DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 1 | /* |
| 2 | * AltiVec optimizations for libjpeg-turbo |
| 3 | * |
| 4 | * Copyright (C) 2014, D. R. Commander. |
| 5 | * Copyright (C) 2014, Jay Foad. |
| 6 | * All rights reserved. |
| 7 | * This software is provided 'as-is', without any express or implied |
| 8 | * warranty. In no event will the authors be held liable for any damages |
| 9 | * arising from the use of this software. |
| 10 | * |
| 11 | * Permission is granted to anyone to use this software for any purpose, |
| 12 | * including commercial applications, and to alter it and redistribute it |
| 13 | * freely, subject to the following restrictions: |
| 14 | * |
| 15 | * 1. The origin of this software must not be misrepresented; you must not |
| 16 | * claim that you wrote the original software. If you use this software |
| 17 | * in a product, an acknowledgment in the product documentation would be |
| 18 | * appreciated but is not required. |
| 19 | * 2. Altered source versions must be plainly marked as such, and must not be |
| 20 | * misrepresented as being the original software. |
| 21 | * 3. This notice may not be removed or altered from any source distribution. |
| 22 | */ |
| 23 | |
| 24 | /* This file is included by jccolor-altivec.c */ |
| 25 | |
| 26 | |
| 27 | void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf, |
| 28 | JSAMPIMAGE output_buf, |
| 29 | JDIMENSION output_row, int num_rows) |
| 30 | { |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 31 | JSAMPROW inptr, outptr0, outptr1, outptr2; |
| 32 | int pitch = img_width * RGB_PIXELSIZE, offset, num_cols; |
| 33 | unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16]; |
| 34 | |
| 35 | __vector unsigned char rgb0, rgb1 = {0}, rgb2 = {0}, rgb3 = {0}, |
| 36 | rgbg0, rgbg1, rgbg2, rgbg3, y, cb, cr; |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 37 | #if RGB_PIXELSIZE == 4 |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 38 | __vector unsigned char rgb4 = {0}; |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 39 | #endif |
| 40 | __vector short rg0, rg1, rg2, rg3, bg0, bg1, bg2, bg3; |
| 41 | __vector unsigned short y01, y23, cr01, cr23, cb01, cb23; |
| 42 | __vector int y0, y1, y2, y3, cr0, cr1, cr2, cr3, cb0, cb1, cb2, cb3; |
| 43 | |
| 44 | /* Constants */ |
| 45 | __vector short pw_f0299_f0337 = { __4X2(F_0_299, F_0_337) }, |
| 46 | pw_f0114_f0250 = { __4X2(F_0_114, F_0_250) }, |
| 47 | pw_mf016_mf033 = { __4X2(-F_0_168, -F_0_331) }, |
| 48 | pw_mf008_mf041 = { __4X2(-F_0_081, -F_0_418) }; |
| 49 | __vector unsigned short pw_f050_f000 = { __4X2(F_0_500, 0) }; |
| 50 | __vector int pd_onehalf = { __4X(ONE_HALF) }, |
| 51 | pd_onehalfm1_cj = { __4X(ONE_HALF - 1 + (CENTERJSAMPLE << SCALEBITS)) }; |
| 52 | __vector unsigned char zero = { __16X(0) }, |
| 53 | shift_pack_index = |
| 54 | { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29}; |
| 55 | |
| 56 | while (--num_rows >= 0) { |
| 57 | inptr = *input_buf++; |
| 58 | outptr0 = output_buf[0][output_row]; |
| 59 | outptr1 = output_buf[1][output_row]; |
| 60 | outptr2 = output_buf[2][output_row]; |
| 61 | output_row++; |
| 62 | |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 63 | for (num_cols = pitch; num_cols > 0; |
| 64 | num_cols -= RGB_PIXELSIZE * 16, inptr += RGB_PIXELSIZE * 16, |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 65 | outptr0 += 16, outptr1 += 16, outptr2 += 16) { |
| 66 | |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 67 | /* Load 16 pixels == 48 or 64 bytes */ |
| 68 | offset = (size_t)inptr & 15; |
| 69 | if (offset) { |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 70 | __vector unsigned char unaligned_shift_index; |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 71 | int bytes = num_cols + offset; |
| 72 | |
| 73 | if (bytes >= (RGB_PIXELSIZE + 1) * 16) { |
| 74 | /* Fast path -- we have enough buffer space to load all vectors. |
| 75 | * Even if we don't need them all, this is faster than narrowing |
| 76 | * down which ones we need. |
| 77 | */ |
| 78 | rgb0 = vec_ld(0, inptr); |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 79 | rgb1 = vec_ld(16, inptr); |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 80 | rgb2 = vec_ld(32, inptr); |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 81 | rgb3 = vec_ld(48, inptr); |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 82 | #if RGB_PIXELSIZE == 4 |
| 83 | rgb4 = vec_ld(64, inptr); |
| 84 | #endif |
| 85 | } else { |
| 86 | if (bytes & 15) { |
| 87 | /* Slow path to prevent buffer overread. Since there is no way to |
| 88 | * read a partial AltiVec register, overread would occur on the |
| 89 | * last chunk of the last image row if the right edge is not on a |
| 90 | * 16-byte boundary. It could also occur on other rows if the |
| 91 | * bytes per row is low enough. Since we can't determine whether |
| 92 | * we're on the last image row, we have to assume every row is the |
| 93 | * last. |
| 94 | */ |
| 95 | memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); |
| 96 | rgb0 = vec_ld(0, tmpbuf); |
| 97 | rgb1 = vec_ld(16, tmpbuf); |
| 98 | rgb2 = vec_ld(32, tmpbuf); |
| 99 | #if RGB_PIXELSIZE == 4 |
| 100 | rgb3 = vec_ld(48, tmpbuf); |
| 101 | #endif |
| 102 | goto start; /* Skip permutation */ |
| 103 | } else { |
| 104 | /* Medium path -- if the right edge is vector-aligned, then we can |
| 105 | * read full vectors (but with a lot of branches.) |
| 106 | */ |
| 107 | rgb0 = vec_ld(0, inptr); |
| 108 | if (bytes > 16) { |
| 109 | rgb1 = vec_ld(16, inptr); |
| 110 | if (bytes > 32) { |
| 111 | rgb2 = vec_ld(32, inptr); |
| 112 | if (bytes > 48) { |
| 113 | rgb3 = vec_ld(48, inptr); |
| 114 | #if RGB_PIXELSIZE == 4 |
| 115 | if (bytes > 64) |
| 116 | rgb4 = vec_ld(64, inptr); |
| 117 | #endif |
| 118 | } |
| 119 | } |
| 120 | } |
| 121 | } |
| 122 | } |
| 123 | |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 124 | unaligned_shift_index = vec_lvsl(0, inptr); |
| 125 | rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index); |
| 126 | rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index); |
| 127 | rgb2 = vec_perm(rgb2, rgb3, unaligned_shift_index); |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 128 | #if RGB_PIXELSIZE == 4 |
| 129 | rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index); |
| 130 | #endif |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 131 | } else { |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 132 | if (num_cols >= RGB_PIXELSIZE * 16) { |
| 133 | /* Fast path */ |
| 134 | rgb0 = vec_ld(0, inptr); |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 135 | rgb1 = vec_ld(16, inptr); |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 136 | rgb2 = vec_ld(32, inptr); |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 137 | #if RGB_PIXELSIZE == 4 |
| 138 | rgb3 = vec_ld(48, inptr); |
| 139 | #endif |
| 140 | } else { |
| 141 | if (num_cols & 15) { |
| 142 | /* Slow path */ |
| 143 | memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); |
| 144 | rgb0 = vec_ld(0, tmpbuf); |
| 145 | rgb1 = vec_ld(16, tmpbuf); |
| 146 | rgb2 = vec_ld(32, tmpbuf); |
| 147 | #if RGB_PIXELSIZE == 4 |
| 148 | rgb3 = vec_ld(48, tmpbuf); |
| 149 | #endif |
| 150 | } else { |
| 151 | /* Medium path */ |
| 152 | rgb0 = vec_ld(0, inptr); |
| 153 | if (num_cols > 16) { |
| 154 | rgb1 = vec_ld(16, inptr); |
| 155 | if (num_cols > 32) { |
| 156 | rgb2 = vec_ld(32, inptr); |
| 157 | #if RGB_PIXELSIZE == 4 |
| 158 | if (num_cols > 48) |
| 159 | rgb3 = vec_ld(48, inptr); |
| 160 | #endif |
| 161 | } |
| 162 | } |
| 163 | } |
| 164 | } |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 165 | } |
| 166 | |
DRC | 2534788 | 2015-01-10 11:32:36 +0000 | [diff] [blame^] | 167 | start: |
| 168 | #if RGB_PIXELSIZE == 3 |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 169 | /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 |
| 170 | * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga |
| 171 | * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf |
| 172 | * |
| 173 | * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3 |
| 174 | * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7 |
| 175 | * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb |
| 176 | * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf |
| 177 | */ |
| 178 | rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX0); |
| 179 | rgbg1 = vec_perm(rgb0, rgb1, (__vector unsigned char)RGBG_INDEX1); |
| 180 | rgbg2 = vec_perm(rgb1, rgb2, (__vector unsigned char)RGBG_INDEX2); |
| 181 | rgbg3 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX3); |
| 182 | #else |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 183 | /* rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3 |
DRC | 8de75d0 | 2015-01-08 10:42:54 +0000 | [diff] [blame] | 184 | * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7 |
| 185 | * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb |
| 186 | * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf |
DRC | 62bae20 | 2014-12-22 13:42:26 +0000 | [diff] [blame] | 187 | * |
| 188 | * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3 |
| 189 | * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7 |
| 190 | * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb |
| 191 | * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf |
| 192 | */ |
| 193 | rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX); |
| 194 | rgbg1 = vec_perm(rgb1, rgb1, (__vector unsigned char)RGBG_INDEX); |
| 195 | rgbg2 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX); |
| 196 | rgbg3 = vec_perm(rgb3, rgb3, (__vector unsigned char)RGBG_INDEX); |
| 197 | #endif |
| 198 | |
| 199 | /* rg0 = R0 G0 R1 G1 R2 G2 R3 G3 |
| 200 | * bg0 = B0 G0 B1 G1 B2 G2 B3 G3 |
| 201 | * ... |
| 202 | * |
| 203 | * NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't |
| 204 | * support unsigned vectors. |
| 205 | */ |
| 206 | rg0 = (__vector signed short)vec_mergeh(zero, rgbg0); |
| 207 | bg0 = (__vector signed short)vec_mergel(zero, rgbg0); |
| 208 | rg1 = (__vector signed short)vec_mergeh(zero, rgbg1); |
| 209 | bg1 = (__vector signed short)vec_mergel(zero, rgbg1); |
| 210 | rg2 = (__vector signed short)vec_mergeh(zero, rgbg2); |
| 211 | bg2 = (__vector signed short)vec_mergel(zero, rgbg2); |
| 212 | rg3 = (__vector signed short)vec_mergeh(zero, rgbg3); |
| 213 | bg3 = (__vector signed short)vec_mergel(zero, rgbg3); |
| 214 | |
| 215 | /* (Original) |
| 216 | * Y = 0.29900 * R + 0.58700 * G + 0.11400 * B |
| 217 | * Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE |
| 218 | * Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE |
| 219 | * |
| 220 | * (This implementation) |
| 221 | * Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G |
| 222 | * Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE |
| 223 | * Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE |
| 224 | */ |
| 225 | |
| 226 | /* Calculate Y values */ |
| 227 | |
| 228 | y0 = vec_msums(rg0, pw_f0299_f0337, pd_onehalf); |
| 229 | y1 = vec_msums(rg1, pw_f0299_f0337, pd_onehalf); |
| 230 | y2 = vec_msums(rg2, pw_f0299_f0337, pd_onehalf); |
| 231 | y3 = vec_msums(rg3, pw_f0299_f0337, pd_onehalf); |
| 232 | y0 = vec_msums(bg0, pw_f0114_f0250, y0); |
| 233 | y1 = vec_msums(bg1, pw_f0114_f0250, y1); |
| 234 | y2 = vec_msums(bg2, pw_f0114_f0250, y2); |
| 235 | y3 = vec_msums(bg3, pw_f0114_f0250, y3); |
| 236 | /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from |
| 237 | * each dword into a new 16-bit vector, which is the equivalent of |
| 238 | * descaling the 32-bit results (right-shifting by 16 bits) and then |
| 239 | * packing them. |
| 240 | */ |
| 241 | y01 = vec_perm((__vector unsigned short)y0, (__vector unsigned short)y1, |
| 242 | shift_pack_index); |
| 243 | y23 = vec_perm((__vector unsigned short)y2, (__vector unsigned short)y3, |
| 244 | shift_pack_index); |
| 245 | y = vec_pack(y01, y23); |
| 246 | vec_st(y, 0, outptr0); |
| 247 | |
| 248 | /* Calculate Cb values */ |
| 249 | cb0 = vec_msums(rg0, pw_mf016_mf033, pd_onehalfm1_cj); |
| 250 | cb1 = vec_msums(rg1, pw_mf016_mf033, pd_onehalfm1_cj); |
| 251 | cb2 = vec_msums(rg2, pw_mf016_mf033, pd_onehalfm1_cj); |
| 252 | cb3 = vec_msums(rg3, pw_mf016_mf033, pd_onehalfm1_cj); |
| 253 | cb0 = (__vector int)vec_msum((__vector unsigned short)bg0, pw_f050_f000, |
| 254 | (__vector unsigned int)cb0); |
| 255 | cb1 = (__vector int)vec_msum((__vector unsigned short)bg1, pw_f050_f000, |
| 256 | (__vector unsigned int)cb1); |
| 257 | cb2 = (__vector int)vec_msum((__vector unsigned short)bg2, pw_f050_f000, |
| 258 | (__vector unsigned int)cb2); |
| 259 | cb3 = (__vector int)vec_msum((__vector unsigned short)bg3, pw_f050_f000, |
| 260 | (__vector unsigned int)cb3); |
| 261 | cb01 = vec_perm((__vector unsigned short)cb0, |
| 262 | (__vector unsigned short)cb1, shift_pack_index); |
| 263 | cb23 = vec_perm((__vector unsigned short)cb2, |
| 264 | (__vector unsigned short)cb3, shift_pack_index); |
| 265 | cb = vec_pack(cb01, cb23); |
| 266 | vec_st(cb, 0, outptr1); |
| 267 | |
| 268 | /* Calculate Cr values */ |
| 269 | cr0 = vec_msums(bg0, pw_mf008_mf041, pd_onehalfm1_cj); |
| 270 | cr1 = vec_msums(bg1, pw_mf008_mf041, pd_onehalfm1_cj); |
| 271 | cr2 = vec_msums(bg2, pw_mf008_mf041, pd_onehalfm1_cj); |
| 272 | cr3 = vec_msums(bg3, pw_mf008_mf041, pd_onehalfm1_cj); |
| 273 | cr0 = (__vector int)vec_msum((__vector unsigned short)rg0, pw_f050_f000, |
| 274 | (__vector unsigned int)cr0); |
| 275 | cr1 = (__vector int)vec_msum((__vector unsigned short)rg1, pw_f050_f000, |
| 276 | (__vector unsigned int)cr1); |
| 277 | cr2 = (__vector int)vec_msum((__vector unsigned short)rg2, pw_f050_f000, |
| 278 | (__vector unsigned int)cr2); |
| 279 | cr3 = (__vector int)vec_msum((__vector unsigned short)rg3, pw_f050_f000, |
| 280 | (__vector unsigned int)cr3); |
| 281 | cr01 = vec_perm((__vector unsigned short)cr0, |
| 282 | (__vector unsigned short)cr1, shift_pack_index); |
| 283 | cr23 = vec_perm((__vector unsigned short)cr2, |
| 284 | (__vector unsigned short)cr3, shift_pack_index); |
| 285 | cr = vec_pack(cr01, cr23); |
| 286 | vec_st(cr, 0, outptr2); |
| 287 | } |
| 288 | } |
| 289 | } |