xref: /third_party/mesa3d/src/panfrost/lib/pan_afbc.c (revision bf215546) 
1/*
2 * Copyright (C) 2019 Collabora, Ltd.
3 *
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:
10 *
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
13 * Software.
14 *
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
25 */
26
27#include "pan_texture.h"
28
29/* Arm FrameBuffer Compression (AFBC) is a lossless compression scheme natively
30 * implemented in Mali GPUs (as well as many display controllers paired with
31 * Mali GPUs, etc). Where possible, Panfrost prefers to use AFBC for both
32 * rendering and texturing. In most cases, this is a performance-win due to a
33 * dramatic reduction in memory bandwidth and cache locality compared to a
34 * linear resources.
35 *
36 * AFBC divides the framebuffer into 16x16 tiles (other sizes possible, TODO:
37 * do we need to support this?). So, the width and height each must be aligned
38 * up to 16 pixels. This is inherently good for performance; note that for a 4
39 * byte-per-pixel format like RGBA8888, that means that rows are 16*4=64 byte
40 * aligned, which is the cache-line size.
41 *
42 * For each AFBC-compressed resource, there is a single contiguous
43 * (CPU/GPU-shared) buffer. This buffer itself is divided into two parts:
44 * header and body, placed immediately after each other.
45 *
46 * The AFBC header contains 16 bytes of metadata per tile.
47 *
48 * The AFBC body is the same size as the original linear resource (padded to
49 * the nearest tile). Although the body comes immediately after the header, it
50 * must also be cache-line aligned, so there can sometimes be a bit of padding
51 * between the header and body.
52 *
53 * As an example, a 64x64 RGBA framebuffer contains 64/16 = 4 tiles horizontally and
54 * 4 tiles vertically. There are 4*4=16 tiles in total, each containing 16
55 * bytes of metadata, so there is a 16*16=256 byte header. 64x64 is already
56 * tile aligned, so the body is 64*64 * 4 bytes per pixel = 16384 bytes of
57 * body.
58 *
59 * From userspace, Panfrost needs to be able to calculate these sizes. It
60 * explicitly does not and can not know the format of the data contained within
61 * this header and body. The GPU has native support for AFBC encode/decode. For
62 * an internal FBO or a framebuffer used for scanout with an AFBC-compatible
63 * winsys/display-controller, the buffer is maintained AFBC throughout flight,
64 * and the driver never needs to know the internal data. For edge cases where
65 * the driver really does need to read/write from the AFBC resource, we
66 * generate a linear staging buffer and use the GPU to blit AFBC<--->linear.
67 */
68
69/* AFBC supports compressing a few canonical formats. Additional formats are
70 * available by using a canonical internal format. Given a PIPE format, find
71 * the canonical AFBC internal format if it exists, or NONE if the format
72 * cannot be compressed. */
73
74enum pipe_format
75panfrost_afbc_format(unsigned arch, enum pipe_format format)
76{
77        /* Don't allow swizzled formats on v7 */
78        switch (format) {
79        case PIPE_FORMAT_B8G8R8A8_UNORM:
80        case PIPE_FORMAT_B8G8R8X8_UNORM:
81        case PIPE_FORMAT_A8R8G8B8_UNORM:
82        case PIPE_FORMAT_X8R8G8B8_UNORM:
83        case PIPE_FORMAT_X8B8G8R8_UNORM:
84        case PIPE_FORMAT_A8B8G8R8_UNORM:
85        case PIPE_FORMAT_B8G8R8_UNORM:
86        case PIPE_FORMAT_B5G6R5_UNORM:
87                if (arch >= 7)
88                        return PIPE_FORMAT_NONE;
89
90                break;
91        default:
92                break;
93        }
94
95        switch (format) {
96        case PIPE_FORMAT_Z16_UNORM:
97                return PIPE_FORMAT_R8G8_UNORM;
98
99        case PIPE_FORMAT_R8G8B8_UNORM:
100        case PIPE_FORMAT_B8G8R8_UNORM:
101                return PIPE_FORMAT_R8G8B8_UNORM;
102
103        case PIPE_FORMAT_R8G8B8A8_UNORM:
104        case PIPE_FORMAT_R8G8B8X8_UNORM:
105        case PIPE_FORMAT_Z24_UNORM_S8_UINT:
106        case PIPE_FORMAT_Z24X8_UNORM:
107        case PIPE_FORMAT_X24S8_UINT:
108        case PIPE_FORMAT_B8G8R8A8_UNORM:
109        case PIPE_FORMAT_B8G8R8X8_UNORM:
110        case PIPE_FORMAT_A8R8G8B8_UNORM:
111        case PIPE_FORMAT_X8R8G8B8_UNORM:
112        case PIPE_FORMAT_X8B8G8R8_UNORM:
113        case PIPE_FORMAT_A8B8G8R8_UNORM:
114                return PIPE_FORMAT_R8G8B8A8_UNORM;
115
116        case PIPE_FORMAT_R5G6B5_UNORM:
117        case PIPE_FORMAT_B5G6R5_UNORM:
118                return PIPE_FORMAT_R5G6B5_UNORM;
119
120        /* TODO: More AFBC formats */
121        default:
122                return PIPE_FORMAT_NONE;
123        }
124}
125
126/* A format may be compressed as AFBC if it has an AFBC internal format */
127
128bool
129panfrost_format_supports_afbc(const struct panfrost_device *dev, enum pipe_format format)
130{
131        return panfrost_afbc_format(dev->arch, format) != PIPE_FORMAT_NONE;
132}
133
134/* The lossless colour transform (AFBC_FORMAT_MOD_YTR) requires RGB. */
135
136bool
137panfrost_afbc_can_ytr(enum pipe_format format)
138{
139        const struct util_format_description *desc =
140                util_format_description(format);
141
142        /* YTR is only defined for RGB(A) */
143        if (desc->nr_channels != 3 && desc->nr_channels != 4)
144                return false;
145
146        /* The fourth channel if it exists doesn't matter */
147        return desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB;
148}
149
150/*
151 * Check if the device supports AFBC with tiled headers (and hence also solid
152 * colour blocks).
153 */
154bool
155panfrost_afbc_can_tile(const struct panfrost_device *dev)
156{
157        return (dev->arch >= 7);
158}
159