil buffer is stored with two
317 * This, together with the fact that stencil buffers are referred to as
318 * being Y-tiled in the PRMs for older hardware implies that the
319 * physical size of a W-tile is actually the same as for a Y-tile.
321 phys_B
= isl_extent2d(128, 32);
325 case ISL_TILING_Ys
: {
326 bool is_Ys
= tiling
== ISL_TILING_Ys
;
329 unsigned width
= 1 << (6 + (ffs(bs
) / 2) + (2 * is_Ys
));
330 unsigned height
= 1 << (6 - (ffs(bs
) / 2) + (2 * is_Ys
));
332 logical_el
= isl_extent2d(width
/ bs
, height
);
333 phys_B
= isl_extent2d(width
, height
);
338 /* HiZ buffers are required to have ISL_FORMAT_HIZ which is an 8x4
339 * 128bpb format. The tiling has the same physical dimensions as
340 * Y-tiling but actually has two HiZ columns per Y-tiled column.
343 logical_el
= isl_extent2d(16, 16);
344 phys_B
= isl_extent2d(128, 32);
348 /* CCS surfaces are required to have one of the GENX_CCS_* formats which
349 * have a block size of 1 or 2 bits per block and each CCS element
350 * corresponds to one cache-line pair in the main surface. From the Sky
351 * Lake PRM Vol. 12 in the section on planes:
353 * "The Color Control Surface (CCS) contains the compression status
354 * of the cache-line pairs. The compression state of the cache-line
355 * pair is specified by 2 bits in the CCS. Each CCS cache-line
356 * represents an area on the main surface of 16x16 sets of 128 byte
357 * Y-tiled cache-line-pairs. CCS is always Y tiled."
359 * The CCS being Y-tiled implies that it's an 8x8 grid of cache-lines.
360 * Since each cache line corresponds to a 16x16 set of cache-line pairs,
361 * that yields total tile area of 128x128 cache-line pairs or CCS
362 * elements. On older hardware, each CCS element is 1 bit and the tile
363 * is 128x256 elements.
365 assert(format_bpb
== 1 || format_bpb
== 2);
366 logical_el
= isl_extent2d(128, 256 / format_bpb
);
367 phys_B
= isl_extent2d(128, 32);
370 case ISL_TILING_GEN12_CCS
:
371 /* From the Bspec, Gen Graphics > Gen12 > Memory Data Formats > Memory
372 * Compression > Memory Compression - Gen12:
374 * 4 bits of auxiliary plane data are required for 2 cachelines of
375 * main surface data. This results in a single cacheline of auxiliary
376 * plane data mapping to 4 4K pages of main surface data for the 4K
377 * pages (tile Y ) and 1 64K Tile Ys page.
379 * The Y-tiled pairing bit of 9 shown in the table below that Bspec
380 * section expresses that the 2 cachelines of main surface data are
381 * horizontally adjacent.
383 * TODO: Handle Ys, Yf and their pairing bits.
385 * Therefore, each CCS cacheline represents a 512Bx32 row area and each
386 * element represents a 32Bx4 row area.
388 assert(format_bpb
== 4);
389 logical_el
= isl_extent2d(16, 8);
390 phys_B
= isl_extent2d(64, 1);
394 unreachable("not reached");
397 *tile_info
= (struct isl_tile_info
) {
399 .format_bpb
= format_bpb
,
400 .logical_extent_el
= logical_el
,
401 .phys_extent_B
= phys_B
,
406 isl_color_value_is_zero(union isl_color_value value
,
407 enum isl_format format
)
409 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
411 #define RETURN_FALSE_IF_NOT_0(c, i) \
412 if (fmtl->channels.c.bits && value.u32[i] != 0) \
415 RETURN_FALSE_IF_NOT_0(r
, 0);
416 RETURN_FALSE_IF_NOT_0(g
, 1);
417 RETURN_FALSE_IF_NOT_0(b
, 2);
418 RETURN_FALSE_IF_NOT_0(a
, 3);
420 #undef RETURN_FALSE_IF_NOT_0
426 isl_color_value_is_zero_one(union isl_color_value value
,
427 enum isl_format format
)
429 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
431 #define RETURN_FALSE_IF_NOT_0_1(c, i, field) \
432 if (fmtl->channels.c.bits && value.field[i] != 0 && value.field[i] != 1) \
435 if (isl_format_has_int_channel(format
)) {
436 RETURN_FALSE_IF_NOT_0_1(r
, 0, u32
);
437 RETURN_FALSE_IF_NOT_0_1(g
, 1, u32
);
438 RETURN_FALSE_IF_NOT_0_1(b
, 2, u32
);
439 RETURN_FALSE_IF_NOT_0_1(a
, 3, u32
);
441 RETURN_FALSE_IF_NOT_0_1(r
, 0, f32
);
442 RETURN_FALSE_IF_NOT_0_1(g
, 1, f32
);
443 RETURN_FALSE_IF_NOT_0_1(b
, 2, f32
);
444 RETURN_FALSE_IF_NOT_0_1(a
, 3, f32
);
447 #undef RETURN_FALSE_IF_NOT_0_1
453 * @param[out] tiling is set only on success
456 isl_surf_choose_tiling(const struct isl_device
*dev
,
457 const struct isl_surf_init_info
*restrict info
,
458 enum isl_tiling
*tiling
)
460 isl_tiling_flags_t tiling_flags
= info
->tiling_flags
;
462 /* HiZ surfaces always use the HiZ tiling */
463 if (info
->usage
& ISL_SURF_USAGE_HIZ_BIT
) {
464 assert(info
->format
== ISL_FORMAT_HIZ
);
465 assert(tiling_flags
== ISL_TILING_HIZ_BIT
);
466 *tiling
= isl_tiling_flag_to_enum(tiling_flags
);
470 /* CCS surfaces always use the CCS tiling */
471 if (info
->usage
& ISL_SURF_USAGE_CCS_BIT
) {
472 assert(isl_format_get_layout(info
->format
)->txc
== ISL_TXC_CCS
);
473 UNUSED
bool ivb_ccs
= ISL_DEV_GEN(dev
) < 12 &&
474 tiling_flags
== ISL_TILING_CCS_BIT
;
475 UNUSED
bool tgl_ccs
= ISL_DEV_GEN(dev
) >= 12 &&
476 tiling_flags
== ISL_TILING_GEN12_CCS_BIT
;
477 assert(ivb_ccs
!= tgl_ccs
);
478 *tiling
= isl_tiling_flag_to_enum(tiling_flags
);
482 if (ISL_DEV_GEN(dev
) >= 6) {
483 isl_gen6_filter_tiling(dev
, info
, &tiling_flags
);
485 isl_gen4_filter_tiling(dev
, info
, &tiling_flags
);
488 #define CHOOSE(__tiling) \
490 if (tiling_flags & (1u << (__tiling))) { \
491 *tiling = (__tiling); \
496 /* Of the tiling modes remaining, choose the one that offers the best
500 if (info
->dim
== ISL_SURF_DIM_1D
) {
501 /* Prefer linear for 1D surfaces because they do not benefit from
502 * tiling. To the contrary, tiling leads to wasted memory and poor
503 * memory locality due to the swizzling and alignment restrictions
504 * required in tiled surfaces.
506 CHOOSE(ISL_TILING_LINEAR
);
509 CHOOSE(ISL_TILING_Ys
);
510 CHOOSE(ISL_TILING_Yf
);
511 CHOOSE(ISL_TILING_Y0
);
512 CHOOSE(ISL_TILING_X
);
513 CHOOSE(ISL_TILING_W
);
514 CHOOSE(ISL_TILING_LINEAR
);
518 /* No tiling mode accomodates the inputs. */
523 isl_choose_msaa_layout(const struct isl_device
*dev
,
524 const struct isl_surf_init_info
*info
,
525 enum isl_tiling tiling
,
526 enum isl_msaa_layout
*msaa_layout
)
528 if (ISL_DEV_GEN(dev
) >= 8) {
529 return isl_gen8_choose_msaa_layout(dev
, info
, tiling
, msaa_layout
);
530 } else if (ISL_DEV_GEN(dev
) >= 7) {
531 return isl_gen7_choose_msaa_layout(dev
, info
, tiling
, msaa_layout
);
532 } else if (ISL_DEV_GEN(dev
) >= 6) {
533 return isl_gen6_choose_msaa_layout(dev
, info
, tiling
, msaa_layout
);
535 return isl_gen4_choose_msaa_layout(dev
, info
, tiling
, msaa_layout
);
540 isl_get_interleaved_msaa_px_size_sa(uint32_t samples
)
542 assert(isl_is_pow2(samples
));
544 /* From the Broadwell PRM >> Volume 5: Memory Views >> Computing Mip Level
547 * If the surface is multisampled and it is a depth or stencil surface
548 * or Multisampled Surface StorageFormat in SURFACE_STATE is
549 * MSFMT_DEPTH_STENCIL, W_L and H_L must be adjusted as follows before
552 return (struct isl_extent2d
) {
553 .width
= 1 << ((ffs(samples
) - 0) / 2),
554 .height
= 1 << ((ffs(samples
) - 1) / 2),
559 isl_msaa_interleaved_scale_px_to_sa(uint32_t samples
,
560 uint32_t *width
, uint32_t *height
)
562 const struct isl_extent2d px_size_sa
=
563 isl_get_interleaved_msaa_px_size_sa(samples
);
566 *width
= isl_align(*width
, 2) * px_size_sa
.width
;
568 *height
= isl_align(*height
, 2) * px_size_sa
.height
;
571 static enum isl_array_pitch_span
572 isl_choose_array_pitch_span(const struct isl_device
*dev
,
573 const struct isl_surf_init_info
*restrict info
,
574 enum isl_dim_layout dim_layout
,
575 const struct isl_extent4d
*phys_level0_sa
)
577 switch (dim_layout
) {
578 case ISL_DIM_LAYOUT_GEN9_1D
:
579 case ISL_DIM_LAYOUT_GEN4_2D
:
580 if (ISL_DEV_GEN(dev
) >= 8) {
581 /* QPitch becomes programmable in Broadwell. So choose the
582 * most compact QPitch possible in order to conserve memory.
584 * From the Broadwell PRM >> Volume 2d: Command Reference: Structures
585 * >> RENDER_SURFACE_STATE Surface QPitch (p325):
587 * - Software must ensure that this field is set to a value
588 * sufficiently large such that the array slices in the surface
589 * do not overlap. Refer to the Memory Data Formats section for
590 * information on how surfaces are stored in memory.
592 * - This field specifies the distance in rows between array
593 * slices. It is used only in the following cases:
595 * - Surface Array is enabled OR
596 * - Number of Mulitsamples is not NUMSAMPLES_1 and
597 * Multisampled Surface Storage Format set to MSFMT_MSS OR
598 * - Surface Type is SURFTYPE_CUBE
600 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
601 } else if (ISL_DEV_GEN(dev
) >= 7) {
602 /* Note that Ivybridge introduces
603 * RENDER_SURFACE_STATE.SurfaceArraySpacing, which provides the
604 * driver more control over the QPitch.
607 if (phys_level0_sa
->array_len
== 1) {
608 /* The hardware will never use the QPitch. So choose the most
609 * compact QPitch possible in order to conserve memory.
611 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
614 if (isl_surf_usage_is_depth_or_stencil(info
->usage
) ||
615 (info
->usage
& ISL_SURF_USAGE_HIZ_BIT
)) {
616 /* From the Ivybridge PRM >> Volume 1 Part 1: Graphics Core >>
617 * Section 6.18.4.7: Surface Arrays (p112):
619 * If Surface Array Spacing is set to ARYSPC_FULL (note that
620 * the depth buffer and stencil buffer have an implied value of
623 return ISL_ARRAY_PITCH_SPAN_FULL
;
626 if (info
->levels
== 1) {
627 /* We are able to set RENDER_SURFACE_STATE.SurfaceArraySpacing
630 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
633 return ISL_ARRAY_PITCH_SPAN_FULL
;
634 } else if ((ISL_DEV_GEN(dev
) == 5 || ISL_DEV_GEN(dev
) == 6) &&
635 ISL_DEV_USE_SEPARATE_STENCIL(dev
) &&
636 isl_surf_usage_is_stencil(info
->usage
)) {
637 /* [ILK-SNB] Errata from the Sandy Bridge PRM >> Volume 4 Part 1:
638 * Graphics Core >> Section 7.18.3.7: Surface Arrays:
640 * The separate stencil buffer does not support mip mapping, thus
641 * the storage for LODs other than LOD 0 is not needed.
643 assert(info
->levels
== 1);
644 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
646 if ((ISL_DEV_GEN(dev
) == 5 || ISL_DEV_GEN(dev
) == 6) &&
647 ISL_DEV_USE_SEPARATE_STENCIL(dev
) &&
648 isl_surf_usage_is_stencil(info
->usage
)) {
649 /* [ILK-SNB] Errata from the Sandy Bridge PRM >> Volume 4 Part 1:
650 * Graphics Core >> Section 7.18.3.7: Surface Arrays:
652 * The separate stencil buffer does not support mip mapping,
653 * thus the storage for LODs other than LOD 0 is not needed.
655 assert(info
->levels
== 1);
656 assert(phys_level0_sa
->array_len
== 1);
657 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
660 if (phys_level0_sa
->array_len
== 1) {
661 /* The hardware will never use the QPitch. So choose the most
662 * compact QPitch possible in order to conserve memory.
664 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
667 return ISL_ARRAY_PITCH_SPAN_FULL
;
670 case ISL_DIM_LAYOUT_GEN4_3D
:
671 /* The hardware will never use the QPitch. So choose the most
672 * compact QPitch possible in order to conserve memory.
674 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
676 case ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
:
677 /* Each array image in the gen6 stencil of HiZ surface is compact in the
678 * sense that every LOD is a compact array of the same size as LOD0.
680 return ISL_ARRAY_PITCH_SPAN_COMPACT
;
683 unreachable("bad isl_dim_layout");
684 return ISL_ARRAY_PITCH_SPAN_FULL
;
688 isl_choose_image_alignment_el(const struct isl_device
*dev
,
689 const struct isl_surf_init_info
*restrict info
,
690 enum isl_tiling tiling
,
691 enum isl_dim_layout dim_layout
,
692 enum isl_msaa_layout msaa_layout
,
693 struct isl_extent3d
*image_align_el
)
695 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
696 if (fmtl
->txc
== ISL_TXC_MCS
) {
697 assert(tiling
== ISL_TILING_Y0
);
700 * IvyBrigde PRM Vol 2, Part 1, "11.7 MCS Buffer for Render Target(s)":
702 * Height, width, and layout of MCS buffer in this case must match with
703 * Render Target height, width, and layout. MCS buffer is tiledY.
705 * To avoid wasting memory, choose the smallest alignment possible:
706 * HALIGN_4 and VALIGN_4.
708 *image_align_el
= isl_extent3d(4, 4, 1);
710 } else if (info
->format
== ISL_FORMAT_HIZ
) {
711 assert(ISL_DEV_GEN(dev
) >= 6);
712 if (ISL_DEV_GEN(dev
) == 6) {
713 /* HiZ surfaces on Sandy Bridge are packed tightly. */
714 *image_align_el
= isl_extent3d(1, 1, 1);
715 } else if (ISL_DEV_GEN(dev
) < 12) {
716 /* On gen7+, HiZ surfaces are always aligned to 16x8 pixels in the
717 * primary surface which works out to 2x2 HiZ elments.
719 *image_align_el
= isl_extent3d(2, 2, 1);
721 /* On gen12+, HiZ surfaces are always aligned to 16x16 pixels in the
722 * primary surface which works out to 2x4 HiZ elments.
725 *image_align_el
= isl_extent3d(2, 4, 1);
730 if (ISL_DEV_GEN(dev
) >= 12) {
731 isl_gen12_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
732 msaa_layout
, image_align_el
);
733 } else if (ISL_DEV_GEN(dev
) >= 9) {
734 isl_gen9_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
735 msaa_layout
, image_align_el
);
736 } else if (ISL_DEV_GEN(dev
) >= 8) {
737 isl_gen8_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
738 msaa_layout
, image_align_el
);
739 } else if (ISL_DEV_GEN(dev
) >= 7) {
740 isl_gen7_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
741 msaa_layout
, image_align_el
);
742 } else if (ISL_DEV_GEN(dev
) >= 6) {
743 isl_gen6_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
744 msaa_layout
, image_align_el
);
746 isl_gen4_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
,
747 msaa_layout
, image_align_el
);
751 static enum isl_dim_layout
752 isl_surf_choose_dim_layout(const struct isl_device
*dev
,
753 enum isl_surf_dim logical_dim
,
754 enum isl_tiling tiling
,
755 isl_surf_usage_flags_t usage
)
757 /* Sandy bridge needs a special layout for HiZ and stencil. */
758 if (ISL_DEV_GEN(dev
) == 6 &&
759 (tiling
== ISL_TILING_W
|| tiling
== ISL_TILING_HIZ
))
760 return ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
;
762 if (ISL_DEV_GEN(dev
) >= 9) {
763 switch (logical_dim
) {
764 case ISL_SURF_DIM_1D
:
765 /* From the Sky Lake PRM Vol. 5, "1D Surfaces":
767 * One-dimensional surfaces use a tiling mode of linear.
768 * Technically, they are not tiled resources, but the Tiled
769 * Resource Mode field in RENDER_SURFACE_STATE is still used to
770 * indicate the alignment requirements for this linear surface
771 * (See 1D Alignment requirements for how 4K and 64KB Tiled
772 * Resource Modes impact alignment). Alternatively, a 1D surface
773 * can be defined as a 2D tiled surface (e.g. TileY or TileX) with
776 * In other words, ISL_DIM_LAYOUT_GEN9_1D is only used for linear
777 * surfaces and, for tiled surfaces, ISL_DIM_LAYOUT_GEN4_2D is used.
779 if (tiling
== ISL_TILING_LINEAR
)
780 return ISL_DIM_LAYOUT_GEN9_1D
;
782 return ISL_DIM_LAYOUT_GEN4_2D
;
783 case ISL_SURF_DIM_2D
:
784 case ISL_SURF_DIM_3D
:
785 return ISL_DIM_LAYOUT_GEN4_2D
;
788 switch (logical_dim
) {
789 case ISL_SURF_DIM_1D
:
790 case ISL_SURF_DIM_2D
:
791 /* From the G45 PRM Vol. 1a, "6.17.4.1 Hardware Cube Map Layout":
793 * The cube face textures are stored in the same way as 3D surfaces
794 * are stored (see section 6.17.5 for details). For cube surfaces,
795 * however, the depth is equal to the number of faces (always 6) and
796 * is not reduced for each MIP.
798 if (ISL_DEV_GEN(dev
) == 4 && (usage
& ISL_SURF_USAGE_CUBE_BIT
))
799 return ISL_DIM_LAYOUT_GEN4_3D
;
801 return ISL_DIM_LAYOUT_GEN4_2D
;
802 case ISL_SURF_DIM_3D
:
803 return ISL_DIM_LAYOUT_GEN4_3D
;
807 unreachable("bad isl_surf_dim");
808 return ISL_DIM_LAYOUT_GEN4_2D
;
812 * Calculate the physical extent of the surface's first level, in units of
816 isl_calc_phys_level0_extent_sa(const struct isl_device
*dev
,
817 const struct isl_surf_init_info
*restrict info
,
818 enum isl_dim_layout dim_layout
,
819 enum isl_tiling tiling
,
820 enum isl_msaa_layout msaa_layout
,
821 struct isl_extent4d
*phys_level0_sa
)
823 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
825 if (isl_format_is_planar(info
->format
))
826 unreachable("Planar formats unsupported");
829 case ISL_SURF_DIM_1D
:
830 assert(info
->height
== 1);
831 assert(info
->depth
== 1);
832 assert(info
->samples
== 1);
834 switch (dim_layout
) {
835 case ISL_DIM_LAYOUT_GEN4_3D
:
836 unreachable("bad isl_dim_layout");
838 case ISL_DIM_LAYOUT_GEN9_1D
:
839 case ISL_DIM_LAYOUT_GEN4_2D
:
840 case ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
:
841 *phys_level0_sa
= (struct isl_extent4d
) {
845 .a
= info
->array_len
,
851 case ISL_SURF_DIM_2D
:
852 if (ISL_DEV_GEN(dev
) == 4 && (info
->usage
& ISL_SURF_USAGE_CUBE_BIT
))
853 assert(dim_layout
== ISL_DIM_LAYOUT_GEN4_3D
);
855 assert(dim_layout
== ISL_DIM_LAYOUT_GEN4_2D
||
856 dim_layout
== ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
);
858 if (tiling
== ISL_TILING_Ys
&& info
->samples
> 1)
859 isl_finishme("%s:%s: multisample TileYs layout", __FILE__
, __func__
);
861 switch (msaa_layout
) {
862 case ISL_MSAA_LAYOUT_NONE
:
863 assert(info
->depth
== 1);
864 assert(info
->samples
== 1);
866 *phys_level0_sa
= (struct isl_extent4d
) {
870 .a
= info
->array_len
,
874 case ISL_MSAA_LAYOUT_ARRAY
:
875 assert(info
->depth
== 1);
876 assert(info
->levels
== 1);
877 assert(isl_format_supports_multisampling(dev
->info
, info
->format
));
878 assert(fmtl
->bw
== 1 && fmtl
->bh
== 1);
880 *phys_level0_sa
= (struct isl_extent4d
) {
884 .a
= info
->array_len
* info
->samples
,
888 case ISL_MSAA_LAYOUT_INTERLEAVED
:
889 assert(info
->depth
== 1);
890 assert(info
->levels
== 1);
891 assert(isl_format_supports_multisampling(dev
->info
, info
->format
));
893 *phys_level0_sa
= (struct isl_extent4d
) {
897 .a
= info
->array_len
,
900 isl_msaa_interleaved_scale_px_to_sa(info
->samples
,
907 case ISL_SURF_DIM_3D
:
908 assert(info
->array_len
== 1);
909 assert(info
->samples
== 1);
912 isl_finishme("%s:%s: compression block with depth > 1",
916 switch (dim_layout
) {
917 case ISL_DIM_LAYOUT_GEN9_1D
:
918 case ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
:
919 unreachable("bad isl_dim_layout");
921 case ISL_DIM_LAYOUT_GEN4_2D
:
922 assert(ISL_DEV_GEN(dev
) >= 9);
924 *phys_level0_sa
= (struct isl_extent4d
) {
932 case ISL_DIM_LAYOUT_GEN4_3D
:
933 assert(ISL_DEV_GEN(dev
) < 9);
934 *phys_level0_sa
= (struct isl_extent4d
) {
947 * Calculate the pitch between physical array slices, in units of rows of
951 isl_calc_array_pitch_el_rows_gen4_2d(
952 const struct isl_device
*dev
,
953 const struct isl_surf_init_info
*restrict info
,
954 const struct isl_tile_info
*tile_info
,
955 const struct isl_extent3d
*image_align_sa
,
956 const struct isl_extent4d
*phys_level0_sa
,
957 enum isl_array_pitch_span array_pitch_span
,
958 const struct isl_extent2d
*phys_slice0_sa
)
960 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
961 uint32_t pitch_sa_rows
= 0;
963 switch (array_pitch_span
) {
964 case ISL_ARRAY_PITCH_SPAN_COMPACT
:
965 pitch_sa_rows
= isl_align_npot(phys_slice0_sa
->h
, image_align_sa
->h
);
967 case ISL_ARRAY_PITCH_SPAN_FULL
: {
968 /* The QPitch equation is found in the Broadwell PRM >> Volume 5:
969 * Memory Views >> Common Surface Formats >> Surface Layout >> 2D
970 * Surfaces >> Surface Arrays.
972 uint32_t H0_sa
= phys_level0_sa
->h
;
973 uint32_t H1_sa
= isl_minify(H0_sa
, 1);
975 uint32_t h0_sa
= isl_align_npot(H0_sa
, image_align_sa
->h
);
976 uint32_t h1_sa
= isl_align_npot(H1_sa
, image_align_sa
->h
);
979 if (ISL_DEV_GEN(dev
) >= 7) {
980 /* The QPitch equation changed slightly in Ivybridge. */
986 pitch_sa_rows
= h0_sa
+ h1_sa
+ (m
* image_align_sa
->h
);
988 if (ISL_DEV_GEN(dev
) == 6 && info
->samples
> 1 &&
989 (info
->height
% 4 == 1)) {
990 /* [SNB] Errata from the Sandy Bridge PRM >> Volume 4 Part 1:
991 * Graphics Core >> Section 7.18.3.7: Surface Arrays:
993 * [SNB] Errata: Sampler MSAA Qpitch will be 4 greater than
994 * the value calculated in the equation above , for every
995 * other odd Surface Height starting from 1 i.e. 1,5,9,13.
997 * XXX(chadv): Is the errata natural corollary of the physical
998 * layout of interleaved samples?
1003 pitch_sa_rows
= isl_align_npot(pitch_sa_rows
, fmtl
->bh
);
1008 assert(pitch_sa_rows
% fmtl
->bh
== 0);
1009 uint32_t pitch_el_rows
= pitch_sa_rows
/ fmtl
->bh
;
1011 if (ISL_DEV_GEN(dev
) >= 9 && ISL_DEV_GEN(dev
) <= 11 &&
1012 fmtl
->txc
== ISL_TXC_CCS
) {
1014 * From the Sky Lake PRM Vol 7, "MCS Buffer for Render Target(s)" (p. 632):
1016 * "Mip-mapped and arrayed surfaces are supported with MCS buffer
1017 * layout with these alignments in the RT space: Horizontal
1018 * Alignment = 128 and Vertical Alignment = 64."
1020 * From the Sky Lake PRM Vol. 2d, "RENDER_SURFACE_STATE" (p. 435):
1022 * "For non-multisampled render target's CCS auxiliary surface,
1023 * QPitch must be computed with Horizontal Alignment = 128 and
1024 * Surface Vertical Alignment = 256. These alignments are only for
1025 * CCS buffer and not for associated render target."
1027 * The first restriction is already handled by isl_choose_image_alignment_el
1028 * but the second restriction, which is an extension of the first, only
1029 * applies to qpitch and must be applied here.
1031 * The second restriction disappears on Gen12.
1033 assert(fmtl
->bh
== 4);
1034 pitch_el_rows
= isl_align(pitch_el_rows
, 256 / 4);
1037 if (ISL_DEV_GEN(dev
) >= 9 &&
1038 info
->dim
== ISL_SURF_DIM_3D
&&
1039 tile_info
->tiling
!= ISL_TILING_LINEAR
) {
1040 /* From the Skylake BSpec >> RENDER_SURFACE_STATE >> Surface QPitch:
1042 * Tile Mode != Linear: This field must be set to an integer multiple
1043 * of the tile height
1045 pitch_el_rows
= isl_align(pitch_el_rows
, tile_info
->logical_extent_el
.height
);
1048 return pitch_el_rows
;
1052 * A variant of isl_calc_phys_slice0_extent_sa() specific to
1053 * ISL_DIM_LAYOUT_GEN4_2D.
1056 isl_calc_phys_slice0_extent_sa_gen4_2d(
1057 const struct isl_device
*dev
,
1058 const struct isl_surf_init_info
*restrict info
,
1059 enum isl_msaa_layout msaa_layout
,
1060 const struct isl_extent3d
*image_align_sa
,
1061 const struct isl_extent4d
*phys_level0_sa
,
1062 struct isl_extent2d
*phys_slice0_sa
)
1064 assert(phys_level0_sa
->depth
== 1);
1066 if (info
->levels
== 1) {
1067 /* Do not pad the surface to the image alignment.
1069 * For tiled surfaces, using a reduced alignment here avoids wasting CPU
1070 * cycles on the below mipmap layout caluclations. Reducing the
1071 * alignment here is safe because we later align the row pitch and array
1072 * pitch to the tile boundary. It is safe even for
1073 * ISL_MSAA_LAYOUT_INTERLEAVED, because phys_level0_sa is already scaled
1074 * to accomodate the interleaved samples.
1076 * For linear surfaces, reducing the alignment here permits us to later
1077 * choose an arbitrary, non-aligned row pitch. If the surface backs
1078 * a VkBuffer, then an arbitrary pitch may be needed to accomodate
1079 * VkBufferImageCopy::bufferRowLength.
1081 *phys_slice0_sa
= (struct isl_extent2d
) {
1082 .w
= phys_level0_sa
->w
,
1083 .h
= phys_level0_sa
->h
,
1088 uint32_t slice_top_w
= 0;
1089 uint32_t slice_bottom_w
= 0;
1090 uint32_t slice_left_h
= 0;
1091 uint32_t slice_right_h
= 0;
1093 uint32_t W0
= phys_level0_sa
->w
;
1094 uint32_t H0
= phys_level0_sa
->h
;
1096 for (uint32_t l
= 0; l
< info
->levels
; ++l
) {
1097 uint32_t W
= isl_minify(W0
, l
);
1098 uint32_t H
= isl_minify(H0
, l
);
1100 uint32_t w
= isl_align_npot(W
, image_align_sa
->w
);
1101 uint32_t h
= isl_align_npot(H
, image_align_sa
->h
);
1107 } else if (l
== 1) {
1110 } else if (l
== 2) {
1111 slice_bottom_w
+= w
;
1118 *phys_slice0_sa
= (struct isl_extent2d
) {
1119 .w
= MAX(slice_top_w
, slice_bottom_w
),
1120 .h
= MAX(slice_left_h
, slice_right_h
),
1125 isl_calc_phys_total_extent_el_gen4_2d(
1126 const struct isl_device
*dev
,
1127 const struct isl_surf_init_info
*restrict info
,
1128 const struct isl_tile_info
*tile_info
,
1129 enum isl_msaa_layout msaa_layout
,
1130 const struct isl_extent3d
*image_align_sa
,
1131 const struct isl_extent4d
*phys_level0_sa
,
1132 enum isl_array_pitch_span array_pitch_span
,
1133 uint32_t *array_pitch_el_rows
,
1134 struct isl_extent2d
*total_extent_el
)
1136 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1138 struct isl_extent2d phys_slice0_sa
;
1139 isl_calc_phys_slice0_extent_sa_gen4_2d(dev
, info
, msaa_layout
,
1140 image_align_sa
, phys_level0_sa
,
1142 *array_pitch_el_rows
=
1143 isl_calc_array_pitch_el_rows_gen4_2d(dev
, info
, tile_info
,
1144 image_align_sa
, phys_level0_sa
,
1147 *total_extent_el
= (struct isl_extent2d
) {
1148 .w
= isl_align_div_npot(phys_slice0_sa
.w
, fmtl
->bw
),
1149 .h
= *array_pitch_el_rows
* (phys_level0_sa
->array_len
- 1) +
1150 isl_align_div_npot(phys_slice0_sa
.h
, fmtl
->bh
),
1155 * A variant of isl_calc_phys_slice0_extent_sa() specific to
1156 * ISL_DIM_LAYOUT_GEN4_3D.
1159 isl_calc_phys_total_extent_el_gen4_3d(
1160 const struct isl_device
*dev
,
1161 const struct isl_surf_init_info
*restrict info
,
1162 const struct isl_extent3d
*image_align_sa
,
1163 const struct isl_extent4d
*phys_level0_sa
,
1164 uint32_t *array_pitch_el_rows
,
1165 struct isl_extent2d
*phys_total_el
)
1167 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1169 assert(info
->samples
== 1);
1171 if (info
->dim
!= ISL_SURF_DIM_3D
) {
1172 /* From the G45 PRM Vol. 1a, "6.17.4.1 Hardware Cube Map Layout":
1174 * The cube face textures are stored in the same way as 3D surfaces
1175 * are stored (see section 6.17.5 for details). For cube surfaces,
1176 * however, the depth is equal to the number of faces (always 6) and
1177 * is not reduced for each MIP.
1179 assert(ISL_DEV_GEN(dev
) == 4);
1180 assert(info
->usage
& ISL_SURF_USAGE_CUBE_BIT
);
1181 assert(phys_level0_sa
->array_len
== 6);
1183 assert(phys_level0_sa
->array_len
== 1);
1186 uint32_t total_w
= 0;
1187 uint32_t total_h
= 0;
1189 uint32_t W0
= phys_level0_sa
->w
;
1190 uint32_t H0
= phys_level0_sa
->h
;
1191 uint32_t D0
= phys_level0_sa
->d
;
1192 uint32_t A0
= phys_level0_sa
->a
;
1194 for (uint32_t l
= 0; l
< info
->levels
; ++l
) {
1195 uint32_t level_w
= isl_align_npot(isl_minify(W0
, l
), image_align_sa
->w
);
1196 uint32_t level_h
= isl_align_npot(isl_minify(H0
, l
), image_align_sa
->h
);
1197 uint32_t level_d
= info
->dim
== ISL_SURF_DIM_3D
? isl_minify(D0
, l
) : A0
;
1199 uint32_t max_layers_horiz
= MIN(level_d
, 1u << l
);
1200 uint32_t max_layers_vert
= isl_align(level_d
, 1u << l
) / (1u << l
);
1202 total_w
= MAX(total_w
, level_w
* max_layers_horiz
);
1203 total_h
+= level_h
* max_layers_vert
;
1206 /* GEN4_3D layouts don't really have an array pitch since each LOD has a
1207 * different number of horizontal and vertical layers. We have to set it
1208 * to something, so at least make it true for LOD0.
1210 *array_pitch_el_rows
=
1211 isl_align_npot(phys_level0_sa
->h
, image_align_sa
->h
) / fmtl
->bw
;
1212 *phys_total_el
= (struct isl_extent2d
) {
1213 .w
= isl_assert_div(total_w
, fmtl
->bw
),
1214 .h
= isl_assert_div(total_h
, fmtl
->bh
),
1219 * A variant of isl_calc_phys_slice0_extent_sa() specific to
1220 * ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ.
1223 isl_calc_phys_total_extent_el_gen6_stencil_hiz(
1224 const struct isl_device
*dev
,
1225 const struct isl_surf_init_info
*restrict info
,
1226 const struct isl_tile_info
*tile_info
,
1227 const struct isl_extent3d
*image_align_sa
,
1228 const struct isl_extent4d
*phys_level0_sa
,
1229 uint32_t *array_pitch_el_rows
,
1230 struct isl_extent2d
*phys_total_el
)
1232 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1234 const struct isl_extent2d tile_extent_sa
= {
1235 .w
= tile_info
->logical_extent_el
.w
* fmtl
->bw
,
1236 .h
= tile_info
->logical_extent_el
.h
* fmtl
->bh
,
1238 /* Tile size is a multiple of image alignment */
1239 assert(tile_extent_sa
.w
% image_align_sa
->w
== 0);
1240 assert(tile_extent_sa
.h
% image_align_sa
->h
== 0);
1242 const uint32_t W0
= phys_level0_sa
->w
;
1243 const uint32_t H0
= phys_level0_sa
->h
;
1245 /* Each image has the same height as LOD0 because the hardware thinks
1246 * everything is LOD0
1248 const uint32_t H
= isl_align(H0
, image_align_sa
->h
) * phys_level0_sa
->a
;
1250 uint32_t total_top_w
= 0;
1251 uint32_t total_bottom_w
= 0;
1252 uint32_t total_h
= 0;
1254 for (uint32_t l
= 0; l
< info
->levels
; ++l
) {
1255 const uint32_t W
= isl_minify(W0
, l
);
1257 const uint32_t w
= isl_align(W
, tile_extent_sa
.w
);
1258 const uint32_t h
= isl_align(H
, tile_extent_sa
.h
);
1263 } else if (l
== 1) {
1267 total_bottom_w
+= w
;
1271 *array_pitch_el_rows
=
1272 isl_assert_div(isl_align(H0
, image_align_sa
->h
), fmtl
->bh
);
1273 *phys_total_el
= (struct isl_extent2d
) {
1274 .w
= isl_assert_div(MAX(total_top_w
, total_bottom_w
), fmtl
->bw
),
1275 .h
= isl_assert_div(total_h
, fmtl
->bh
),
1280 * A variant of isl_calc_phys_slice0_extent_sa() specific to
1281 * ISL_DIM_LAYOUT_GEN9_1D.
1284 isl_calc_phys_total_extent_el_gen9_1d(
1285 const struct isl_device
*dev
,
1286 const struct isl_surf_init_info
*restrict info
,
1287 const struct isl_extent3d
*image_align_sa
,
1288 const struct isl_extent4d
*phys_level0_sa
,
1289 uint32_t *array_pitch_el_rows
,
1290 struct isl_extent2d
*phys_total_el
)
1292 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1294 assert(phys_level0_sa
->height
== 1);
1295 assert(phys_level0_sa
->depth
== 1);
1296 assert(info
->samples
== 1);
1297 assert(image_align_sa
->w
>= fmtl
->bw
);
1299 uint32_t slice_w
= 0;
1300 const uint32_t W0
= phys_level0_sa
->w
;
1302 for (uint32_t l
= 0; l
< info
->levels
; ++l
) {
1303 uint32_t W
= isl_minify(W0
, l
);
1304 uint32_t w
= isl_align_npot(W
, image_align_sa
->w
);
1309 *array_pitch_el_rows
= 1;
1310 *phys_total_el
= (struct isl_extent2d
) {
1311 .w
= isl_assert_div(slice_w
, fmtl
->bw
),
1312 .h
= phys_level0_sa
->array_len
,
1317 * Calculate the two-dimensional total physical extent of the surface, in
1318 * units of surface elements.
1321 isl_calc_phys_total_extent_el(const struct isl_device
*dev
,
1322 const struct isl_surf_init_info
*restrict info
,
1323 const struct isl_tile_info
*tile_info
,
1324 enum isl_dim_layout dim_layout
,
1325 enum isl_msaa_layout msaa_layout
,
1326 const struct isl_extent3d
*image_align_sa
,
1327 const struct isl_extent4d
*phys_level0_sa
,
1328 enum isl_array_pitch_span array_pitch_span
,
1329 uint32_t *array_pitch_el_rows
,
1330 struct isl_extent2d
*total_extent_el
)
1332 switch (dim_layout
) {
1333 case ISL_DIM_LAYOUT_GEN9_1D
:
1334 assert(array_pitch_span
== ISL_ARRAY_PITCH_SPAN_COMPACT
);
1335 isl_calc_phys_total_extent_el_gen9_1d(dev
, info
,
1336 image_align_sa
, phys_level0_sa
,
1337 array_pitch_el_rows
,
1340 case ISL_DIM_LAYOUT_GEN4_2D
:
1341 isl_calc_phys_total_extent_el_gen4_2d(dev
, info
, tile_info
, msaa_layout
,
1342 image_align_sa
, phys_level0_sa
,
1344 array_pitch_el_rows
,
1347 case ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
:
1348 assert(array_pitch_span
== ISL_ARRAY_PITCH_SPAN_COMPACT
);
1349 isl_calc_phys_total_extent_el_gen6_stencil_hiz(dev
, info
, tile_info
,
1352 array_pitch_el_rows
,
1355 case ISL_DIM_LAYOUT_GEN4_3D
:
1356 assert(array_pitch_span
== ISL_ARRAY_PITCH_SPAN_COMPACT
);
1357 isl_calc_phys_total_extent_el_gen4_3d(dev
, info
,
1358 image_align_sa
, phys_level0_sa
,
1359 array_pitch_el_rows
,
1364 unreachable("invalid value for dim_layout");
1368 isl_calc_row_pitch_alignment(const struct isl_device
*dev
,
1369 const struct isl_surf_init_info
*surf_info
,
1370 const struct isl_tile_info
*tile_info
)
1372 if (tile_info
->tiling
!= ISL_TILING_LINEAR
) {
1373 /* According to BSpec: 44930, Gen12's CCS-compressed surface pitches must
1374 * be 512B-aligned. CCS is only support on Y tilings.
1376 * Only consider 512B alignment when :
1377 * - AUX is not explicitly disabled
1378 * - the caller has specified no pitch
1380 * isl_surf_get_ccs_surf() will check that the main surface alignment
1381 * matches CCS expectations.
1383 if (ISL_DEV_GEN(dev
) >= 12 &&
1384 isl_format_supports_ccs_e(dev
->info
, surf_info
->format
) &&
1385 tile_info
->tiling
!= ISL_TILING_X
&&
1386 !(surf_info
->usage
& ISL_SURF_USAGE_DISABLE_AUX_BIT
) &&
1387 surf_info
->row_pitch_B
== 0) {
1388 return isl_align(tile_info
->phys_extent_B
.width
, 512);
1391 return tile_info
->phys_extent_B
.width
;
1394 /* From the Broadwel PRM >> Volume 2d: Command Reference: Structures >>
1395 * RENDER_SURFACE_STATE Surface Pitch (p349):
1397 * - For linear render target surfaces and surfaces accessed with the
1398 * typed data port messages, the pitch must be a multiple of the
1399 * element size for non-YUV surface formats. Pitch must be
1400 * a multiple of 2 * element size for YUV surface formats.
1402 * - [Requirements for SURFTYPE_BUFFER and SURFTYPE_STRBUF, which we
1403 * ignore because isl doesn't do buffers.]
1405 * - For other linear surfaces, the pitch can be any multiple of
1408 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf_info
->format
);
1409 const uint32_t bs
= fmtl
->bpb
/ 8;
1412 if (surf_info
->usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) {
1413 if (isl_format_is_yuv(surf_info
->format
)) {
1422 /* From the Broadwell PRM >> Volume 2c: Command Reference: Registers >>
1423 * PRI_STRIDE Stride (p1254):
1425 * "When using linear memory, this must be at least 64 byte aligned."
1427 if (surf_info
->usage
& ISL_SURF_USAGE_DISPLAY_BIT
)
1428 alignment
= isl_align(alignment
, 64);
1434 isl_calc_linear_min_row_pitch(const struct isl_device
*dev
,
1435 const struct isl_surf_init_info
*info
,
1436 const struct isl_extent2d
*phys_total_el
,
1437 uint32_t alignment_B
)
1439 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1440 const uint32_t bs
= fmtl
->bpb
/ 8;
1442 return isl_align_npot(bs
* phys_total_el
->w
, alignment_B
);
1446 isl_calc_tiled_min_row_pitch(const struct isl_device
*dev
,
1447 const struct isl_surf_init_info
*surf_info
,
1448 const struct isl_tile_info
*tile_info
,
1449 const struct isl_extent2d
*phys_total_el
,
1450 uint32_t alignment_B
)
1452 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf_info
->format
);
1454 assert(fmtl
->bpb
% tile_info
->format_bpb
== 0);
1456 const uint32_t tile_el_scale
= fmtl
->bpb
/ tile_info
->format_bpb
;
1457 const uint32_t total_w_tl
=
1458 isl_align_div(phys_total_el
->w
* tile_el_scale
,
1459 tile_info
->logical_extent_el
.width
);
1461 /* In some cases the alignment of the pitch might be > to the tile size
1462 * (for example Gen12 CCS requires 512B alignment while the tile's width
1463 * can be 128B), so align the row pitch to the alignment.
1465 assert(alignment_B
>= tile_info
->phys_extent_B
.width
);
1466 return isl_align(total_w_tl
* tile_info
->phys_extent_B
.width
, alignment_B
);
1470 isl_calc_min_row_pitch(const struct isl_device
*dev
,
1471 const struct isl_surf_init_info
*surf_info
,
1472 const struct isl_tile_info
*tile_info
,
1473 const struct isl_extent2d
*phys_total_el
,
1474 uint32_t alignment_B
)
1476 if (tile_info
->tiling
== ISL_TILING_LINEAR
) {
1477 return isl_calc_linear_min_row_pitch(dev
, surf_info
, phys_total_el
,
1480 return isl_calc_tiled_min_row_pitch(dev
, surf_info
, tile_info
,
1481 phys_total_el
, alignment_B
);
1486 * Is `pitch` in the valid range for a hardware bitfield, if the bitfield's
1487 * size is `bits` bits?
1489 * Hardware pitch fields are offset by 1. For example, if the size of
1490 * RENDER_SURFACE_STATE::SurfacePitch is B bits, then the range of valid
1491 * pitches is [1, 2^b] inclusive. If the surface pitch is N, then
1492 * RENDER_SURFACE_STATE::SurfacePitch must be set to N-1.
1495 pitch_in_range(uint32_t n
, uint32_t bits
)
1498 return likely(bits
!= 0 && 1 <= n
&& n
<= (1 << bits
));
1502 isl_calc_row_pitch(const struct isl_device
*dev
,
1503 const struct isl_surf_init_info
*surf_info
,
1504 const struct isl_tile_info
*tile_info
,
1505 enum isl_dim_layout dim_layout
,
1506 const struct isl_extent2d
*phys_total_el
,
1507 uint32_t *out_row_pitch_B
)
1509 uint32_t alignment_B
=
1510 isl_calc_row_pitch_alignment(dev
, surf_info
, tile_info
);
1512 const uint32_t min_row_pitch_B
=
1513 isl_calc_min_row_pitch(dev
, surf_info
, tile_info
, phys_total_el
,
1516 if (surf_info
->row_pitch_B
!= 0) {
1517 if (surf_info
->row_pitch_B
< min_row_pitch_B
)
1520 if (surf_info
->row_pitch_B
% alignment_B
!= 0)
1524 const uint32_t row_pitch_B
=
1525 surf_info
->row_pitch_B
!= 0 ? surf_info
->row_pitch_B
: min_row_pitch_B
;
1527 const uint32_t row_pitch_tl
= row_pitch_B
/ tile_info
->phys_extent_B
.width
;
1529 if (row_pitch_B
== 0)
1532 if (dim_layout
== ISL_DIM_LAYOUT_GEN9_1D
) {
1533 /* SurfacePitch is ignored for this layout. */
1537 if ((surf_info
->usage
& (ISL_SURF_USAGE_RENDER_TARGET_BIT
|
1538 ISL_SURF_USAGE_TEXTURE_BIT
|
1539 ISL_SURF_USAGE_STORAGE_BIT
)) &&
1540 !pitch_in_range(row_pitch_B
, RENDER_SURFACE_STATE_SurfacePitch_bits(dev
->info
)))
1543 if ((surf_info
->usage
& (ISL_SURF_USAGE_CCS_BIT
|
1544 ISL_SURF_USAGE_MCS_BIT
)) &&
1545 !pitch_in_range(row_pitch_tl
, RENDER_SURFACE_STATE_AuxiliarySurfacePitch_bits(dev
->info
)))
1548 if ((surf_info
->usage
& ISL_SURF_USAGE_DEPTH_BIT
) &&
1549 !pitch_in_range(row_pitch_B
, _3DSTATE_DEPTH_BUFFER_SurfacePitch_bits(dev
->info
)))
1552 if ((surf_info
->usage
& ISL_SURF_USAGE_HIZ_BIT
) &&
1553 !pitch_in_range(row_pitch_B
, _3DSTATE_HIER_DEPTH_BUFFER_SurfacePitch_bits(dev
->info
)))
1556 const uint32_t stencil_pitch_bits
= dev
->use_separate_stencil
?
1557 _3DSTATE_STENCIL_BUFFER_SurfacePitch_bits(dev
->info
) :
1558 _3DSTATE_DEPTH_BUFFER_SurfacePitch_bits(dev
->info
);
1560 if ((surf_info
->usage
& ISL_SURF_USAGE_STENCIL_BIT
) &&
1561 !pitch_in_range(row_pitch_B
, stencil_pitch_bits
))
1565 *out_row_pitch_B
= row_pitch_B
;
1570 isl_surf_init_s(const struct isl_device
*dev
,
1571 struct isl_surf
*surf
,
1572 const struct isl_surf_init_info
*restrict info
)
1574 const struct isl_format_layout
*fmtl
= isl_format_get_layout(info
->format
);
1576 const struct isl_extent4d logical_level0_px
= {
1580 .a
= info
->array_len
,
1583 enum isl_tiling tiling
;
1584 if (!isl_surf_choose_tiling(dev
, info
, &tiling
))
1587 struct isl_tile_info tile_info
;
1588 isl_tiling_get_info(tiling
, fmtl
->bpb
, &tile_info
);
1590 const enum isl_dim_layout dim_layout
=
1591 isl_surf_choose_dim_layout(dev
, info
->dim
, tiling
, info
->usage
);
1593 enum isl_msaa_layout msaa_layout
;
1594 if (!isl_choose_msaa_layout(dev
, info
, tiling
, &msaa_layout
))
1597 struct isl_extent3d image_align_el
;
1598 isl_choose_image_alignment_el(dev
, info
, tiling
, dim_layout
, msaa_layout
,
1601 struct isl_extent3d image_align_sa
=
1602 isl_extent3d_el_to_sa(info
->format
, image_align_el
);
1604 struct isl_extent4d phys_level0_sa
;
1605 isl_calc_phys_level0_extent_sa(dev
, info
, dim_layout
, tiling
, msaa_layout
,
1608 enum isl_array_pitch_span array_pitch_span
=
1609 isl_choose_array_pitch_span(dev
, info
, dim_layout
, &phys_level0_sa
);
1611 uint32_t array_pitch_el_rows
;
1612 struct isl_extent2d phys_total_el
;
1613 isl_calc_phys_total_extent_el(dev
, info
, &tile_info
,
1614 dim_layout
, msaa_layout
,
1615 &image_align_sa
, &phys_level0_sa
,
1616 array_pitch_span
, &array_pitch_el_rows
,
1619 uint32_t row_pitch_B
;
1620 if (!isl_calc_row_pitch(dev
, info
, &tile_info
, dim_layout
,
1621 &phys_total_el
, &row_pitch_B
))
1624 uint32_t base_alignment_B
;
1626 if (tiling
== ISL_TILING_LINEAR
) {
1627 size_B
= (uint64_t) row_pitch_B
* phys_total_el
.h
;
1629 /* From the Broadwell PRM Vol 2d, RENDER_SURFACE_STATE::SurfaceBaseAddress:
1631 * "The Base Address for linear render target surfaces and surfaces
1632 * accessed with the typed surface read/write data port messages must
1633 * be element-size aligned, for non-YUV surface formats, or a
1634 * multiple of 2 element-sizes for YUV surface formats. Other linear
1635 * surfaces have no alignment requirements (byte alignment is
1638 base_alignment_B
= MAX(1, info
->min_alignment_B
);
1639 if (info
->usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) {
1640 if (isl_format_is_yuv(info
->format
)) {
1641 base_alignment_B
= MAX(base_alignment_B
, fmtl
->bpb
/ 4);
1643 base_alignment_B
= MAX(base_alignment_B
, fmtl
->bpb
/ 8);
1646 base_alignment_B
= isl_round_up_to_power_of_two(base_alignment_B
);
1648 /* From the Skylake PRM Vol 2c, PLANE_STRIDE::Stride:
1650 * "For Linear memory, this field specifies the stride in chunks of
1651 * 64 bytes (1 cache line)."
1653 if (isl_surf_usage_is_display(info
->usage
))
1654 base_alignment_B
= MAX(base_alignment_B
, 64);
1656 const uint32_t total_h_tl
=
1657 isl_align_div(phys_total_el
.h
, tile_info
.logical_extent_el
.height
);
1659 size_B
= (uint64_t) total_h_tl
* tile_info
.phys_extent_B
.height
* row_pitch_B
;
1661 const uint32_t tile_size_B
= tile_info
.phys_extent_B
.width
*
1662 tile_info
.phys_extent_B
.height
;
1663 assert(isl_is_pow2(info
->min_alignment_B
) && isl_is_pow2(tile_size_B
));
1664 base_alignment_B
= MAX(info
->min_alignment_B
, tile_size_B
);
1666 /* The diagram in the Bspec section Memory Compression - Gen12, shows
1667 * that the CCS is indexed in 256B chunks. However, the
1668 * PLANE_AUX_DIST::Auxiliary Surface Distance field is in units of 4K
1669 * pages. We currently don't assign the usage field like we do for main
1670 * surfaces, so just use 4K for now.
1672 if (tiling
== ISL_TILING_GEN12_CCS
)
1673 base_alignment_B
= MAX(base_alignment_B
, 4096);
1675 /* Gen12+ requires that images be 64K-aligned if they're going to used
1676 * with CCS. This is because the Aux translation table maps main
1677 * surface addresses to aux addresses at a 64K (in the main surface)
1678 * granularity. Because we don't know for sure in ISL if a surface will
1679 * use CCS, we have to guess based on the DISABLE_AUX usage bit. The
1680 * one thing we do know is that we haven't enable CCS on linear images
1681 * yet so we can avoid the extra alignment there.
1683 if (ISL_DEV_GEN(dev
) >= 12 &&
1684 !(info
->usage
& ISL_SURF_USAGE_DISABLE_AUX_BIT
)) {
1685 base_alignment_B
= MAX(base_alignment_B
, 64 * 1024);
1689 if (ISL_DEV_GEN(dev
) < 9) {
1690 /* From the Broadwell PRM Vol 5, Surface Layout:
1692 * "In addition to restrictions on maximum height, width, and depth,
1693 * surfaces are also restricted to a maximum size in bytes. This
1694 * maximum is 2 GB for all products and all surface types."
1696 * This comment is applicable to all Pre-gen9 platforms.
1698 if (size_B
> (uint64_t) 1 << 31)
1700 } else if (ISL_DEV_GEN(dev
) < 11) {
1701 /* From the Skylake PRM Vol 5, Maximum Surface Size in Bytes:
1702 * "In addition to restrictions on maximum height, width, and depth,
1703 * surfaces are also restricted to a maximum size of 2^38 bytes.
1704 * All pixels within the surface must be contained within 2^38 bytes
1705 * of the base address."
1707 if (size_B
> (uint64_t) 1 << 38)
1710 /* gen11+ platforms raised this limit to 2^44 bytes. */
1711 if (size_B
> (uint64_t) 1 << 44)
1715 *surf
= (struct isl_surf
) {
1717 .dim_layout
= dim_layout
,
1718 .msaa_layout
= msaa_layout
,
1720 .format
= info
->format
,
1722 .levels
= info
->levels
,
1723 .samples
= info
->samples
,
1725 .image_alignment_el
= image_align_el
,
1726 .logical_level0_px
= logical_level0_px
,
1727 .phys_level0_sa
= phys_level0_sa
,
1730 .alignment_B
= base_alignment_B
,
1731 .row_pitch_B
= row_pitch_B
,
1732 .array_pitch_el_rows
= array_pitch_el_rows
,
1733 .array_pitch_span
= array_pitch_span
,
1735 .usage
= info
->usage
,
1742 isl_surf_get_tile_info(const struct isl_surf
*surf
,
1743 struct isl_tile_info
*tile_info
)
1745 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf
->format
);
1746 isl_tiling_get_info(surf
->tiling
, fmtl
->bpb
, tile_info
);
1750 isl_surf_get_hiz_surf(const struct isl_device
*dev
,
1751 const struct isl_surf
*surf
,
1752 struct isl_surf
*hiz_surf
)
1754 assert(ISL_DEV_GEN(dev
) >= 5 && ISL_DEV_USE_SEPARATE_STENCIL(dev
));
1756 if (!isl_surf_usage_is_depth(surf
->usage
))
1759 /* HiZ only works with Y-tiled depth buffers */
1760 if (!isl_tiling_is_any_y(surf
->tiling
))
1763 /* On SNB+, compressed depth buffers cannot be interleaved with stencil. */
1764 switch (surf
->format
) {
1765 case ISL_FORMAT_R24_UNORM_X8_TYPELESS
:
1766 if (isl_surf_usage_is_depth_and_stencil(surf
->usage
)) {
1767 assert(ISL_DEV_GEN(dev
) == 5);
1768 unreachable("This should work, but is untested");
1771 case ISL_FORMAT_R16_UNORM
:
1772 case ISL_FORMAT_R32_FLOAT
:
1774 case ISL_FORMAT_R32_FLOAT_X8X24_TYPELESS
:
1775 if (ISL_DEV_GEN(dev
) == 5) {
1776 assert(isl_surf_usage_is_depth_and_stencil(surf
->usage
));
1777 unreachable("This should work, but is untested");
1784 /* Multisampled depth is always interleaved */
1785 assert(surf
->msaa_layout
== ISL_MSAA_LAYOUT_NONE
||
1786 surf
->msaa_layout
== ISL_MSAA_LAYOUT_INTERLEAVED
);
1788 /* From the Broadwell PRM Vol. 7, "Hierarchical Depth Buffer":
1790 * "The Surface Type, Height, Width, Depth, Minimum Array Element, Render
1791 * Target View Extent, and Depth Coordinate Offset X/Y of the
1792 * hierarchical depth buffer are inherited from the depth buffer. The
1793 * height and width of the hierarchical depth buffer that must be
1794 * allocated are computed by the following formulas, where HZ is the
1795 * hierarchical depth buffer and Z is the depth buffer. The Z_Height,
1796 * Z_Width, and Z_Depth values given in these formulas are those present
1797 * in 3DSTATE_DEPTH_BUFFER incremented by one.
1799 * "The value of Z_Height and Z_Width must each be multiplied by 2 before
1800 * being applied to the table below if Number of Multisamples is set to
1801 * NUMSAMPLES_4. The value of Z_Height must be multiplied by 2 and
1802 * Z_Width must be multiplied by 4 before being applied to the table
1803 * below if Number of Multisamples is set to NUMSAMPLES_8."
1805 * In the Sky Lake PRM, the second paragraph is replaced with this:
1807 * "The Z_Height and Z_Width values must equal those present in
1808 * 3DSTATE_DEPTH_BUFFER incremented by one."
1810 * In other words, on Sandy Bridge through Broadwell, each 128-bit HiZ
1811 * block corresponds to a region of 8x4 samples in the primary depth
1812 * surface. On Sky Lake, on the other hand, each HiZ block corresponds to
1813 * a region of 8x4 pixels in the primary depth surface regardless of the
1814 * number of samples. The dimensions of a HiZ block in both pixels and
1815 * samples are given in the table below:
1817 * | SNB - BDW | SKL+
1818 * ------+-----------+-------------
1819 * 1x | 8 x 4 sa | 8 x 4 sa
1820 * MSAA | 8 x 4 px | 8 x 4 px
1821 * ------+-----------+-------------
1822 * 2x | 8 x 4 sa | 16 x 4 sa
1823 * MSAA | 4 x 4 px | 8 x 4 px
1824 * ------+-----------+-------------
1825 * 4x | 8 x 4 sa | 16 x 8 sa
1826 * MSAA | 4 x 2 px | 8 x 4 px
1827 * ------+-----------+-------------
1828 * 8x | 8 x 4 sa | 32 x 8 sa
1829 * MSAA | 2 x 2 px | 8 x 4 px
1830 * ------+-----------+-------------
1831 * 16x | N/A | 32 x 16 sa
1832 * MSAA | N/A | 8 x 4 px
1833 * ------+-----------+-------------
1835 * There are a number of different ways that this discrepency could be
1836 * handled. The way we have chosen is to simply make MSAA HiZ have the
1837 * same number of samples as the parent surface pre-Sky Lake and always be
1838 * single-sampled on Sky Lake and above. Since the block sizes of
1839 * compressed formats are given in samples, this neatly handles everything
1840 * without the need for additional HiZ formats with different block sizes
1843 const unsigned samples
= ISL_DEV_GEN(dev
) >= 9 ? 1 : surf
->samples
;
1845 return isl_surf_init(dev
, hiz_surf
,
1847 .format
= ISL_FORMAT_HIZ
,
1848 .width
= surf
->logical_level0_px
.width
,
1849 .height
= surf
->logical_level0_px
.height
,
1850 .depth
= surf
->logical_level0_px
.depth
,
1851 .levels
= surf
->levels
,
1852 .array_len
= surf
->logical_level0_px
.array_len
,
1854 .usage
= ISL_SURF_USAGE_HIZ_BIT
,
1855 .tiling_flags
= ISL_TILING_HIZ_BIT
);
1859 isl_surf_get_mcs_surf(const struct isl_device
*dev
,
1860 const struct isl_surf
*surf
,
1861 struct isl_surf
*mcs_surf
)
1863 /* It must be multisampled with an array layout */
1864 if (surf
->msaa_layout
!= ISL_MSAA_LAYOUT_ARRAY
)
1867 if (mcs_surf
->size_B
> 0)
1870 /* The following are true of all multisampled surfaces */
1871 assert(surf
->samples
> 1);
1872 assert(surf
->dim
== ISL_SURF_DIM_2D
);
1873 assert(surf
->levels
== 1);
1874 assert(surf
->logical_level0_px
.depth
== 1);
1876 /* From the Ivy Bridge PRM, Vol4 Part1 p77 ("MCS Enable"):
1878 * This field must be set to 0 for all SINT MSRTs when all RT channels
1881 * In practice this means that we have to disable MCS for all signed
1882 * integer MSAA buffers. The alternative, to disable MCS only when one
1883 * of the render target channels is disabled, is impractical because it
1884 * would require converting between CMS and UMS MSAA layouts on the fly,
1885 * which is expensive.
1887 if (ISL_DEV_GEN(dev
) == 7 && isl_format_has_sint_channel(surf
->format
))
1890 /* The "Auxiliary Surface Pitch" field in RENDER_SURFACE_STATE is only 9
1891 * bits which means the maximum pitch of a compression surface is 512
1892 * tiles or 64KB (since MCS is always Y-tiled). Since a 16x MCS buffer is
1893 * 64bpp, this gives us a maximum width of 8192 pixels. We can create
1894 * larger multisampled surfaces, we just can't compress them. For 2x, 4x,
1895 * and 8x, we have enough room for the full 16k supported by the hardware.
1897 if (surf
->samples
== 16 && surf
->logical_level0_px
.width
> 8192)
1900 enum isl_format mcs_format
;
1901 switch (surf
->samples
) {
1902 case 2: mcs_format
= ISL_FORMAT_MCS_2X
; break;
1903 case 4: mcs_format
= ISL_FORMAT_MCS_4X
; break;
1904 case 8: mcs_format
= ISL_FORMAT_MCS_8X
; break;
1905 case 16: mcs_format
= ISL_FORMAT_MCS_16X
; break;
1907 unreachable("Invalid sample count");
1910 return isl_surf_init(dev
, mcs_surf
,
1911 .dim
= ISL_SURF_DIM_2D
,
1912 .format
= mcs_format
,
1913 .width
= surf
->logical_level0_px
.width
,
1914 .height
= surf
->logical_level0_px
.height
,
1917 .array_len
= surf
->logical_level0_px
.array_len
,
1918 .samples
= 1, /* MCS surfaces are really single-sampled */
1919 .usage
= ISL_SURF_USAGE_MCS_BIT
,
1920 .tiling_flags
= ISL_TILING_Y0_BIT
);
1924 isl_surf_supports_ccs(const struct isl_device
*dev
,
1925 const struct isl_surf
*surf
)
1927 /* CCS support does not exist prior to Gen7 */
1928 if (ISL_DEV_GEN(dev
) <= 6)
1931 if (surf
->usage
& ISL_SURF_USAGE_DISABLE_AUX_BIT
)
1934 if (isl_format_is_compressed(surf
->format
))
1937 if (!isl_is_pow2(isl_format_get_layout(surf
->format
)->bpb
))
1940 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
1941 * Target(s)", beneath the "Fast Color Clear" bullet (p326):
1943 * - Support is limited to tiled render targets.
1945 * From the Skylake documentation, it is made clear that X-tiling is no
1948 * - MCS and Lossless compression is supported for
1949 * TiledY/TileYs/TileYf non-MSRTs only.
1951 * From the BSpec (44930) for Gen12:
1953 * Linear CCS is only allowed for Untyped Buffers but only via HDC
1954 * Data-Port messages.
1956 * We never use untyped messages on surfaces created by ISL on Gen9+ so
1957 * this means linear is out on Gen12+ as well.
1959 if (surf
->tiling
== ISL_TILING_LINEAR
)
1962 if (ISL_DEV_GEN(dev
) >= 12) {
1963 if (isl_surf_usage_is_stencil(surf
->usage
) && surf
->samples
> 1)
1966 /* [TGL+] CCS can only be added to a non-D16-formatted depth buffer if
1967 * it has HiZ. If not for GEN:BUG:1406512483 "deprecate compression
1968 * enable states", D16 would be supported. Supporting D16 requires being
1969 * able to specify that the control surface is present and
1970 * simultaneously disabling compression. The above bug makes it so that
1971 * it's not possible to specify this configuration.
1973 * Note: ISL Doesn't currently support depth CCS without HiZ at all.
1975 if (isl_surf_usage_is_depth(surf
->usage
) &&
1976 surf
->format
== ISL_FORMAT_R16_UNORM
) {
1980 /* On Gen12, 8BPP surfaces cannot be compressed if any level is not
1981 * 32Bx4row-aligned. For now, just reject the cases where alignment
1984 if (isl_format_get_layout(surf
->format
)->bpb
== 8 && surf
->levels
>= 3) {
1985 isl_finishme("%s:%s: CCS for 8BPP textures with 3+ miplevels is "
1986 "disabled, but support for more levels is possible.",
1987 __FILE__
, __func__
);
1991 /* On Gen12, all CCS-compressed surface pitches must be multiples of
1994 if (surf
->row_pitch_B
% 512 != 0)
1997 /* According to GEN:BUG:1406738321, 3D textures need a blit to a new
1998 * surface in order to perform a resolve. For now, just disable CCS.
2000 if (surf
->dim
== ISL_SURF_DIM_3D
) {
2001 isl_finishme("%s:%s: CCS for 3D textures is disabled, but a workaround"
2002 " is available.", __FILE__
, __func__
);
2006 /* GEN:BUG:1207137018
2008 * TODO: implement following workaround currently covered by the
2009 * restriction above. If following conditions are met:
2011 * - RENDER_SURFACE_STATE.Surface Type == 3D
2012 * - RENDER_SURFACE_STATE.Auxiliary Surface Mode != AUX_NONE
2013 * - RENDER_SURFACE_STATE.Tiled ResourceMode is TYF or TYS
2015 * Set the value of RENDER_SURFACE_STATE.Mip Tail Start LOD to a mip
2016 * that larger than those present in the surface (i.e. 15)
2019 /* TODO: Handle the other tiling formats */
2020 if (surf
->tiling
!= ISL_TILING_Y0
)
2023 /* ISL_DEV_GEN(dev) < 12 */
2024 if (surf
->samples
> 1)
2027 /* CCS is only for color images on Gen7-11 */
2028 if (isl_surf_usage_is_depth_or_stencil(surf
->usage
))
2031 /* The PRM doesn't say this explicitly, but fast-clears don't appear to
2032 * work for 3D textures until gen9 where the layout of 3D textures
2033 * changes to match 2D array textures.
2035 if (ISL_DEV_GEN(dev
) <= 8 && surf
->dim
!= ISL_SURF_DIM_2D
)
2038 /* From the HSW PRM Volume 7: 3D-Media-GPGPU, page 652 (Color Clear of
2039 * Non-MultiSampler Render Target Restrictions):
2041 * "Support is for non-mip-mapped and non-array surface types only."
2043 * This restriction is lifted on gen8+. Technically, it may be possible
2044 * to create a CCS for an arrayed or mipmapped image and only enable
2045 * CCS_D when rendering to the base slice. However, there is no
2046 * documentation tell us what the hardware would do in that case or what
2047 * it does if you walk off the bases slice. (Does it ignore CCS or does
2048 * it start scribbling over random memory?) We play it safe and just
2049 * follow the docs and don't allow CCS_D for arrayed or mip-mapped
2052 if (ISL_DEV_GEN(dev
) <= 7 &&
2053 (surf
->levels
> 1 || surf
->logical_level0_px
.array_len
> 1))
2056 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
2057 * Target(s)", beneath the "Fast Color Clear" bullet (p326):
2059 * - MCS buffer for non-MSRT is supported only for RT formats 32bpp,
2060 * 64bpp, and 128bpp.
2062 if (isl_format_get_layout(surf
->format
)->bpb
< 32)
2065 /* From the Skylake documentation, it is made clear that X-tiling is no
2068 * - MCS and Lossless compression is supported for
2069 * TiledY/TileYs/TileYf non-MSRTs only.
2071 if (ISL_DEV_GEN(dev
) >= 9 && !isl_tiling_is_any_y(surf
->tiling
))
2079 isl_surf_get_ccs_surf(const struct isl_device
*dev
,
2080 const struct isl_surf
*surf
,
2081 struct isl_surf
*aux_surf
,
2082 struct isl_surf
*extra_aux_surf
,
2083 uint32_t row_pitch_B
)
2087 /* An uninitialized surface is needed to get a CCS surface. */
2088 if (aux_surf
->size_B
> 0 &&
2089 (extra_aux_surf
== NULL
|| extra_aux_surf
->size_B
> 0)) {
2093 /* A surface can't have two CCS surfaces. */
2094 if (aux_surf
->usage
& ISL_SURF_USAGE_CCS_BIT
)
2097 if (!isl_surf_supports_ccs(dev
, surf
))
2100 if (ISL_DEV_GEN(dev
) >= 12) {
2101 enum isl_format ccs_format
;
2102 switch (isl_format_get_layout(surf
->format
)->bpb
) {
2103 case 8: ccs_format
= ISL_FORMAT_GEN12_CCS_8BPP_Y0
; break;
2104 case 16: ccs_format
= ISL_FORMAT_GEN12_CCS_16BPP_Y0
; break;
2105 case 32: ccs_format
= ISL_FORMAT_GEN12_CCS_32BPP_Y0
; break;
2106 case 64: ccs_format
= ISL_FORMAT_GEN12_CCS_64BPP_Y0
; break;
2107 case 128: ccs_format
= ISL_FORMAT_GEN12_CCS_128BPP_Y0
; break;
2112 /* On Gen12, the CCS is a scaled-down version of the main surface. We
2113 * model this as the CCS compressing a 2D-view of the entire surface.
2115 struct isl_surf
*ccs_surf
=
2116 aux_surf
->size_B
> 0 ? extra_aux_surf
: aux_surf
;
2118 isl_surf_init(dev
, ccs_surf
,
2119 .dim
= ISL_SURF_DIM_2D
,
2120 .format
= ccs_format
,
2121 .width
= isl_surf_get_row_pitch_el(surf
),
2122 .height
= surf
->size_B
/ surf
->row_pitch_B
,
2127 .row_pitch_B
= row_pitch_B
,
2128 .usage
= ISL_SURF_USAGE_CCS_BIT
,
2129 .tiling_flags
= ISL_TILING_GEN12_CCS_BIT
);
2130 assert(!ok
|| ccs_surf
->size_B
== surf
->size_B
/ 256);
2133 enum isl_format ccs_format
;
2134 if (ISL_DEV_GEN(dev
) >= 9) {
2135 switch (isl_format_get_layout(surf
->format
)->bpb
) {
2136 case 32: ccs_format
= ISL_FORMAT_GEN9_CCS_32BPP
; break;
2137 case 64: ccs_format
= ISL_FORMAT_GEN9_CCS_64BPP
; break;
2138 case 128: ccs_format
= ISL_FORMAT_GEN9_CCS_128BPP
; break;
2139 default: unreachable("Unsupported CCS format");
2142 } else if (surf
->tiling
== ISL_TILING_Y0
) {
2143 switch (isl_format_get_layout(surf
->format
)->bpb
) {
2144 case 32: ccs_format
= ISL_FORMAT_GEN7_CCS_32BPP_Y
; break;
2145 case 64: ccs_format
= ISL_FORMAT_GEN7_CCS_64BPP_Y
; break;
2146 case 128: ccs_format
= ISL_FORMAT_GEN7_CCS_128BPP_Y
; break;
2147 default: unreachable("Unsupported CCS format");
2149 } else if (surf
->tiling
== ISL_TILING_X
) {
2150 switch (isl_format_get_layout(surf
->format
)->bpb
) {
2151 case 32: ccs_format
= ISL_FORMAT_GEN7_CCS_32BPP_X
; break;
2152 case 64: ccs_format
= ISL_FORMAT_GEN7_CCS_64BPP_X
; break;
2153 case 128: ccs_format
= ISL_FORMAT_GEN7_CCS_128BPP_X
; break;
2154 default: unreachable("Unsupported CCS format");
2157 unreachable("Invalid tiling format");
2160 return isl_surf_init(dev
, aux_surf
,
2162 .format
= ccs_format
,
2163 .width
= surf
->logical_level0_px
.width
,
2164 .height
= surf
->logical_level0_px
.height
,
2165 .depth
= surf
->logical_level0_px
.depth
,
2166 .levels
= surf
->levels
,
2167 .array_len
= surf
->logical_level0_px
.array_len
,
2169 .row_pitch_B
= row_pitch_B
,
2170 .usage
= ISL_SURF_USAGE_CCS_BIT
,
2171 .tiling_flags
= ISL_TILING_CCS_BIT
);
2175 #define isl_genX_call(dev, func, ...) \
2176 switch (ISL_DEV_GEN(dev)) { \
2178 /* G45 surface state is the same as gen5 */ \
2179 if (ISL_DEV_IS_G4X(dev)) { \
2180 isl_gen5_##func(__VA_ARGS__); \
2182 isl_gen4_##func(__VA_ARGS__); \
2186 isl_gen5_##func(__VA_ARGS__); \
2189 isl_gen6_##func(__VA_ARGS__); \
2192 if (ISL_DEV_IS_HASWELL(dev)) { \
2193 isl_gen75_##func(__VA_ARGS__); \
2195 isl_gen7_##func(__VA_ARGS__); \
2199 isl_gen8_##func(__VA_ARGS__); \
2202 isl_gen9_##func(__VA_ARGS__); \
2205 isl_gen10_##func(__VA_ARGS__); \
2208 isl_gen11_##func(__VA_ARGS__); \
2211 isl_gen12_##func(__VA_ARGS__); \
2214 assert(!"Unknown hardware generation"); \
2218 isl_surf_fill_state_s(const struct isl_device
*dev
, void *state
,
2219 const struct isl_surf_fill_state_info
*restrict info
)
2222 isl_surf_usage_flags_t _base_usage
=
2223 info
->view
->usage
& (ISL_SURF_USAGE_RENDER_TARGET_BIT
|
2224 ISL_SURF_USAGE_TEXTURE_BIT
|
2225 ISL_SURF_USAGE_STORAGE_BIT
);
2226 /* They may only specify one of the above bits at a time */
2227 assert(__builtin_popcount(_base_usage
) == 1);
2228 /* The only other allowed bit is ISL_SURF_USAGE_CUBE_BIT */
2229 assert((info
->view
->usage
& ~ISL_SURF_USAGE_CUBE_BIT
) == _base_usage
);
2232 if (info
->surf
->dim
== ISL_SURF_DIM_3D
) {
2233 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2234 info
->surf
->logical_level0_px
.depth
);
2236 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2237 info
->surf
->logical_level0_px
.array_len
);
2240 isl_genX_call(dev
, surf_fill_state_s
, dev
, state
, info
);
2244 isl_buffer_fill_state_s(const struct isl_device
*dev
, void *state
,
2245 const struct isl_buffer_fill_state_info
*restrict info
)
2247 isl_genX_call(dev
, buffer_fill_state_s
, dev
, state
, info
);
2251 isl_null_fill_state(const struct isl_device
*dev
, void *state
,
2252 struct isl_extent3d size
)
2254 isl_genX_call(dev
, null_fill_state
, state
, size
);
2258 isl_emit_depth_stencil_hiz_s(const struct isl_device
*dev
, void *batch
,
2259 const struct isl_depth_stencil_hiz_emit_info
*restrict info
)
2261 if (info
->depth_surf
&& info
->stencil_surf
) {
2262 if (!dev
->info
->has_hiz_and_separate_stencil
) {
2263 assert(info
->depth_surf
== info
->stencil_surf
);
2264 assert(info
->depth_address
== info
->stencil_address
);
2266 assert(info
->depth_surf
->dim
== info
->stencil_surf
->dim
);
2269 if (info
->depth_surf
) {
2270 assert((info
->depth_surf
->usage
& ISL_SURF_USAGE_DEPTH_BIT
));
2271 if (info
->depth_surf
->dim
== ISL_SURF_DIM_3D
) {
2272 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2273 info
->depth_surf
->logical_level0_px
.depth
);
2275 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2276 info
->depth_surf
->logical_level0_px
.array_len
);
2280 if (info
->stencil_surf
) {
2281 assert((info
->stencil_surf
->usage
& ISL_SURF_USAGE_STENCIL_BIT
));
2282 if (info
->stencil_surf
->dim
== ISL_SURF_DIM_3D
) {
2283 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2284 info
->stencil_surf
->logical_level0_px
.depth
);
2286 assert(info
->view
->base_array_layer
+ info
->view
->array_len
<=
2287 info
->stencil_surf
->logical_level0_px
.array_len
);
2291 isl_genX_call(dev
, emit_depth_stencil_hiz_s
, dev
, batch
, info
);
2295 * A variant of isl_surf_get_image_offset_sa() specific to
2296 * ISL_DIM_LAYOUT_GEN4_2D.
2299 get_image_offset_sa_gen4_2d(const struct isl_surf
*surf
,
2300 uint32_t level
, uint32_t logical_array_layer
,
2301 uint32_t *x_offset_sa
,
2302 uint32_t *y_offset_sa
)
2304 assert(level
< surf
->levels
);
2305 if (surf
->dim
== ISL_SURF_DIM_3D
)
2306 assert(logical_array_layer
< surf
->logical_level0_px
.depth
);
2308 assert(logical_array_layer
< surf
->logical_level0_px
.array_len
);
2310 const struct isl_extent3d image_align_sa
=
2311 isl_surf_get_image_alignment_sa(surf
);
2313 const uint32_t W0
= surf
->phys_level0_sa
.width
;
2314 const uint32_t H0
= surf
->phys_level0_sa
.height
;
2316 const uint32_t phys_layer
= logical_array_layer
*
2317 (surf
->msaa_layout
== ISL_MSAA_LAYOUT_ARRAY
? surf
->samples
: 1);
2320 uint32_t y
= phys_layer
* isl_surf_get_array_pitch_sa_rows(surf
);
2322 for (uint32_t l
= 0; l
< level
; ++l
) {
2324 uint32_t W
= isl_minify(W0
, l
);
2325 x
+= isl_align_npot(W
, image_align_sa
.w
);
2327 uint32_t H
= isl_minify(H0
, l
);
2328 y
+= isl_align_npot(H
, image_align_sa
.h
);
2337 * A variant of isl_surf_get_image_offset_sa() specific to
2338 * ISL_DIM_LAYOUT_GEN4_3D.
2341 get_image_offset_sa_gen4_3d(const struct isl_surf
*surf
,
2342 uint32_t level
, uint32_t logical_z_offset_px
,
2343 uint32_t *x_offset_sa
,
2344 uint32_t *y_offset_sa
)
2346 assert(level
< surf
->levels
);
2347 if (surf
->dim
== ISL_SURF_DIM_3D
) {
2348 assert(surf
->phys_level0_sa
.array_len
== 1);
2349 assert(logical_z_offset_px
< isl_minify(surf
->phys_level0_sa
.depth
, level
));
2351 assert(surf
->dim
== ISL_SURF_DIM_2D
);
2352 assert(surf
->usage
& ISL_SURF_USAGE_CUBE_BIT
);
2353 assert(surf
->phys_level0_sa
.array_len
== 6);
2354 assert(logical_z_offset_px
< surf
->phys_level0_sa
.array_len
);
2357 const struct isl_extent3d image_align_sa
=
2358 isl_surf_get_image_alignment_sa(surf
);
2360 const uint32_t W0
= surf
->phys_level0_sa
.width
;
2361 const uint32_t H0
= surf
->phys_level0_sa
.height
;
2362 const uint32_t D0
= surf
->phys_level0_sa
.depth
;
2363 const uint32_t AL
= surf
->phys_level0_sa
.array_len
;
2368 for (uint32_t l
= 0; l
< level
; ++l
) {
2369 const uint32_t level_h
= isl_align_npot(isl_minify(H0
, l
), image_align_sa
.h
);
2370 const uint32_t level_d
=
2371 isl_align_npot(surf
->dim
== ISL_SURF_DIM_3D
? isl_minify(D0
, l
) : AL
,
2373 const uint32_t max_layers_vert
= isl_align(level_d
, 1u << l
) / (1u << l
);
2375 y
+= level_h
* max_layers_vert
;
2378 const uint32_t level_w
= isl_align_npot(isl_minify(W0
, level
), image_align_sa
.w
);
2379 const uint32_t level_h
= isl_align_npot(isl_minify(H0
, level
), image_align_sa
.h
);
2380 const uint32_t level_d
=
2381 isl_align_npot(surf
->dim
== ISL_SURF_DIM_3D
? isl_minify(D0
, level
) : AL
,
2384 const uint32_t max_layers_horiz
= MIN(level_d
, 1u << level
);
2386 x
+= level_w
* (logical_z_offset_px
% max_layers_horiz
);
2387 y
+= level_h
* (logical_z_offset_px
/ max_layers_horiz
);
2394 get_image_offset_sa_gen6_stencil_hiz(const struct isl_surf
*surf
,
2396 uint32_t logical_array_layer
,
2397 uint32_t *x_offset_sa
,
2398 uint32_t *y_offset_sa
)
2400 assert(level
< surf
->levels
);
2401 assert(surf
->logical_level0_px
.depth
== 1);
2402 assert(logical_array_layer
< surf
->logical_level0_px
.array_len
);
2404 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf
->format
);
2406 const struct isl_extent3d image_align_sa
=
2407 isl_surf_get_image_alignment_sa(surf
);
2409 struct isl_tile_info tile_info
;
2410 isl_tiling_get_info(surf
->tiling
, fmtl
->bpb
, &tile_info
);
2411 const struct isl_extent2d tile_extent_sa
= {
2412 .w
= tile_info
.logical_extent_el
.w
* fmtl
->bw
,
2413 .h
= tile_info
.logical_extent_el
.h
* fmtl
->bh
,
2415 /* Tile size is a multiple of image alignment */
2416 assert(tile_extent_sa
.w
% image_align_sa
.w
== 0);
2417 assert(tile_extent_sa
.h
% image_align_sa
.h
== 0);
2419 const uint32_t W0
= surf
->phys_level0_sa
.w
;
2420 const uint32_t H0
= surf
->phys_level0_sa
.h
;
2422 /* Each image has the same height as LOD0 because the hardware thinks
2423 * everything is LOD0
2425 const uint32_t H
= isl_align(H0
, image_align_sa
.h
);
2427 /* Quick sanity check for consistency */
2428 if (surf
->phys_level0_sa
.array_len
> 1)
2429 assert(surf
->array_pitch_el_rows
== isl_assert_div(H
, fmtl
->bh
));
2431 uint32_t x
= 0, y
= 0;
2432 for (uint32_t l
= 0; l
< level
; ++l
) {
2433 const uint32_t W
= isl_minify(W0
, l
);
2435 const uint32_t w
= isl_align(W
, tile_extent_sa
.w
);
2436 const uint32_t h
= isl_align(H
* surf
->phys_level0_sa
.a
,
2446 y
+= H
* logical_array_layer
;
2453 * A variant of isl_surf_get_image_offset_sa() specific to
2454 * ISL_DIM_LAYOUT_GEN9_1D.
2457 get_image_offset_sa_gen9_1d(const struct isl_surf
*surf
,
2458 uint32_t level
, uint32_t layer
,
2459 uint32_t *x_offset_sa
,
2460 uint32_t *y_offset_sa
)
2462 assert(level
< surf
->levels
);
2463 assert(layer
< surf
->phys_level0_sa
.array_len
);
2464 assert(surf
->phys_level0_sa
.height
== 1);
2465 assert(surf
->phys_level0_sa
.depth
== 1);
2466 assert(surf
->samples
== 1);
2468 const uint32_t W0
= surf
->phys_level0_sa
.width
;
2469 const struct isl_extent3d image_align_sa
=
2470 isl_surf_get_image_alignment_sa(surf
);
2474 for (uint32_t l
= 0; l
< level
; ++l
) {
2475 uint32_t W
= isl_minify(W0
, l
);
2476 uint32_t w
= isl_align_npot(W
, image_align_sa
.w
);
2482 *y_offset_sa
= layer
* isl_surf_get_array_pitch_sa_rows(surf
);
2486 * Calculate the offset, in units of surface samples, to a subimage in the
2489 * @invariant level < surface levels
2490 * @invariant logical_array_layer < logical array length of surface
2491 * @invariant logical_z_offset_px < logical depth of surface at level
2494 isl_surf_get_image_offset_sa(const struct isl_surf
*surf
,
2496 uint32_t logical_array_layer
,
2497 uint32_t logical_z_offset_px
,
2498 uint32_t *x_offset_sa
,
2499 uint32_t *y_offset_sa
)
2501 assert(level
< surf
->levels
);
2502 assert(logical_array_layer
< surf
->logical_level0_px
.array_len
);
2503 assert(logical_z_offset_px
2504 < isl_minify(surf
->logical_level0_px
.depth
, level
));
2506 switch (surf
->dim_layout
) {
2507 case ISL_DIM_LAYOUT_GEN9_1D
:
2508 get_image_offset_sa_gen9_1d(surf
, level
, logical_array_layer
,
2509 x_offset_sa
, y_offset_sa
);
2511 case ISL_DIM_LAYOUT_GEN4_2D
:
2512 get_image_offset_sa_gen4_2d(surf
, level
, logical_array_layer
2513 + logical_z_offset_px
,
2514 x_offset_sa
, y_offset_sa
);
2516 case ISL_DIM_LAYOUT_GEN4_3D
:
2517 get_image_offset_sa_gen4_3d(surf
, level
, logical_array_layer
+
2518 logical_z_offset_px
,
2519 x_offset_sa
, y_offset_sa
);
2521 case ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ
:
2522 get_image_offset_sa_gen6_stencil_hiz(surf
, level
, logical_array_layer
+
2523 logical_z_offset_px
,
2524 x_offset_sa
, y_offset_sa
);
2528 unreachable("not reached");
2533 isl_surf_get_image_offset_el(const struct isl_surf
*surf
,
2535 uint32_t logical_array_layer
,
2536 uint32_t logical_z_offset_px
,
2537 uint32_t *x_offset_el
,
2538 uint32_t *y_offset_el
)
2540 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf
->format
);
2542 assert(level
< surf
->levels
);
2543 assert(logical_array_layer
< surf
->logical_level0_px
.array_len
);
2544 assert(logical_z_offset_px
2545 < isl_minify(surf
->logical_level0_px
.depth
, level
));
2547 uint32_t x_offset_sa
, y_offset_sa
;
2548 isl_surf_get_image_offset_sa(surf
, level
,
2549 logical_array_layer
,
2550 logical_z_offset_px
,
2554 *x_offset_el
= x_offset_sa
/ fmtl
->bw
;
2555 *y_offset_el
= y_offset_sa
/ fmtl
->bh
;
2559 isl_surf_get_image_offset_B_tile_sa(const struct isl_surf
*surf
,
2561 uint32_t logical_array_layer
,
2562 uint32_t logical_z_offset_px
,
2564 uint32_t *x_offset_sa
,
2565 uint32_t *y_offset_sa
)
2567 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf
->format
);
2569 uint32_t total_x_offset_el
, total_y_offset_el
;
2570 isl_surf_get_image_offset_el(surf
, level
, logical_array_layer
,
2571 logical_z_offset_px
,
2573 &total_y_offset_el
);
2575 uint32_t x_offset_el
, y_offset_el
;
2576 isl_tiling_get_intratile_offset_el(surf
->tiling
, fmtl
->bpb
,
2585 *x_offset_sa
= x_offset_el
* fmtl
->bw
;
2587 assert(x_offset_el
== 0);
2591 *y_offset_sa
= y_offset_el
* fmtl
->bh
;
2593 assert(y_offset_el
== 0);
2598 isl_surf_get_image_range_B_tile(const struct isl_surf
*surf
,
2600 uint32_t logical_array_layer
,
2601 uint32_t logical_z_offset_px
,
2602 uint32_t *start_tile_B
,
2603 uint32_t *end_tile_B
)
2605 uint32_t start_x_offset_el
, start_y_offset_el
;
2606 isl_surf_get_image_offset_el(surf
, level
, logical_array_layer
,
2607 logical_z_offset_px
,
2609 &start_y_offset_el
);
2611 /* Compute the size of the subimage in surface elements */
2612 const uint32_t subimage_w_sa
= isl_minify(surf
->phys_level0_sa
.w
, level
);
2613 const uint32_t subimage_h_sa
= isl_minify(surf
->phys_level0_sa
.h
, level
);
2614 const struct isl_format_layout
*fmtl
= isl_format_get_layout(surf
->format
);
2615 const uint32_t subimage_w_el
= isl_align_div_npot(subimage_w_sa
, fmtl
->bw
);
2616 const uint32_t subimage_h_el
= isl_align_div_npot(subimage_h_sa
, fmtl
->bh
);
2618 /* Find the last pixel */
2619 uint32_t end_x_offset_el
= start_x_offset_el
+ subimage_w_el
- 1;
2620 uint32_t end_y_offset_el
= start_y_offset_el
+ subimage_h_el
- 1;
2622 UNUSED
uint32_t x_offset_el
, y_offset_el
;
2623 isl_tiling_get_intratile_offset_el(surf
->tiling
, fmtl
->bpb
,
2631 isl_tiling_get_intratile_offset_el(surf
->tiling
, fmtl
->bpb
,
2639 /* We want the range we return to be exclusive but the tile containing the
2640 * last pixel (what we just calculated) is inclusive. Add one.
2644 assert(*end_tile_B
<= surf
->size_B
);
2648 isl_surf_get_image_surf(const struct isl_device
*dev
,
2649 const struct isl_surf
*surf
,
2651 uint32_t logical_array_layer
,
2652 uint32_t logical_z_offset_px
,
2653 struct isl_surf
*image_surf
,
2655 uint32_t *x_offset_sa
,
2656 uint32_t *y_offset_sa
)
2658 isl_surf_get_image_offset_B_tile_sa(surf
,
2660 logical_array_layer
,
2661 logical_z_offset_px
,
2666 /* Even for cube maps there will be only single face, therefore drop the
2667 * corresponding flag if present.
2669 const isl_surf_usage_flags_t usage
=
2670 surf
->usage
& (~ISL_SURF_USAGE_CUBE_BIT
);
2673 ok
= isl_surf_init(dev
, image_surf
,
2674 .dim
= ISL_SURF_DIM_2D
,
2675 .format
= surf
->format
,
2676 .width
= isl_minify(surf
->logical_level0_px
.w
, level
),
2677 .height
= isl_minify(surf
->logical_level0_px
.h
, level
),
2681 .samples
= surf
->samples
,
2682 .row_pitch_B
= surf
->row_pitch_B
,
2684 .tiling_flags
= (1 << surf
->tiling
));
2689 isl_tiling_get_intratile_offset_el(enum isl_tiling tiling
,
2691 uint32_t row_pitch_B
,
2692 uint32_t total_x_offset_el
,
2693 uint32_t total_y_offset_el
,
2694 uint32_t *base_address_offset
,
2695 uint32_t *x_offset_el
,
2696 uint32_t *y_offset_el
)
2698 if (tiling
== ISL_TILING_LINEAR
) {
2699 assert(bpb
% 8 == 0);
2700 *base_address_offset
= total_y_offset_el
* row_pitch_B
+
2701 total_x_offset_el
* (bpb
/ 8);
2707 struct isl_tile_info tile_info
;
2708 isl_tiling_get_info(tiling
, bpb
, &tile_info
);
2710 assert(row_pitch_B
% tile_info
.phys_extent_B
.width
== 0);
2712 /* For non-power-of-two formats, we need the address to be both tile and
2713 * element-aligned. The easiest way to achieve this is to work with a tile
2714 * that is three times as wide as the regular tile.
2716 * The tile info returned by get_tile_info has a logical size that is an
2717 * integer number of tile_info.format_bpb size elements. To scale the
2718 * tile, we scale up the physical width and then treat the logical tile
2719 * size as if it has bpb size elements.
2721 const uint32_t tile_el_scale
= bpb
/ tile_info
.format_bpb
;
2722 tile_info
.phys_extent_B
.width
*= tile_el_scale
;
2724 /* Compute the offset into the tile */
2725 *x_offset_el
= total_x_offset_el
% tile_info
.logical_extent_el
.w
;
2726 *y_offset_el
= total_y_offset_el
% tile_info
.logical_extent_el
.h
;
2728 /* Compute the offset of the tile in units of whole tiles */
2729 uint32_t x_offset_tl
= total_x_offset_el
/ tile_info
.logical_extent_el
.w
;
2730 uint32_t y_offset_tl
= total_y_offset_el
/ tile_info
.logical_extent_el
.h
;
2732 *base_address_offset
=
2733 y_offset_tl
* tile_info
.phys_extent_B
.h
* row_pitch_B
+
2734 x_offset_tl
* tile_info
.phys_extent_B
.h
* tile_info
.phys_extent_B
.w
;
2738 isl_surf_get_depth_format(const struct isl_device
*dev
,
2739 const struct isl_surf
*surf
)
2741 /* Support for separate stencil buffers began in gen5. Support for
2742 * interleaved depthstencil buffers ceased in gen7. The intermediate gens,
2743 * those that supported separate and interleaved stencil, were gen5 and
2746 * For a list of all available formats, see the Sandybridge PRM >> Volume
2747 * 2 Part 1: 3D/Media - 3D Pipeline >> 3DSTATE_DEPTH_BUFFER >> Surface
2751 bool has_stencil
= surf
->usage
& ISL_SURF_USAGE_STENCIL_BIT
;
2753 assert(surf
->usage
& ISL_SURF_USAGE_DEPTH_BIT
);
2756 assert(ISL_DEV_GEN(dev
) < 7);
2758 switch (surf
->format
) {
2760 unreachable("bad isl depth format");
2761 case ISL_FORMAT_R32_FLOAT_X8X24_TYPELESS
:
2762 assert(ISL_DEV_GEN(dev
) < 7);
2763 return 0; /* D32_FLOAT_S8X24_UINT */
2764 case ISL_FORMAT_R32_FLOAT
:
2765 assert(!has_stencil
);
2766 return 1; /* D32_FLOAT */
2767 case ISL_FORMAT_R24_UNORM_X8_TYPELESS
:
2769 assert(ISL_DEV_GEN(dev
) < 7);
2770 return 2; /* D24_UNORM_S8_UINT */
2772 assert(ISL_DEV_GEN(dev
) >= 5);
2773 return 3; /* D24_UNORM_X8_UINT */
2775 case ISL_FORMAT_R16_UNORM
:
2776 assert(!has_stencil
);
2777 return 5; /* D16_UNORM */
2782 isl_swizzle_supports_rendering(const struct gen_device_info
*devinfo
,
2783 struct isl_swizzle swizzle
)
2785 if (devinfo
->is_haswell
) {
2786 /* From the Haswell PRM,
2787 * RENDER_SURFACE_STATE::Shader Channel Select Red
2789 * "The Shader channel selects also define which shader channels are
2790 * written to which surface channel. If the Shader channel select is
2791 * SCS_ZERO or SCS_ONE then it is not written to the surface. If the
2792 * shader channel select is SCS_RED it is written to the surface red
2793 * channel and so on. If more than one shader channel select is set
2794 * to the same surface channel only the first shader channel in RGBA
2795 * order will be written."
2798 } else if (devinfo
->gen
<= 7) {
2799 /* Ivy Bridge and early doesn't have any swizzling */
2800 return isl_swizzle_is_identity(swizzle
);
2802 /* From the Sky Lake PRM Vol. 2d,
2803 * RENDER_SURFACE_STATE::Shader Channel Select Red
2805 * "For Render Target, Red, Green and Blue Shader Channel Selects
2806 * MUST be such that only valid components can be swapped i.e. only
2807 * change the order of components in the pixel. Any other values for
2808 * these Shader Channel Select fields are not valid for Render
2809 * Targets. This also means that there MUST not be multiple shader
2810 * channels mapped to the same RT channel."
2812 * From the Sky Lake PRM Vol. 2d,
2813 * RENDER_SURFACE_STATE::Shader Channel Select Alpha
2815 * "For Render Target, this field MUST be programmed to
2816 * value = SCS_ALPHA."
2818 return (swizzle
.r
== ISL_CHANNEL_SELECT_RED
||
2819 swizzle
.r
== ISL_CHANNEL_SELECT_GREEN
||
2820 swizzle
.r
== ISL_CHANNEL_SELECT_BLUE
) &&
2821 (swizzle
.g
== ISL_CHANNEL_SELECT_RED
||
2822 swizzle
.g
== ISL_CHANNEL_SELECT_GREEN
||
2823 swizzle
.g
== ISL_CHANNEL_SELECT_BLUE
) &&
2824 (swizzle
.b
== ISL_CHANNEL_SELECT_RED
||
2825 swizzle
.b
== ISL_CHANNEL_SELECT_GREEN
||
2826 swizzle
.b
== ISL_CHANNEL_SELECT_BLUE
) &&
2827 swizzle
.r
!= swizzle
.g
&&
2828 swizzle
.r
!= swizzle
.b
&&
2829 swizzle
.g
!= swizzle
.b
&&
2830 swizzle
.a
== ISL_CHANNEL_SELECT_ALPHA
;
2834 static enum isl_channel_select
2835 swizzle_select(enum isl_channel_select chan
, struct isl_swizzle swizzle
)
2838 case ISL_CHANNEL_SELECT_ZERO
:
2839 case ISL_CHANNEL_SELECT_ONE
:
2841 case ISL_CHANNEL_SELECT_RED
:
2843 case ISL_CHANNEL_SELECT_GREEN
:
2845 case ISL_CHANNEL_SELECT_BLUE
:
2847 case ISL_CHANNEL_SELECT_ALPHA
:
2850 unreachable("Invalid swizzle component");
2855 * Returns the single swizzle that is equivalent to applying the two given
2856 * swizzles in sequence.
2859 isl_swizzle_compose(struct isl_swizzle first
, struct isl_swizzle second
)
2861 return (struct isl_swizzle
) {
2862 .r
= swizzle_select(first
.r
, second
),
2863 .g
= swizzle_select(first
.g
, second
),
2864 .b
= swizzle_select(first
.b
, second
),
2865 .a
= swizzle_select(first
.a
, second
),
2870 * Returns a swizzle that is the pseudo-inverse of this swizzle.
2873 isl_swizzle_invert(struct isl_swizzle swizzle
)
2875 /* Default to zero for channels which do not show up in the swizzle */
2876 enum isl_channel_select chans
[4] = {
2877 ISL_CHANNEL_SELECT_ZERO
,
2878 ISL_CHANNEL_SELECT_ZERO
,
2879 ISL_CHANNEL_SELECT_ZERO
,
2880 ISL_CHANNEL_SELECT_ZERO
,
2883 /* We go in ABGR order so that, if there are any duplicates, the first one
2884 * is taken if you look at it in RGBA order. This is what Haswell hardware
2885 * does for render target swizzles.
2887 if ((unsigned)(swizzle
.a
- ISL_CHANNEL_SELECT_RED
) < 4)
2888 chans
[swizzle
.a
- ISL_CHANNEL_SELECT_RED
] = ISL_CHANNEL_SELECT_ALPHA
;
2889 if ((unsigned)(swizzle
.b
- ISL_CHANNEL_SELECT_RED
) < 4)
2890 chans
[swizzle
.b
- ISL_CHANNEL_SELECT_RED
] = ISL_CHANNEL_SELECT_BLUE
;
2891 if ((unsigned)(swizzle
.g
- ISL_CHANNEL_SELECT_RED
) < 4)
2892 chans
[swizzle
.g
- ISL_CHANNEL_SELECT_RED
] = ISL_CHANNEL_SELECT_GREEN
;
2893 if ((unsigned)(swizzle
.r
- ISL_CHANNEL_SELECT_RED
) < 4)
2894 chans
[swizzle
.r
- ISL_CHANNEL_SELECT_RED
] = ISL_CHANNEL_SELECT_RED
;
2896 return (struct isl_swizzle
) { chans
[0], chans
[1], chans
[2], chans
[3] };
2899 /** Applies an inverse swizzle to a color value */
2900 union isl_color_value
2901 isl_color_value_swizzle_inv(union isl_color_value src
,
2902 struct isl_swizzle swizzle
)
2904 union isl_color_value dst
= { .u32
= { 0, } };
2906 /* We assign colors in ABGR order so that the first one will be taken in
2907 * RGBA precedence order. According to the PRM docs for shader channel
2908 * select, this matches Haswell hardware behavior.
2910 if ((unsigned)(swizzle
.a
- ISL_CHANNEL_SELECT_RED
) < 4)
2911 dst
.u32
[swizzle
.a
- ISL_CHANNEL_SELECT_RED
] = src
.u32
[3];
2912 if ((unsigned)(swizzle
.b
- ISL_CHANNEL_SELECT_RED
) < 4)
2913 dst
.u32
[swizzle
.b
- ISL_CHANNEL_SELECT_RED
] = src
.u32
[2];
2914 if ((unsigned)(swizzle
.g
- ISL_CHANNEL_SELECT_RED
) < 4)
2915 dst
.u32
[swizzle
.g
- ISL_CHANNEL_SELECT_RED
] = src
.u32
[1];
2916 if ((unsigned)(swizzle
.r
- ISL_CHANNEL_SELECT_RED
) < 4)
2917 dst
.u32
[swizzle
.r
- ISL_CHANNEL_SELECT_RED
] = src
.u32
[0];
2923 isl_format_get_aux_map_encoding(enum isl_format format
)
2926 case ISL_FORMAT_R32G32B32A32_FLOAT
: return 0x11;
2927 case ISL_FORMAT_R32G32B32X32_FLOAT
: return 0x11;
2928 case ISL_FORMAT_R32G32B32A32_SINT
: return 0x12;
2929 case ISL_FORMAT_R32G32B32A32_UINT
: return 0x13;
2930 case ISL_FORMAT_R16G16B16A16_UNORM
: return 0x14;
2931 case ISL_FORMAT_R16G16B16A16_SNORM
: return 0x15;
2932 case ISL_FORMAT_R16G16B16A16_SINT
: return 0x16;
2933 case ISL_FORMAT_R16G16B16A16_UINT
: return 0x17;
2934 case ISL_FORMAT_R16G16B16A16_FLOAT
: return 0x10;
2935 case ISL_FORMAT_R16G16B16X16_FLOAT
: return 0x10;
2936 case ISL_FORMAT_R32G32_FLOAT
: return 0x11;
2937 case ISL_FORMAT_R32G32_SINT
: return 0x12;
2938 case ISL_FORMAT_R32G32_UINT
: return 0x13;
2939 case ISL_FORMAT_B8G8R8A8_UNORM
: return 0xA;
2940 case ISL_FORMAT_B8G8R8X8_UNORM
: return 0xA;
2941 case ISL_FORMAT_B8G8R8A8_UNORM_SRGB
: return 0xA;
2942 case ISL_FORMAT_B8G8R8X8_UNORM_SRGB
: return 0xA;
2943 case ISL_FORMAT_R10G10B10A2_UNORM
: return 0x18;
2944 case ISL_FORMAT_R10G10B10A2_UNORM_SRGB
: return 0x18;
2945 case ISL_FORMAT_R10G10B10_FLOAT_A2_UNORM
: return 0x19;
2946 case ISL_FORMAT_R10G10B10A2_UINT
: return 0x1A;
2947 case ISL_FORMAT_R8G8B8A8_UNORM
: return 0xA;
2948 case ISL_FORMAT_R8G8B8A8_UNORM_SRGB
: return 0xA;
2949 case ISL_FORMAT_R8G8B8A8_SNORM
: return 0x1B;
2950 case ISL_FORMAT_R8G8B8A8_SINT
: return 0x1C;
2951 case ISL_FORMAT_R8G8B8A8_UINT
: return 0x1D;
2952 case ISL_FORMAT_R16G16_UNORM
: return 0x14;
2953 case ISL_FORMAT_R16G16_SNORM
: return 0x15;
2954 case ISL_FORMAT_R16G16_SINT
: return 0x16;
2955 case ISL_FORMAT_R16G16_UINT
: return 0x17;
2956 case ISL_FORMAT_R16G16_FLOAT
: return 0x10;
2957 case ISL_FORMAT_B10G10R10A2_UNORM
: return 0x18;
2958 case ISL_FORMAT_B10G10R10A2_UNORM_SRGB
: return 0x18;
2959 case ISL_FORMAT_R11G11B10_FLOAT
: return 0x1E;
2960 case ISL_FORMAT_R32_SINT
: return 0x12;
2961 case ISL_FORMAT_R32_UINT
: return 0x13;
2962 case ISL_FORMAT_R32_FLOAT
: return 0x11;
2963 case ISL_FORMAT_R24_UNORM_X8_TYPELESS
: return 0x11;
2964 case ISL_FORMAT_B5G6R5_UNORM
: return 0xA;
2965 case ISL_FORMAT_B5G6R5_UNORM_SRGB
: return 0xA;
2966 case ISL_FORMAT_B5G5R5A1_UNORM
: return 0xA;
2967 case ISL_FORMAT_B5G5R5A1_UNORM_SRGB
: return 0xA;
2968 case ISL_FORMAT_B4G4R4A4_UNORM
: return 0xA;
2969 case ISL_FORMAT_B4G4R4A4_UNORM_SRGB
: return 0xA;
2970 case ISL_FORMAT_R8G8_UNORM
: return 0xA;
2971 case ISL_FORMAT_R8G8_SNORM
: return 0x1B;
2972 case ISL_FORMAT_R8G8_SINT
: return 0x1C;
2973 case ISL_FORMAT_R8G8_UINT
: return 0x1D;
2974 case ISL_FORMAT_R16_UNORM
: return 0x14;
2975 case ISL_FORMAT_R16_SNORM
: return 0x15;
2976 case ISL_FORMAT_R16_SINT
: return 0x16;
2977 case ISL_FORMAT_R16_UINT
: return 0x17;
2978 case ISL_FORMAT_R16_FLOAT
: return 0x10;
2979 case ISL_FORMAT_B5G5R5X1_UNORM
: return 0xA;
2980 case ISL_FORMAT_B5G5R5X1_UNORM_SRGB
: return 0xA;
2981 case ISL_FORMAT_A1B5G5R5_UNORM
: return 0xA;
2982 case ISL_FORMAT_A4B4G4R4_UNORM
: return 0xA;
2983 case ISL_FORMAT_R8_UNORM
: return 0xA;
2984 case ISL_FORMAT_R8_SNORM
: return 0x1B;
2985 case ISL_FORMAT_R8_SINT
: return 0x1C;
2986 case ISL_FORMAT_R8_UINT
: return 0x1D;
2987 case ISL_FORMAT_A8_UNORM
: return 0xA;
2989 unreachable("Unsupported aux-map format!");