diff options
Diffstat (limited to 'src/video_core/utils.h')
| -rw-r--r-- | src/video_core/utils.h | 164 |
1 files changed, 0 insertions, 164 deletions
diff --git a/src/video_core/utils.h b/src/video_core/utils.h deleted file mode 100644 index e0a14d48f..000000000 --- a/src/video_core/utils.h +++ /dev/null | |||
| @@ -1,164 +0,0 @@ | |||
| 1 | // Copyright 2014 Citra Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #pragma once | ||
| 6 | |||
| 7 | #include "common/common_types.h" | ||
| 8 | |||
| 9 | namespace VideoCore { | ||
| 10 | |||
| 11 | // 8x8 Z-Order coordinate from 2D coordinates | ||
| 12 | static inline u32 MortonInterleave(u32 x, u32 y) { | ||
| 13 | static const u32 xlut[] = {0x00, 0x01, 0x04, 0x05, 0x10, 0x11, 0x14, 0x15}; | ||
| 14 | static const u32 ylut[] = {0x00, 0x02, 0x08, 0x0a, 0x20, 0x22, 0x28, 0x2a}; | ||
| 15 | return xlut[x % 8] + ylut[y % 8]; | ||
| 16 | } | ||
| 17 | |||
| 18 | /** | ||
| 19 | * Calculates the offset of the position of the pixel in Morton order | ||
| 20 | */ | ||
| 21 | static inline u32 GetMortonOffset(u32 x, u32 y, u32 bytes_per_pixel) { | ||
| 22 | // Images are split into 8x8 tiles. Each tile is composed of four 4x4 subtiles each | ||
| 23 | // of which is composed of four 2x2 subtiles each of which is composed of four texels. | ||
| 24 | // Each structure is embedded into the next-bigger one in a diagonal pattern, e.g. | ||
| 25 | // texels are laid out in a 2x2 subtile like this: | ||
| 26 | // 2 3 | ||
| 27 | // 0 1 | ||
| 28 | // | ||
| 29 | // The full 8x8 tile has the texels arranged like this: | ||
| 30 | // | ||
| 31 | // 42 43 46 47 58 59 62 63 | ||
| 32 | // 40 41 44 45 56 57 60 61 | ||
| 33 | // 34 35 38 39 50 51 54 55 | ||
| 34 | // 32 33 36 37 48 49 52 53 | ||
| 35 | // 10 11 14 15 26 27 30 31 | ||
| 36 | // 08 09 12 13 24 25 28 29 | ||
| 37 | // 02 03 06 07 18 19 22 23 | ||
| 38 | // 00 01 04 05 16 17 20 21 | ||
| 39 | // | ||
| 40 | // This pattern is what's called Z-order curve, or Morton order. | ||
| 41 | |||
| 42 | const unsigned int block_height = 8; | ||
| 43 | const unsigned int coarse_x = x & ~7; | ||
| 44 | |||
| 45 | u32 i = VideoCore::MortonInterleave(x, y); | ||
| 46 | |||
| 47 | const unsigned int offset = coarse_x * block_height; | ||
| 48 | |||
| 49 | return (i + offset) * bytes_per_pixel; | ||
| 50 | } | ||
| 51 | |||
| 52 | static inline u32 MortonInterleave128(u32 x, u32 y) { | ||
| 53 | // 128x128 Z-Order coordinate from 2D coordinates | ||
| 54 | static constexpr u32 xlut[] = { | ||
| 55 | 0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042, | ||
| 56 | 0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809, | ||
| 57 | 0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000, | ||
| 58 | 0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043, | ||
| 59 | 0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a, | ||
| 60 | 0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001, | ||
| 61 | 0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048, | ||
| 62 | 0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b, | ||
| 63 | 0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002, | ||
| 64 | 0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049, | ||
| 65 | 0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840, | ||
| 66 | 0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003, | ||
| 67 | 0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a, | ||
| 68 | 0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841, | ||
| 69 | 0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008, | ||
| 70 | 0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b, | ||
| 71 | 0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842, | ||
| 72 | 0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009, | ||
| 73 | 0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800, | ||
| 74 | 0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843, | ||
| 75 | 0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a, | ||
| 76 | 0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801, | ||
| 77 | 0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848, | ||
| 78 | 0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b, | ||
| 79 | 0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802, | ||
| 80 | 0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849, | ||
| 81 | 0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040, | ||
| 82 | 0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803, | ||
| 83 | 0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a, | ||
| 84 | 0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041, | ||
| 85 | 0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808, | ||
| 86 | 0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b, | ||
| 87 | 0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042, | ||
| 88 | 0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809, | ||
| 89 | 0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b, | ||
| 90 | }; | ||
| 91 | static constexpr u32 ylut[] = { | ||
| 92 | 0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090, | ||
| 93 | 0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124, | ||
| 94 | 0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200, | ||
| 95 | 0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294, | ||
| 96 | 0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330, | ||
| 97 | 0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404, | ||
| 98 | 0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0, | ||
| 99 | 0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534, | ||
| 100 | 0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610, | ||
| 101 | 0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4, | ||
| 102 | 0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780, | ||
| 103 | 0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014, | ||
| 104 | 0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0, | ||
| 105 | 0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184, | ||
| 106 | 0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220, | ||
| 107 | 0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4, | ||
| 108 | 0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390, | ||
| 109 | 0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424, | ||
| 110 | 0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500, | ||
| 111 | 0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594, | ||
| 112 | 0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630, | ||
| 113 | 0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704, | ||
| 114 | 0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0, | ||
| 115 | 0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034, | ||
| 116 | 0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110, | ||
| 117 | 0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4, | ||
| 118 | 0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280, | ||
| 119 | 0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314, | ||
| 120 | 0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0, | ||
| 121 | 0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484, | ||
| 122 | 0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520, | ||
| 123 | 0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4, | ||
| 124 | 0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690, | ||
| 125 | 0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724, | ||
| 126 | 0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4, | ||
| 127 | }; | ||
| 128 | return xlut[x % 128] + ylut[y % 128]; | ||
| 129 | } | ||
| 130 | |||
| 131 | static inline u32 GetMortonOffset128(u32 x, u32 y, u32 bytes_per_pixel) { | ||
| 132 | // Calculates the offset of the position of the pixel in Morton order | ||
| 133 | // Framebuffer images are split into 128x128 tiles. | ||
| 134 | |||
| 135 | const unsigned int block_height = 128; | ||
| 136 | const unsigned int coarse_x = x & ~127; | ||
| 137 | |||
| 138 | u32 i = MortonInterleave128(x, y); | ||
| 139 | |||
| 140 | const unsigned int offset = coarse_x * block_height; | ||
| 141 | |||
| 142 | return (i + offset) * bytes_per_pixel; | ||
| 143 | } | ||
| 144 | |||
| 145 | static inline void MortonCopyPixels128(u32 width, u32 height, u32 bytes_per_pixel, | ||
| 146 | u32 gl_bytes_per_pixel, u8* morton_data, u8* gl_data, | ||
| 147 | bool morton_to_gl) { | ||
| 148 | u8* data_ptrs[2]; | ||
| 149 | for (unsigned y = 0; y < height; ++y) { | ||
| 150 | for (unsigned x = 0; x < width; ++x) { | ||
| 151 | const u32 coarse_y = y & ~127; | ||
| 152 | u32 morton_offset = | ||
| 153 | GetMortonOffset128(x, y, bytes_per_pixel) + coarse_y * width * bytes_per_pixel; | ||
| 154 | u32 gl_pixel_index = (x + y * width) * gl_bytes_per_pixel; | ||
| 155 | |||
| 156 | data_ptrs[morton_to_gl] = morton_data + morton_offset; | ||
| 157 | data_ptrs[!morton_to_gl] = &gl_data[gl_pixel_index]; | ||
| 158 | |||
| 159 | memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel); | ||
| 160 | } | ||
| 161 | } | ||
| 162 | } | ||
| 163 | |||
| 164 | } // namespace VideoCore | ||