1 /* 2 * AltiVec optimizations for libjpeg-turbo 3 * 4 * Copyright (C) 2015, D. R. Commander. All Rights Reserved. 5 * 6 * This software is provided 'as-is', without any express or implied 7 * warranty. In no event will the authors be held liable for any damages 8 * arising from the use of this software. 9 * 10 * Permission is granted to anyone to use this software for any purpose, 11 * including commercial applications, and to alter it and redistribute it 12 * freely, subject to the following restrictions: 13 * 14 * 1. The origin of this software must not be misrepresented; you must not 15 * claim that you wrote the original software. If you use this software 16 * in a product, an acknowledgment in the product documentation would be 17 * appreciated but is not required. 18 * 2. Altered source versions must be plainly marked as such, and must not be 19 * misrepresented as being the original software. 20 * 3. This notice may not be removed or altered from any source distribution. 21 */ 22 23 /* This file is included by jdcolor-altivec.c */ 24 25 26 void jsimd_ycc_rgb_convert_altivec (JDIMENSION out_width, JSAMPIMAGE input_buf, 27 JDIMENSION input_row, 28 JSAMPARRAY output_buf, int num_rows) 29 { 30 JSAMPROW outptr, inptr0, inptr1, inptr2; 31 int pitch = out_width * RGB_PIXELSIZE, num_cols; 32 #if __BIG_ENDIAN__ 33 int offset; 34 #endif 35 unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16]; 36 37 __vector unsigned char rgb0, rgb1, rgb2, rgbx0, rgbx1, rgbx2, rgbx3, 38 y, cb, cr; 39 #if __BIG_ENDIAN__ 40 __vector unsigned char edgel, edgeh, edges, out0, out1, out2, out3; 41 #if RGB_PIXELSIZE == 4 42 __vector unsigned char out4; 43 #endif 44 #endif 45 #if RGB_PIXELSIZE == 4 46 __vector unsigned char rgb3; 47 #endif 48 __vector short rg0, rg1, rg2, rg3, bx0, bx1, bx2, bx3, yl, yh, cbl, cbh, 49 crl, crh, rl, rh, gl, gh, bl, bh, g0w, g1w, g2w, g3w; 50 __vector int g0, g1, g2, g3; 51 52 /* Constants 53 * NOTE: The >> 1 is to compensate for the fact that vec_madds() returns 17 54 * high-order bits, not 16. 55 */ 56 __vector short pw_f0402 = { __8X(F_0_402 >> 1) }, 57 pw_mf0228 = { __8X(-F_0_228 >> 1) }, 58 pw_mf0344_f0285 = { __4X2(-F_0_344, F_0_285) }, 59 pw_one = { __8X(1) }, pw_255 = { __8X(255) }, 60 pw_cj = { __8X(CENTERJSAMPLE) }; 61 __vector int pd_onehalf = { __4X(ONE_HALF) }; 62 __vector unsigned char pb_zero = { __16X(0) }, 63 #if __BIG_ENDIAN__ 64 shift_pack_index = {0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29}; 65 #else 66 shift_pack_index = {2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31}; 67 #endif 68 69 while (--num_rows >= 0) { 70 inptr0 = input_buf[0][input_row]; 71 inptr1 = input_buf[1][input_row]; 72 inptr2 = input_buf[2][input_row]; 73 input_row++; 74 outptr = *output_buf++; 75 76 for (num_cols = pitch; num_cols > 0; 77 num_cols -= RGB_PIXELSIZE * 16, outptr += RGB_PIXELSIZE * 16, 78 inptr0 += 16, inptr1 += 16, inptr2 += 16) { 79 80 y = vec_ld(0, inptr0); 81 /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't 82 * support unsigned vectors. 83 */ 84 yl = (__vector signed short)VEC_UNPACKHU(y); 85 yh = (__vector signed short)VEC_UNPACKLU(y); 86 87 cb = vec_ld(0, inptr1); 88 cbl = (__vector signed short)VEC_UNPACKHU(cb); 89 cbh = (__vector signed short)VEC_UNPACKLU(cb); 90 cbl = vec_sub(cbl, pw_cj); 91 cbh = vec_sub(cbh, pw_cj); 92 93 cr = vec_ld(0, inptr2); 94 crl = (__vector signed short)VEC_UNPACKHU(cr); 95 crh = (__vector signed short)VEC_UNPACKLU(cr); 96 crl = vec_sub(crl, pw_cj); 97 crh = vec_sub(crh, pw_cj); 98 99 /* (Original) 100 * R = Y + 1.40200 * Cr 101 * G = Y - 0.34414 * Cb - 0.71414 * Cr 102 * B = Y + 1.77200 * Cb 103 * 104 * (This implementation) 105 * R = Y + 0.40200 * Cr + Cr 106 * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 107 * B = Y - 0.22800 * Cb + Cb + Cb 108 */ 109 bl = vec_add(cbl, cbl); 110 bh = vec_add(cbh, cbh); 111 bl = vec_madds(bl, pw_mf0228, pw_one); 112 bh = vec_madds(bh, pw_mf0228, pw_one); 113 bl = vec_sra(bl, (__vector unsigned short)pw_one); 114 bh = vec_sra(bh, (__vector unsigned short)pw_one); 115 bl = vec_add(bl, cbl); 116 bh = vec_add(bh, cbh); 117 bl = vec_add(bl, cbl); 118 bh = vec_add(bh, cbh); 119 bl = vec_add(bl, yl); 120 bh = vec_add(bh, yh); 121 122 rl = vec_add(crl, crl); 123 rh = vec_add(crh, crh); 124 rl = vec_madds(rl, pw_f0402, pw_one); 125 rh = vec_madds(rh, pw_f0402, pw_one); 126 rl = vec_sra(rl, (__vector unsigned short)pw_one); 127 rh = vec_sra(rh, (__vector unsigned short)pw_one); 128 rl = vec_add(rl, crl); 129 rh = vec_add(rh, crh); 130 rl = vec_add(rl, yl); 131 rh = vec_add(rh, yh); 132 133 g0w = vec_mergeh(cbl, crl); 134 g1w = vec_mergel(cbl, crl); 135 g0 = vec_msums(g0w, pw_mf0344_f0285, pd_onehalf); 136 g1 = vec_msums(g1w, pw_mf0344_f0285, pd_onehalf); 137 g2w = vec_mergeh(cbh, crh); 138 g3w = vec_mergel(cbh, crh); 139 g2 = vec_msums(g2w, pw_mf0344_f0285, pd_onehalf); 140 g3 = vec_msums(g3w, pw_mf0344_f0285, pd_onehalf); 141 /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from 142 * each dword into a new 16-bit vector, which is the equivalent of 143 * descaling the 32-bit results (right-shifting by 16 bits) and then 144 * packing them. 145 */ 146 gl = vec_perm((__vector short)g0, (__vector short)g1, shift_pack_index); 147 gh = vec_perm((__vector short)g2, (__vector short)g3, shift_pack_index); 148 gl = vec_sub(gl, crl); 149 gh = vec_sub(gh, crh); 150 gl = vec_add(gl, yl); 151 gh = vec_add(gh, yh); 152 153 rg0 = vec_mergeh(rl, gl); 154 bx0 = vec_mergeh(bl, pw_255); 155 rg1 = vec_mergel(rl, gl); 156 bx1 = vec_mergel(bl, pw_255); 157 rg2 = vec_mergeh(rh, gh); 158 bx2 = vec_mergeh(bh, pw_255); 159 rg3 = vec_mergel(rh, gh); 160 bx3 = vec_mergel(bh, pw_255); 161 162 rgbx0 = vec_packsu(rg0, bx0); 163 rgbx1 = vec_packsu(rg1, bx1); 164 rgbx2 = vec_packsu(rg2, bx2); 165 rgbx3 = vec_packsu(rg3, bx3); 166 167 #if RGB_PIXELSIZE == 3 168 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 169 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 170 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 171 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 172 * 173 * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 174 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga 175 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf 176 */ 177 rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0); 178 rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1); 179 rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2); 180 #else 181 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 182 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 183 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 184 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 185 * 186 * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3 187 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7 188 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb 189 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf 190 */ 191 rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX); 192 rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX); 193 rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX); 194 rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX); 195 #endif 196 197 #if __BIG_ENDIAN__ 198 offset = (size_t)outptr & 15; 199 if (offset) { 200 __vector unsigned char unaligned_shift_index; 201 int bytes = num_cols + offset; 202 203 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) { 204 /* Slow path to prevent buffer overwrite. Since there is no way to 205 * write a partial AltiVec register, overwrite would occur on the 206 * last chunk of the last image row if the right edge is not on a 207 * 16-byte boundary. It could also occur on other rows if the bytes 208 * per row is low enough. Since we can't determine whether we're on 209 * the last image row, we have to assume every row is the last. 210 */ 211 vec_st(rgb0, 0, tmpbuf); 212 vec_st(rgb1, 16, tmpbuf); 213 vec_st(rgb2, 32, tmpbuf); 214 #if RGB_PIXELSIZE == 4 215 vec_st(rgb3, 48, tmpbuf); 216 #endif 217 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 218 } else { 219 /* Fast path */ 220 unaligned_shift_index = vec_lvsl(0, outptr); 221 edgel = vec_ld(0, outptr); 222 edgeh = vec_ld(min(num_cols - 1, RGB_PIXELSIZE * 16), outptr); 223 edges = vec_perm(edgeh, edgel, unaligned_shift_index); 224 unaligned_shift_index = vec_lvsr(0, outptr); 225 out0 = vec_perm(edges, rgb0, unaligned_shift_index); 226 out1 = vec_perm(rgb0, rgb1, unaligned_shift_index); 227 out2 = vec_perm(rgb1, rgb2, unaligned_shift_index); 228 #if RGB_PIXELSIZE == 4 229 out3 = vec_perm(rgb2, rgb3, unaligned_shift_index); 230 out4 = vec_perm(rgb3, edges, unaligned_shift_index); 231 #else 232 out3 = vec_perm(rgb2, edges, unaligned_shift_index); 233 #endif 234 vec_st(out0, 0, outptr); 235 if (bytes > 16) 236 vec_st(out1, 16, outptr); 237 if (bytes > 32) 238 vec_st(out2, 32, outptr); 239 if (bytes > 48) 240 vec_st(out3, 48, outptr); 241 #if RGB_PIXELSIZE == 4 242 if (bytes > 64) 243 vec_st(out4, 64, outptr); 244 #endif 245 } 246 } else { 247 #endif /* __BIG_ENDIAN__ */ 248 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) { 249 /* Slow path */ 250 VEC_ST(rgb0, 0, tmpbuf); 251 VEC_ST(rgb1, 16, tmpbuf); 252 VEC_ST(rgb2, 32, tmpbuf); 253 #if RGB_PIXELSIZE == 4 254 VEC_ST(rgb3, 48, tmpbuf); 255 #endif 256 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 257 } else { 258 /* Fast path */ 259 VEC_ST(rgb0, 0, outptr); 260 if (num_cols > 16) 261 VEC_ST(rgb1, 16, outptr); 262 if (num_cols > 32) 263 VEC_ST(rgb2, 32, outptr); 264 #if RGB_PIXELSIZE == 4 265 if (num_cols > 48) 266 VEC_ST(rgb3, 48, outptr); 267 #endif 268 } 269 #if __BIG_ENDIAN__ 270 } 271 #endif 272 } 273 } 274 } 275
