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 jcgray-altivec.c */
25
26
jsimd_rgb_gray_convert_altivec(JDIMENSION img_width,JSAMPARRAY input_buf,JSAMPIMAGE output_buf,JDIMENSION output_row,int num_rows)27 void jsimd_rgb_gray_convert_altivec(JDIMENSION img_width, JSAMPARRAY input_buf,
28 JSAMPIMAGE output_buf,
29 JDIMENSION output_row, int num_rows)
30 {
31 JSAMPROW inptr, outptr;
32 int pitch = img_width * RGB_PIXELSIZE, num_cols;
33 #if __BIG_ENDIAN__
34 int offset;
35 unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];
36 #endif
37
38 __vector unsigned char rgb0, rgb1 = { 0 }, rgb2 = { 0 },
39 rgbg0, rgbg1, rgbg2, rgbg3, y;
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;
48 __vector int y0, y1, y2, y3;
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 __vector int pd_onehalf = { __4X(ONE_HALF) };
54 __vector unsigned char pb_zero = { __16X(0) },
55 #if __BIG_ENDIAN__
56 shift_pack_index =
57 { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 };
58 #else
59 shift_pack_index =
60 { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 };
61 #endif
62
63 while (--num_rows >= 0) {
64 inptr = *input_buf++;
65 outptr = output_buf[0][output_row];
66 output_row++;
67
68 for (num_cols = pitch; num_cols > 0;
69 num_cols -= RGB_PIXELSIZE * 16, inptr += RGB_PIXELSIZE * 16,
70 outptr += 16) {
71
72 #if __BIG_ENDIAN__
73 /* Load 16 pixels == 48 or 64 bytes */
74 offset = (size_t)inptr & 15;
75 if (offset) {
76 __vector unsigned char unaligned_shift_index;
77 int bytes = num_cols + offset;
78
79 if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
80 /* Slow path to prevent buffer overread. Since there is no way to
81 * read a partial AltiVec register, overread would occur on the last
82 * chunk of the last image row if the right edge is not on a 16-byte
83 * boundary. It could also occur on other rows if the bytes per row
84 * is low enough. Since we can't determine whether we're on the last
85 * image row, we have to assume every row is the last.
86 */
87 memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
88 rgb0 = vec_ld(0, tmpbuf);
89 rgb1 = vec_ld(16, tmpbuf);
90 rgb2 = vec_ld(32, tmpbuf);
91 #if RGB_PIXELSIZE == 4
92 rgb3 = vec_ld(48, tmpbuf);
93 #endif
94 } else {
95 /* Fast path */
96 rgb0 = vec_ld(0, inptr);
97 if (bytes > 16)
98 rgb1 = vec_ld(16, inptr);
99 if (bytes > 32)
100 rgb2 = vec_ld(32, inptr);
101 if (bytes > 48)
102 rgb3 = vec_ld(48, inptr);
103 #if RGB_PIXELSIZE == 4
104 if (bytes > 64)
105 rgb4 = vec_ld(64, inptr);
106 #endif
107 unaligned_shift_index = vec_lvsl(0, inptr);
108 rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index);
109 rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index);
110 rgb2 = vec_perm(rgb2, rgb3, unaligned_shift_index);
111 #if RGB_PIXELSIZE == 4
112 rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index);
113 #endif
114 }
115 } else {
116 if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
117 /* Slow path */
118 memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
119 rgb0 = vec_ld(0, tmpbuf);
120 rgb1 = vec_ld(16, tmpbuf);
121 rgb2 = vec_ld(32, tmpbuf);
122 #if RGB_PIXELSIZE == 4
123 rgb3 = vec_ld(48, tmpbuf);
124 #endif
125 } else {
126 /* Fast path */
127 rgb0 = vec_ld(0, inptr);
128 if (num_cols > 16)
129 rgb1 = vec_ld(16, inptr);
130 if (num_cols > 32)
131 rgb2 = vec_ld(32, inptr);
132 #if RGB_PIXELSIZE == 4
133 if (num_cols > 48)
134 rgb3 = vec_ld(48, inptr);
135 #endif
136 }
137 }
138 #else
139 /* Little endian */
140 rgb0 = vec_vsx_ld(0, inptr);
141 if (num_cols > 16)
142 rgb1 = vec_vsx_ld(16, inptr);
143 if (num_cols > 32)
144 rgb2 = vec_vsx_ld(32, inptr);
145 #if RGB_PIXELSIZE == 4
146 if (num_cols > 48)
147 rgb3 = vec_vsx_ld(48, inptr);
148 #endif
149 #endif
150
151 #if RGB_PIXELSIZE == 3
152 /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
153 * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
154 * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
155 *
156 * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
157 * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
158 * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
159 * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
160 */
161 rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX0);
162 rgbg1 = vec_perm(rgb0, rgb1, (__vector unsigned char)RGBG_INDEX1);
163 rgbg2 = vec_perm(rgb1, rgb2, (__vector unsigned char)RGBG_INDEX2);
164 rgbg3 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX3);
165 #else
166 /* rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
167 * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
168 * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
169 * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
170 *
171 * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
172 * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
173 * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
174 * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
175 */
176 rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX);
177 rgbg1 = vec_perm(rgb1, rgb1, (__vector unsigned char)RGBG_INDEX);
178 rgbg2 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX);
179 rgbg3 = vec_perm(rgb3, rgb3, (__vector unsigned char)RGBG_INDEX);
180 #endif
181
182 /* rg0 = R0 G0 R1 G1 R2 G2 R3 G3
183 * bg0 = B0 G0 B1 G1 B2 G2 B3 G3
184 * ...
185 *
186 * NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
187 * support unsigned vectors.
188 */
189 rg0 = (__vector signed short)VEC_UNPACKHU(rgbg0);
190 bg0 = (__vector signed short)VEC_UNPACKLU(rgbg0);
191 rg1 = (__vector signed short)VEC_UNPACKHU(rgbg1);
192 bg1 = (__vector signed short)VEC_UNPACKLU(rgbg1);
193 rg2 = (__vector signed short)VEC_UNPACKHU(rgbg2);
194 bg2 = (__vector signed short)VEC_UNPACKLU(rgbg2);
195 rg3 = (__vector signed short)VEC_UNPACKHU(rgbg3);
196 bg3 = (__vector signed short)VEC_UNPACKLU(rgbg3);
197
198 /* (Original)
199 * Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
200 *
201 * (This implementation)
202 * Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
203 */
204
205 /* Calculate Y values */
206
207 y0 = vec_msums(rg0, pw_f0299_f0337, pd_onehalf);
208 y1 = vec_msums(rg1, pw_f0299_f0337, pd_onehalf);
209 y2 = vec_msums(rg2, pw_f0299_f0337, pd_onehalf);
210 y3 = vec_msums(rg3, pw_f0299_f0337, pd_onehalf);
211 y0 = vec_msums(bg0, pw_f0114_f0250, y0);
212 y1 = vec_msums(bg1, pw_f0114_f0250, y1);
213 y2 = vec_msums(bg2, pw_f0114_f0250, y2);
214 y3 = vec_msums(bg3, pw_f0114_f0250, y3);
215 /* Clever way to avoid 4 shifts + 2 packs. This packs the high word from
216 * each dword into a new 16-bit vector, which is the equivalent of
217 * descaling the 32-bit results (right-shifting by 16 bits) and then
218 * packing them.
219 */
220 yl = vec_perm((__vector unsigned short)y0, (__vector unsigned short)y1,
221 shift_pack_index);
222 yh = vec_perm((__vector unsigned short)y2, (__vector unsigned short)y3,
223 shift_pack_index);
224 y = vec_pack(yl, yh);
225 vec_st(y, 0, outptr);
226 }
227 }
228 }
229