jdcolext-altivec.c (10386B)
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, JSAMPARRAY output_buf, 28 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 = 65 { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 }; 66 #else 67 shift_pack_index = 68 { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 }; 69 #endif 70 71 while (--num_rows >= 0) { 72 inptr0 = input_buf[0][input_row]; 73 inptr1 = input_buf[1][input_row]; 74 inptr2 = input_buf[2][input_row]; 75 input_row++; 76 outptr = *output_buf++; 77 78 for (num_cols = pitch; num_cols > 0; 79 num_cols -= RGB_PIXELSIZE * 16, outptr += RGB_PIXELSIZE * 16, 80 inptr0 += 16, inptr1 += 16, inptr2 += 16) { 81 82 y = vec_ld(0, inptr0); 83 /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't 84 * support unsigned vectors. 85 */ 86 yl = (__vector signed short)VEC_UNPACKHU(y); 87 yh = (__vector signed short)VEC_UNPACKLU(y); 88 89 cb = vec_ld(0, inptr1); 90 cbl = (__vector signed short)VEC_UNPACKHU(cb); 91 cbh = (__vector signed short)VEC_UNPACKLU(cb); 92 cbl = vec_sub(cbl, pw_cj); 93 cbh = vec_sub(cbh, pw_cj); 94 95 cr = vec_ld(0, inptr2); 96 crl = (__vector signed short)VEC_UNPACKHU(cr); 97 crh = (__vector signed short)VEC_UNPACKLU(cr); 98 crl = vec_sub(crl, pw_cj); 99 crh = vec_sub(crh, pw_cj); 100 101 /* (Original) 102 * R = Y + 1.40200 * Cr 103 * G = Y - 0.34414 * Cb - 0.71414 * Cr 104 * B = Y + 1.77200 * Cb 105 * 106 * (This implementation) 107 * R = Y + 0.40200 * Cr + Cr 108 * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 109 * B = Y - 0.22800 * Cb + Cb + Cb 110 */ 111 bl = vec_add(cbl, cbl); 112 bh = vec_add(cbh, cbh); 113 bl = vec_madds(bl, pw_mf0228, pw_one); 114 bh = vec_madds(bh, pw_mf0228, pw_one); 115 bl = vec_sra(bl, (__vector unsigned short)pw_one); 116 bh = vec_sra(bh, (__vector unsigned short)pw_one); 117 bl = vec_add(bl, cbl); 118 bh = vec_add(bh, cbh); 119 bl = vec_add(bl, cbl); 120 bh = vec_add(bh, cbh); 121 bl = vec_add(bl, yl); 122 bh = vec_add(bh, yh); 123 124 rl = vec_add(crl, crl); 125 rh = vec_add(crh, crh); 126 rl = vec_madds(rl, pw_f0402, pw_one); 127 rh = vec_madds(rh, pw_f0402, pw_one); 128 rl = vec_sra(rl, (__vector unsigned short)pw_one); 129 rh = vec_sra(rh, (__vector unsigned short)pw_one); 130 rl = vec_add(rl, crl); 131 rh = vec_add(rh, crh); 132 rl = vec_add(rl, yl); 133 rh = vec_add(rh, yh); 134 135 g0w = vec_mergeh(cbl, crl); 136 g1w = vec_mergel(cbl, crl); 137 g0 = vec_msums(g0w, pw_mf0344_f0285, pd_onehalf); 138 g1 = vec_msums(g1w, pw_mf0344_f0285, pd_onehalf); 139 g2w = vec_mergeh(cbh, crh); 140 g3w = vec_mergel(cbh, crh); 141 g2 = vec_msums(g2w, pw_mf0344_f0285, pd_onehalf); 142 g3 = vec_msums(g3w, pw_mf0344_f0285, pd_onehalf); 143 /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from 144 * each dword into a new 16-bit vector, which is the equivalent of 145 * descaling the 32-bit results (right-shifting by 16 bits) and then 146 * packing them. 147 */ 148 gl = vec_perm((__vector short)g0, (__vector short)g1, shift_pack_index); 149 gh = vec_perm((__vector short)g2, (__vector short)g3, shift_pack_index); 150 gl = vec_sub(gl, crl); 151 gh = vec_sub(gh, crh); 152 gl = vec_add(gl, yl); 153 gh = vec_add(gh, yh); 154 155 rg0 = vec_mergeh(rl, gl); 156 bx0 = vec_mergeh(bl, pw_255); 157 rg1 = vec_mergel(rl, gl); 158 bx1 = vec_mergel(bl, pw_255); 159 rg2 = vec_mergeh(rh, gh); 160 bx2 = vec_mergeh(bh, pw_255); 161 rg3 = vec_mergel(rh, gh); 162 bx3 = vec_mergel(bh, pw_255); 163 164 rgbx0 = vec_packsu(rg0, bx0); 165 rgbx1 = vec_packsu(rg1, bx1); 166 rgbx2 = vec_packsu(rg2, bx2); 167 rgbx3 = vec_packsu(rg3, bx3); 168 169 #if RGB_PIXELSIZE == 3 170 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 171 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 172 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 173 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 174 * 175 * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 176 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga 177 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf 178 */ 179 rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0); 180 rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1); 181 rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2); 182 #else 183 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 184 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 185 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 186 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 187 * 188 * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3 189 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7 190 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb 191 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf 192 */ 193 rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX); 194 rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX); 195 rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX); 196 rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX); 197 #endif 198 199 #if __BIG_ENDIAN__ 200 offset = (size_t)outptr & 15; 201 if (offset) { 202 __vector unsigned char unaligned_shift_index; 203 int bytes = num_cols + offset; 204 205 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) { 206 /* Slow path to prevent buffer overwrite. Since there is no way to 207 * write a partial AltiVec register, overwrite would occur on the 208 * last chunk of the last image row if the right edge is not on a 209 * 16-byte boundary. It could also occur on other rows if the bytes 210 * per row is low enough. Since we can't determine whether we're on 211 * the last image row, we have to assume every row is the last. 212 */ 213 vec_st(rgb0, 0, tmpbuf); 214 vec_st(rgb1, 16, tmpbuf); 215 vec_st(rgb2, 32, tmpbuf); 216 #if RGB_PIXELSIZE == 4 217 vec_st(rgb3, 48, tmpbuf); 218 #endif 219 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 220 } else { 221 /* Fast path */ 222 unaligned_shift_index = vec_lvsl(0, outptr); 223 edgel = vec_ld(0, outptr); 224 edgeh = vec_ld(min(num_cols - 1, RGB_PIXELSIZE * 16), outptr); 225 edges = vec_perm(edgeh, edgel, unaligned_shift_index); 226 unaligned_shift_index = vec_lvsr(0, outptr); 227 out0 = vec_perm(edges, rgb0, unaligned_shift_index); 228 out1 = vec_perm(rgb0, rgb1, unaligned_shift_index); 229 out2 = vec_perm(rgb1, rgb2, unaligned_shift_index); 230 #if RGB_PIXELSIZE == 4 231 out3 = vec_perm(rgb2, rgb3, unaligned_shift_index); 232 out4 = vec_perm(rgb3, edges, unaligned_shift_index); 233 #else 234 out3 = vec_perm(rgb2, edges, unaligned_shift_index); 235 #endif 236 vec_st(out0, 0, outptr); 237 if (bytes > 16) 238 vec_st(out1, 16, outptr); 239 if (bytes > 32) 240 vec_st(out2, 32, outptr); 241 if (bytes > 48) 242 vec_st(out3, 48, outptr); 243 #if RGB_PIXELSIZE == 4 244 if (bytes > 64) 245 vec_st(out4, 64, outptr); 246 #endif 247 } 248 } else { 249 #endif /* __BIG_ENDIAN__ */ 250 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) { 251 /* Slow path */ 252 VEC_ST(rgb0, 0, tmpbuf); 253 VEC_ST(rgb1, 16, tmpbuf); 254 VEC_ST(rgb2, 32, tmpbuf); 255 #if RGB_PIXELSIZE == 4 256 VEC_ST(rgb3, 48, tmpbuf); 257 #endif 258 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 259 } else { 260 /* Fast path */ 261 VEC_ST(rgb0, 0, outptr); 262 if (num_cols > 16) 263 VEC_ST(rgb1, 16, outptr); 264 if (num_cols > 32) 265 VEC_ST(rgb2, 32, outptr); 266 #if RGB_PIXELSIZE == 4 267 if (num_cols > 48) 268 VEC_ST(rgb3, 48, outptr); 269 #endif 270 } 271 #if __BIG_ENDIAN__ 272 } 273 #endif 274 } 275 } 276 }