Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(171)

Side by Side Diff: source/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.c

Issue 812033011: libvpx: Pull from upstream (Closed) Base URL: svn://svn.chromium.org/chrome/trunk/deps/third_party/libvpx/
Patch Set: Created 5 years, 11 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch | Annotate | Revision Log
OLDNEW
(Empty)
1 /*
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <stddef.h>
12 #include <arm_neon.h>
13
14 #include "./vpx_config.h"
15 #include "vpx_ports/mem.h"
16
17 void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
18 uint8_t *dst, ptrdiff_t dst_stride,
19 const int16_t *filter_x, int x_step_q4,
20 const int16_t *filter_y, int y_step_q4,
21 int w, int h);
22 void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
23 uint8_t *dst, ptrdiff_t dst_stride,
24 const int16_t *filter_x, int x_step_q4,
25 const int16_t *filter_y, int y_step_q4,
26 int w, int h);
27
28 static INLINE int32x4_t MULTIPLY_BY_Q0(
29 int16x4_t dsrc0,
30 int16x4_t dsrc1,
31 int16x4_t dsrc2,
32 int16x4_t dsrc3,
33 int16x4_t dsrc4,
34 int16x4_t dsrc5,
35 int16x4_t dsrc6,
36 int16x4_t dsrc7,
37 int16x8_t q0s16) {
38 int32x4_t qdst;
39 int16x4_t d0s16, d1s16;
40
41 d0s16 = vget_low_s16(q0s16);
42 d1s16 = vget_high_s16(q0s16);
43
44 qdst = vmull_lane_s16(dsrc0, d0s16, 0);
45 qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
46 qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
47 qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
48 qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
49 qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
50 qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
51 qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
52 return qdst;
53 }
54
55 void vp9_convolve8_horiz_neon(
56 uint8_t *src,
57 ptrdiff_t src_stride,
58 uint8_t *dst,
59 ptrdiff_t dst_stride,
60 const int16_t *filter_x,
61 int x_step_q4,
62 const int16_t *filter_y, // unused
63 int y_step_q4, // unused
64 int w,
65 int h) {
66 int width;
67 uint8_t *s, *d, *psrc, *pdst;
68 uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
69 uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
70 uint8x16_t q12u8, q13u8, q14u8, q15u8;
71 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
72 int16x4_t d24s16, d25s16, d26s16, d27s16;
73 uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
74 int16x8_t q0s16;
75 uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
76 int32x4_t q1s32, q2s32, q14s32, q15s32;
77 uint16x8x2_t q0x2u16;
78 uint8x8x2_t d0x2u8, d1x2u8;
79 uint32x2x2_t d0x2u32;
80 uint16x4x2_t d0x2u16, d1x2u16;
81 uint32x4x2_t q0x2u32;
82
83 if (x_step_q4 != 16) {
84 vp9_convolve8_horiz_c(src, src_stride, dst, dst_stride,
85 filter_x, x_step_q4,
86 filter_y, y_step_q4, w, h);
87 return;
88 }
89
90 q0s16 = vld1q_s16(filter_x);
91
92 src -= 3; // adjust for taps
93 for (; h > 0; h -= 4,
94 src += src_stride * 4,
95 dst += dst_stride * 4) { // loop_horiz_v
96 s = src;
97 d24u8 = vld1_u8(s);
98 s += src_stride;
99 d25u8 = vld1_u8(s);
100 s += src_stride;
101 d26u8 = vld1_u8(s);
102 s += src_stride;
103 d27u8 = vld1_u8(s);
104
105 q12u8 = vcombine_u8(d24u8, d25u8);
106 q13u8 = vcombine_u8(d26u8, d27u8);
107
108 q0x2u16 = vtrnq_u16(vreinterpretq_u16_u8(q12u8),
109 vreinterpretq_u16_u8(q13u8));
110 d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
111 d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
112 d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
113 d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
114 d0x2u8 = vtrn_u8(d24u8, d25u8);
115 d1x2u8 = vtrn_u8(d26u8, d27u8);
116
117 __builtin_prefetch(src + src_stride * 4);
118 __builtin_prefetch(src + src_stride * 5);
119 __builtin_prefetch(src + src_stride * 6);
120
121 q8u16 = vmovl_u8(d0x2u8.val[0]);
122 q9u16 = vmovl_u8(d0x2u8.val[1]);
123 q10u16 = vmovl_u8(d1x2u8.val[0]);
124 q11u16 = vmovl_u8(d1x2u8.val[1]);
125
126 d16u16 = vget_low_u16(q8u16);
127 d17u16 = vget_high_u16(q8u16);
128 d18u16 = vget_low_u16(q9u16);
129 d19u16 = vget_high_u16(q9u16);
130 q8u16 = vcombine_u16(d16u16, d18u16); // vswp 17 18
131 q9u16 = vcombine_u16(d17u16, d19u16);
132
133 d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
134 d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16)); // vmov 23 21
135 for (width = w, psrc = src + 7, pdst = dst;
136 width > 0;
137 width -= 4, psrc += 4, pdst += 4) { // loop_horiz
138 s = psrc;
139 d28u32 = vld1_dup_u32((const uint32_t *)s);
140 s += src_stride;
141 d29u32 = vld1_dup_u32((const uint32_t *)s);
142 s += src_stride;
143 d31u32 = vld1_dup_u32((const uint32_t *)s);
144 s += src_stride;
145 d30u32 = vld1_dup_u32((const uint32_t *)s);
146
147 __builtin_prefetch(psrc + 64);
148
149 d0x2u16 = vtrn_u16(vreinterpret_u16_u32(d28u32),
150 vreinterpret_u16_u32(d31u32));
151 d1x2u16 = vtrn_u16(vreinterpret_u16_u32(d29u32),
152 vreinterpret_u16_u32(d30u32));
153 d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]), // d28
154 vreinterpret_u8_u16(d1x2u16.val[0])); // d29
155 d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]), // d31
156 vreinterpret_u8_u16(d1x2u16.val[1])); // d30
157
158 __builtin_prefetch(psrc + 64 + src_stride);
159
160 q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
161 q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
162 q0x2u32 = vtrnq_u32(vreinterpretq_u32_u8(q14u8),
163 vreinterpretq_u32_u8(q15u8));
164
165 d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
166 d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
167 q12u16 = vmovl_u8(d28u8);
168 q13u16 = vmovl_u8(d29u8);
169
170 __builtin_prefetch(psrc + 64 + src_stride * 2);
171
172 d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
173 d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
174 d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
175 d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
176 d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
177 d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
178 d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
179 d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
180 d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
181
182 q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16,
183 d18s16, d19s16, d23s16, d24s16, q0s16);
184 q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16,
185 d19s16, d23s16, d24s16, d26s16, q0s16);
186 q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16,
187 d23s16, d24s16, d26s16, d27s16, q0s16);
188 q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16,
189 d24s16, d26s16, d27s16, d25s16, q0s16);
190
191 __builtin_prefetch(psrc + 60 + src_stride * 3);
192
193 d2u16 = vqrshrun_n_s32(q1s32, 7);
194 d3u16 = vqrshrun_n_s32(q2s32, 7);
195 d4u16 = vqrshrun_n_s32(q14s32, 7);
196 d5u16 = vqrshrun_n_s32(q15s32, 7);
197
198 q1u16 = vcombine_u16(d2u16, d3u16);
199 q2u16 = vcombine_u16(d4u16, d5u16);
200
201 d2u8 = vqmovn_u16(q1u16);
202 d3u8 = vqmovn_u16(q2u16);
203
204 d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8),
205 vreinterpret_u16_u8(d3u8));
206 d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
207 vreinterpret_u32_u16(d0x2u16.val[1]));
208 d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
209 vreinterpret_u8_u32(d0x2u32.val[1]));
210
211 d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
212 d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
213
214 d = pdst;
215 vst1_lane_u32((uint32_t *)d, d2u32, 0);
216 d += dst_stride;
217 vst1_lane_u32((uint32_t *)d, d3u32, 0);
218 d += dst_stride;
219 vst1_lane_u32((uint32_t *)d, d2u32, 1);
220 d += dst_stride;
221 vst1_lane_u32((uint32_t *)d, d3u32, 1);
222
223 q8u16 = q9u16;
224 d20s16 = d23s16;
225 q11u16 = q12u16;
226 q9u16 = q13u16;
227 d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
228 }
229 }
230 return;
231 }
232
233 void vp9_convolve8_vert_neon(
234 uint8_t *src,
235 ptrdiff_t src_stride,
236 uint8_t *dst,
237 ptrdiff_t dst_stride,
238 const int16_t *filter_x, // unused
239 int x_step_q4, // unused
240 const int16_t *filter_y,
241 int y_step_q4,
242 int w,
243 int h) {
244 int height;
245 uint8_t *s, *d;
246 uint32x2_t d2u32, d3u32;
247 uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
248 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
249 int16x4_t d24s16, d25s16, d26s16, d27s16;
250 uint16x4_t d2u16, d3u16, d4u16, d5u16;
251 int16x8_t q0s16;
252 uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
253 int32x4_t q1s32, q2s32, q14s32, q15s32;
254
255 if (y_step_q4 != 16) {
256 vp9_convolve8_vert_c(src, src_stride, dst, dst_stride,
257 filter_x, x_step_q4,
258 filter_y, y_step_q4, w, h);
259 return;
260 }
261
262 src -= src_stride * 3;
263 q0s16 = vld1q_s16(filter_y);
264 for (; w > 0; w -= 4, src += 4, dst += 4) { // loop_vert_h
265 s = src;
266 d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
267 s += src_stride;
268 d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
269 s += src_stride;
270 d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
271 s += src_stride;
272 d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
273 s += src_stride;
274 d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
275 s += src_stride;
276 d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
277 s += src_stride;
278 d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
279 s += src_stride;
280
281 q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
282 q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
283 q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
284 q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
285
286 d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
287 d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
288 d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
289 d = dst;
290 for (height = h; height > 0; height -= 4) { // loop_vert
291 d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
292 s += src_stride;
293 d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
294 s += src_stride;
295 d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
296 s += src_stride;
297 d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
298 s += src_stride;
299
300 q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
301 q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
302
303 d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
304 d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
305 d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
306 d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
307 d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
308 d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
309 d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
310 d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
311
312 __builtin_prefetch(d);
313 __builtin_prefetch(d + dst_stride);
314 q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16,
315 d20s16, d21s16, d22s16, d24s16, q0s16);
316 __builtin_prefetch(d + dst_stride * 2);
317 __builtin_prefetch(d + dst_stride * 3);
318 q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16,
319 d21s16, d22s16, d24s16, d26s16, q0s16);
320 __builtin_prefetch(s);
321 __builtin_prefetch(s + src_stride);
322 q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16,
323 d22s16, d24s16, d26s16, d27s16, q0s16);
324 __builtin_prefetch(s + src_stride * 2);
325 __builtin_prefetch(s + src_stride * 3);
326 q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16,
327 d24s16, d26s16, d27s16, d25s16, q0s16);
328
329 d2u16 = vqrshrun_n_s32(q1s32, 7);
330 d3u16 = vqrshrun_n_s32(q2s32, 7);
331 d4u16 = vqrshrun_n_s32(q14s32, 7);
332 d5u16 = vqrshrun_n_s32(q15s32, 7);
333
334 q1u16 = vcombine_u16(d2u16, d3u16);
335 q2u16 = vcombine_u16(d4u16, d5u16);
336
337 d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
338 d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
339
340 vst1_lane_u32((uint32_t *)d, d2u32, 0);
341 d += dst_stride;
342 vst1_lane_u32((uint32_t *)d, d2u32, 1);
343 d += dst_stride;
344 vst1_lane_u32((uint32_t *)d, d3u32, 0);
345 d += dst_stride;
346 vst1_lane_u32((uint32_t *)d, d3u32, 1);
347 d += dst_stride;
348
349 q8u16 = q10u16;
350 d18s16 = d22s16;
351 d19s16 = d24s16;
352 q10u16 = q13u16;
353 d22s16 = d25s16;
354 }
355 }
356 return;
357 }
OLDNEW
« no previous file with comments | « source/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.asm ('k') | source/libvpx/vp9/common/arm/neon/vp9_convolve8_neon_asm.asm » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698