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