swscale_altivec.c
Go to the documentation of this file.
1 /*
2  * AltiVec-enhanced yuv2yuvX
3  *
4  * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
5  * based on the equivalent C code in swscale.c
6  *
7  * This file is part of Libav.
8  *
9  * Libav is free software; you can redistribute it and/or
10  * modify it under the terms of the GNU Lesser General Public
11  * License as published by the Free Software Foundation; either
12  * version 2.1 of the License, or (at your option) any later version.
13  *
14  * Libav is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17  * Lesser General Public License for more details.
18  *
19  * You should have received a copy of the GNU Lesser General Public
20  * License along with Libav; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22  */
23 
24 #include <inttypes.h>
25 #include "config.h"
26 #include "libswscale/swscale.h"
28 #include "libavutil/cpu.h"
29 #include "yuv2rgb_altivec.h"
30 
31 #define vzero vec_splat_s32(0)
32 
33 static inline void
34 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW)
35 {
36  register int i;
37  vector unsigned int altivec_vectorShiftInt19 =
38  vec_add(vec_splat_u32(10), vec_splat_u32(9));
39  if ((uintptr_t)dest % 16) {
40  /* badly aligned store, we force store alignment */
41  /* and will handle load misalignment on val w/ vec_perm */
42  vector unsigned char perm1;
43  vector signed int v1;
44  for (i = 0 ; (i < dstW) &&
45  (((uintptr_t)dest + i) % 16) ; i++) {
46  int t = val[i] >> 19;
47  dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
48  }
49  perm1 = vec_lvsl(i << 2, val);
50  v1 = vec_ld(i << 2, val);
51  for ( ; i < (dstW - 15); i+=16) {
52  int offset = i << 2;
53  vector signed int v2 = vec_ld(offset + 16, val);
54  vector signed int v3 = vec_ld(offset + 32, val);
55  vector signed int v4 = vec_ld(offset + 48, val);
56  vector signed int v5 = vec_ld(offset + 64, val);
57  vector signed int v12 = vec_perm(v1, v2, perm1);
58  vector signed int v23 = vec_perm(v2, v3, perm1);
59  vector signed int v34 = vec_perm(v3, v4, perm1);
60  vector signed int v45 = vec_perm(v4, v5, perm1);
61 
62  vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
63  vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
64  vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
65  vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
66  vector unsigned short vs1 = vec_packsu(vA, vB);
67  vector unsigned short vs2 = vec_packsu(vC, vD);
68  vector unsigned char vf = vec_packsu(vs1, vs2);
69  vec_st(vf, i, dest);
70  v1 = v5;
71  }
72  } else { // dest is properly aligned, great
73  for (i = 0; i < (dstW - 15); i+=16) {
74  int offset = i << 2;
75  vector signed int v1 = vec_ld(offset, val);
76  vector signed int v2 = vec_ld(offset + 16, val);
77  vector signed int v3 = vec_ld(offset + 32, val);
78  vector signed int v4 = vec_ld(offset + 48, val);
79  vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
80  vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
81  vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
82  vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
83  vector unsigned short vs1 = vec_packsu(v5, v6);
84  vector unsigned short vs2 = vec_packsu(v7, v8);
85  vector unsigned char vf = vec_packsu(vs1, vs2);
86  vec_st(vf, i, dest);
87  }
88  }
89  for ( ; i < dstW ; i++) {
90  int t = val[i] >> 19;
91  dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
92  }
93 }
94 
95 //FIXME remove the usage of scratch buffers.
96 static void
97 yuv2planeX_altivec(const int16_t *filter, int filterSize,
98  const int16_t **src, uint8_t *dest, int dstW,
99  const uint8_t *dither, int offset)
100 {
101  register int i, j;
102  {
103  DECLARE_ALIGNED(16, int, val)[dstW];
104 
105  for (i=0; i<dstW; i++)
106  val[i] = dither[(i + offset) & 7] << 12;
107 
108  for (j = 0; j < filterSize; j++) {
109  vector signed short l1, vLumFilter = vec_ld(j << 1, filter);
110  vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter);
111  vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
112  vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
113 
114  perm = vec_lvsl(0, src[j]);
115  l1 = vec_ld(0, src[j]);
116 
117  for (i = 0; i < (dstW - 7); i+=8) {
118  int offset = i << 2;
119  vector signed short l2 = vec_ld((i << 1) + 16, src[j]);
120 
121  vector signed int v1 = vec_ld(offset, val);
122  vector signed int v2 = vec_ld(offset + 16, val);
123 
124  vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
125 
126  vector signed int i1 = vec_mule(vLumFilter, ls);
127  vector signed int i2 = vec_mulo(vLumFilter, ls);
128 
129  vector signed int vf1 = vec_mergeh(i1, i2);
130  vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
131 
132  vector signed int vo1 = vec_add(v1, vf1);
133  vector signed int vo2 = vec_add(v2, vf2);
134 
135  vec_st(vo1, offset, val);
136  vec_st(vo2, offset + 16, val);
137 
138  l1 = l2;
139  }
140  for ( ; i < dstW; i++) {
141  val[i] += src[j][i] * filter[j];
142  }
143  }
144  altivec_packIntArrayToCharArray(val, dest, dstW);
145  }
146 }
147 
148 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
149  const uint8_t *src, const int16_t *filter,
150  const int32_t *filterPos, int filterSize)
151 {
152  register int i;
153  DECLARE_ALIGNED(16, int, tempo)[4];
154 
155  if (filterSize % 4) {
156  for (i=0; i<dstW; i++) {
157  register int j;
158  register int srcPos = filterPos[i];
159  register int val = 0;
160  for (j=0; j<filterSize; j++) {
161  val += ((int)src[srcPos + j])*filter[filterSize*i + j];
162  }
163  dst[i] = FFMIN(val>>7, (1<<15)-1);
164  }
165  }
166  else
167  switch (filterSize) {
168  case 4:
169  {
170  for (i=0; i<dstW; i++) {
171  register int srcPos = filterPos[i];
172 
173  vector unsigned char src_v0 = vec_ld(srcPos, src);
174  vector unsigned char src_v1, src_vF;
175  vector signed short src_v, filter_v;
176  vector signed int val_vEven, val_s;
177  if ((((uintptr_t)src + srcPos) % 16) > 12) {
178  src_v1 = vec_ld(srcPos + 16, src);
179  }
180  src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
181 
182  src_v = // vec_unpackh sign-extends...
183  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
184  // now put our elements in the even slots
185  src_v = vec_mergeh(src_v, (vector signed short)vzero);
186 
187  filter_v = vec_ld(i << 3, filter);
188  // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
189 
190  // The neat trick: We only care for half the elements,
191  // high or low depending on (i<<3)%16 (it's 0 or 8 here),
192  // and we're going to use vec_mule, so we choose
193  // carefully how to "unpack" the elements into the even slots.
194  if ((i << 3) % 16)
195  filter_v = vec_mergel(filter_v, (vector signed short)vzero);
196  else
197  filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
198 
199  val_vEven = vec_mule(src_v, filter_v);
200  val_s = vec_sums(val_vEven, vzero);
201  vec_st(val_s, 0, tempo);
202  dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
203  }
204  }
205  break;
206 
207  case 8:
208  {
209  for (i=0; i<dstW; i++) {
210  register int srcPos = filterPos[i];
211 
212  vector unsigned char src_v0 = vec_ld(srcPos, src);
213  vector unsigned char src_v1, src_vF;
214  vector signed short src_v, filter_v;
215  vector signed int val_v, val_s;
216  if ((((uintptr_t)src + srcPos) % 16) > 8) {
217  src_v1 = vec_ld(srcPos + 16, src);
218  }
219  src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
220 
221  src_v = // vec_unpackh sign-extends...
222  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
223  filter_v = vec_ld(i << 4, filter);
224  // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
225 
226  val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
227  val_s = vec_sums(val_v, vzero);
228  vec_st(val_s, 0, tempo);
229  dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
230  }
231  }
232  break;
233 
234  case 16:
235  {
236  for (i=0; i<dstW; i++) {
237  register int srcPos = filterPos[i];
238 
239  vector unsigned char src_v0 = vec_ld(srcPos, src);
240  vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
241  vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
242 
243  vector signed short src_vA = // vec_unpackh sign-extends...
244  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
245  vector signed short src_vB = // vec_unpackh sign-extends...
246  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
247 
248  vector signed short filter_v0 = vec_ld(i << 5, filter);
249  vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
250  // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
251 
252  vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
253  vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
254 
255  vector signed int val_s = vec_sums(val_v, vzero);
256 
257  vec_st(val_s, 0, tempo);
258  dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
259  }
260  }
261  break;
262 
263  default:
264  {
265  for (i=0; i<dstW; i++) {
266  register int j;
267  register int srcPos = filterPos[i];
268 
269  vector signed int val_s, val_v = (vector signed int)vzero;
270  vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
271  vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
272 
273  vector unsigned char src_v0 = vec_ld(srcPos, src);
274  vector unsigned char permS = vec_lvsl(srcPos, src);
275 
276  for (j = 0 ; j < filterSize - 15; j += 16) {
277  vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
278  vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
279 
280  vector signed short src_vA = // vec_unpackh sign-extends...
281  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
282  vector signed short src_vB = // vec_unpackh sign-extends...
283  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
284 
285  vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
286  vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
287  vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
288  vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
289 
290  vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
291  val_v = vec_msums(src_vB, filter_v1, val_acc);
292 
293  filter_v0R = filter_v2R;
294  src_v0 = src_v1;
295  }
296 
297  if (j < filterSize-7) {
298  // loading src_v0 is useless, it's already done above
299  //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
300  vector unsigned char src_v1, src_vF;
301  vector signed short src_v, filter_v1R, filter_v;
302  if ((((uintptr_t)src + srcPos) % 16) > 8) {
303  src_v1 = vec_ld(srcPos + j + 16, src);
304  }
305  src_vF = vec_perm(src_v0, src_v1, permS);
306 
307  src_v = // vec_unpackh sign-extends...
308  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
309  // loading filter_v0R is useless, it's already done above
310  //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
311  filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
312  filter_v = vec_perm(filter_v0R, filter_v1R, permF);
313 
314  val_v = vec_msums(src_v, filter_v, val_v);
315  }
316 
317  val_s = vec_sums(val_v, vzero);
318 
319  vec_st(val_s, 0, tempo);
320  dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
321  }
322 
323  }
324  }
325 }
326 
328 {
329  enum PixelFormat dstFormat = c->dstFormat;
330 
332  return;
333 
334  if (c->srcBpc == 8 && c->dstBpc <= 10) {
336  }
337  if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
338  dstFormat != PIX_FMT_NV12 && dstFormat != PIX_FMT_NV21 &&
339  !c->alpPixBuf) {
341  }
342 
343  /* The following list of supported dstFormat values should
344  * match what's found in the body of ff_yuv2packedX_altivec() */
345  if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
346  switch (c->dstFormat) {
347  case PIX_FMT_ABGR: c->yuv2packedX = ff_yuv2abgr_X_altivec; break;
348  case PIX_FMT_BGRA: c->yuv2packedX = ff_yuv2bgra_X_altivec; break;
349  case PIX_FMT_ARGB: c->yuv2packedX = ff_yuv2argb_X_altivec; break;
350  case PIX_FMT_RGBA: c->yuv2packedX = ff_yuv2rgba_X_altivec; break;
351  case PIX_FMT_BGR24: c->yuv2packedX = ff_yuv2bgr24_X_altivec; break;
352  case PIX_FMT_RGB24: c->yuv2packedX = ff_yuv2rgb24_X_altivec; break;
353  }
354  }
355 }