1 /*
2  * AltiVec optimizations for libjpeg-turbo
3  *
4  * Copyright (C) 2014-2015, D. R. Commander.
5  * Copyright (C) 2014, Jay Foad.
6  * All rights reserved.
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,
28                                      JSAMPARRAY input_buf,
29                                      JSAMPIMAGE output_buf,
30                                      JDIMENSION output_row, int num_rows)
31 {
32   JSAMPROW inptr, outptr;
33   int pitch = img_width * RGB_PIXELSIZE, num_cols;
34 #if __BIG_ENDIAN__
35   int offset;
36   unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];
37 #endif
38 
39   __vector unsigned char rgb0, rgb1 = {0}, rgb2 = {0},
40     rgbg0, rgbg1, rgbg2, rgbg3, y;
41 #if __BIG_ENDIAN__ || RGB_PIXELSIZE == 4
42   __vector unsigned char rgb3 = {0};
43 #endif
44 #if __BIG_ENDIAN__ && RGB_PIXELSIZE == 4
45   __vector unsigned char rgb4 = {0};
46 #endif
47   __vector short rg0, rg1, rg2, rg3, bg0, bg1, bg2, bg3;
48   __vector unsigned short yl, yh;
49   __vector int y0, y1, y2, y3;
50 
51   /* Constants */
52   __vector short pw_f0299_f0337 = { __4X2(F_0_299, F_0_337) },
53     pw_f0114_f0250 = { __4X2(F_0_114, F_0_250) };
54   __vector int pd_onehalf = { __4X(ONE_HALF) };
55   __vector unsigned char pb_zero = { __16X(0) },
56 #if __BIG_ENDIAN__
57     shift_pack_index = {0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29};
58 #else
59     shift_pack_index = {2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31};
60 #endif
61 
62   while (--num_rows >= 0) {
63     inptr = *input_buf++;
64     outptr = output_buf[0][output_row];
65     output_row++;
66 
67     for (num_cols = pitch; num_cols > 0;
68          num_cols -= RGB_PIXELSIZE * 16, inptr += RGB_PIXELSIZE * 16,
69          outptr += 16) {
70 
71 #if __BIG_ENDIAN__
72       /* Load 16 pixels == 48 or 64 bytes */
73       offset = (size_t)inptr & 15;
74       if (offset) {
75         __vector unsigned char unaligned_shift_index;
76         int bytes = num_cols + offset;
77 
78         if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
79           /* Slow path to prevent buffer overread.  Since there is no way to
80            * read a partial AltiVec register, overread would occur on the last
81            * chunk of the last image row if the right edge is not on a 16-byte
82            * boundary.  It could also occur on other rows if the bytes per row
83            * is low enough.  Since we can't determine whether we're on the last
84            * image row, we have to assume every row is the last.
85            */
86           memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
87           rgb0 = vec_ld(0, tmpbuf);
88           rgb1 = vec_ld(16, tmpbuf);
89           rgb2 = vec_ld(32, tmpbuf);
90 #if RGB_PIXELSIZE == 4
91           rgb3 = vec_ld(48, tmpbuf);
92 #endif
93         } else {
94           /* Fast path */
95           rgb0 = vec_ld(0, inptr);
96           if (bytes > 16)
97             rgb1 = vec_ld(16, inptr);
98           if (bytes > 32)
99             rgb2 = vec_ld(32, inptr);
100           if (bytes > 48)
101             rgb3 = vec_ld(48, inptr);
102 #if RGB_PIXELSIZE == 4
103           if (bytes > 64)
104             rgb4 = vec_ld(64, inptr);
105 #endif
106           unaligned_shift_index = vec_lvsl(0, inptr);
107           rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index);
108           rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index);
109           rgb2 = vec_perm(rgb2, rgb3, unaligned_shift_index);
110 #if RGB_PIXELSIZE == 4
111           rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index);
112 #endif
113         }
114       } else {
115         if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
116           /* Slow path */
117           memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
118           rgb0 = vec_ld(0, tmpbuf);
119           rgb1 = vec_ld(16, tmpbuf);
120           rgb2 = vec_ld(32, tmpbuf);
121 #if RGB_PIXELSIZE == 4
122           rgb3 = vec_ld(48, tmpbuf);
123 #endif
124         } else {
125           /* Fast path */
126           rgb0 = vec_ld(0, inptr);
127           if (num_cols > 16)
128             rgb1 = vec_ld(16, inptr);
129           if (num_cols > 32)
130             rgb2 = vec_ld(32, inptr);
131 #if RGB_PIXELSIZE == 4
132           if (num_cols > 48)
133             rgb3 = vec_ld(48, inptr);
134 #endif
135         }
136       }
137 #else
138       /* Little endian */
139       rgb0 = vec_vsx_ld(0, inptr);
140       if (num_cols > 16)
141         rgb1 = vec_vsx_ld(16, inptr);
142       if (num_cols > 32)
143         rgb2 = vec_vsx_ld(32, inptr);
144 #if RGB_PIXELSIZE == 4
145       if (num_cols > 48)
146         rgb3 = vec_vsx_ld(48, inptr);
147 #endif
148 #endif
149 
150 #if RGB_PIXELSIZE == 3
151       /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
152        * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
153        * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
154        *
155        * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
156        * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
157        * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
158        * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
159        */
160       rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX0);
161       rgbg1 = vec_perm(rgb0, rgb1, (__vector unsigned char)RGBG_INDEX1);
162       rgbg2 = vec_perm(rgb1, rgb2, (__vector unsigned char)RGBG_INDEX2);
163       rgbg3 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX3);
164 #else
165       /* rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
166        * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
167        * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
168        * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
169        *
170        * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
171        * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
172        * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
173        * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
174        */
175       rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX);
176       rgbg1 = vec_perm(rgb1, rgb1, (__vector unsigned char)RGBG_INDEX);
177       rgbg2 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX);
178       rgbg3 = vec_perm(rgb3, rgb3, (__vector unsigned char)RGBG_INDEX);
179 #endif
180 
181       /* rg0 = R0 G0 R1 G1 R2 G2 R3 G3
182        * bg0 = B0 G0 B1 G1 B2 G2 B3 G3
183        * ...
184        *
185        * NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
186        * support unsigned vectors.
187        */
188       rg0 = (__vector signed short)VEC_UNPACKHU(rgbg0);
189       bg0 = (__vector signed short)VEC_UNPACKLU(rgbg0);
190       rg1 = (__vector signed short)VEC_UNPACKHU(rgbg1);
191       bg1 = (__vector signed short)VEC_UNPACKLU(rgbg1);
192       rg2 = (__vector signed short)VEC_UNPACKHU(rgbg2);
193       bg2 = (__vector signed short)VEC_UNPACKLU(rgbg2);
194       rg3 = (__vector signed short)VEC_UNPACKHU(rgbg3);
195       bg3 = (__vector signed short)VEC_UNPACKLU(rgbg3);
196 
197       /* (Original)
198        * Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
199        *
200        * (This implementation)
201        * Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
202        */
203 
204       /* Calculate Y values */
205 
206       y0 = vec_msums(rg0, pw_f0299_f0337, pd_onehalf);
207       y1 = vec_msums(rg1, pw_f0299_f0337, pd_onehalf);
208       y2 = vec_msums(rg2, pw_f0299_f0337, pd_onehalf);
209       y3 = vec_msums(rg3, pw_f0299_f0337, pd_onehalf);
210       y0 = vec_msums(bg0, pw_f0114_f0250, y0);
211       y1 = vec_msums(bg1, pw_f0114_f0250, y1);
212       y2 = vec_msums(bg2, pw_f0114_f0250, y2);
213       y3 = vec_msums(bg3, pw_f0114_f0250, y3);
214       /* Clever way to avoid 4 shifts + 2 packs.  This packs the high word from
215        * each dword into a new 16-bit vector, which is the equivalent of
216        * descaling the 32-bit results (right-shifting by 16 bits) and then
217        * packing them.
218        */
219       yl = vec_perm((__vector unsigned short)y0, (__vector unsigned short)y1,
220                     shift_pack_index);
221       yh = vec_perm((__vector unsigned short)y2, (__vector unsigned short)y3,
222                     shift_pack_index);
223       y = vec_pack(yl, yh);
224       vec_st(y, 0, outptr);
225     }
226   }
227 }
228