2 * AltiVec-enhanced yuv2yuvX
4 * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
5 * based on the equivalent C code in swscale.c
7 * This file is part of FFmpeg.
9 * FFmpeg 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.
14 * FFmpeg 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.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 #define vzero vec_splat_s32(0)
27 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW)
30 vector unsigned int altivec_vectorShiftInt19 =
31 vec_add(vec_splat_u32(10), vec_splat_u32(9));
32 if ((unsigned long)dest % 16) {
33 /* badly aligned store, we force store alignment */
34 /* and will handle load misalignment on val w/ vec_perm */
35 vector unsigned char perm1;
37 for (i = 0 ; (i < dstW) &&
38 (((unsigned long)dest + i) % 16) ; i++) {
40 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
42 perm1 = vec_lvsl(i << 2, val);
43 v1 = vec_ld(i << 2, val);
44 for ( ; i < (dstW - 15); i+=16) {
46 vector signed int v2 = vec_ld(offset + 16, val);
47 vector signed int v3 = vec_ld(offset + 32, val);
48 vector signed int v4 = vec_ld(offset + 48, val);
49 vector signed int v5 = vec_ld(offset + 64, val);
50 vector signed int v12 = vec_perm(v1, v2, perm1);
51 vector signed int v23 = vec_perm(v2, v3, perm1);
52 vector signed int v34 = vec_perm(v3, v4, perm1);
53 vector signed int v45 = vec_perm(v4, v5, perm1);
55 vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
56 vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
57 vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
58 vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
59 vector unsigned short vs1 = vec_packsu(vA, vB);
60 vector unsigned short vs2 = vec_packsu(vC, vD);
61 vector unsigned char vf = vec_packsu(vs1, vs2);
65 } else { // dest is properly aligned, great
66 for (i = 0; i < (dstW - 15); i+=16) {
68 vector signed int v1 = vec_ld(offset, val);
69 vector signed int v2 = vec_ld(offset + 16, val);
70 vector signed int v3 = vec_ld(offset + 32, val);
71 vector signed int v4 = vec_ld(offset + 48, val);
72 vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
73 vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
74 vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
75 vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
76 vector unsigned short vs1 = vec_packsu(v5, v6);
77 vector unsigned short vs2 = vec_packsu(v7, v8);
78 vector unsigned char vf = vec_packsu(vs1, vs2);
82 for ( ; i < dstW ; i++) {
84 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
89 yuv2yuvX_altivec_real(const int16_t *lumFilter, const int16_t **lumSrc, int lumFilterSize,
90 const int16_t *chrFilter, const int16_t **chrSrc, int chrFilterSize,
91 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW)
93 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
96 DECLARE_ALIGNED(16, int, val)[dstW];
98 for (i = 0; i < (dstW -7); i+=4) {
99 vec_st(vini, i << 2, val);
101 for (; i < dstW; i++) {
105 for (j = 0; j < lumFilterSize; j++) {
106 vector signed short l1, vLumFilter = vec_ld(j << 1, lumFilter);
107 vector unsigned char perm, perm0 = vec_lvsl(j << 1, lumFilter);
108 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
109 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
111 perm = vec_lvsl(0, lumSrc[j]);
112 l1 = vec_ld(0, lumSrc[j]);
114 for (i = 0; i < (dstW - 7); i+=8) {
116 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
118 vector signed int v1 = vec_ld(offset, val);
119 vector signed int v2 = vec_ld(offset + 16, val);
121 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
123 vector signed int i1 = vec_mule(vLumFilter, ls);
124 vector signed int i2 = vec_mulo(vLumFilter, ls);
126 vector signed int vf1 = vec_mergeh(i1, i2);
127 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
129 vector signed int vo1 = vec_add(v1, vf1);
130 vector signed int vo2 = vec_add(v2, vf2);
132 vec_st(vo1, offset, val);
133 vec_st(vo2, offset + 16, val);
137 for ( ; i < dstW; i++) {
138 val[i] += lumSrc[j][i] * lumFilter[j];
141 altivec_packIntArrayToCharArray(val, dest, dstW);
144 DECLARE_ALIGNED(16, int, u)[chrDstW];
145 DECLARE_ALIGNED(16, int, v)[chrDstW];
147 for (i = 0; i < (chrDstW -7); i+=4) {
148 vec_st(vini, i << 2, u);
149 vec_st(vini, i << 2, v);
151 for (; i < chrDstW; i++) {
156 for (j = 0; j < chrFilterSize; j++) {
157 vector signed short l1, l1_V, vChrFilter = vec_ld(j << 1, chrFilter);
158 vector unsigned char perm, perm0 = vec_lvsl(j << 1, chrFilter);
159 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
160 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
162 perm = vec_lvsl(0, chrSrc[j]);
163 l1 = vec_ld(0, chrSrc[j]);
164 l1_V = vec_ld(2048 << 1, chrSrc[j]);
166 for (i = 0; i < (chrDstW - 7); i+=8) {
168 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]);
169 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]);
171 vector signed int v1 = vec_ld(offset, u);
172 vector signed int v2 = vec_ld(offset + 16, u);
173 vector signed int v1_V = vec_ld(offset, v);
174 vector signed int v2_V = vec_ld(offset + 16, v);
176 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7]
177 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055]
179 vector signed int i1 = vec_mule(vChrFilter, ls);
180 vector signed int i2 = vec_mulo(vChrFilter, ls);
181 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
182 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
184 vector signed int vf1 = vec_mergeh(i1, i2);
185 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
186 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
187 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
189 vector signed int vo1 = vec_add(v1, vf1);
190 vector signed int vo2 = vec_add(v2, vf2);
191 vector signed int vo1_V = vec_add(v1_V, vf1_V);
192 vector signed int vo2_V = vec_add(v2_V, vf2_V);
194 vec_st(vo1, offset, u);
195 vec_st(vo2, offset + 16, u);
196 vec_st(vo1_V, offset, v);
197 vec_st(vo2_V, offset + 16, v);
202 for ( ; i < chrDstW; i++) {
203 u[i] += chrSrc[j][i] * chrFilter[j];
204 v[i] += chrSrc[j][i + 2048] * chrFilter[j];
207 altivec_packIntArrayToCharArray(u, uDest, chrDstW);
208 altivec_packIntArrayToCharArray(v, vDest, chrDstW);
212 static inline void hScale_altivec_real(int16_t *dst, int dstW,
213 const uint8_t *src, int srcW,
214 int xInc, const int16_t *filter,
215 const int16_t *filterPos, int filterSize)
218 DECLARE_ALIGNED(16, int, tempo)[4];
220 if (filterSize % 4) {
221 for (i=0; i<dstW; i++) {
223 register int srcPos = filterPos[i];
224 register int val = 0;
225 for (j=0; j<filterSize; j++) {
226 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
228 dst[i] = FFMIN(val>>7, (1<<15)-1);
232 switch (filterSize) {
235 for (i=0; i<dstW; i++) {
236 register int srcPos = filterPos[i];
238 vector unsigned char src_v0 = vec_ld(srcPos, src);
239 vector unsigned char src_v1, src_vF;
240 vector signed short src_v, filter_v;
241 vector signed int val_vEven, val_s;
242 if ((((int)src + srcPos)% 16) > 12) {
243 src_v1 = vec_ld(srcPos + 16, src);
245 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
247 src_v = // vec_unpackh sign-extends...
248 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
249 // now put our elements in the even slots
250 src_v = vec_mergeh(src_v, (vector signed short)vzero);
252 filter_v = vec_ld(i << 3, filter);
253 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
255 // The neat trick: We only care for half the elements,
256 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
257 // and we're going to use vec_mule, so we choose
258 // carefully how to "unpack" the elements into the even slots.
260 filter_v = vec_mergel(filter_v, (vector signed short)vzero);
262 filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
264 val_vEven = vec_mule(src_v, filter_v);
265 val_s = vec_sums(val_vEven, vzero);
266 vec_st(val_s, 0, tempo);
267 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
274 for (i=0; i<dstW; i++) {
275 register int srcPos = filterPos[i];
277 vector unsigned char src_v0 = vec_ld(srcPos, src);
278 vector unsigned char src_v1, src_vF;
279 vector signed short src_v, filter_v;
280 vector signed int val_v, val_s;
281 if ((((int)src + srcPos)% 16) > 8) {
282 src_v1 = vec_ld(srcPos + 16, src);
284 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
286 src_v = // vec_unpackh sign-extends...
287 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
288 filter_v = vec_ld(i << 4, filter);
289 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
291 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
292 val_s = vec_sums(val_v, vzero);
293 vec_st(val_s, 0, tempo);
294 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
301 for (i=0; i<dstW; i++) {
302 register int srcPos = filterPos[i];
304 vector unsigned char src_v0 = vec_ld(srcPos, src);
305 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
306 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
308 vector signed short src_vA = // vec_unpackh sign-extends...
309 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
310 vector signed short src_vB = // vec_unpackh sign-extends...
311 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
313 vector signed short filter_v0 = vec_ld(i << 5, filter);
314 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
315 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
317 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
318 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
320 vector signed int val_s = vec_sums(val_v, vzero);
322 vec_st(val_s, 0, tempo);
323 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
330 for (i=0; i<dstW; i++) {
332 register int srcPos = filterPos[i];
334 vector signed int val_s, val_v = (vector signed int)vzero;
335 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
336 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
338 vector unsigned char src_v0 = vec_ld(srcPos, src);
339 vector unsigned char permS = vec_lvsl(srcPos, src);
341 for (j = 0 ; j < filterSize - 15; j += 16) {
342 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
343 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
345 vector signed short src_vA = // vec_unpackh sign-extends...
346 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
347 vector signed short src_vB = // vec_unpackh sign-extends...
348 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
350 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
351 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
352 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
353 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
355 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
356 val_v = vec_msums(src_vB, filter_v1, val_acc);
358 filter_v0R = filter_v2R;
362 if (j < filterSize-7) {
363 // loading src_v0 is useless, it's already done above
364 //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
365 vector unsigned char src_v1, src_vF;
366 vector signed short src_v, filter_v1R, filter_v;
367 if ((((int)src + srcPos)% 16) > 8) {
368 src_v1 = vec_ld(srcPos + j + 16, src);
370 src_vF = vec_perm(src_v0, src_v1, permS);
372 src_v = // vec_unpackh sign-extends...
373 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
374 // loading filter_v0R is useless, it's already done above
375 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
376 filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
377 filter_v = vec_perm(filter_v0R, filter_v1R, permF);
379 val_v = vec_msums(src_v, filter_v, val_v);
382 val_s = vec_sums(val_v, vzero);
384 vec_st(val_s, 0, tempo);
385 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
392 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, const uint8_t* src[], int srcStride[], int srcSliceY,
393 int srcSliceH, uint8_t* dstParam[], int dstStride_a[])
395 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
396 // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]);
397 const uint8_t *ysrc = src[0];
398 const uint8_t *usrc = src[1];
399 const uint8_t *vsrc = src[2];
400 const int width = c->srcW;
401 const int height = srcSliceH;
402 const int lumStride = srcStride[0];
403 const int chromStride = srcStride[1];
404 const int dstStride = dstStride_a[0];
405 const vector unsigned char yperm = vec_lvsl(0, ysrc);
406 const int vertLumPerChroma = 2;
407 register unsigned int y;
410 yv12toyuy2(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride);
414 /* This code assumes:
416 1) dst is 16 bytes-aligned
417 2) dstStride is a multiple of 16
418 3) width is a multiple of 16
419 4) lum & chrom stride are multiples of 8
422 for (y=0; y<height; y++) {
424 for (i = 0; i < width - 31; i+= 32) {
425 const unsigned int j = i >> 1;
426 vector unsigned char v_yA = vec_ld(i, ysrc);
427 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
428 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
429 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
430 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
431 vector unsigned char v_uA = vec_ld(j, usrc);
432 vector unsigned char v_uB = vec_ld(j + 16, usrc);
433 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
434 vector unsigned char v_vA = vec_ld(j, vsrc);
435 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
436 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
437 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
438 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
439 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
440 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
441 vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b);
442 vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b);
443 vec_st(v_yuy2_0, (i << 1), dst);
444 vec_st(v_yuy2_1, (i << 1) + 16, dst);
445 vec_st(v_yuy2_2, (i << 1) + 32, dst);
446 vec_st(v_yuy2_3, (i << 1) + 48, dst);
449 const unsigned int j = i >> 1;
450 vector unsigned char v_y1 = vec_ld(i, ysrc);
451 vector unsigned char v_u = vec_ld(j, usrc);
452 vector unsigned char v_v = vec_ld(j, vsrc);
453 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
454 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
455 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
456 vec_st(v_yuy2_0, (i << 1), dst);
457 vec_st(v_yuy2_1, (i << 1) + 16, dst);
459 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
470 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, const uint8_t* src[], int srcStride[], int srcSliceY,
471 int srcSliceH, uint8_t* dstParam[], int dstStride_a[])
473 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
474 // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]);
475 const uint8_t *ysrc = src[0];
476 const uint8_t *usrc = src[1];
477 const uint8_t *vsrc = src[2];
478 const int width = c->srcW;
479 const int height = srcSliceH;
480 const int lumStride = srcStride[0];
481 const int chromStride = srcStride[1];
482 const int dstStride = dstStride_a[0];
483 const int vertLumPerChroma = 2;
484 const vector unsigned char yperm = vec_lvsl(0, ysrc);
485 register unsigned int y;
488 yv12touyvy(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride);
492 /* This code assumes:
494 1) dst is 16 bytes-aligned
495 2) dstStride is a multiple of 16
496 3) width is a multiple of 16
497 4) lum & chrom stride are multiples of 8
500 for (y=0; y<height; y++) {
502 for (i = 0; i < width - 31; i+= 32) {
503 const unsigned int j = i >> 1;
504 vector unsigned char v_yA = vec_ld(i, ysrc);
505 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
506 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
507 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
508 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
509 vector unsigned char v_uA = vec_ld(j, usrc);
510 vector unsigned char v_uB = vec_ld(j + 16, usrc);
511 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
512 vector unsigned char v_vA = vec_ld(j, vsrc);
513 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
514 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
515 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
516 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
517 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
518 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
519 vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2);
520 vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2);
521 vec_st(v_uyvy_0, (i << 1), dst);
522 vec_st(v_uyvy_1, (i << 1) + 16, dst);
523 vec_st(v_uyvy_2, (i << 1) + 32, dst);
524 vec_st(v_uyvy_3, (i << 1) + 48, dst);
527 const unsigned int j = i >> 1;
528 vector unsigned char v_y1 = vec_ld(i, ysrc);
529 vector unsigned char v_u = vec_ld(j, usrc);
530 vector unsigned char v_v = vec_ld(j, vsrc);
531 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
532 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
533 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
534 vec_st(v_uyvy_0, (i << 1), dst);
535 vec_st(v_uyvy_1, (i << 1) + 16, dst);
537 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {