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 jdmerge-altivec.c */ 24 25 26 void jsimd_h2v1_merged_upsample_altivec(JDIMENSION output_width, 27 JSAMPIMAGE input_buf, 28 JDIMENSION in_row_group_ctr, 29 JSAMPARRAY output_buf) 30 { 31 JSAMPROW outptr, inptr0, inptr1, inptr2; 32 int pitch = output_width * RGB_PIXELSIZE, num_cols, yloop; 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, rgb2, rgbx0, rgbx1, rgbx2, rgbx3, 39 y, cb, cr; 40 #if __BIG_ENDIAN__ 41 __vector unsigned char edgel, edgeh, edges, out0, out1, out2, out3; 42 #if RGB_PIXELSIZE == 4 43 __vector unsigned char out4; 44 #endif 45 #endif 46 #if RGB_PIXELSIZE == 4 47 __vector unsigned char rgb3; 48 #endif 49 __vector short rg0, rg1, rg2, rg3, bx0, bx1, bx2, bx3, ye, yo, cbl, cbh, 50 crl, crh, r_yl, r_yh, g_yl, g_yh, b_yl, b_yh, g_y0w, g_y1w, g_y2w, g_y3w, 51 rl, rh, gl, gh, bl, bh, re, ro, ge, go, be, bo; 52 __vector int g_y0, g_y1, g_y2, g_y3; 53 54 /* Constants 55 * NOTE: The >> 1 is to compensate for the fact that vec_madds() returns 17 56 * high-order bits, not 16. 57 */ 58 __vector short pw_f0402 = { __8X(F_0_402 >> 1) }, 59 pw_mf0228 = { __8X(-F_0_228 >> 1) }, 60 pw_mf0344_f0285 = { __4X2(-F_0_344, F_0_285) }, 61 pw_one = { __8X(1) }, pw_255 = { __8X(255) }, 62 pw_cj = { __8X(CENTERJSAMPLE) }; 63 __vector int pd_onehalf = { __4X(ONE_HALF) }; 64 __vector unsigned char pb_zero = { __16X(0) }, 65 #if __BIG_ENDIAN__ 66 shift_pack_index = 67 { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 }, 68 even_index = 69 { 0, 16, 0, 18, 0, 20, 0, 22, 0, 24, 0, 26, 0, 28, 0, 30 }, 70 odd_index = 71 { 0, 17, 0, 19, 0, 21, 0, 23, 0, 25, 0, 27, 0, 29, 0, 31 }; 72 #else 73 shift_pack_index = 74 { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 }, 75 even_index = 76 { 16, 0, 18, 0, 20, 0, 22, 0, 24, 0, 26, 0, 28, 0, 30, 0 }, 77 odd_index = 78 { 17, 0, 19, 0, 21, 0, 23, 0, 25, 0, 27, 0, 29, 0, 31, 0 }; 79 #endif 80 81 inptr0 = input_buf[0][in_row_group_ctr]; 82 inptr1 = input_buf[1][in_row_group_ctr]; 83 inptr2 = input_buf[2][in_row_group_ctr]; 84 outptr = output_buf[0]; 85 86 for (num_cols = pitch; num_cols > 0; inptr1 += 16, inptr2 += 16) { 87 88 cb = vec_ld(0, inptr1); 89 /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't 90 * support unsigned vectors. 91 */ 92 cbl = (__vector signed short)VEC_UNPACKHU(cb); 93 cbh = (__vector signed short)VEC_UNPACKLU(cb); 94 cbl = vec_sub(cbl, pw_cj); 95 cbh = vec_sub(cbh, pw_cj); 96 97 cr = vec_ld(0, inptr2); 98 crl = (__vector signed short)VEC_UNPACKHU(cr); 99 crh = (__vector signed short)VEC_UNPACKLU(cr); 100 crl = vec_sub(crl, pw_cj); 101 crh = vec_sub(crh, pw_cj); 102 103 /* (Original) 104 * R = Y + 1.40200 * Cr 105 * G = Y - 0.34414 * Cb - 0.71414 * Cr 106 * B = Y + 1.77200 * Cb 107 * 108 * (This implementation) 109 * R = Y + 0.40200 * Cr + Cr 110 * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 111 * B = Y - 0.22800 * Cb + Cb + Cb 112 */ 113 b_yl = vec_add(cbl, cbl); 114 b_yh = vec_add(cbh, cbh); 115 b_yl = vec_madds(b_yl, pw_mf0228, pw_one); 116 b_yh = vec_madds(b_yh, pw_mf0228, pw_one); 117 b_yl = vec_sra(b_yl, (__vector unsigned short)pw_one); 118 b_yh = vec_sra(b_yh, (__vector unsigned short)pw_one); 119 b_yl = vec_add(b_yl, cbl); 120 b_yh = vec_add(b_yh, cbh); 121 b_yl = vec_add(b_yl, cbl); 122 b_yh = vec_add(b_yh, cbh); 123 124 r_yl = vec_add(crl, crl); 125 r_yh = vec_add(crh, crh); 126 r_yl = vec_madds(r_yl, pw_f0402, pw_one); 127 r_yh = vec_madds(r_yh, pw_f0402, pw_one); 128 r_yl = vec_sra(r_yl, (__vector unsigned short)pw_one); 129 r_yh = vec_sra(r_yh, (__vector unsigned short)pw_one); 130 r_yl = vec_add(r_yl, crl); 131 r_yh = vec_add(r_yh, crh); 132 133 g_y0w = vec_mergeh(cbl, crl); 134 g_y1w = vec_mergel(cbl, crl); 135 g_y0 = vec_msums(g_y0w, pw_mf0344_f0285, pd_onehalf); 136 g_y1 = vec_msums(g_y1w, pw_mf0344_f0285, pd_onehalf); 137 g_y2w = vec_mergeh(cbh, crh); 138 g_y3w = vec_mergel(cbh, crh); 139 g_y2 = vec_msums(g_y2w, pw_mf0344_f0285, pd_onehalf); 140 g_y3 = vec_msums(g_y3w, 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 g_yl = vec_perm((__vector short)g_y0, (__vector short)g_y1, 147 shift_pack_index); 148 g_yh = vec_perm((__vector short)g_y2, (__vector short)g_y3, 149 shift_pack_index); 150 g_yl = vec_sub(g_yl, crl); 151 g_yh = vec_sub(g_yh, crh); 152 153 for (yloop = 0; yloop < 2 && num_cols > 0; yloop++, 154 num_cols -= RGB_PIXELSIZE * 16, 155 outptr += RGB_PIXELSIZE * 16, inptr0 += 16) { 156 157 y = vec_ld(0, inptr0); 158 ye = (__vector signed short)vec_perm(pb_zero, y, even_index); 159 yo = (__vector signed short)vec_perm(pb_zero, y, odd_index); 160 161 if (yloop == 0) { 162 be = vec_add(b_yl, ye); 163 bo = vec_add(b_yl, yo); 164 re = vec_add(r_yl, ye); 165 ro = vec_add(r_yl, yo); 166 ge = vec_add(g_yl, ye); 167 go = vec_add(g_yl, yo); 168 } else { 169 be = vec_add(b_yh, ye); 170 bo = vec_add(b_yh, yo); 171 re = vec_add(r_yh, ye); 172 ro = vec_add(r_yh, yo); 173 ge = vec_add(g_yh, ye); 174 go = vec_add(g_yh, yo); 175 } 176 177 rl = vec_mergeh(re, ro); 178 rh = vec_mergel(re, ro); 179 gl = vec_mergeh(ge, go); 180 gh = vec_mergel(ge, go); 181 bl = vec_mergeh(be, bo); 182 bh = vec_mergel(be, bo); 183 184 rg0 = vec_mergeh(rl, gl); 185 bx0 = vec_mergeh(bl, pw_255); 186 rg1 = vec_mergel(rl, gl); 187 bx1 = vec_mergel(bl, pw_255); 188 rg2 = vec_mergeh(rh, gh); 189 bx2 = vec_mergeh(bh, pw_255); 190 rg3 = vec_mergel(rh, gh); 191 bx3 = vec_mergel(bh, pw_255); 192 193 rgbx0 = vec_packsu(rg0, bx0); 194 rgbx1 = vec_packsu(rg1, bx1); 195 rgbx2 = vec_packsu(rg2, bx2); 196 rgbx3 = vec_packsu(rg3, bx3); 197 198 #if RGB_PIXELSIZE == 3 199 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 200 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 201 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 202 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 203 * 204 * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 205 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga 206 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf 207 */ 208 rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0); 209 rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1); 210 rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2); 211 #else 212 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3 213 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7 214 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb 215 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf 216 * 217 * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3 218 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7 219 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb 220 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf 221 */ 222 rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX); 223 rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX); 224 rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX); 225 rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX); 226 #endif 227 228 #if __BIG_ENDIAN__ 229 offset = (size_t)outptr & 15; 230 if (offset) { 231 __vector unsigned char unaligned_shift_index; 232 int bytes = num_cols + offset; 233 234 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) { 235 /* Slow path to prevent buffer overwrite. Since there is no way to 236 * write a partial AltiVec register, overwrite would occur on the 237 * last chunk of the last image row if the right edge is not on a 238 * 16-byte boundary. It could also occur on other rows if the bytes 239 * per row is low enough. Since we can't determine whether we're on 240 * the last image row, we have to assume every row is the last. 241 */ 242 vec_st(rgb0, 0, tmpbuf); 243 vec_st(rgb1, 16, tmpbuf); 244 vec_st(rgb2, 32, tmpbuf); 245 #if RGB_PIXELSIZE == 4 246 vec_st(rgb3, 48, tmpbuf); 247 #endif 248 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 249 } else { 250 /* Fast path */ 251 unaligned_shift_index = vec_lvsl(0, outptr); 252 edgel = vec_ld(0, outptr); 253 edgeh = vec_ld(min(num_cols - 1, RGB_PIXELSIZE * 16), outptr); 254 edges = vec_perm(edgeh, edgel, unaligned_shift_index); 255 unaligned_shift_index = vec_lvsr(0, outptr); 256 out0 = vec_perm(edges, rgb0, unaligned_shift_index); 257 out1 = vec_perm(rgb0, rgb1, unaligned_shift_index); 258 out2 = vec_perm(rgb1, rgb2, unaligned_shift_index); 259 #if RGB_PIXELSIZE == 4 260 out3 = vec_perm(rgb2, rgb3, unaligned_shift_index); 261 out4 = vec_perm(rgb3, edges, unaligned_shift_index); 262 #else 263 out3 = vec_perm(rgb2, edges, unaligned_shift_index); 264 #endif 265 vec_st(out0, 0, outptr); 266 if (bytes > 16) 267 vec_st(out1, 16, outptr); 268 if (bytes > 32) 269 vec_st(out2, 32, outptr); 270 if (bytes > 48) 271 vec_st(out3, 48, outptr); 272 #if RGB_PIXELSIZE == 4 273 if (bytes > 64) 274 vec_st(out4, 64, outptr); 275 #endif 276 } 277 } else { 278 #endif /* __BIG_ENDIAN__ */ 279 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) { 280 /* Slow path */ 281 VEC_ST(rgb0, 0, tmpbuf); 282 VEC_ST(rgb1, 16, tmpbuf); 283 VEC_ST(rgb2, 32, tmpbuf); 284 #if RGB_PIXELSIZE == 4 285 VEC_ST(rgb3, 48, tmpbuf); 286 #endif 287 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16)); 288 } else { 289 /* Fast path */ 290 VEC_ST(rgb0, 0, outptr); 291 if (num_cols > 16) 292 VEC_ST(rgb1, 16, outptr); 293 if (num_cols > 32) 294 VEC_ST(rgb2, 32, outptr); 295 #if RGB_PIXELSIZE == 4 296 if (num_cols > 48) 297 VEC_ST(rgb3, 48, outptr); 298 #endif 299 } 300 #if __BIG_ENDIAN__ 301 } 302 #endif 303 } 304 } 305 } 306 307 308 void jsimd_h2v2_merged_upsample_altivec(JDIMENSION output_width, 309 JSAMPIMAGE input_buf, 310 JDIMENSION in_row_group_ctr, 311 JSAMPARRAY output_buf) 312 { 313 JSAMPROW inptr, outptr; 314 315 inptr = input_buf[0][in_row_group_ctr]; 316 outptr = output_buf[0]; 317 318 input_buf[0][in_row_group_ctr] = input_buf[0][in_row_group_ctr * 2]; 319 jsimd_h2v1_merged_upsample_altivec(output_width, input_buf, in_row_group_ctr, 320 output_buf); 321 322 input_buf[0][in_row_group_ctr] = input_buf[0][in_row_group_ctr * 2 + 1]; 323 output_buf[0] = output_buf[1]; 324 jsimd_h2v1_merged_upsample_altivec(output_width, input_buf, in_row_group_ctr, 325 output_buf); 326 327 input_buf[0][in_row_group_ctr] = inptr; 328 output_buf[0] = outptr; 329 } 330