freedreno: reduce extra height alignment in a6xx layout
[mesa.git] / src / freedreno / fdl / fd6_layout.c
1 /*
2 * Copyright (C) 2018 Rob Clark <robclark@freedesktop.org>
3 * Copyright © 2018-2019 Google, Inc.
4 *
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
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
14 * Software.
15 *
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 NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 * Authors:
25 * Rob Clark <robclark@freedesktop.org>
26 */
27
28 #include <stdio.h>
29
30 #include "freedreno_layout.h"
31
32 /* indexed by cpp, including msaa 2x and 4x:
33 * TODO:
34 * cpp=1 UBWC needs testing at larger texture sizes
35 * missing UBWC blockwidth/blockheight for npot+64 cpp
36 * missing 96/128 CPP for 8x MSAA with 32_32_32/32_32_32_32
37 */
38 static const struct tile_alignment {
39 unsigned basealign;
40 unsigned pitchalign;
41 unsigned heightalign;
42 /* UBWC block width/height. Used in size alignment, and calculating a
43 * descriptor's FLAG_BUFFER_LOG2W/H for mipmapping.
44 */
45 uint8_t ubwc_blockwidth;
46 uint8_t ubwc_blockheight;
47 } tile_alignment[] = {
48 [1] = { 64, 128, 32, 16, 4 },
49 [2] = { 128, 128, 16, 16, 4 },
50 [3] = { 256, 64, 32 },
51 [4] = { 256, 64, 16, 16, 4 },
52 [6] = { 256, 64, 16 },
53 [8] = { 256, 64, 16, 8, 4, },
54 [12] = { 256, 64, 16 },
55 [16] = { 256, 64, 16, 4, 4, },
56 [24] = { 256, 64, 16 },
57 [32] = { 256, 64, 16, 4, 2 },
58 [48] = { 256, 64, 16 },
59 [64] = { 256, 64, 16 },
60
61 /* special cases for r8g8: */
62 [0] = { 256, 64, 32, 16, 8 },
63 };
64
65 #define RGB_TILE_WIDTH_ALIGNMENT 64
66 #define RGB_TILE_HEIGHT_ALIGNMENT 16
67 #define UBWC_PLANE_SIZE_ALIGNMENT 4096
68
69 static const struct tile_alignment *
70 fdl6_tile_alignment(struct fdl_layout *layout)
71 {
72 debug_assert(layout->cpp < ARRAY_SIZE(tile_alignment));
73
74 if ((layout->cpp == 2) && (util_format_get_nr_components(layout->format) == 2))
75 return &tile_alignment[0];
76 else
77 return &tile_alignment[layout->cpp];
78 }
79
80 static int
81 fdl6_pitchalign(struct fdl_layout *layout, int level)
82 {
83 uint32_t pitchalign = 64;
84 if (fdl_tile_mode(layout, level))
85 pitchalign = fdl6_tile_alignment(layout)->pitchalign;
86
87 return pitchalign;
88 }
89
90 /* NOTE: good way to test this is: (for example)
91 * piglit/bin/texelFetch fs sampler3D 100x100x8
92 */
93 void
94 fdl6_layout(struct fdl_layout *layout,
95 enum pipe_format format, uint32_t nr_samples,
96 uint32_t width0, uint32_t height0, uint32_t depth0,
97 uint32_t mip_levels, uint32_t array_size, bool is_3d)
98 {
99 assert(nr_samples > 0);
100 layout->width0 = width0;
101 layout->height0 = height0;
102 layout->depth0 = depth0;
103
104 layout->cpp = util_format_get_blocksize(format);
105 layout->cpp *= nr_samples;
106 layout->cpp_shift = ffs(layout->cpp) - 1;
107
108 layout->format = format;
109 layout->nr_samples = nr_samples;
110 layout->layer_first = !is_3d;
111
112 if (depth0 > 1)
113 layout->ubwc = false;
114 if (tile_alignment[layout->cpp].ubwc_blockwidth == 0)
115 layout->ubwc = false;
116
117 const struct tile_alignment *ta = fdl6_tile_alignment(layout);
118
119 /* in layer_first layout, the level (slice) contains just one
120 * layer (since in fact the layer contains the slices)
121 */
122 uint32_t layers_in_level = layout->layer_first ? 1 : array_size;
123
124 debug_assert(ta->pitchalign);
125
126 if (layout->tile_mode) {
127 layout->base_align = ta->basealign;
128 } else {
129 layout->base_align = 64;
130 }
131
132 uint32_t pitch0 = util_align_npot(width0, fdl6_pitchalign(layout, 0));
133
134 uint32_t ubwc_width0 = width0;
135 uint32_t ubwc_height0 = height0;
136 uint32_t ubwc_tile_height_alignment = RGB_TILE_HEIGHT_ALIGNMENT;
137 if (mip_levels > 1) {
138 /* With mipmapping enabled, UBWC layout is power-of-two sized,
139 * specified in log2 width/height in the descriptors. The height
140 * alignment is 64 for mipmapping, but for buffer sharing (always
141 * single level) other participants expect 16.
142 */
143 ubwc_width0 = util_next_power_of_two(width0);
144 ubwc_height0 = util_next_power_of_two(height0);
145 ubwc_tile_height_alignment = 64;
146 }
147 ubwc_width0 = align(DIV_ROUND_UP(ubwc_width0, ta->ubwc_blockwidth),
148 RGB_TILE_WIDTH_ALIGNMENT);
149 ubwc_height0 = align(DIV_ROUND_UP(ubwc_height0,
150 ta->ubwc_blockheight),
151 ubwc_tile_height_alignment);
152
153 for (uint32_t level = 0; level < mip_levels; level++) {
154 uint32_t depth = u_minify(depth0, level);
155 struct fdl_slice *slice = &layout->slices[level];
156 struct fdl_slice *ubwc_slice = &layout->ubwc_slices[level];
157 uint32_t tile_mode = fdl_tile_mode(layout, level);
158 uint32_t height;
159
160 /* tiled levels of 3D textures are rounded up to PoT dimensions: */
161 if (is_3d && tile_mode) {
162 height = u_minify(util_next_power_of_two(height0), level);
163 } else {
164 height = u_minify(height0, level);
165 }
166
167 uint32_t nblocksy = util_format_get_nblocksy(format, height);
168 if (tile_mode)
169 nblocksy = align(nblocksy, ta->heightalign);
170
171 /* The blits used for mem<->gmem work at a granularity of
172 * 16x4, which can cause faults due to over-fetch on the
173 * last level. The simple solution is to over-allocate a
174 * bit the last level to ensure any over-fetch is harmless.
175 * The pitch is already sufficiently aligned, but height
176 * may not be. note this only matters if last level is linear
177 */
178 if (level == mip_levels - 1)
179 height = align(nblocksy, 4);
180
181 uint32_t nblocksx =
182 util_align_npot(util_format_get_nblocksx(format, u_minify(pitch0, level)),
183 fdl6_pitchalign(layout, level));
184
185 slice->offset = layout->size;
186 uint32_t blocks = nblocksx * nblocksy;
187
188 slice->pitch = nblocksx * layout->cpp;
189
190 /* 1d array and 2d array textures must all have the same layer size
191 * for each miplevel on a6xx. 3d textures can have different layer
192 * sizes for high levels, but the hw auto-sizer is buggy (or at least
193 * different than what this code does), so as soon as the layer size
194 * range gets into range, we stop reducing it.
195 */
196 if (is_3d) {
197 if (level < 1 || layout->slices[level - 1].size0 > 0xf000) {
198 slice->size0 = align(blocks * layout->cpp, 4096);
199 } else {
200 slice->size0 = layout->slices[level - 1].size0;
201 }
202 } else {
203 slice->size0 = blocks * layout->cpp;
204 }
205
206 layout->size += slice->size0 * depth * layers_in_level;
207
208 if (layout->ubwc) {
209 /* with UBWC every level is aligned to 4K */
210 layout->size = align(layout->size, 4096);
211
212 uint32_t meta_pitch = align(u_minify(ubwc_width0, level),
213 RGB_TILE_WIDTH_ALIGNMENT);
214 uint32_t meta_height = align(u_minify(ubwc_height0, level),
215 ubwc_tile_height_alignment);
216
217 ubwc_slice->size0 = align(meta_pitch * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
218 ubwc_slice->pitch = meta_pitch;
219 ubwc_slice->offset = layout->ubwc_layer_size;
220 layout->ubwc_layer_size += ubwc_slice->size0;
221 }
222 }
223
224 if (layout->layer_first) {
225 layout->layer_size = align(layout->size, 4096);
226 layout->size = layout->layer_size * array_size;
227 }
228
229 /* Place the UBWC slices before the uncompressed slices, because the
230 * kernel expects UBWC to be at the start of the buffer. In the HW, we
231 * get to program the UBWC and non-UBWC offset/strides
232 * independently.
233 */
234 if (layout->ubwc) {
235 for (uint32_t level = 0; level < mip_levels; level++)
236 layout->slices[level].offset += layout->ubwc_layer_size * array_size;
237 layout->size += layout->ubwc_layer_size * array_size;
238 }
239 }
240
241 void
242 fdl6_get_ubwc_blockwidth(struct fdl_layout *layout,
243 uint32_t *blockwidth, uint32_t *blockheight)
244 {
245 const struct tile_alignment *ta = fdl6_tile_alignment(layout);
246 *blockwidth = ta->ubwc_blockwidth;
247 *blockheight = ta->ubwc_blockheight;
248 }