ARM: Add 'neon_composite_over_n_8888_0565_ca' fast path
[profile/ivi/pixman.git] / pixman / pixman-arm-neon.c
1 /*
2  * Copyright © 2009 ARM Ltd, Movial Creative Technologies Oy
3  *
4  * Permission to use, copy, modify, distribute, and sell this software and its
5  * documentation for any purpose is hereby granted without fee, provided that
6  * the above copyright notice appear in all copies and that both that
7  * copyright notice and this permission notice appear in supporting
8  * documentation, and that the name of ARM Ltd not be used in
9  * advertising or publicity pertaining to distribution of the software without
10  * specific, written prior permission.  ARM Ltd makes no
11  * representations about the suitability of this software for any purpose.  It
12  * is provided "as is" without express or implied warranty.
13  *
14  * THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS
15  * SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND
16  * FITNESS, IN NO EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY
17  * SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
18  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN
19  * AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
20  * OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS
21  * SOFTWARE.
22  *
23  * Author:  Ian Rickards (ian.rickards@arm.com)
24  * Author:  Jonathan Morton (jonathan.morton@movial.com)
25  * Author:  Markku Vire (markku.vire@movial.com)
26  *
27  */
28
29 #ifdef HAVE_CONFIG_H
30 #include <config.h>
31 #endif
32
33 #include <string.h>
34 #include "pixman-private.h"
35 #include "pixman-arm-common.h"
36
37 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_8888_8888,
38                                    uint32_t, 1, uint32_t, 1)
39 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_x888_8888,
40                                    uint32_t, 1, uint32_t, 1)
41 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0565_0565,
42                                    uint16_t, 1, uint16_t, 1)
43 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0888_0888,
44                                    uint8_t, 3, uint8_t, 3)
45 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_8888_0565,
46                                    uint32_t, 1, uint16_t, 1)
47 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0565_8888,
48                                    uint16_t, 1, uint32_t, 1)
49 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0888_8888_rev,
50                                    uint8_t, 3, uint32_t, 1)
51 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0888_0565_rev,
52                                    uint8_t, 3, uint16_t, 1)
53 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_pixbuf_8888,
54                                    uint32_t, 1, uint32_t, 1)
55 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_rpixbuf_8888,
56                                    uint32_t, 1, uint32_t, 1)
57 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, add_8_8,
58                                    uint8_t, 1, uint8_t, 1)
59 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, add_8888_8888,
60                                    uint32_t, 1, uint32_t, 1)
61 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, over_8888_0565,
62                                    uint32_t, 1, uint16_t, 1)
63 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, over_8888_8888,
64                                    uint32_t, 1, uint32_t, 1)
65 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, out_reverse_8_0565,
66                                    uint8_t, 1, uint16_t, 1)
67
68 PIXMAN_ARM_BIND_FAST_PATH_N_DST (SKIP_ZERO_SRC, neon, over_n_0565,
69                                  uint16_t, 1)
70 PIXMAN_ARM_BIND_FAST_PATH_N_DST (SKIP_ZERO_SRC, neon, over_n_8888,
71                                  uint32_t, 1)
72 PIXMAN_ARM_BIND_FAST_PATH_N_DST (SKIP_ZERO_SRC, neon, over_reverse_n_8888,
73                                  uint32_t, 1)
74 PIXMAN_ARM_BIND_FAST_PATH_N_DST (0, neon, in_n_8,
75                                  uint8_t, 1)
76
77 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, over_n_8_0565,
78                                       uint8_t, 1, uint16_t, 1)
79 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, over_n_8_8888,
80                                       uint8_t, 1, uint32_t, 1)
81 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, over_n_8888_8888_ca,
82                                       uint32_t, 1, uint32_t, 1)
83 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, over_n_8888_0565_ca,
84                                       uint32_t, 1, uint16_t, 1)
85 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, over_n_8_8,
86                                       uint8_t, 1, uint8_t, 1)
87 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, add_n_8_8,
88                                       uint8_t, 1, uint8_t, 1)
89 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (SKIP_ZERO_SRC, neon, add_n_8_8888,
90                                       uint8_t, 1, uint32_t, 1)
91
92 PIXMAN_ARM_BIND_FAST_PATH_SRC_N_DST (SKIP_ZERO_MASK, neon, over_8888_n_8888,
93                                      uint32_t, 1, uint32_t, 1)
94 PIXMAN_ARM_BIND_FAST_PATH_SRC_N_DST (SKIP_ZERO_MASK, neon, over_8888_n_0565,
95                                      uint32_t, 1, uint16_t, 1)
96 PIXMAN_ARM_BIND_FAST_PATH_SRC_N_DST (SKIP_ZERO_MASK, neon, over_0565_n_0565,
97                                      uint16_t, 1, uint16_t, 1)
98 PIXMAN_ARM_BIND_FAST_PATH_SRC_N_DST (SKIP_ZERO_MASK, neon, add_8888_n_8888,
99                                      uint32_t, 1, uint32_t, 1)
100
101 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_8_8_8,
102                                         uint8_t, 1, uint8_t, 1, uint8_t, 1)
103 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_0565_8_0565,
104                                         uint16_t, 1, uint8_t, 1, uint16_t, 1)
105 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_8888_8_8888,
106                                         uint32_t, 1, uint8_t, 1, uint32_t, 1)
107 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_8888_8888_8888,
108                                         uint32_t, 1, uint32_t, 1, uint32_t, 1)
109 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_8888,
110                                         uint32_t, 1, uint8_t, 1, uint32_t, 1)
111 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8888_8888,
112                                         uint32_t, 1, uint32_t, 1, uint32_t, 1)
113 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_0565,
114                                         uint32_t, 1, uint8_t, 1, uint16_t, 1)
115 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_0565_8_0565,
116                                         uint16_t, 1, uint8_t, 1, uint16_t, 1)
117
118 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 8888_8888, OVER,
119                                         uint32_t, uint32_t)
120 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 8888_0565, OVER,
121                                         uint32_t, uint16_t)
122 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 8888_0565, SRC,
123                                         uint32_t, uint16_t)
124 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 0565_8888, SRC,
125                                         uint16_t, uint32_t)
126
127 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_A8_DST (SKIP_ZERO_SRC, neon, 8888_8_0565,
128                                            OVER, uint32_t, uint16_t)
129 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_A8_DST (SKIP_ZERO_SRC, neon, 0565_8_0565,
130                                            OVER, uint16_t, uint16_t)
131
132 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 8888_8888, SRC,
133                                          uint32_t, uint32_t)
134 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 8888_0565, SRC,
135                                          uint32_t, uint16_t)
136 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 0565_x888, SRC,
137                                          uint16_t, uint32_t)
138 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 0565_0565, SRC,
139                                          uint16_t, uint16_t)
140
141 void
142 pixman_composite_src_n_8_asm_neon (int32_t   w,
143                                    int32_t   h,
144                                    uint8_t  *dst,
145                                    int32_t   dst_stride,
146                                    uint8_t   src);
147
148 void
149 pixman_composite_src_n_0565_asm_neon (int32_t   w,
150                                       int32_t   h,
151                                       uint16_t *dst,
152                                       int32_t   dst_stride,
153                                       uint16_t  src);
154
155 void
156 pixman_composite_src_n_8888_asm_neon (int32_t   w,
157                                       int32_t   h,
158                                       uint32_t *dst,
159                                       int32_t   dst_stride,
160                                       uint32_t  src);
161
162 static pixman_bool_t
163 pixman_fill_neon (uint32_t *bits,
164                   int       stride,
165                   int       bpp,
166                   int       x,
167                   int       y,
168                   int       width,
169                   int       height,
170                   uint32_t  _xor)
171 {
172     /* stride is always multiple of 32bit units in pixman */
173     uint32_t byte_stride = stride * sizeof(uint32_t);
174
175     switch (bpp)
176     {
177     case 8:
178         pixman_composite_src_n_8_asm_neon (
179                 width,
180                 height,
181                 (uint8_t *)(((char *) bits) + y * byte_stride + x),
182                 byte_stride,
183                 _xor & 0xff);
184         return TRUE;
185     case 16:
186         pixman_composite_src_n_0565_asm_neon (
187                 width,
188                 height,
189                 (uint16_t *)(((char *) bits) + y * byte_stride + x * 2),
190                 byte_stride / 2,
191                 _xor & 0xffff);
192         return TRUE;
193     case 32:
194         pixman_composite_src_n_8888_asm_neon (
195                 width,
196                 height,
197                 (uint32_t *)(((char *) bits) + y * byte_stride + x * 4),
198                 byte_stride / 4,
199                 _xor);
200         return TRUE;
201     default:
202         return FALSE;
203     }
204 }
205
206 static pixman_bool_t
207 pixman_blt_neon (uint32_t *src_bits,
208                  uint32_t *dst_bits,
209                  int       src_stride,
210                  int       dst_stride,
211                  int       src_bpp,
212                  int       dst_bpp,
213                  int       src_x,
214                  int       src_y,
215                  int       dst_x,
216                  int       dst_y,
217                  int       width,
218                  int       height)
219 {
220     if (src_bpp != dst_bpp)
221         return FALSE;
222
223     switch (src_bpp)
224     {
225     case 16:
226         pixman_composite_src_0565_0565_asm_neon (
227                 width, height,
228                 (uint16_t *)(((char *) dst_bits) +
229                 dst_y * dst_stride * 4 + dst_x * 2), dst_stride * 2,
230                 (uint16_t *)(((char *) src_bits) +
231                 src_y * src_stride * 4 + src_x * 2), src_stride * 2);
232         return TRUE;
233     case 32:
234         pixman_composite_src_8888_8888_asm_neon (
235                 width, height,
236                 (uint32_t *)(((char *) dst_bits) +
237                 dst_y * dst_stride * 4 + dst_x * 4), dst_stride,
238                 (uint32_t *)(((char *) src_bits) +
239                 src_y * src_stride * 4 + src_x * 4), src_stride);
240         return TRUE;
241     default:
242         return FALSE;
243     }
244 }
245
246 static const pixman_fast_path_t arm_neon_fast_paths[] =
247 {
248     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     r5g6b5,   neon_composite_src_0565_0565),
249     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     b5g6r5,   neon_composite_src_0565_0565),
250     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
251     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
252     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
253     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
254     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     a8r8g8b8, neon_composite_src_0565_8888),
255     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     x8r8g8b8, neon_composite_src_0565_8888),
256     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     a8b8g8r8, neon_composite_src_0565_8888),
257     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     x8b8g8r8, neon_composite_src_0565_8888),
258     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
259     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
260     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
261     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
262     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     a8r8g8b8, neon_composite_src_8888_8888),
263     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     a8b8g8r8, neon_composite_src_8888_8888),
264     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
265     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     a8b8g8r8, neon_composite_src_x888_8888),
266     PIXMAN_STD_FAST_PATH (SRC,  r8g8b8,   null,     r8g8b8,   neon_composite_src_0888_0888),
267     PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     x8r8g8b8, neon_composite_src_0888_8888_rev),
268     PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     r5g6b5,   neon_composite_src_0888_0565_rev),
269     PIXMAN_STD_FAST_PATH (SRC,  pixbuf,   pixbuf,   a8r8g8b8, neon_composite_src_pixbuf_8888),
270     PIXMAN_STD_FAST_PATH (SRC,  pixbuf,   pixbuf,   a8b8g8r8, neon_composite_src_rpixbuf_8888),
271     PIXMAN_STD_FAST_PATH (SRC,  rpixbuf,  rpixbuf,  a8r8g8b8, neon_composite_src_rpixbuf_8888),
272     PIXMAN_STD_FAST_PATH (SRC,  rpixbuf,  rpixbuf,  a8b8g8r8, neon_composite_src_pixbuf_8888),
273     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       a8,       neon_composite_over_n_8_8),
274     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       r5g6b5,   neon_composite_over_n_8_0565),
275     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       b5g6r5,   neon_composite_over_n_8_0565),
276     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       a8r8g8b8, neon_composite_over_n_8_8888),
277     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       x8r8g8b8, neon_composite_over_n_8_8888),
278     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       a8b8g8r8, neon_composite_over_n_8_8888),
279     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       x8b8g8r8, neon_composite_over_n_8_8888),
280     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     r5g6b5,   neon_composite_over_n_0565),
281     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     a8r8g8b8, neon_composite_over_n_8888),
282     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     x8r8g8b8, neon_composite_over_n_8888),
283     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8r8g8b8, a8r8g8b8, neon_composite_over_n_8888_8888_ca),
284     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8r8g8b8, x8r8g8b8, neon_composite_over_n_8888_8888_ca),
285     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8b8g8r8, a8b8g8r8, neon_composite_over_n_8888_8888_ca),
286     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8b8g8r8, x8b8g8r8, neon_composite_over_n_8888_8888_ca),
287     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8r8g8b8, r5g6b5,   neon_composite_over_n_8888_0565_ca),
288     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8b8g8r8, b5g6r5,   neon_composite_over_n_8888_0565_ca),
289     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, solid,    a8r8g8b8, neon_composite_over_8888_n_8888),
290     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, solid,    x8r8g8b8, neon_composite_over_8888_n_8888),
291     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, solid,    r5g6b5,   neon_composite_over_8888_n_0565),
292     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, solid,    b5g6r5,   neon_composite_over_8888_n_0565),
293     PIXMAN_STD_FAST_PATH (OVER, r5g6b5,   solid,    r5g6b5,   neon_composite_over_0565_n_0565),
294     PIXMAN_STD_FAST_PATH (OVER, b5g6r5,   solid,    b5g6r5,   neon_composite_over_0565_n_0565),
295     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       a8r8g8b8, neon_composite_over_8888_8_8888),
296     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       x8r8g8b8, neon_composite_over_8888_8_8888),
297     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       a8b8g8r8, neon_composite_over_8888_8_8888),
298     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       x8b8g8r8, neon_composite_over_8888_8_8888),
299     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       r5g6b5,   neon_composite_over_8888_8_0565),
300     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       b5g6r5,   neon_composite_over_8888_8_0565),
301     PIXMAN_STD_FAST_PATH (OVER, r5g6b5,   a8,       r5g6b5,   neon_composite_over_0565_8_0565),
302     PIXMAN_STD_FAST_PATH (OVER, b5g6r5,   a8,       b5g6r5,   neon_composite_over_0565_8_0565),
303     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_over_8888_8888_8888),
304     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     r5g6b5,   neon_composite_over_8888_0565),
305     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     b5g6r5,   neon_composite_over_8888_0565),
306     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     a8r8g8b8, neon_composite_over_8888_8888),
307     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     x8r8g8b8, neon_composite_over_8888_8888),
308     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     a8b8g8r8, neon_composite_over_8888_8888),
309     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     x8b8g8r8, neon_composite_over_8888_8888),
310     PIXMAN_STD_FAST_PATH (OVER, x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
311     PIXMAN_STD_FAST_PATH (OVER, x8b8g8r8, null,     a8b8g8r8, neon_composite_src_x888_8888),
312     PIXMAN_STD_FAST_PATH (ADD,  solid,    a8,       a8,       neon_composite_add_n_8_8),
313     PIXMAN_STD_FAST_PATH (ADD,  solid,    a8,       a8r8g8b8, neon_composite_add_n_8_8888),
314     PIXMAN_STD_FAST_PATH (ADD,  solid,    a8,       a8b8g8r8, neon_composite_add_n_8_8888),
315     PIXMAN_STD_FAST_PATH (ADD,  a8,       a8,       a8,       neon_composite_add_8_8_8),
316     PIXMAN_STD_FAST_PATH (ADD,  r5g6b5,   a8,       r5g6b5,   neon_composite_add_0565_8_0565),
317     PIXMAN_STD_FAST_PATH (ADD,  b5g6r5,   a8,       b5g6r5,   neon_composite_add_0565_8_0565),
318     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, a8,       a8r8g8b8, neon_composite_add_8888_8_8888),
319     PIXMAN_STD_FAST_PATH (ADD,  a8b8g8r8, a8,       a8b8g8r8, neon_composite_add_8888_8_8888),
320     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_add_8888_8888_8888),
321     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, solid,    a8r8g8b8, neon_composite_add_8888_n_8888),
322     PIXMAN_STD_FAST_PATH (ADD,  a8b8g8r8, solid,    a8b8g8r8, neon_composite_add_8888_n_8888),
323     PIXMAN_STD_FAST_PATH (ADD,  a8,       null,     a8,       neon_composite_add_8_8),
324     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, null,     a8r8g8b8, neon_composite_add_8888_8888),
325     PIXMAN_STD_FAST_PATH (ADD,  a8b8g8r8, null,     a8b8g8r8, neon_composite_add_8888_8888),
326     PIXMAN_STD_FAST_PATH (IN,   solid,    null,     a8,       neon_composite_in_n_8),
327     PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8r8g8b8, neon_composite_over_reverse_n_8888),
328     PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8b8g8r8, neon_composite_over_reverse_n_8888),
329     PIXMAN_STD_FAST_PATH (OUT_REVERSE,  a8,    null, r5g6b5,   neon_composite_out_reverse_8_0565),
330     PIXMAN_STD_FAST_PATH (OUT_REVERSE,  a8,    null, b5g6r5,   neon_composite_out_reverse_8_0565),
331
332     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, neon_8888_8888),
333     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, a8b8g8r8, neon_8888_8888),
334     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, x8r8g8b8, neon_8888_8888),
335     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, x8b8g8r8, neon_8888_8888),
336
337     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, r5g6b5, neon_8888_0565),
338     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, b5g6r5, neon_8888_0565),
339
340     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, a8r8g8b8, r5g6b5, neon_8888_0565),
341     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, x8r8g8b8, r5g6b5, neon_8888_0565),
342     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, a8b8g8r8, b5g6r5, neon_8888_0565),
343     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, x8b8g8r8, b5g6r5, neon_8888_0565),
344
345     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, b5g6r5, x8b8g8r8, neon_0565_8888),
346     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (SRC, r5g6b5, x8r8g8b8, neon_0565_8888),
347     /* Note: NONE repeat is not supported yet */
348     SIMPLE_NEAREST_FAST_PATH_COVER (SRC, r5g6b5, a8r8g8b8, neon_0565_8888),
349     SIMPLE_NEAREST_FAST_PATH_COVER (SRC, b5g6r5, a8b8g8r8, neon_0565_8888),
350     SIMPLE_NEAREST_FAST_PATH_PAD (SRC, r5g6b5, a8r8g8b8, neon_0565_8888),
351     SIMPLE_NEAREST_FAST_PATH_PAD (SRC, b5g6r5, a8b8g8r8, neon_0565_8888),
352
353     PIXMAN_ARM_SIMPLE_NEAREST_A8_MASK_FAST_PATH (OVER, a8r8g8b8, r5g6b5, neon_8888_8_0565),
354     PIXMAN_ARM_SIMPLE_NEAREST_A8_MASK_FAST_PATH (OVER, a8b8g8r8, b5g6r5, neon_8888_8_0565),
355
356     PIXMAN_ARM_SIMPLE_NEAREST_A8_MASK_FAST_PATH (OVER, r5g6b5, r5g6b5, neon_0565_8_0565),
357     PIXMAN_ARM_SIMPLE_NEAREST_A8_MASK_FAST_PATH (OVER, b5g6r5, b5g6r5, neon_0565_8_0565),
358
359     SIMPLE_BILINEAR_FAST_PATH (SRC, a8r8g8b8, a8r8g8b8, neon_8888_8888),
360     SIMPLE_BILINEAR_FAST_PATH (SRC, a8r8g8b8, x8r8g8b8, neon_8888_8888),
361     SIMPLE_BILINEAR_FAST_PATH (SRC, x8r8g8b8, x8r8g8b8, neon_8888_8888),
362
363     SIMPLE_BILINEAR_FAST_PATH (SRC, a8r8g8b8, r5g6b5, neon_8888_0565),
364     SIMPLE_BILINEAR_FAST_PATH (SRC, x8r8g8b8, r5g6b5, neon_8888_0565),
365
366     SIMPLE_BILINEAR_FAST_PATH (SRC, r5g6b5, x8r8g8b8, neon_0565_x888),
367     SIMPLE_BILINEAR_FAST_PATH (SRC, r5g6b5, r5g6b5, neon_0565_0565),
368
369     { PIXMAN_OP_NONE },
370 };
371
372 static pixman_bool_t
373 arm_neon_blt (pixman_implementation_t *imp,
374               uint32_t *               src_bits,
375               uint32_t *               dst_bits,
376               int                      src_stride,
377               int                      dst_stride,
378               int                      src_bpp,
379               int                      dst_bpp,
380               int                      src_x,
381               int                      src_y,
382               int                      dst_x,
383               int                      dst_y,
384               int                      width,
385               int                      height)
386 {
387     if (!pixman_blt_neon (
388             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
389             src_x, src_y, dst_x, dst_y, width, height))
390
391     {
392         return _pixman_implementation_blt (
393             imp->delegate,
394             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
395             src_x, src_y, dst_x, dst_y, width, height);
396     }
397
398     return TRUE;
399 }
400
401 static pixman_bool_t
402 arm_neon_fill (pixman_implementation_t *imp,
403                uint32_t *               bits,
404                int                      stride,
405                int                      bpp,
406                int                      x,
407                int                      y,
408                int                      width,
409                int                      height,
410                uint32_t xor)
411 {
412     if (pixman_fill_neon (bits, stride, bpp, x, y, width, height, xor))
413         return TRUE;
414
415     return _pixman_implementation_fill (
416         imp->delegate, bits, stride, bpp, x, y, width, height, xor);
417 }
418
419 #define BIND_COMBINE_U(name)                                             \
420 void                                                                     \
421 pixman_composite_scanline_##name##_mask_asm_neon (int32_t         w,     \
422                                                   const uint32_t *dst,   \
423                                                   const uint32_t *src,   \
424                                                   const uint32_t *mask); \
425                                                                          \
426 void                                                                     \
427 pixman_composite_scanline_##name##_asm_neon (int32_t         w,          \
428                                              const uint32_t *dst,        \
429                                              const uint32_t *src);       \
430                                                                          \
431 static void                                                              \
432 neon_combine_##name##_u (pixman_implementation_t *imp,                   \
433                          pixman_op_t              op,                    \
434                          uint32_t *               dest,                  \
435                          const uint32_t *         src,                   \
436                          const uint32_t *         mask,                  \
437                          int                      width)                 \
438 {                                                                        \
439     if (mask)                                                            \
440         pixman_composite_scanline_##name##_mask_asm_neon (width, dest,   \
441                                                           src, mask);    \
442     else                                                                 \
443         pixman_composite_scanline_##name##_asm_neon (width, dest, src);  \
444 }
445
446 BIND_COMBINE_U (over)
447 BIND_COMBINE_U (add)
448 BIND_COMBINE_U (out_reverse)
449
450 pixman_implementation_t *
451 _pixman_implementation_create_arm_neon (pixman_implementation_t *fallback)
452 {
453     pixman_implementation_t *imp =
454         _pixman_implementation_create (fallback, arm_neon_fast_paths);
455
456     imp->combine_32[PIXMAN_OP_OVER] = neon_combine_over_u;
457     imp->combine_32[PIXMAN_OP_ADD] = neon_combine_add_u;
458     imp->combine_32[PIXMAN_OP_OUT_REVERSE] = neon_combine_out_reverse_u;
459
460     imp->blt = arm_neon_blt;
461     imp->fill = arm_neon_fill;
462
463     return imp;
464 }