2 * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
3 * Copyright 2010 Marek Olšák <maraeo@gmail.com>
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE. */
24 #include "r300_texture_desc.h"
25 #include "r300_context.h"
27 #include "util/u_format.h"
29 /* Returns the number of pixels that the texture should be aligned to
30 * in the given dimension. */
31 unsigned r300_get_pixel_alignment(enum pipe_format format
,
33 enum radeon_bo_layout microtile
,
34 enum radeon_bo_layout macrotile
,
35 enum r300_dim dim
, boolean is_rs690
)
37 static const unsigned table
[2][5][3][2] =
40 /* Macro: linear linear linear
41 Micro: linear tiled square-tiled */
42 {{ 32, 1}, { 8, 4}, { 0, 0}}, /* 8 bits per pixel */
43 {{ 16, 1}, { 8, 2}, { 4, 4}}, /* 16 bits per pixel */
44 {{ 8, 1}, { 4, 2}, { 0, 0}}, /* 32 bits per pixel */
45 {{ 4, 1}, { 2, 2}, { 0, 0}}, /* 64 bits per pixel */
46 {{ 2, 1}, { 0, 0}, { 0, 0}} /* 128 bits per pixel */
49 /* Macro: tiled tiled tiled
50 Micro: linear tiled square-tiled */
51 {{256, 8}, {64, 32}, { 0, 0}}, /* 8 bits per pixel */
52 {{128, 8}, {64, 16}, {32, 32}}, /* 16 bits per pixel */
53 {{ 64, 8}, {32, 16}, { 0, 0}}, /* 32 bits per pixel */
54 {{ 32, 8}, {16, 16}, { 0, 0}}, /* 64 bits per pixel */
55 {{ 16, 8}, { 0, 0}, { 0, 0}} /* 128 bits per pixel */
60 unsigned pixsize
= util_format_get_blocksize(format
);
62 assert(macrotile
<= RADEON_LAYOUT_TILED
);
63 assert(microtile
<= RADEON_LAYOUT_SQUARETILED
);
64 assert(pixsize
<= 16);
65 assert(dim
<= DIM_HEIGHT
);
67 tile
= table
[macrotile
][util_logbase2(pixsize
)][microtile
][dim
];
68 if (macrotile
== 0 && is_rs690
&& dim
== DIM_WIDTH
) {
71 h_tile
= table
[macrotile
][util_logbase2(pixsize
)][microtile
][DIM_HEIGHT
];
72 align
= 64 / (pixsize
* h_tile
);
81 /* Return true if macrotiling should be enabled on the miplevel. */
82 static boolean
r300_texture_macro_switch(struct r300_resource
*tex
,
87 unsigned tile
, texdim
;
89 if (tex
->b
.b
.nr_samples
> 1) {
93 tile
= r300_get_pixel_alignment(tex
->b
.b
.format
, tex
->b
.b
.nr_samples
,
94 tex
->tex
.microtile
, RADEON_LAYOUT_TILED
, dim
, 0);
95 if (dim
== DIM_WIDTH
) {
96 texdim
= u_minify(tex
->tex
.width0
, level
);
98 texdim
= u_minify(tex
->tex
.height0
, level
);
101 /* See TX_FILTER1_n.MACRO_SWITCH. */
103 return texdim
>= tile
;
105 return texdim
> tile
;
110 * Return the stride, in bytes, of the texture image of the given texture
111 * at the given level.
113 static unsigned r300_texture_get_stride(struct r300_screen
*screen
,
114 struct r300_resource
*tex
,
117 unsigned tile_width
, width
, stride
;
118 boolean is_rs690
= (screen
->caps
.family
== CHIP_FAMILY_RS600
||
119 screen
->caps
.family
== CHIP_FAMILY_RS690
||
120 screen
->caps
.family
== CHIP_FAMILY_RS740
);
122 if (tex
->tex
.stride_in_bytes_override
)
123 return tex
->tex
.stride_in_bytes_override
;
125 /* Check the level. */
126 if (level
> tex
->b
.b
.last_level
) {
127 SCREEN_DBG(screen
, DBG_TEX
, "%s: level (%u) > last_level (%u)\n",
128 __FUNCTION__
, level
, tex
->b
.b
.last_level
);
132 width
= u_minify(tex
->tex
.width0
, level
);
134 if (util_format_is_plain(tex
->b
.b
.format
)) {
135 tile_width
= r300_get_pixel_alignment(tex
->b
.b
.format
,
138 tex
->tex
.macrotile
[level
],
139 DIM_WIDTH
, is_rs690
);
140 width
= align(width
, tile_width
);
142 stride
= util_format_get_stride(tex
->b
.b
.format
, width
);
143 /* The alignment to 32 bytes is sort of implied by the layout... */
146 return align(util_format_get_stride(tex
->b
.b
.format
, width
), is_rs690
? 64 : 32);
150 static unsigned r300_texture_get_nblocksy(struct r300_resource
*tex
,
152 boolean
*out_aligned_for_cbzb
)
154 unsigned height
, tile_height
;
156 height
= u_minify(tex
->tex
.height0
, level
);
158 /* Mipmapped and 3D textures must have their height aligned to POT. */
159 if ((tex
->b
.b
.target
!= PIPE_TEXTURE_1D
&&
160 tex
->b
.b
.target
!= PIPE_TEXTURE_2D
&&
161 tex
->b
.b
.target
!= PIPE_TEXTURE_RECT
) ||
162 tex
->b
.b
.last_level
!= 0) {
163 height
= util_next_power_of_two(height
);
166 if (util_format_is_plain(tex
->b
.b
.format
)) {
167 tile_height
= r300_get_pixel_alignment(tex
->b
.b
.format
,
170 tex
->tex
.macrotile
[level
],
172 height
= align(height
, tile_height
);
174 /* See if the CBZB clear can be used on the buffer,
175 * taking the texture size into account. */
176 if (out_aligned_for_cbzb
) {
177 if (tex
->tex
.macrotile
[level
]) {
178 /* When clearing, the layer (width*height) is horizontally split
179 * into two, and the upper and lower halves are cleared by the CB
180 * and ZB units, respectively. Therefore, the number of macrotiles
181 * in the Y direction must be even. */
183 /* Align the height so that there is an even number of macrotiles.
184 * Do so for 3 or more macrotiles in the Y direction. */
185 if (level
== 0 && tex
->b
.b
.last_level
== 0 &&
186 (tex
->b
.b
.target
== PIPE_TEXTURE_1D
||
187 tex
->b
.b
.target
== PIPE_TEXTURE_2D
||
188 tex
->b
.b
.target
== PIPE_TEXTURE_RECT
) &&
189 height
>= tile_height
* 3) {
190 height
= align(height
, tile_height
* 2);
193 *out_aligned_for_cbzb
= height
% (tile_height
* 2) == 0;
195 *out_aligned_for_cbzb
= FALSE
;
200 return util_format_get_nblocksy(tex
->b
.b
.format
, height
);
203 /* Get a width in pixels from a stride in bytes. */
204 unsigned r300_stride_to_width(enum pipe_format format
,
205 unsigned stride_in_bytes
)
207 return (stride_in_bytes
/ util_format_get_blocksize(format
)) *
208 util_format_get_blockwidth(format
);
211 static void r300_setup_miptree(struct r300_screen
*screen
,
212 struct r300_resource
*tex
,
213 boolean align_for_cbzb
)
215 struct pipe_resource
*base
= &tex
->b
.b
;
216 unsigned stride
, size
, layer_size
, nblocksy
, i
;
217 boolean rv350_mode
= screen
->caps
.family
>= CHIP_FAMILY_R350
;
218 boolean aligned_for_cbzb
;
220 tex
->tex
.size_in_bytes
= 0;
222 SCREEN_DBG(screen
, DBG_TEXALLOC
,
223 "r300: Making miptree for texture, format %s\n",
224 util_format_short_name(base
->format
));
226 for (i
= 0; i
<= base
->last_level
; i
++) {
227 /* Let's see if this miplevel can be macrotiled. */
228 tex
->tex
.macrotile
[i
] =
229 (tex
->tex
.macrotile
[0] == RADEON_LAYOUT_TILED
&&
230 r300_texture_macro_switch(tex
, i
, rv350_mode
, DIM_WIDTH
) &&
231 r300_texture_macro_switch(tex
, i
, rv350_mode
, DIM_HEIGHT
)) ?
232 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
234 stride
= r300_texture_get_stride(screen
, tex
, i
);
236 /* Compute the number of blocks in Y, see if the CBZB clear can be
237 * used on the texture. */
238 aligned_for_cbzb
= FALSE
;
239 if (align_for_cbzb
&& tex
->tex
.cbzb_allowed
[i
])
240 nblocksy
= r300_texture_get_nblocksy(tex
, i
, &aligned_for_cbzb
);
242 nblocksy
= r300_texture_get_nblocksy(tex
, i
, NULL
);
244 layer_size
= stride
* nblocksy
;
246 if (base
->nr_samples
> 1) {
247 layer_size
*= base
->nr_samples
;
250 if (base
->target
== PIPE_TEXTURE_CUBE
)
251 size
= layer_size
* 6;
253 size
= layer_size
* u_minify(tex
->tex
.depth0
, i
);
255 tex
->tex
.offset_in_bytes
[i
] = tex
->tex
.size_in_bytes
;
256 tex
->tex
.size_in_bytes
= tex
->tex
.offset_in_bytes
[i
] + size
;
257 tex
->tex
.layer_size_in_bytes
[i
] = layer_size
;
258 tex
->tex
.stride_in_bytes
[i
] = stride
;
259 tex
->tex
.cbzb_allowed
[i
] = tex
->tex
.cbzb_allowed
[i
] && aligned_for_cbzb
;
261 SCREEN_DBG(screen
, DBG_TEXALLOC
, "r300: Texture miptree: Level %d "
262 "(%dx%dx%d px, pitch %d bytes) %d bytes total, macrotiled %s\n",
263 i
, u_minify(tex
->tex
.width0
, i
), u_minify(tex
->tex
.height0
, i
),
264 u_minify(tex
->tex
.depth0
, i
), stride
, tex
->tex
.size_in_bytes
,
265 tex
->tex
.macrotile
[i
] ? "TRUE" : "FALSE");
269 static void r300_setup_flags(struct r300_resource
*tex
)
271 tex
->tex
.uses_stride_addressing
=
272 !util_is_power_of_two(tex
->b
.b
.width0
) ||
273 (tex
->tex
.stride_in_bytes_override
&&
274 r300_stride_to_width(tex
->b
.b
.format
,
275 tex
->tex
.stride_in_bytes_override
) != tex
->b
.b
.width0
);
278 tex
->tex
.uses_stride_addressing
||
279 !util_is_power_of_two(tex
->b
.b
.height0
) ||
280 !util_is_power_of_two(tex
->b
.b
.depth0
);
283 static void r300_setup_cbzb_flags(struct r300_screen
*rscreen
,
284 struct r300_resource
*tex
)
287 boolean first_level_valid
;
289 bpp
= util_format_get_blocksizebits(tex
->b
.b
.format
);
291 /* 1) The texture must be point-sampled,
292 * 2) The depth must be 16 or 32 bits.
293 * 3) If the midpoint ZB offset is not aligned to 2048, it returns garbage
294 * with certain texture sizes. Macrotiling ensures the alignment. */
295 first_level_valid
= tex
->b
.b
.nr_samples
<= 1 &&
296 (bpp
== 16 || bpp
== 32) &&
297 tex
->tex
.macrotile
[0];
299 if (SCREEN_DBG_ON(rscreen
, DBG_NO_CBZB
))
300 first_level_valid
= FALSE
;
302 for (i
= 0; i
<= tex
->b
.b
.last_level
; i
++)
303 tex
->tex
.cbzb_allowed
[i
] = first_level_valid
&& tex
->tex
.macrotile
[i
];
306 static unsigned r300_pixels_to_dwords(unsigned stride
,
308 unsigned xblock
, unsigned yblock
)
310 return (util_align_npot(stride
, xblock
) * align(height
, yblock
)) / (xblock
* yblock
);
313 static void r300_setup_hyperz_properties(struct r300_screen
*screen
,
314 struct r300_resource
*tex
)
316 /* The tile size of 1 DWORD in ZMASK RAM is:
318 * GPU Pipes 4x4 mode 8x8 mode
319 * ------------------------------------------
320 * R580 4P/1Z 32x32 64x64
321 * RV570 3P/1Z 48x16 96x32
322 * RV530 1P/2Z 32x16 64x32
325 static unsigned zmask_blocks_x_per_dw
[4] = {4, 8, 12, 8};
326 static unsigned zmask_blocks_y_per_dw
[4] = {4, 4, 4, 8};
328 /* In HIZ RAM, one dword is always 8x8 pixels (each byte is 4x4 pixels),
329 * but the blocks have very weird ordering.
331 * With 2 pipes and an image of size 8xY, where Y >= 1,
332 * clearing 4 dwords clears blocks like this:
336 * where numbers correspond to dword indices. The blocks are interleaved
337 * in the X direction, so the alignment must be 4x1 blocks (32x8 pixels).
339 * With 4 pipes and an image of size 8xY, where Y >= 4,
340 * clearing 8 dwords clears blocks like this:
345 * where numbers correspond to dword indices. The blocks are interleaved
346 * in both directions, so the alignment must be 4x4 blocks (32x32 pixels)
348 static unsigned hiz_align_x
[4] = {8, 32, 48, 32};
349 static unsigned hiz_align_y
[4] = {8, 8, 8, 32};
351 if (util_format_is_depth_or_stencil(tex
->b
.b
.format
) &&
352 util_format_get_blocksizebits(tex
->b
.b
.format
) == 32 &&
353 tex
->tex
.microtile
) {
356 if (screen
->caps
.family
== CHIP_FAMILY_RV530
) {
357 pipes
= screen
->info
.r300_num_z_pipes
;
359 pipes
= screen
->info
.r300_num_gb_pipes
;
362 for (i
= 0; i
<= tex
->b
.b
.last_level
; i
++) {
363 unsigned zcomp_numdw
, zcompsize
, hiz_numdw
, stride
, height
;
365 stride
= r300_stride_to_width(tex
->b
.b
.format
,
366 tex
->tex
.stride_in_bytes
[i
]);
367 stride
= align(stride
, 16);
368 height
= u_minify(tex
->b
.b
.height0
, i
);
370 /* The 8x8 compression mode needs macrotiling. */
371 zcompsize
= screen
->caps
.z_compress
== R300_ZCOMP_8X8
&&
372 tex
->tex
.macrotile
[i
] &&
373 tex
->b
.b
.nr_samples
<= 1 ? 8 : 4;
375 /* Get the ZMASK buffer size in dwords. */
376 zcomp_numdw
= r300_pixels_to_dwords(stride
, height
,
377 zmask_blocks_x_per_dw
[pipes
-1] * zcompsize
,
378 zmask_blocks_y_per_dw
[pipes
-1] * zcompsize
);
380 /* Check whether we have enough ZMASK memory. */
381 if (util_format_get_blocksizebits(tex
->b
.b
.format
) == 32 &&
382 zcomp_numdw
<= screen
->caps
.zmask_ram
* pipes
) {
383 tex
->tex
.zmask_dwords
[i
] = zcomp_numdw
;
384 tex
->tex
.zcomp8x8
[i
] = zcompsize
== 8;
386 tex
->tex
.zmask_stride_in_pixels
[i
] =
387 util_align_npot(stride
, zmask_blocks_x_per_dw
[pipes
-1] * zcompsize
);
389 tex
->tex
.zmask_dwords
[i
] = 0;
390 tex
->tex
.zcomp8x8
[i
] = FALSE
;
391 tex
->tex
.zmask_stride_in_pixels
[i
] = 0;
395 stride
= util_align_npot(stride
, hiz_align_x
[pipes
-1]);
396 height
= align(height
, hiz_align_y
[pipes
-1]);
398 /* Get the HIZ buffer size in dwords. */
399 hiz_numdw
= (stride
* height
) / (8*8 * pipes
);
401 /* Check whether we have enough HIZ memory. */
402 if (hiz_numdw
<= screen
->caps
.hiz_ram
* pipes
) {
403 tex
->tex
.hiz_dwords
[i
] = hiz_numdw
;
404 tex
->tex
.hiz_stride_in_pixels
[i
] = stride
;
406 tex
->tex
.hiz_dwords
[i
] = 0;
407 tex
->tex
.hiz_stride_in_pixels
[i
] = 0;
413 static void r300_setup_tiling(struct r300_screen
*screen
,
414 struct r300_resource
*tex
)
416 enum pipe_format format
= tex
->b
.b
.format
;
417 boolean rv350_mode
= screen
->caps
.family
>= CHIP_FAMILY_R350
;
418 boolean is_zb
= util_format_is_depth_or_stencil(format
);
419 boolean dbg_no_tiling
= SCREEN_DBG_ON(screen
, DBG_NO_TILING
);
421 if (tex
->b
.b
.nr_samples
> 1) {
422 tex
->tex
.microtile
= RADEON_LAYOUT_TILED
;
423 tex
->tex
.macrotile
[0] = RADEON_LAYOUT_TILED
;
427 tex
->tex
.microtile
= RADEON_LAYOUT_LINEAR
;
428 tex
->tex
.macrotile
[0] = RADEON_LAYOUT_LINEAR
;
430 if (!util_format_is_plain(format
)) {
434 /* If height == 1, disable microtiling except for zbuffer. */
435 if (!is_zb
&& (tex
->b
.b
.height0
== 1 || dbg_no_tiling
)) {
439 /* Set microtiling. */
440 switch (util_format_get_blocksize(format
)) {
444 tex
->tex
.microtile
= RADEON_LAYOUT_TILED
;
448 tex
->tex
.microtile
= RADEON_LAYOUT_SQUARETILED
;
456 /* Set macrotiling. */
457 if (r300_texture_macro_switch(tex
, 0, rv350_mode
, DIM_WIDTH
) &&
458 r300_texture_macro_switch(tex
, 0, rv350_mode
, DIM_HEIGHT
)) {
459 tex
->tex
.macrotile
[0] = RADEON_LAYOUT_TILED
;
463 static void r300_tex_print_info(struct r300_resource
*tex
,
467 "r300: %s: Macro: %s, Micro: %s, Pitch: %i, Dim: %ix%ix%i, "
468 "LastLevel: %i, Size: %i, Format: %s\n",
470 tex
->tex
.macrotile
[0] ? "YES" : " NO",
471 tex
->tex
.microtile
? "YES" : " NO",
472 r300_stride_to_width(tex
->b
.b
.format
, tex
->tex
.stride_in_bytes
[0]),
473 tex
->b
.b
.width0
, tex
->b
.b
.height0
, tex
->b
.b
.depth0
,
474 tex
->b
.b
.last_level
, tex
->tex
.size_in_bytes
,
475 util_format_short_name(tex
->b
.b
.format
));
478 void r300_texture_desc_init(struct r300_screen
*rscreen
,
479 struct r300_resource
*tex
,
480 const struct pipe_resource
*base
)
482 tex
->b
.b
.target
= base
->target
;
483 tex
->b
.b
.format
= base
->format
;
484 tex
->b
.b
.width0
= base
->width0
;
485 tex
->b
.b
.height0
= base
->height0
;
486 tex
->b
.b
.depth0
= base
->depth0
;
487 tex
->b
.b
.array_size
= base
->array_size
;
488 tex
->b
.b
.last_level
= base
->last_level
;
489 tex
->b
.b
.nr_samples
= base
->nr_samples
;
490 tex
->tex
.width0
= base
->width0
;
491 tex
->tex
.height0
= base
->height0
;
492 tex
->tex
.depth0
= base
->depth0
;
494 r300_setup_flags(tex
);
496 /* Align a 3D NPOT texture to POT. */
497 if (base
->target
== PIPE_TEXTURE_3D
&& tex
->tex
.is_npot
) {
498 tex
->tex
.width0
= util_next_power_of_two(tex
->tex
.width0
);
499 tex
->tex
.height0
= util_next_power_of_two(tex
->tex
.height0
);
500 tex
->tex
.depth0
= util_next_power_of_two(tex
->tex
.depth0
);
504 if (tex
->tex
.microtile
== RADEON_LAYOUT_UNKNOWN
) {
505 r300_setup_tiling(rscreen
, tex
);
508 r300_setup_cbzb_flags(rscreen
, tex
);
510 /* Setup the miptree description. */
511 r300_setup_miptree(rscreen
, tex
, TRUE
);
512 /* If the required buffer size is larger than the given max size,
513 * try again without the alignment for the CBZB clear. */
514 if (tex
->buf
&& tex
->tex
.size_in_bytes
> tex
->buf
->size
) {
515 r300_setup_miptree(rscreen
, tex
, FALSE
);
517 /* Make sure the buffer we got is large enough. */
518 if (tex
->tex
.size_in_bytes
> tex
->buf
->size
) {
520 "r300: I got a pre-allocated buffer to use it as a texture "
521 "storage, but the buffer is too small. I'll use the buffer "
522 "anyway, because I can't crash here, but it's dangerous. "
523 "This can be a DDX bug. Got: %iB, Need: %iB, Info:\n",
524 tex
->buf
->size
, tex
->tex
.size_in_bytes
);
525 r300_tex_print_info(tex
, "texture_desc_init");
526 /* Ooops, what now. Apps will break if we fail this,
527 * so just pretend everything's okay. */
531 r300_setup_hyperz_properties(rscreen
, tex
);
533 if (SCREEN_DBG_ON(rscreen
, DBG_TEX
))
534 r300_tex_print_info(tex
, "texture_desc_init");
537 unsigned r300_texture_get_offset(struct r300_resource
*tex
,
538 unsigned level
, unsigned layer
)
540 unsigned offset
= tex
->tex
.offset_in_bytes
[level
];
542 switch (tex
->b
.b
.target
) {
543 case PIPE_TEXTURE_3D
:
544 case PIPE_TEXTURE_CUBE
:
545 return offset
+ layer
* tex
->tex
.layer_size_in_bytes
[level
];