freedreno: Move the layout debug under FD_MESA_DEBUG=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 {
39 unsigned pitchalign;
40 unsigned heightalign;
41 uint8_t ubwc_blockwidth;
42 uint8_t ubwc_blockheight;
43 } tile_alignment[] = {
44 [1] = { 128, 32, 16, 4 },
45 [2] = { 128, 16, 16, 4 },
46 [3] = { 64, 32 },
47 [4] = { 64, 16, 16, 4 },
48 [6] = { 64, 16 },
49 [8] = { 64, 16, 8, 4, },
50 [12] = { 64, 16 },
51 [16] = { 64, 16, 4, 4, },
52 [24] = { 64, 16 },
53 [32] = { 64, 16, 4, 2 },
54 [48] = { 64, 16 },
55 [64] = { 64, 16 },
56
57 /* special cases for r8g8: */
58 [0] = { 64, 32, 16, 4 },
59 };
60
61 #define RGB_TILE_WIDTH_ALIGNMENT 64
62 #define RGB_TILE_HEIGHT_ALIGNMENT 16
63 #define UBWC_PLANE_SIZE_ALIGNMENT 4096
64
65 /* NOTE: good way to test this is: (for example)
66 * piglit/bin/texelFetch fs sampler3D 100x100x8
67 */
68 void
69 fdl6_layout(struct fdl_layout *layout,
70 enum pipe_format format, uint32_t nr_samples,
71 uint32_t width0, uint32_t height0, uint32_t depth0,
72 uint32_t mip_levels, uint32_t array_size, bool is_3d, bool ubwc)
73 {
74 assert(nr_samples > 0);
75 layout->width0 = width0;
76 layout->height0 = height0;
77 layout->depth0 = depth0;
78
79 layout->cpp = util_format_get_blocksize(format);
80 layout->cpp *= nr_samples;
81 layout->format = format;
82 layout->nr_samples = nr_samples;
83
84 const struct util_format_description *format_desc =
85 util_format_description(format);
86 uint32_t depth = depth0;
87 /* linear dimensions: */
88 uint32_t lwidth = width0;
89 uint32_t lheight = height0;
90 /* tile_mode dimensions: */
91 uint32_t twidth = util_next_power_of_two(lwidth);
92 uint32_t theight = util_next_power_of_two(lheight);
93 int ta = layout->cpp;
94
95 /* The z16/r16 formats seem to not play by the normal tiling rules: */
96 if ((layout->cpp == 2) && (util_format_get_nr_components(format) == 2))
97 ta = 0;
98
99 uint32_t alignment;
100 if (is_3d) {
101 layout->layer_first = false;
102 alignment = 4096;
103 } else {
104 layout->layer_first = true;
105 alignment = 1;
106 }
107 /* in layer_first layout, the level (slice) contains just one
108 * layer (since in fact the layer contains the slices)
109 */
110 uint32_t layers_in_level = layout->layer_first ? 1 : array_size;
111
112 debug_assert(ta < ARRAY_SIZE(tile_alignment));
113 debug_assert(tile_alignment[ta].pitchalign);
114
115 for (uint32_t level = 0; level < mip_levels; level++) {
116 struct fdl_slice *slice = &layout->slices[level];
117 struct fdl_slice *ubwc_slice = &layout->ubwc_slices[level];
118 uint32_t tile_mode = (ubwc ?
119 layout->tile_mode : fdl_tile_mode(layout, level));
120 uint32_t width, height;
121
122 /* tiled levels of 3D textures are rounded up to PoT dimensions: */
123 if (is_3d && tile_mode) {
124 width = twidth;
125 height = theight;
126 } else {
127 width = lwidth;
128 height = lheight;
129 }
130 uint32_t aligned_height = height;
131 uint32_t pitchalign;
132
133 if (tile_mode) {
134 pitchalign = tile_alignment[ta].pitchalign;
135 aligned_height = align(aligned_height,
136 tile_alignment[ta].heightalign);
137 } else {
138 pitchalign = 64;
139 }
140
141 /* The blits used for mem<->gmem work at a granularity of
142 * 32x32, which can cause faults due to over-fetch on the
143 * last level. The simple solution is to over-allocate a
144 * bit the last level to ensure any over-fetch is harmless.
145 * The pitch is already sufficiently aligned, but height
146 * may not be:
147 */
148 if (level == mip_levels - 1)
149 aligned_height = align(aligned_height, 32);
150
151 if (format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC)
152 slice->pitch =
153 util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
154 else
155 slice->pitch = align(width, pitchalign);
156
157 slice->offset = layout->size;
158 uint32_t blocks = util_format_get_nblocks(format,
159 slice->pitch, aligned_height);
160
161 /* 1d array and 2d array textures must all have the same layer size
162 * for each miplevel on a6xx. 3d textures can have different layer
163 * sizes for high levels, but the hw auto-sizer is buggy (or at least
164 * different than what this code does), so as soon as the layer size
165 * range gets into range, we stop reducing it.
166 */
167 if (is_3d) {
168 if (level < 1 || layout->slices[level - 1].size0 > 0xf000) {
169 slice->size0 = align(blocks * layout->cpp, alignment);
170 } else {
171 slice->size0 = layout->slices[level - 1].size0;
172 }
173 } else {
174 slice->size0 = align(blocks * layout->cpp, alignment);
175 }
176
177 layout->size += slice->size0 * depth * layers_in_level;
178
179 if (ubwc) {
180 /* with UBWC every level is aligned to 4K */
181 layout->size = align(layout->size, 4096);
182
183 uint32_t block_width = tile_alignment[ta].ubwc_blockwidth;
184 uint32_t block_height = tile_alignment[ta].ubwc_blockheight;
185 uint32_t meta_pitch = align(DIV_ROUND_UP(width, block_width),
186 RGB_TILE_WIDTH_ALIGNMENT);
187 uint32_t meta_height = align(DIV_ROUND_UP(height, block_height),
188 RGB_TILE_HEIGHT_ALIGNMENT);
189
190 /* it looks like mipmaps need alignment to power of two
191 * TODO: needs testing with large npot textures
192 * (needed for the first level?)
193 */
194 if (mip_levels > 1) {
195 meta_pitch = util_next_power_of_two(meta_pitch);
196 meta_height = util_next_power_of_two(meta_height);
197 }
198
199 ubwc_slice->size0 = align(meta_pitch * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
200 ubwc_slice->pitch = meta_pitch;
201 ubwc_slice->offset = layout->ubwc_size;
202 layout->ubwc_size += ubwc_slice->size0;
203 }
204
205 depth = u_minify(depth, 1);
206 lwidth = u_minify(lwidth, 1);
207 lheight = u_minify(lheight, 1);
208 twidth = u_minify(twidth, 1);
209 theight = u_minify(theight, 1);
210 }
211
212 if (layout->layer_first) {
213 layout->layer_size = align(layout->size, 4096);
214 layout->size = layout->layer_size * array_size;
215 }
216
217 /* Place the UBWC slices before the uncompressed slices, because the
218 * kernel expects UBWC to be at the start of the buffer. In the HW, we
219 * get to program the UBWC and non-UBWC offset/strides
220 * independently.
221 */
222 if (ubwc) {
223 for (uint32_t level = 0; level < mip_levels; level++)
224 layout->slices[level].offset += layout->ubwc_size * array_size;
225 layout->size += layout->ubwc_size * array_size;
226 }
227 }
228
229 void
230 fdl6_get_ubwc_blockwidth(struct fdl_layout *layout,
231 uint32_t *blockwidth, uint32_t *blockheight)
232 {
233 *blockwidth = tile_alignment[layout->cpp].ubwc_blockwidth;
234 *blockheight = tile_alignment[layout->cpp].ubwc_blockheight;
235 }