2 AltiVec-enhanced yuv2yuvX
4 Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
5 based on the equivalent C code in "postproc/swscale.c"
8 This program is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 2 of the License, or
11 (at your option) any later version.
13 This program is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with this program; if not, write to the Free Software
20 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
29 static const vector signed int vzero =
30 (const vector signed int)AVV(0, 0, 0, 0);
31 static const vector unsigned int altivec_vectorShiftInt19 =
32 (const vector unsigned int)AVV(19, 19, 19, 19);
35 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) {
37 if ((unsigned long)dest % 16) {
38 /* badly aligned store, we force store alignement */
39 /* and will handle load misalignement on val w/ vec_perm */
40 for (i = 0 ; (i < dstW) &&
41 (((unsigned long)dest + i) % 16) ; i++) {
43 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
45 vector unsigned char perm1 = vec_lvsl(i << 2, val);
46 vector signed int v1 = vec_ld(i << 2, val);
47 for ( ; i < (dstW - 15); i+=16) {
49 vector signed int v2 = vec_ld(offset + 16, val);
50 vector signed int v3 = vec_ld(offset + 32, val);
51 vector signed int v4 = vec_ld(offset + 48, val);
52 vector signed int v5 = vec_ld(offset + 64, val);
53 vector signed int v12 = vec_perm(v1,v2,perm1);
54 vector signed int v23 = vec_perm(v2,v3,perm1);
55 vector signed int v34 = vec_perm(v3,v4,perm1);
56 vector signed int v45 = vec_perm(v4,v5,perm1);
58 vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
59 vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
60 vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
61 vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
62 vector unsigned short vs1 = vec_packsu(vA, vB);
63 vector unsigned short vs2 = vec_packsu(vC, vD);
64 vector unsigned char vf = vec_packsu(vs1, vs2);
68 } else { // dest is properly aligned, great
69 for (i = 0; i < (dstW - 15); i+=16) {
71 vector signed int v1 = vec_ld(offset, val);
72 vector signed int v2 = vec_ld(offset + 16, val);
73 vector signed int v3 = vec_ld(offset + 32, val);
74 vector signed int v4 = vec_ld(offset + 48, val);
75 vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
76 vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
77 vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
78 vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
79 vector unsigned short vs1 = vec_packsu(v5, v6);
80 vector unsigned short vs2 = vec_packsu(v7, v8);
81 vector unsigned char vf = vec_packsu(vs1, vs2);
85 for ( ; i < dstW ; i++) {
87 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
92 yuv2yuvX_altivec_real(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
93 int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
94 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW)
96 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
99 int __attribute__ ((aligned (16))) val[dstW];
101 for (i = 0; i < (dstW -7); i+=4) {
102 vec_st(vini, i << 2, val);
104 for (; i < dstW; i++) {
108 for (j = 0; j < lumFilterSize; j++) {
109 vector signed short vLumFilter = vec_ld(j << 1, lumFilter);
110 vector unsigned char perm0 = vec_lvsl(j << 1, lumFilter);
111 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
112 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
114 vector unsigned char perm = vec_lvsl(0, lumSrc[j]);
115 vector signed short l1 = vec_ld(0, lumSrc[j]);
117 for (i = 0; i < (dstW - 7); i+=8) {
119 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
121 vector signed int v1 = vec_ld(offset, val);
122 vector signed int v2 = vec_ld(offset + 16, val);
124 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
126 vector signed int i1 = vec_mule(vLumFilter, ls);
127 vector signed int i2 = vec_mulo(vLumFilter, ls);
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]
132 vector signed int vo1 = vec_add(v1, vf1);
133 vector signed int vo2 = vec_add(v2, vf2);
135 vec_st(vo1, offset, val);
136 vec_st(vo2, offset + 16, val);
140 for ( ; i < dstW; i++) {
141 val[i] += lumSrc[j][i] * lumFilter[j];
144 altivec_packIntArrayToCharArray(val,dest,dstW);
147 int __attribute__ ((aligned (16))) u[chrDstW];
148 int __attribute__ ((aligned (16))) v[chrDstW];
150 for (i = 0; i < (chrDstW -7); i+=4) {
151 vec_st(vini, i << 2, u);
152 vec_st(vini, i << 2, v);
154 for (; i < chrDstW; i++) {
159 for (j = 0; j < chrFilterSize; j++) {
160 vector signed short vChrFilter = vec_ld(j << 1, chrFilter);
161 vector unsigned char perm0 = vec_lvsl(j << 1, chrFilter);
162 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
163 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
165 vector unsigned char perm = vec_lvsl(0, chrSrc[j]);
166 vector signed short l1 = vec_ld(0, chrSrc[j]);
167 vector signed short l1_V = vec_ld(2048 << 1, chrSrc[j]);
169 for (i = 0; i < (chrDstW - 7); i+=8) {
171 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]);
172 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]);
174 vector signed int v1 = vec_ld(offset, u);
175 vector signed int v2 = vec_ld(offset + 16, u);
176 vector signed int v1_V = vec_ld(offset, v);
177 vector signed int v2_V = vec_ld(offset + 16, v);
179 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7]
180 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055]
182 vector signed int i1 = vec_mule(vChrFilter, ls);
183 vector signed int i2 = vec_mulo(vChrFilter, ls);
184 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
185 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
187 vector signed int vf1 = vec_mergeh(i1, i2);
188 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
189 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
190 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
192 vector signed int vo1 = vec_add(v1, vf1);
193 vector signed int vo2 = vec_add(v2, vf2);
194 vector signed int vo1_V = vec_add(v1_V, vf1_V);
195 vector signed int vo2_V = vec_add(v2_V, vf2_V);
197 vec_st(vo1, offset, u);
198 vec_st(vo2, offset + 16, u);
199 vec_st(vo1_V, offset, v);
200 vec_st(vo2_V, offset + 16, v);
205 for ( ; i < chrDstW; i++) {
206 u[i] += chrSrc[j][i] * chrFilter[j];
207 v[i] += chrSrc[j][i + 2048] * chrFilter[j];
210 altivec_packIntArrayToCharArray(u,uDest,chrDstW);
211 altivec_packIntArrayToCharArray(v,vDest,chrDstW);
215 static inline void hScale_altivec_real(int16_t *dst, int dstW, uint8_t *src, int srcW, int xInc, int16_t *filter, int16_t *filterPos, int filterSize) {
217 int __attribute__ ((aligned (16))) tempo[4];
219 if (filterSize % 4) {
220 for(i=0; i<dstW; i++) {
222 register int srcPos = filterPos[i];
223 register int val = 0;
224 for(j=0; j<filterSize; j++) {
225 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
227 dst[i] = MIN(MAX(0, val>>7), (1<<15)-1);
231 switch (filterSize) {
234 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;
240 if ((((int)src + srcPos)% 16) > 12) {
241 src_v1 = vec_ld(srcPos + 16, src);
243 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
245 vector signed short src_v = // vec_unpackh sign-extends...
246 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
247 // now put our elements in the even slots
248 src_v = vec_mergeh(src_v, (vector signed short)vzero);
250 vector signed short filter_v = vec_ld(i << 3, filter);
251 // the 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2)
253 // the neat trick : we only care for half the elements,
254 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
255 // and we're going to use vec_mule, so we chose
256 // carefully how to "unpack" the elements into the even slots
258 filter_v = vec_mergel(filter_v,(vector signed short)vzero);
260 filter_v = vec_mergeh(filter_v,(vector signed short)vzero);
262 vector signed int val_vEven = vec_mule(src_v, filter_v);
263 vector signed int val_s = vec_sums(val_vEven, vzero);
264 vec_st(val_s, 0, tempo);
265 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
272 for(i=0; i<dstW; i++) {
273 register int srcPos = filterPos[i];
275 vector unsigned char src_v0 = vec_ld(srcPos, src);
276 vector unsigned char src_v1;
277 if ((((int)src + srcPos)% 16) > 8) {
278 src_v1 = vec_ld(srcPos + 16, src);
280 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
282 vector signed short src_v = // vec_unpackh sign-extends...
283 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
284 vector signed short filter_v = vec_ld(i << 4, filter);
285 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
287 vector signed int val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
288 vector signed int val_s = vec_sums(val_v, vzero);
289 vec_st(val_s, 0, tempo);
290 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
297 for(i=0; i<dstW; i++) {
298 register int srcPos = filterPos[i];
300 vector unsigned char src_v0 = vec_ld(srcPos, src);
301 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
302 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
304 vector signed short src_vA = // vec_unpackh sign-extends...
305 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
306 vector signed short src_vB = // vec_unpackh sign-extends...
307 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
309 vector signed short filter_v0 = vec_ld(i << 5, filter);
310 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
311 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
313 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
314 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
316 vector signed int val_s = vec_sums(val_v, vzero);
318 vec_st(val_s, 0, tempo);
319 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
326 for(i=0; i<dstW; i++) {
328 register int srcPos = filterPos[i];
330 vector signed int val_v = (vector signed int)vzero;
331 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
332 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
334 vector unsigned char src_v0 = vec_ld(srcPos, src);
335 vector unsigned char permS = vec_lvsl(srcPos, src);
337 for (j = 0 ; j < filterSize - 15; j += 16) {
338 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
339 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
341 vector signed short src_vA = // vec_unpackh sign-extends...
342 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
343 vector signed short src_vB = // vec_unpackh sign-extends...
344 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
346 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
347 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
348 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
349 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
351 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
352 val_v = vec_msums(src_vB, filter_v1, val_acc);
354 filter_v0R = filter_v2R;
358 if (j < (filterSize-7)) {
359 // loading src_v0 is useless, it's already done above
360 //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
361 vector unsigned char src_v1;
362 if ((((int)src + srcPos)% 16) > 8) {
363 src_v1 = vec_ld(srcPos + j + 16, src);
365 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
367 vector signed short src_v = // vec_unpackh sign-extends...
368 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
369 // loading filter_v0R is useless, it's already done above
370 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
371 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
372 vector signed short filter_v = vec_perm(filter_v0R, filter_v1R, permF);
374 val_v = vec_msums(src_v, filter_v, val_v);
377 vector signed int val_s = vec_sums(val_v, vzero);
379 vec_st(val_s, 0, tempo);
380 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
387 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
388 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
389 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
390 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
391 uint8_t *ysrc = src[0];
392 uint8_t *usrc = src[1];
393 uint8_t *vsrc = src[2];
394 const int width = c->srcW;
395 const int height = srcSliceH;
396 const int lumStride = srcStride[0];
397 const int chromStride = srcStride[1];
398 const int dstStride = dstStride_a[0];
399 const vector unsigned char yperm = vec_lvsl(0, ysrc);
400 const int vertLumPerChroma = 2;
401 register unsigned int y;
405 1) dst is 16 bytes-aligned
406 2) dstStride is a multiple of 16
407 3) width is a multiple of 16
408 4) lum&chrom stride are multiple of 8
411 for(y=0; y<height; y++)
414 for (i = 0; i < width - 31; i+= 32) {
415 const unsigned int j = i >> 1;
416 vector unsigned char v_yA = vec_ld(i, ysrc);
417 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
418 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
419 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
420 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
421 vector unsigned char v_uA = vec_ld(j, usrc);
422 vector unsigned char v_uB = vec_ld(j + 16, usrc);
423 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
424 vector unsigned char v_vA = vec_ld(j, vsrc);
425 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
426 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
427 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
428 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
429 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
430 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
431 vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b);
432 vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b);
433 vec_st(v_yuy2_0, (i << 1), dst);
434 vec_st(v_yuy2_1, (i << 1) + 16, dst);
435 vec_st(v_yuy2_2, (i << 1) + 32, dst);
436 vec_st(v_yuy2_3, (i << 1) + 48, dst);
439 const unsigned int j = i >> 1;
440 vector unsigned char v_y1 = vec_ld(i, ysrc);
441 vector unsigned char v_u = vec_ld(j, usrc);
442 vector unsigned char v_v = vec_ld(j, vsrc);
443 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
444 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
445 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
446 vec_st(v_yuy2_0, (i << 1), dst);
447 vec_st(v_yuy2_1, (i << 1) + 16, dst);
449 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
461 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
462 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
463 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
464 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
465 uint8_t *ysrc = src[0];
466 uint8_t *usrc = src[1];
467 uint8_t *vsrc = src[2];
468 const int width = c->srcW;
469 const int height = srcSliceH;
470 const int lumStride = srcStride[0];
471 const int chromStride = srcStride[1];
472 const int dstStride = dstStride_a[0];
473 const int vertLumPerChroma = 2;
474 const vector unsigned char yperm = vec_lvsl(0, ysrc);
475 register unsigned int y;
479 1) dst is 16 bytes-aligned
480 2) dstStride is a multiple of 16
481 3) width is a multiple of 16
482 4) lum&chrom stride are multiple of 8
485 for(y=0; y<height; y++)
488 for (i = 0; i < width - 31; i+= 32) {
489 const unsigned int j = i >> 1;
490 vector unsigned char v_yA = vec_ld(i, ysrc);
491 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
492 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
493 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
494 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
495 vector unsigned char v_uA = vec_ld(j, usrc);
496 vector unsigned char v_uB = vec_ld(j + 16, usrc);
497 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
498 vector unsigned char v_vA = vec_ld(j, vsrc);
499 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
500 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
501 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
502 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
503 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
504 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
505 vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2);
506 vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2);
507 vec_st(v_uyvy_0, (i << 1), dst);
508 vec_st(v_uyvy_1, (i << 1) + 16, dst);
509 vec_st(v_uyvy_2, (i << 1) + 32, dst);
510 vec_st(v_uyvy_3, (i << 1) + 48, dst);
513 const unsigned int j = i >> 1;
514 vector unsigned char v_y1 = vec_ld(i, ysrc);
515 vector unsigned char v_u = vec_ld(j, usrc);
516 vector unsigned char v_v = vec_ld(j, vsrc);
517 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
518 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
519 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
520 vec_st(v_uyvy_0, (i << 1), dst);
521 vec_st(v_uyvy_1, (i << 1) + 16, dst);
523 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )