freedreno/fdl: Align after dividing by block size
[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 {
39 unsigned basealign;
40 unsigned pitchalign;
41 unsigned heightalign;
42 uint8_t ubwc_blockwidth;
43 uint8_t ubwc_blockheight;
44 } tile_alignment[] = {
45 [1] = { 64, 128, 32, 16, 4 },
46 [2] = { 128, 128, 16, 16, 4 },
47 [3] = { 256, 64, 32 },
48 [4] = { 256, 64, 16, 16, 4 },
49 [6] = { 256, 64, 16 },
50 [8] = { 256, 64, 16, 8, 4, },
51 [12] = { 256, 64, 16 },
52 [16] = { 256, 64, 16, 4, 4, },
53 [24] = { 256, 64, 16 },
54 [32] = { 256, 64, 16, 4, 2 },
55 [48] = { 256, 64, 16 },
56 [64] = { 256, 64, 16 },
57
58 /* special cases for r8g8: */
59 [0] = { 256, 64, 32, 16, 4 },
60 };
61
62 #define RGB_TILE_WIDTH_ALIGNMENT 64
63 #define RGB_TILE_HEIGHT_ALIGNMENT 16
64 #define UBWC_PLANE_SIZE_ALIGNMENT 4096
65
66 static int
67 fdl6_pitchalign(struct fdl_layout *layout, int ta, int level)
68 {
69 uint32_t pitchalign = 64;
70 if (fdl_tile_mode(layout, level))
71 pitchalign = tile_alignment[ta].pitchalign;
72
73 return pitchalign;
74 }
75
76 /* NOTE: good way to test this is: (for example)
77 * piglit/bin/texelFetch fs sampler3D 100x100x8
78 */
79 void
80 fdl6_layout(struct fdl_layout *layout,
81 enum pipe_format format, uint32_t nr_samples,
82 uint32_t width0, uint32_t height0, uint32_t depth0,
83 uint32_t mip_levels, uint32_t array_size, bool is_3d)
84 {
85 assert(nr_samples > 0);
86 layout->width0 = width0;
87 layout->height0 = height0;
88 layout->depth0 = depth0;
89
90 layout->cpp = util_format_get_blocksize(format);
91 layout->cpp *= nr_samples;
92 layout->cpp_shift = ffs(layout->cpp) - 1;
93
94 layout->format = format;
95 layout->nr_samples = nr_samples;
96 layout->layer_first = !is_3d;
97
98 if (depth0 > 1)
99 layout->ubwc = false;
100 if (tile_alignment[layout->cpp].ubwc_blockwidth == 0)
101 layout->ubwc = false;
102
103 int ta = layout->cpp;
104
105 /* The z16/r16 formats seem to not play by the normal tiling rules: */
106 if ((layout->cpp == 2) && (util_format_get_nr_components(format) == 2))
107 ta = 0;
108
109 /* in layer_first layout, the level (slice) contains just one
110 * layer (since in fact the layer contains the slices)
111 */
112 uint32_t layers_in_level = layout->layer_first ? 1 : array_size;
113
114 debug_assert(ta < ARRAY_SIZE(tile_alignment));
115 debug_assert(tile_alignment[ta].pitchalign);
116
117 if (layout->tile_mode) {
118 layout->base_align = tile_alignment[ta].basealign;
119 } else {
120 layout->base_align = 64;
121 }
122
123 uint32_t pitch0 = util_align_npot(width0, fdl6_pitchalign(layout, ta, 0));
124
125 for (uint32_t level = 0; level < mip_levels; level++) {
126 uint32_t depth = u_minify(depth0, level);
127 struct fdl_slice *slice = &layout->slices[level];
128 struct fdl_slice *ubwc_slice = &layout->ubwc_slices[level];
129 uint32_t tile_mode = fdl_tile_mode(layout, level);
130 uint32_t width, height;
131
132 /* tiled levels of 3D textures are rounded up to PoT dimensions: */
133 if (is_3d && tile_mode) {
134 width = u_minify(util_next_power_of_two(width0), level);
135 height = u_minify(util_next_power_of_two(height0), level);
136 } else {
137 width = u_minify(width0, level);
138 height = u_minify(height0, level);
139 }
140
141 uint32_t nblocksy = util_format_get_nblocksy(format, height);
142 if (tile_mode)
143 nblocksy = align(nblocksy, tile_alignment[ta].heightalign);
144
145 /* The blits used for mem<->gmem work at a granularity of
146 * 32x32, which can cause faults due to over-fetch on the
147 * last level. The simple solution is to over-allocate a
148 * bit the last level to ensure any over-fetch is harmless.
149 * The pitch is already sufficiently aligned, but height
150 * may not be:
151 */
152 if (level == mip_levels - 1)
153 nblocksy = align(nblocksy, 32);
154
155 uint32_t nblocksx =
156 util_align_npot(util_format_get_nblocksx(format, u_minify(pitch0, level)),
157 fdl6_pitchalign(layout, ta, level));
158
159 slice->offset = layout->size;
160 uint32_t blocks = nblocksx * nblocksy;
161
162 slice->pitch = nblocksx * layout->cpp;
163
164 /* 1d array and 2d array textures must all have the same layer size
165 * for each miplevel on a6xx. 3d textures can have different layer
166 * sizes for high levels, but the hw auto-sizer is buggy (or at least
167 * different than what this code does), so as soon as the layer size
168 * range gets into range, we stop reducing it.
169 */
170 if (is_3d) {
171 if (level < 1 || layout->slices[level - 1].size0 > 0xf000) {
172 slice->size0 = align(blocks * layout->cpp, 4096);
173 } else {
174 slice->size0 = layout->slices[level - 1].size0;
175 }
176 } else {
177 slice->size0 = blocks * layout->cpp;
178 }
179
180 layout->size += slice->size0 * depth * layers_in_level;
181
182 if (layout->ubwc) {
183 /* with UBWC every level is aligned to 4K */
184 layout->size = align(layout->size, 4096);
185
186 uint32_t block_width = tile_alignment[ta].ubwc_blockwidth;
187 uint32_t block_height = tile_alignment[ta].ubwc_blockheight;
188 uint32_t meta_pitch = align(DIV_ROUND_UP(width, block_width),
189 RGB_TILE_WIDTH_ALIGNMENT);
190 uint32_t meta_height = align(DIV_ROUND_UP(height, block_height),
191 RGB_TILE_HEIGHT_ALIGNMENT);
192
193 /* it looks like mipmaps need alignment to power of two
194 * TODO: needs testing with large npot textures
195 * (needed for the first level?)
196 */
197 if (mip_levels > 1) {
198 meta_pitch = util_next_power_of_two(meta_pitch);
199 meta_height = util_next_power_of_two(meta_height);
200 }
201
202 ubwc_slice->size0 = align(meta_pitch * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
203 ubwc_slice->pitch = meta_pitch;
204 ubwc_slice->offset = layout->ubwc_layer_size;
205 layout->ubwc_layer_size += ubwc_slice->size0;
206 }
207 }
208
209 if (layout->layer_first) {
210 layout->layer_size = align(layout->size, 4096);
211 layout->size = layout->layer_size * array_size;
212 }
213
214 /* Place the UBWC slices before the uncompressed slices, because the
215 * kernel expects UBWC to be at the start of the buffer. In the HW, we
216 * get to program the UBWC and non-UBWC offset/strides
217 * independently.
218 */
219 if (layout->ubwc) {
220 for (uint32_t level = 0; level < mip_levels; level++)
221 layout->slices[level].offset += layout->ubwc_layer_size * array_size;
222 layout->size += layout->ubwc_layer_size * array_size;
223 }
224 }
225
226 void
227 fdl6_get_ubwc_blockwidth(struct fdl_layout *layout,
228 uint32_t *blockwidth, uint32_t *blockheight)
229 {
230 *blockwidth = tile_alignment[layout->cpp].ubwc_blockwidth;
231 *blockheight = tile_alignment[layout->cpp].ubwc_blockheight;
232 }