ARM: NEON optimization for scaled over_8888_0565 with nearest filter
[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, add_8_8,
56                                    uint8_t, 1, uint8_t, 1)
57 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, add_8888_8888,
58                                    uint32_t, 1, uint32_t, 1)
59 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, over_8888_0565,
60                                    uint32_t, 1, uint16_t, 1)
61 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, over_8888_8888,
62                                    uint32_t, 1, uint32_t, 1)
63 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, out_reverse_8_0565,
64                                    uint8_t, 1, uint16_t, 1)
65
66 PIXMAN_ARM_BIND_FAST_PATH_N_DST (neon, over_n_0565,
67                                  uint16_t, 1)
68 PIXMAN_ARM_BIND_FAST_PATH_N_DST (neon, over_n_8888,
69                                  uint32_t, 1)
70 PIXMAN_ARM_BIND_FAST_PATH_N_DST (neon, over_reverse_n_8888,
71                                  uint32_t, 1)
72
73 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (neon, over_n_8_0565,
74                                       uint8_t, 1, uint16_t, 1)
75 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (neon, over_n_8_8888,
76                                       uint8_t, 1, uint32_t, 1)
77 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (neon, over_n_8888_8888_ca,
78                                       uint32_t, 1, uint32_t, 1)
79 PIXMAN_ARM_BIND_FAST_PATH_N_MASK_DST (neon, add_n_8_8,
80                                       uint8_t, 1, uint8_t, 1)
81
82 PIXMAN_ARM_BIND_FAST_PATH_SRC_N_DST (neon, over_8888_n_8888,
83                                      uint32_t, 1, uint32_t, 1)
84
85 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_8_8_8,
86                                         uint8_t, 1, uint8_t, 1, uint8_t, 1)
87 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_0565_8_0565,
88                                         uint16_t, 1, uint8_t, 1, uint16_t, 1)
89 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, add_8888_8888_8888,
90                                         uint32_t, 1, uint32_t, 1, uint32_t, 1)
91 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_8888,
92                                         uint32_t, 1, uint8_t, 1, uint32_t, 1)
93 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8888_8888,
94                                         uint32_t, 1, uint32_t, 1, uint32_t, 1)
95 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_0565,
96                                         uint32_t, 1, uint8_t, 1, uint16_t, 1)
97 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_0565_8_0565,
98                                         uint16_t, 1, uint8_t, 1, uint16_t, 1)
99
100 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 8888_8888, OVER,
101                                         uint32_t, uint32_t)
102 PIXMAN_ARM_BIND_SCALED_NEAREST_SRC_DST (neon, 8888_0565, OVER,
103                                         uint32_t, uint16_t)
104
105 void
106 pixman_composite_src_n_8_asm_neon (int32_t   w,
107                                    int32_t   h,
108                                    uint8_t  *dst,
109                                    int32_t   dst_stride,
110                                    uint8_t   src);
111
112 void
113 pixman_composite_src_n_0565_asm_neon (int32_t   w,
114                                       int32_t   h,
115                                       uint16_t *dst,
116                                       int32_t   dst_stride,
117                                       uint16_t  src);
118
119 void
120 pixman_composite_src_n_8888_asm_neon (int32_t   w,
121                                       int32_t   h,
122                                       uint32_t *dst,
123                                       int32_t   dst_stride,
124                                       uint32_t  src);
125
126 static pixman_bool_t
127 pixman_fill_neon (uint32_t *bits,
128                   int       stride,
129                   int       bpp,
130                   int       x,
131                   int       y,
132                   int       width,
133                   int       height,
134                   uint32_t  _xor)
135 {
136     /* stride is always multiple of 32bit units in pixman */
137     uint32_t byte_stride = stride * sizeof(uint32_t);
138
139     switch (bpp)
140     {
141     case 8:
142         pixman_composite_src_n_8_asm_neon (
143                 width,
144                 height,
145                 (uint8_t *)(((char *) bits) + y * byte_stride + x),
146                 byte_stride,
147                 _xor & 0xff);
148         return TRUE;
149     case 16:
150         pixman_composite_src_n_0565_asm_neon (
151                 width,
152                 height,
153                 (uint16_t *)(((char *) bits) + y * byte_stride + x * 2),
154                 byte_stride / 2,
155                 _xor & 0xffff);
156         return TRUE;
157     case 32:
158         pixman_composite_src_n_8888_asm_neon (
159                 width,
160                 height,
161                 (uint32_t *)(((char *) bits) + y * byte_stride + x * 4),
162                 byte_stride / 4,
163                 _xor);
164         return TRUE;
165     default:
166         return FALSE;
167     }
168 }
169
170 static pixman_bool_t
171 pixman_blt_neon (uint32_t *src_bits,
172                  uint32_t *dst_bits,
173                  int       src_stride,
174                  int       dst_stride,
175                  int       src_bpp,
176                  int       dst_bpp,
177                  int       src_x,
178                  int       src_y,
179                  int       dst_x,
180                  int       dst_y,
181                  int       width,
182                  int       height)
183 {
184     if (src_bpp != dst_bpp)
185         return FALSE;
186
187     switch (src_bpp)
188     {
189     case 16:
190         pixman_composite_src_0565_0565_asm_neon (
191                 width, height,
192                 (uint16_t *)(((char *) dst_bits) +
193                 dst_y * dst_stride * 4 + dst_x * 2), dst_stride * 2,
194                 (uint16_t *)(((char *) src_bits) +
195                 src_y * src_stride * 4 + src_x * 2), src_stride * 2);
196         return TRUE;
197     case 32:
198         pixman_composite_src_8888_8888_asm_neon (
199                 width, height,
200                 (uint32_t *)(((char *) dst_bits) +
201                 dst_y * dst_stride * 4 + dst_x * 4), dst_stride,
202                 (uint32_t *)(((char *) src_bits) +
203                 src_y * src_stride * 4 + src_x * 4), src_stride);
204         return TRUE;
205     default:
206         return FALSE;
207     }
208 }
209
210 static const pixman_fast_path_t arm_neon_fast_paths[] =
211 {
212     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     r5g6b5,   neon_composite_src_0565_0565),
213     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     b5g6r5,   neon_composite_src_0565_0565),
214     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
215     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
216     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
217     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
218     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     a8r8g8b8, neon_composite_src_0565_8888),
219     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     x8r8g8b8, neon_composite_src_0565_8888),
220     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     a8b8g8r8, neon_composite_src_0565_8888),
221     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     x8b8g8r8, neon_composite_src_0565_8888),
222     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
223     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
224     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
225     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
226     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     a8r8g8b8, neon_composite_src_8888_8888),
227     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     a8b8g8r8, neon_composite_src_8888_8888),
228     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
229     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     a8b8g8r8, neon_composite_src_x888_8888),
230     PIXMAN_STD_FAST_PATH (SRC,  r8g8b8,   null,     r8g8b8,   neon_composite_src_0888_0888),
231     PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     x8r8g8b8, neon_composite_src_0888_8888_rev),
232     PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     r5g6b5,   neon_composite_src_0888_0565_rev),
233     PIXMAN_STD_FAST_PATH (SRC,  pixbuf,   pixbuf,   a8r8g8b8, neon_composite_src_pixbuf_8888),
234     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       r5g6b5,   neon_composite_over_n_8_0565),
235     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       b5g6r5,   neon_composite_over_n_8_0565),
236     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       a8r8g8b8, neon_composite_over_n_8_8888),
237     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       x8r8g8b8, neon_composite_over_n_8_8888),
238     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       a8b8g8r8, neon_composite_over_n_8_8888),
239     PIXMAN_STD_FAST_PATH (OVER, solid,    a8,       x8b8g8r8, neon_composite_over_n_8_8888),
240     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     r5g6b5,   neon_composite_over_n_0565),
241     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     a8r8g8b8, neon_composite_over_n_8888),
242     PIXMAN_STD_FAST_PATH (OVER, solid,    null,     x8r8g8b8, neon_composite_over_n_8888),
243     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8r8g8b8, a8r8g8b8, neon_composite_over_n_8888_8888_ca),
244     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8r8g8b8, x8r8g8b8, neon_composite_over_n_8888_8888_ca),
245     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8b8g8r8, a8b8g8r8, neon_composite_over_n_8888_8888_ca),
246     PIXMAN_STD_FAST_PATH_CA (OVER, solid, a8b8g8r8, x8b8g8r8, neon_composite_over_n_8888_8888_ca),
247     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, solid,    a8r8g8b8, neon_composite_over_8888_n_8888),
248     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, solid,    x8r8g8b8, neon_composite_over_8888_n_8888),
249     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       a8r8g8b8, neon_composite_over_8888_8_8888),
250     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       x8r8g8b8, neon_composite_over_8888_8_8888),
251     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       a8b8g8r8, neon_composite_over_8888_8_8888),
252     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       x8b8g8r8, neon_composite_over_8888_8_8888),
253     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       r5g6b5,   neon_composite_over_8888_8_0565),
254     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       b5g6r5,   neon_composite_over_8888_8_0565),
255     PIXMAN_STD_FAST_PATH (OVER, r5g6b5,   a8,       r5g6b5,   neon_composite_over_0565_8_0565),
256     PIXMAN_STD_FAST_PATH (OVER, b5g6r5,   a8,       b5g6r5,   neon_composite_over_0565_8_0565),
257     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_over_8888_8888_8888),
258     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     r5g6b5,   neon_composite_over_8888_0565),
259     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     b5g6r5,   neon_composite_over_8888_0565),
260     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     a8r8g8b8, neon_composite_over_8888_8888),
261     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     x8r8g8b8, neon_composite_over_8888_8888),
262     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     a8b8g8r8, neon_composite_over_8888_8888),
263     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     x8b8g8r8, neon_composite_over_8888_8888),
264     PIXMAN_STD_FAST_PATH (OVER, x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
265     PIXMAN_STD_FAST_PATH (OVER, x8b8g8r8, null,     a8b8g8r8, neon_composite_src_x888_8888),
266     PIXMAN_STD_FAST_PATH (ADD,  solid,    a8,       a8,       neon_composite_add_n_8_8),
267     PIXMAN_STD_FAST_PATH (ADD,  a8,       a8,       a8,       neon_composite_add_8_8_8),
268     PIXMAN_STD_FAST_PATH (ADD,  r5g6b5,   a8,       r5g6b5,   neon_composite_add_0565_8_0565),
269     PIXMAN_STD_FAST_PATH (ADD,  b5g6r5,   a8,       b5g6r5,   neon_composite_add_0565_8_0565),
270     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_add_8888_8888_8888),
271     PIXMAN_STD_FAST_PATH (ADD,  a8,       null,     a8,       neon_composite_add_8_8),
272     PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, null,     a8r8g8b8, neon_composite_add_8888_8888),
273     PIXMAN_STD_FAST_PATH (ADD,  a8b8g8r8, null,     a8b8g8r8, neon_composite_add_8888_8888),
274     PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8r8g8b8, neon_composite_over_reverse_n_8888),
275     PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8b8g8r8, neon_composite_over_reverse_n_8888),
276     PIXMAN_STD_FAST_PATH (OUT_REVERSE,  a8,    null, r5g6b5,   neon_composite_out_reverse_8_0565),
277     PIXMAN_STD_FAST_PATH (OUT_REVERSE,  a8,    null, b5g6r5,   neon_composite_out_reverse_8_0565),
278
279     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, neon_8888_8888),
280     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, a8b8g8r8, neon_8888_8888),
281     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, x8r8g8b8, neon_8888_8888),
282     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, x8b8g8r8, neon_8888_8888),
283
284     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8r8g8b8, r5g6b5, neon_8888_0565),
285     PIXMAN_ARM_SIMPLE_NEAREST_FAST_PATH (OVER, a8b8g8r8, b5g6r5, neon_8888_0565),
286
287     { PIXMAN_OP_NONE },
288 };
289
290 static pixman_bool_t
291 arm_neon_blt (pixman_implementation_t *imp,
292               uint32_t *               src_bits,
293               uint32_t *               dst_bits,
294               int                      src_stride,
295               int                      dst_stride,
296               int                      src_bpp,
297               int                      dst_bpp,
298               int                      src_x,
299               int                      src_y,
300               int                      dst_x,
301               int                      dst_y,
302               int                      width,
303               int                      height)
304 {
305     if (!pixman_blt_neon (
306             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
307             src_x, src_y, dst_x, dst_y, width, height))
308
309     {
310         return _pixman_implementation_blt (
311             imp->delegate,
312             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
313             src_x, src_y, dst_x, dst_y, width, height);
314     }
315
316     return TRUE;
317 }
318
319 static pixman_bool_t
320 arm_neon_fill (pixman_implementation_t *imp,
321                uint32_t *               bits,
322                int                      stride,
323                int                      bpp,
324                int                      x,
325                int                      y,
326                int                      width,
327                int                      height,
328                uint32_t xor)
329 {
330     if (pixman_fill_neon (bits, stride, bpp, x, y, width, height, xor))
331         return TRUE;
332
333     return _pixman_implementation_fill (
334         imp->delegate, bits, stride, bpp, x, y, width, height, xor);
335 }
336
337 #define BIND_COMBINE_U(name)                                             \
338 void                                                                     \
339 pixman_composite_scanline_##name##_mask_asm_neon (int32_t         w,     \
340                                                   const uint32_t *dst,   \
341                                                   const uint32_t *src,   \
342                                                   const uint32_t *mask); \
343                                                                          \
344 void                                                                     \
345 pixman_composite_scanline_##name##_asm_neon (int32_t         w,          \
346                                              const uint32_t *dst,        \
347                                              const uint32_t *src);       \
348                                                                          \
349 static void                                                              \
350 neon_combine_##name##_u (pixman_implementation_t *imp,                   \
351                          pixman_op_t              op,                    \
352                          uint32_t *               dest,                  \
353                          const uint32_t *         src,                   \
354                          const uint32_t *         mask,                  \
355                          int                      width)                 \
356 {                                                                        \
357     if (mask)                                                            \
358         pixman_composite_scanline_##name##_mask_asm_neon (width, dest,   \
359                                                           src, mask);    \
360     else                                                                 \
361         pixman_composite_scanline_##name##_asm_neon (width, dest, src);  \
362 }
363
364 BIND_COMBINE_U (over)
365 BIND_COMBINE_U (add)
366 BIND_COMBINE_U (out_reverse)
367
368 pixman_implementation_t *
369 _pixman_implementation_create_arm_neon (void)
370 {
371 #ifdef USE_ARM_SIMD
372     pixman_implementation_t *fallback = _pixman_implementation_create_arm_simd ();
373 #else
374     pixman_implementation_t *fallback = _pixman_implementation_create_fast_path ();
375 #endif
376     pixman_implementation_t *imp =
377         _pixman_implementation_create (fallback, arm_neon_fast_paths);
378
379     imp->combine_32[PIXMAN_OP_OVER] = neon_combine_over_u;
380     imp->combine_32[PIXMAN_OP_ADD] = neon_combine_add_u;
381     imp->combine_32[PIXMAN_OP_OUT_REVERSE] = neon_combine_out_reverse_u;
382
383     imp->blt = arm_neon_blt;
384     imp->fill = arm_neon_fill;
385
386     return imp;
387 }