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 <[email protected]>
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
54 * and 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 static enum pipe_format
unswizzled_format(enum pipe_format format)70 unswizzled_format(enum pipe_format format)
71 {
72 switch (format) {
73 case PIPE_FORMAT_A8_UNORM:
74 case PIPE_FORMAT_L8_UNORM:
75 case PIPE_FORMAT_I8_UNORM:
76 return PIPE_FORMAT_R8_UNORM;
77
78 case PIPE_FORMAT_L8A8_UNORM:
79 return PIPE_FORMAT_R8G8_UNORM;
80
81 case PIPE_FORMAT_B8G8R8_UNORM:
82 return PIPE_FORMAT_R8G8B8_UNORM;
83
84 case PIPE_FORMAT_R8G8B8X8_UNORM:
85 case PIPE_FORMAT_B8G8R8A8_UNORM:
86 case PIPE_FORMAT_B8G8R8X8_UNORM:
87 case PIPE_FORMAT_A8R8G8B8_UNORM:
88 case PIPE_FORMAT_X8R8G8B8_UNORM:
89 case PIPE_FORMAT_X8B8G8R8_UNORM:
90 case PIPE_FORMAT_A8B8G8R8_UNORM:
91 return PIPE_FORMAT_R8G8B8A8_UNORM;
92
93 case PIPE_FORMAT_B5G6R5_UNORM:
94 return PIPE_FORMAT_R5G6B5_UNORM;
95
96 case PIPE_FORMAT_B5G5R5A1_UNORM:
97 return PIPE_FORMAT_R5G5B5A1_UNORM;
98
99 case PIPE_FORMAT_R10G10B10X2_UNORM:
100 case PIPE_FORMAT_B10G10R10A2_UNORM:
101 case PIPE_FORMAT_B10G10R10X2_UNORM:
102 return PIPE_FORMAT_R10G10B10A2_UNORM;
103
104 case PIPE_FORMAT_A4B4G4R4_UNORM:
105 case PIPE_FORMAT_B4G4R4A4_UNORM:
106 return PIPE_FORMAT_R4G4B4A4_UNORM;
107
108 default:
109 return format;
110 }
111 }
112
113 /* AFBC supports compressing a few canonical formats. Additional formats are
114 * available by using a canonical internal format. Given a PIPE format, find
115 * the canonical AFBC internal format if it exists, or NONE if the format
116 * cannot be compressed. */
117
118 enum pan_afbc_mode
panfrost_afbc_format(unsigned arch,enum pipe_format format)119 panfrost_afbc_format(unsigned arch, enum pipe_format format)
120 {
121 /* sRGB does not change the pixel format itself, only the
122 * interpretation. The interpretation is handled by conversion hardware
123 * independent to the compression hardware, so we can compress sRGB
124 * formats by using the corresponding linear format.
125 */
126 format = util_format_linear(format);
127
128 /* Luminance-alpha not supported for AFBC on v7+ */
129 switch (format) {
130 case PIPE_FORMAT_A8_UNORM:
131 case PIPE_FORMAT_L8_UNORM:
132 case PIPE_FORMAT_I8_UNORM:
133 case PIPE_FORMAT_L8A8_UNORM:
134 if (arch >= 7)
135 return PAN_AFBC_MODE_INVALID;
136 else
137 break;
138 default:
139 break;
140 }
141
142 /* We handle swizzling orthogonally to AFBC */
143 format = unswizzled_format(format);
144
145 /* clang-format off */
146 switch (format) {
147 case PIPE_FORMAT_R8_UNORM: return PAN_AFBC_MODE_R8;
148 case PIPE_FORMAT_R8G8_UNORM: return PAN_AFBC_MODE_R8G8;
149 case PIPE_FORMAT_R8G8B8_UNORM: return PAN_AFBC_MODE_R8G8B8;
150 case PIPE_FORMAT_R8G8B8A8_UNORM: return PAN_AFBC_MODE_R8G8B8A8;
151 case PIPE_FORMAT_R5G6B5_UNORM: return PAN_AFBC_MODE_R5G6B5;
152 case PIPE_FORMAT_R5G5B5A1_UNORM: return PAN_AFBC_MODE_R5G5B5A1;
153 case PIPE_FORMAT_R10G10B10A2_UNORM: return PAN_AFBC_MODE_R10G10B10A2;
154 case PIPE_FORMAT_R4G4B4A4_UNORM: return PAN_AFBC_MODE_R4G4B4A4;
155 case PIPE_FORMAT_Z16_UNORM: return PAN_AFBC_MODE_R8G8;
156
157 case PIPE_FORMAT_Z24_UNORM_S8_UINT: return PAN_AFBC_MODE_R8G8B8A8;
158 case PIPE_FORMAT_Z24X8_UNORM: return PAN_AFBC_MODE_R8G8B8A8;
159 case PIPE_FORMAT_X24S8_UINT: return PAN_AFBC_MODE_R8G8B8A8;
160
161 default: return PAN_AFBC_MODE_INVALID;
162 }
163 /* clang-format on */
164 }
165
166 /* The lossless colour transform (AFBC_FORMAT_MOD_YTR) requires RGB. */
167
168 bool
panfrost_afbc_can_ytr(enum pipe_format format)169 panfrost_afbc_can_ytr(enum pipe_format format)
170 {
171 const struct util_format_description *desc = util_format_description(format);
172
173 /* YTR is only defined for RGB(A) */
174 if (desc->nr_channels != 3 && desc->nr_channels != 4)
175 return false;
176
177 /* The fourth channel if it exists doesn't matter */
178 return desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB;
179 }
180
181 /* Only support packing for RGB formats for now. */
182
183 bool
panfrost_afbc_can_pack(enum pipe_format format)184 panfrost_afbc_can_pack(enum pipe_format format)
185 {
186 const struct util_format_description *desc = util_format_description(format);
187
188 return desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB;
189 }
190