blob: 9cdcd02fd58582a6ff2d7b6f4e52545a94953e33 [file] [log] [blame]
DRCac4daa72015-01-10 22:56:26 +00001/*
2 * AltiVec optimizations for libjpeg-turbo
3 *
4 * Copyright (C) 2015, D. R. Commander.
5 * All rights reserved.
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
26void 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, offset, num_cols;
32 unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];
33
34 __vector unsigned char rgb0, rgb1, rgb2, rgbx0, rgbx1, rgbx2, rgbx3,
35 y, cb, cr, edgel, edgeh, edges, out0, out1, out2, out3;
36#if RGB_PIXELSIZE == 4
37 __vector unsigned char rgb3, out4;
38#endif
39 __vector short rg0, rg1, rg2, rg3, bx0, bx1, bx2, bx3, yl, yh, cbl, cbh,
DRCa6a24c22015-01-13 10:00:12 +000040 crl, crh, rl, rh, gl, gh, bl, bh, g0w, g1w, g2w, g3w;
DRCac4daa72015-01-10 22:56:26 +000041 __vector int g0, g1, g2, g3;
42
43 /* Constants
44 * NOTE: The >> 1 is to compensate for the fact that vec_madds() returns 17
45 * high-order bits, not 16.
46 */
47 __vector short pw_f0402 = { __8X(F_0_402 >> 1) },
48 pw_mf0228 = { __8X(-F_0_228 >> 1) },
49 pw_mf0344_f0285 = { __4X2(-F_0_344, F_0_285) },
DRCa6a24c22015-01-13 10:00:12 +000050 pw_one = { __8X(1) }, pw_255 = { __8X(255) },
DRCac4daa72015-01-10 22:56:26 +000051 pw_cj = { __8X(CENTERJSAMPLE) };
52 __vector int pd_onehalf = { __4X(ONE_HALF) };
DRCa6a24c22015-01-13 10:00:12 +000053 __vector unsigned char pb_zero = { __16X(0) },
DRCac4daa72015-01-10 22:56:26 +000054 shift_pack_index =
55 { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};
56
57 while (--num_rows >= 0) {
58 inptr0 = input_buf[0][input_row];
59 inptr1 = input_buf[1][input_row];
60 inptr2 = input_buf[2][input_row];
61 input_row++;
62 outptr = *output_buf++;
63
64 for (num_cols = pitch; num_cols > 0;
65 num_cols -= RGB_PIXELSIZE * 16, outptr += RGB_PIXELSIZE * 16,
66 inptr0 += 16, inptr1 += 16, inptr2 += 16) {
67
68 y = vec_ld(0, inptr0);
69 /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
70 * support unsigned vectors.
71 */
DRCa6a24c22015-01-13 10:00:12 +000072 yl = (__vector signed short)vec_mergeh(pb_zero, y);
73 yh = (__vector signed short)vec_mergel(pb_zero, y);
DRCac4daa72015-01-10 22:56:26 +000074
75 cb = vec_ld(0, inptr1);
DRCa6a24c22015-01-13 10:00:12 +000076 cbl = (__vector signed short)vec_mergeh(pb_zero, cb);
77 cbh = (__vector signed short)vec_mergel(pb_zero, cb);
DRCac4daa72015-01-10 22:56:26 +000078 cbl = vec_sub(cbl, pw_cj);
79 cbh = vec_sub(cbh, pw_cj);
80
81 cr = vec_ld(0, inptr2);
DRCa6a24c22015-01-13 10:00:12 +000082 crl = (__vector signed short)vec_mergeh(pb_zero, cr);
83 crh = (__vector signed short)vec_mergel(pb_zero, cr);
DRCac4daa72015-01-10 22:56:26 +000084 crl = vec_sub(crl, pw_cj);
85 crh = vec_sub(crh, pw_cj);
86
87 /* (Original)
88 * R = Y + 1.40200 * Cr
89 * G = Y - 0.34414 * Cb - 0.71414 * Cr
90 * B = Y + 1.77200 * Cb
91 *
92 * (This implementation)
93 * R = Y + 0.40200 * Cr + Cr
94 * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
95 * B = Y - 0.22800 * Cb + Cb + Cb
96 */
97 bl = vec_add(cbl, cbl);
98 bh = vec_add(cbh, cbh);
99 bl = vec_madds(bl, pw_mf0228, pw_one);
100 bh = vec_madds(bh, pw_mf0228, pw_one);
101 bl = vec_sra(bl, (__vector unsigned short)pw_one);
102 bh = vec_sra(bh, (__vector unsigned short)pw_one);
103 bl = vec_add(bl, cbl);
104 bh = vec_add(bh, cbh);
105 bl = vec_add(bl, cbl);
106 bh = vec_add(bh, cbh);
107 bl = vec_add(bl, yl);
108 bh = vec_add(bh, yh);
109
110 rl = vec_add(crl, crl);
111 rh = vec_add(crh, crh);
112 rl = vec_madds(rl, pw_f0402, pw_one);
113 rh = vec_madds(rh, pw_f0402, pw_one);
114 rl = vec_sra(rl, (__vector unsigned short)pw_one);
115 rh = vec_sra(rh, (__vector unsigned short)pw_one);
116 rl = vec_add(rl, crl);
117 rh = vec_add(rh, crh);
118 rl = vec_add(rl, yl);
119 rh = vec_add(rh, yh);
120
DRCa6a24c22015-01-13 10:00:12 +0000121 g0w = vec_mergeh(cbl, crl);
122 g1w = vec_mergel(cbl, crl);
123 g0 = vec_msums(g0w, pw_mf0344_f0285, pd_onehalf);
124 g1 = vec_msums(g1w, pw_mf0344_f0285, pd_onehalf);
125 g2w = vec_mergeh(cbh, crh);
126 g3w = vec_mergel(cbh, crh);
127 g2 = vec_msums(g2w, pw_mf0344_f0285, pd_onehalf);
128 g3 = vec_msums(g3w, pw_mf0344_f0285, pd_onehalf);
DRCac4daa72015-01-10 22:56:26 +0000129 /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from
130 * each dword into a new 16-bit vector, which is the equivalent of
131 * descaling the 32-bit results (right-shifting by 16 bits) and then
132 * packing them.
133 */
DRC51eba062015-01-11 07:38:35 +0000134 gl = vec_perm((__vector short)g0, (__vector short)g1, shift_pack_index);
135 gh = vec_perm((__vector short)g2, (__vector short)g3, shift_pack_index);
DRCac4daa72015-01-10 22:56:26 +0000136 gl = vec_sub(gl, crl);
137 gh = vec_sub(gh, crh);
138 gl = vec_add(gl, yl);
139 gh = vec_add(gh, yh);
140
141 rg0 = vec_mergeh(rl, gl);
142 bx0 = vec_mergeh(bl, pw_255);
143 rg1 = vec_mergel(rl, gl);
144 bx1 = vec_mergel(bl, pw_255);
145 rg2 = vec_mergeh(rh, gh);
146 bx2 = vec_mergeh(bh, pw_255);
147 rg3 = vec_mergel(rh, gh);
148 bx3 = vec_mergel(bh, pw_255);
149
150 rgbx0 = vec_packsu(rg0, bx0);
151 rgbx1 = vec_packsu(rg1, bx1);
152 rgbx2 = vec_packsu(rg2, bx2);
153 rgbx3 = vec_packsu(rg3, bx3);
154
155#if RGB_PIXELSIZE == 3
156 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
157 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
158 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
159 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
160 *
161 * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
162 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
163 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
164 */
165 rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0);
166 rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1);
167 rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2);
168#else
169 /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
170 * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
171 * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
172 * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
173 *
174 * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
175 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
176 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
177 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
178 */
179 rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX);
180 rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX);
181 rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX);
182 rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX);
183#endif
184
185 offset = (size_t)outptr & 15;
186 if (offset) {
187 __vector unsigned char unaligned_shift_index;
188 int bytes = num_cols + offset;
189
190 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
191 /* Slow path to prevent buffer overwrite. Since there is no way to
192 * write a partial AltiVec register, overwrite would occur on the
193 * last chunk of the last image row if the right edge is not on a
194 * 16-byte boundary. It could also occur on other rows if the bytes
195 * per row is low enough. Since we can't determine whether we're on
196 * the last image row, we have to assume every row is the last.
197 */
198 vec_st(rgb0, 0, tmpbuf);
199 vec_st(rgb1, 16, tmpbuf);
200 vec_st(rgb2, 32, tmpbuf);
201#if RGB_PIXELSIZE == 4
202 vec_st(rgb3, 48, tmpbuf);
203#endif
204 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
205 } else {
206 /* Fast path */
207 unaligned_shift_index = vec_lvsl(0, outptr);
208 edgel = vec_ld(0, outptr);
209 edgeh = vec_ld(min(num_cols, RGB_PIXELSIZE * 16) + offset, outptr);
210 edges = vec_perm(edgeh, edgel, unaligned_shift_index);
211 unaligned_shift_index = vec_lvsr(0, outptr);
212 out0 = vec_perm(edges, rgb0, unaligned_shift_index);
213 out1 = vec_perm(rgb0, rgb1, unaligned_shift_index);
214 out2 = vec_perm(rgb1, rgb2, unaligned_shift_index);
215#if RGB_PIXELSIZE == 4
216 out3 = vec_perm(rgb2, rgb3, unaligned_shift_index);
217 out4 = vec_perm(rgb3, edges, unaligned_shift_index);
218#else
219 out3 = vec_perm(rgb2, edges, unaligned_shift_index);
220#endif
221 vec_st(out0, 0, outptr);
222 if (bytes > 16)
223 vec_st(out1, 16, outptr);
224 if (bytes > 32)
225 vec_st(out2, 32, outptr);
226 if (bytes > 48)
227 vec_st(out3, 48, outptr);
228#if RGB_PIXELSIZE == 4
229 if (bytes > 64)
230 vec_st(out4, 64, outptr);
231#endif
232 }
233 } else {
234 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
235 /* Slow path */
236 vec_st(rgb0, 0, tmpbuf);
237 vec_st(rgb1, 16, tmpbuf);
238 vec_st(rgb2, 32, tmpbuf);
239#if RGB_PIXELSIZE == 4
240 vec_st(rgb3, 48, tmpbuf);
241#endif
242 memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
243 } else {
244 /* Fast path */
245 vec_st(rgb0, 0, outptr);
246 if (num_cols > 16)
247 vec_st(rgb1, 16, outptr);
248 if (num_cols > 32)
249 vec_st(rgb2, 32, outptr);
250#if RGB_PIXELSIZE == 4
251 if (num_cols > 48)
252 vec_st(rgb3, 48, outptr);
253#endif
254 }
255 }
256 }
257 }
258}