jccolext-altivec.c (11551B)
1 /* 2 * AltiVec optimizations for libjpeg-turbo 3 * 4 * Copyright (C) 2014-2015, D. R. Commander. All Rights Reserved. 5 * Copyright (C) 2014, Jay Foad. All Rights Reserved. 6 * 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 { 31 JSAMPROW inptr, outptr0, outptr1, outptr2; 32 int pitch = img_width * RGB_PIXELSIZE, num_cols; 33 #if __BIG_ENDIAN__ 34 int offset; 35 #endif 36 unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16]; 37 38 __vector unsigned char rgb0, rgb1 = { 0 }, rgb2 = { 0 }, 39 rgbg0, rgbg1, rgbg2, rgbg3, y, cb, cr; 40 #if __BIG_ENDIAN__ || RGB_PIXELSIZE == 4 41 __vector unsigned char rgb3 = { 0 }; 42 #endif 43 #if __BIG_ENDIAN__ && RGB_PIXELSIZE == 4 44 __vector unsigned char rgb4 = { 0 }; 45 #endif 46 __vector short rg0, rg1, rg2, rg3, bg0, bg1, bg2, bg3; 47 __vector unsigned short yl, yh, crl, crh, cbl, cbh; 48 __vector int y0, y1, y2, y3, cr0, cr1, cr2, cr3, cb0, cb1, cb2, cb3; 49 50 /* Constants */ 51 __vector short pw_f0299_f0337 = { __4X2(F_0_299, F_0_337) }, 52 pw_f0114_f0250 = { __4X2(F_0_114, F_0_250) }, 53 pw_mf016_mf033 = { __4X2(-F_0_168, -F_0_331) }, 54 pw_mf008_mf041 = { __4X2(-F_0_081, -F_0_418) }; 55 __vector unsigned short pw_f050_f000 = { __4X2(F_0_500, 0) }; 56 __vector int pd_onehalf = { __4X(ONE_HALF) }, 57 pd_onehalfm1_cj = { __4X(ONE_HALF - 1 + (CENTERJSAMPLE << SCALEBITS)) }; 58 __vector unsigned char pb_zero = { __16X(0) }, 59 #if __BIG_ENDIAN__ 60 shift_pack_index = 61 { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 }; 62 #else 63 shift_pack_index = 64 { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 }; 65 #endif 66 67 while (--num_rows >= 0) { 68 inptr = *input_buf++; 69 outptr0 = output_buf[0][output_row]; 70 outptr1 = output_buf[1][output_row]; 71 outptr2 = output_buf[2][output_row]; 72 output_row++; 73 74 for (num_cols = pitch; num_cols > 0; 75 num_cols -= RGB_PIXELSIZE * 16, inptr += RGB_PIXELSIZE * 16, 76 outptr0 += 16, outptr1 += 16, outptr2 += 16) { 77 78 #if __BIG_ENDIAN__ 79 /* Load 16 pixels == 48 or 64 bytes */ 80 offset = (size_t)inptr & 15; 81 if (offset) { 82 __vector unsigned char unaligned_shift_index; 83 int bytes = num_cols + offset; 84 85 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) { 86 /* Slow path to prevent buffer overread. Since there is no way to 87 * read a partial AltiVec register, overread would occur on the last 88 * chunk of the last image row if the right edge is not on a 16-byte 89 * boundary. It could also occur on other rows if the bytes per row 90 * is low enough. Since we can't determine whether we're on the last 91 * image row, we have to assume every row is the last. 92 */ 93 memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); 94 rgb0 = vec_ld(0, tmpbuf); 95 rgb1 = vec_ld(16, tmpbuf); 96 rgb2 = vec_ld(32, tmpbuf); 97 #if RGB_PIXELSIZE == 4 98 rgb3 = vec_ld(48, tmpbuf); 99 #endif 100 } else { 101 /* Fast path */ 102 rgb0 = vec_ld(0, inptr); 103 if (bytes > 16) 104 rgb1 = vec_ld(16, inptr); 105 if (bytes > 32) 106 rgb2 = vec_ld(32, inptr); 107 if (bytes > 48) 108 rgb3 = vec_ld(48, inptr); 109 #if RGB_PIXELSIZE == 4 110 if (bytes > 64) 111 rgb4 = vec_ld(64, inptr); 112 #endif 113 unaligned_shift_index = vec_lvsl(0, inptr); 114 rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index); 115 rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index); 116 rgb2 = vec_perm(rgb2, rgb3, unaligned_shift_index); 117 #if RGB_PIXELSIZE == 4 118 rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index); 119 #endif 120 } 121 } else { 122 #endif /* __BIG_ENDIAN__ */ 123 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) { 124 /* Slow path */ 125 memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); 126 rgb0 = VEC_LD(0, tmpbuf); 127 rgb1 = VEC_LD(16, tmpbuf); 128 rgb2 = VEC_LD(32, tmpbuf); 129 #if RGB_PIXELSIZE == 4 130 rgb3 = VEC_LD(48, tmpbuf); 131 #endif 132 } else { 133 /* Fast path */ 134 rgb0 = VEC_LD(0, inptr); 135 if (num_cols > 16) 136 rgb1 = VEC_LD(16, inptr); 137 if (num_cols > 32) 138 rgb2 = VEC_LD(32, inptr); 139 #if RGB_PIXELSIZE == 4 140 if (num_cols > 48) 141 rgb3 = VEC_LD(48, inptr); 142 #endif 143 } 144 #if __BIG_ENDIAN__ 145 } 146 #endif 147 148 #if RGB_PIXELSIZE == 3 149 /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 150 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga 151 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf 152 * 153 * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3 154 * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7 155 * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb 156 * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf 157 */ 158 rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX0); 159 rgbg1 = vec_perm(rgb0, rgb1, (__vector unsigned char)RGBG_INDEX1); 160 rgbg2 = vec_perm(rgb1, rgb2, (__vector unsigned char)RGBG_INDEX2); 161 rgbg3 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX3); 162 #else 163 /* rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3 164 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7 165 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb 166 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf 167 * 168 * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3 169 * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7 170 * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb 171 * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf 172 */ 173 rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX); 174 rgbg1 = vec_perm(rgb1, rgb1, (__vector unsigned char)RGBG_INDEX); 175 rgbg2 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX); 176 rgbg3 = vec_perm(rgb3, rgb3, (__vector unsigned char)RGBG_INDEX); 177 #endif 178 179 /* rg0 = R0 G0 R1 G1 R2 G2 R3 G3 180 * bg0 = B0 G0 B1 G1 B2 G2 B3 G3 181 * ... 182 * 183 * NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't 184 * support unsigned vectors. 185 */ 186 rg0 = (__vector signed short)VEC_UNPACKHU(rgbg0); 187 bg0 = (__vector signed short)VEC_UNPACKLU(rgbg0); 188 rg1 = (__vector signed short)VEC_UNPACKHU(rgbg1); 189 bg1 = (__vector signed short)VEC_UNPACKLU(rgbg1); 190 rg2 = (__vector signed short)VEC_UNPACKHU(rgbg2); 191 bg2 = (__vector signed short)VEC_UNPACKLU(rgbg2); 192 rg3 = (__vector signed short)VEC_UNPACKHU(rgbg3); 193 bg3 = (__vector signed short)VEC_UNPACKLU(rgbg3); 194 195 /* (Original) 196 * Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 197 * Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 198 * Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 199 * 200 * (This implementation) 201 * Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 202 * Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 203 * Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 204 */ 205 206 /* Calculate Y values */ 207 208 y0 = vec_msums(rg0, pw_f0299_f0337, pd_onehalf); 209 y1 = vec_msums(rg1, pw_f0299_f0337, pd_onehalf); 210 y2 = vec_msums(rg2, pw_f0299_f0337, pd_onehalf); 211 y3 = vec_msums(rg3, pw_f0299_f0337, pd_onehalf); 212 y0 = vec_msums(bg0, pw_f0114_f0250, y0); 213 y1 = vec_msums(bg1, pw_f0114_f0250, y1); 214 y2 = vec_msums(bg2, pw_f0114_f0250, y2); 215 y3 = vec_msums(bg3, pw_f0114_f0250, y3); 216 /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from 217 * each dword into a new 16-bit vector, which is the equivalent of 218 * descaling the 32-bit results (right-shifting by 16 bits) and then 219 * packing them. 220 */ 221 yl = vec_perm((__vector unsigned short)y0, (__vector unsigned short)y1, 222 shift_pack_index); 223 yh = vec_perm((__vector unsigned short)y2, (__vector unsigned short)y3, 224 shift_pack_index); 225 y = vec_pack(yl, yh); 226 vec_st(y, 0, outptr0); 227 228 /* Calculate Cb values */ 229 cb0 = vec_msums(rg0, pw_mf016_mf033, pd_onehalfm1_cj); 230 cb1 = vec_msums(rg1, pw_mf016_mf033, pd_onehalfm1_cj); 231 cb2 = vec_msums(rg2, pw_mf016_mf033, pd_onehalfm1_cj); 232 cb3 = vec_msums(rg3, pw_mf016_mf033, pd_onehalfm1_cj); 233 cb0 = (__vector int)vec_msum((__vector unsigned short)bg0, pw_f050_f000, 234 (__vector unsigned int)cb0); 235 cb1 = (__vector int)vec_msum((__vector unsigned short)bg1, pw_f050_f000, 236 (__vector unsigned int)cb1); 237 cb2 = (__vector int)vec_msum((__vector unsigned short)bg2, pw_f050_f000, 238 (__vector unsigned int)cb2); 239 cb3 = (__vector int)vec_msum((__vector unsigned short)bg3, pw_f050_f000, 240 (__vector unsigned int)cb3); 241 cbl = vec_perm((__vector unsigned short)cb0, 242 (__vector unsigned short)cb1, shift_pack_index); 243 cbh = vec_perm((__vector unsigned short)cb2, 244 (__vector unsigned short)cb3, shift_pack_index); 245 cb = vec_pack(cbl, cbh); 246 vec_st(cb, 0, outptr1); 247 248 /* Calculate Cr values */ 249 cr0 = vec_msums(bg0, pw_mf008_mf041, pd_onehalfm1_cj); 250 cr1 = vec_msums(bg1, pw_mf008_mf041, pd_onehalfm1_cj); 251 cr2 = vec_msums(bg2, pw_mf008_mf041, pd_onehalfm1_cj); 252 cr3 = vec_msums(bg3, pw_mf008_mf041, pd_onehalfm1_cj); 253 cr0 = (__vector int)vec_msum((__vector unsigned short)rg0, pw_f050_f000, 254 (__vector unsigned int)cr0); 255 cr1 = (__vector int)vec_msum((__vector unsigned short)rg1, pw_f050_f000, 256 (__vector unsigned int)cr1); 257 cr2 = (__vector int)vec_msum((__vector unsigned short)rg2, pw_f050_f000, 258 (__vector unsigned int)cr2); 259 cr3 = (__vector int)vec_msum((__vector unsigned short)rg3, pw_f050_f000, 260 (__vector unsigned int)cr3); 261 crl = vec_perm((__vector unsigned short)cr0, 262 (__vector unsigned short)cr1, shift_pack_index); 263 crh = vec_perm((__vector unsigned short)cr2, 264 (__vector unsigned short)cr3, shift_pack_index); 265 cr = vec_pack(crl, crh); 266 vec_st(cr, 0, outptr2); 267 } 268 } 269 }