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., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
29 #define vzero vec_splat_s32(0)
32 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) {
34 vector unsigned int altivec_vectorShiftInt19 =
35 vec_add(vec_splat_u32(10),vec_splat_u32(9));
36 if ((unsigned long)dest % 16) {
37 /* badly aligned store, we force store alignement */
38 /* and will handle load misalignement on val w/ vec_perm */
39 for (i = 0 ; (i < dstW) &&
40 (((unsigned long)dest + i) % 16) ; i++) {
42 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
44 vector unsigned char perm1 = vec_lvsl(i << 2, val);
45 vector signed int v1 = vec_ld(i << 2, val);
46 for ( ; i < (dstW - 15); i+=16) {
48 vector signed int v2 = vec_ld(offset + 16, val);
49 vector signed int v3 = vec_ld(offset + 32, val);
50 vector signed int v4 = vec_ld(offset + 48, val);
51 vector signed int v5 = vec_ld(offset + 64, val);
52 vector signed int v12 = vec_perm(v1,v2,perm1);
53 vector signed int v23 = vec_perm(v2,v3,perm1);
54 vector signed int v34 = vec_perm(v3,v4,perm1);
55 vector signed int v45 = vec_perm(v4,v5,perm1);
57 vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
58 vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
59 vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
60 vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
61 vector unsigned short vs1 = vec_packsu(vA, vB);
62 vector unsigned short vs2 = vec_packsu(vC, vD);
63 vector unsigned char vf = vec_packsu(vs1, vs2);
67 } else { // dest is properly aligned, great
68 for (i = 0; i < (dstW - 15); i+=16) {
70 vector signed int v1 = vec_ld(offset, val);
71 vector signed int v2 = vec_ld(offset + 16, val);
72 vector signed int v3 = vec_ld(offset + 32, val);
73 vector signed int v4 = vec_ld(offset + 48, val);
74 vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
75 vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
76 vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
77 vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
78 vector unsigned short vs1 = vec_packsu(v5, v6);
79 vector unsigned short vs2 = vec_packsu(v7, v8);
80 vector unsigned char vf = vec_packsu(vs1, vs2);
84 for ( ; i < dstW ; i++) {
86 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
91 yuv2yuvX_altivec_real(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
92 int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
93 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW)
95 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
98 int __attribute__ ((aligned (16))) val[dstW];
100 for (i = 0; i < (dstW -7); i+=4) {
101 vec_st(vini, i << 2, val);
103 for (; i < dstW; i++) {
107 for (j = 0; j < lumFilterSize; j++) {
108 vector signed short vLumFilter = vec_ld(j << 1, lumFilter);
109 vector unsigned char perm0 = vec_lvsl(j << 1, lumFilter);
110 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
111 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
113 vector unsigned char perm = vec_lvsl(0, lumSrc[j]);
114 vector signed short l1 = vec_ld(0, lumSrc[j]);
116 for (i = 0; i < (dstW - 7); i+=8) {
118 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
120 vector signed int v1 = vec_ld(offset, val);
121 vector signed int v2 = vec_ld(offset + 16, val);
123 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
125 vector signed int i1 = vec_mule(vLumFilter, ls);
126 vector signed int i2 = vec_mulo(vLumFilter, ls);
128 vector signed int vf1 = vec_mergeh(i1, i2);
129 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
131 vector signed int vo1 = vec_add(v1, vf1);
132 vector signed int vo2 = vec_add(v2, vf2);
134 vec_st(vo1, offset, val);
135 vec_st(vo2, offset + 16, val);
139 for ( ; i < dstW; i++) {
140 val[i] += lumSrc[j][i] * lumFilter[j];
143 altivec_packIntArrayToCharArray(val,dest,dstW);
146 int __attribute__ ((aligned (16))) u[chrDstW];
147 int __attribute__ ((aligned (16))) v[chrDstW];
149 for (i = 0; i < (chrDstW -7); i+=4) {
150 vec_st(vini, i << 2, u);
151 vec_st(vini, i << 2, v);
153 for (; i < chrDstW; i++) {
158 for (j = 0; j < chrFilterSize; j++) {
159 vector signed short vChrFilter = vec_ld(j << 1, chrFilter);
160 vector unsigned char perm0 = vec_lvsl(j << 1, chrFilter);
161 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
162 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
164 vector unsigned char perm = vec_lvsl(0, chrSrc[j]);
165 vector signed short l1 = vec_ld(0, chrSrc[j]);
166 vector signed short l1_V = vec_ld(2048 << 1, chrSrc[j]);
168 for (i = 0; i < (chrDstW - 7); i+=8) {
170 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]);
171 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]);
173 vector signed int v1 = vec_ld(offset, u);
174 vector signed int v2 = vec_ld(offset + 16, u);
175 vector signed int v1_V = vec_ld(offset, v);
176 vector signed int v2_V = vec_ld(offset + 16, v);
178 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7]
179 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055]
181 vector signed int i1 = vec_mule(vChrFilter, ls);
182 vector signed int i2 = vec_mulo(vChrFilter, ls);
183 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
184 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
186 vector signed int vf1 = vec_mergeh(i1, i2);
187 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
188 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
189 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
191 vector signed int vo1 = vec_add(v1, vf1);
192 vector signed int vo2 = vec_add(v2, vf2);
193 vector signed int vo1_V = vec_add(v1_V, vf1_V);
194 vector signed int vo2_V = vec_add(v2_V, vf2_V);
196 vec_st(vo1, offset, u);
197 vec_st(vo2, offset + 16, u);
198 vec_st(vo1_V, offset, v);
199 vec_st(vo2_V, offset + 16, v);
204 for ( ; i < chrDstW; i++) {
205 u[i] += chrSrc[j][i] * chrFilter[j];
206 v[i] += chrSrc[j][i + 2048] * chrFilter[j];
209 altivec_packIntArrayToCharArray(u,uDest,chrDstW);
210 altivec_packIntArrayToCharArray(v,vDest,chrDstW);
214 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) {
216 int __attribute__ ((aligned (16))) tempo[4];
218 if (filterSize % 4) {
219 for(i=0; i<dstW; i++) {
221 register int srcPos = filterPos[i];
222 register int val = 0;
223 for(j=0; j<filterSize; j++) {
224 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
226 dst[i] = MIN(MAX(0, val>>7), (1<<15)-1);
230 switch (filterSize) {
233 for(i=0; i<dstW; i++) {
235 register int srcPos = filterPos[i];
237 vector unsigned char src_v0 = vec_ld(srcPos, src);
238 vector unsigned char src_v1;
239 if ((((int)src + srcPos)% 16) > 12) {
240 src_v1 = vec_ld(srcPos + 16, src);
242 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
244 vector signed short src_v = // vec_unpackh sign-extends...
245 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
246 // now put our elements in the even slots
247 src_v = vec_mergeh(src_v, (vector signed short)vzero);
249 vector signed short filter_v = vec_ld(i << 3, filter);
250 // the 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2)
252 // the neat trick : we only care for half the elements,
253 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
254 // and we're going to use vec_mule, so we chose
255 // carefully how to "unpack" the elements into the even slots
257 filter_v = vec_mergel(filter_v,(vector signed short)vzero);
259 filter_v = vec_mergeh(filter_v,(vector signed short)vzero);
261 vector signed int val_vEven = vec_mule(src_v, filter_v);
262 vector signed int val_s = vec_sums(val_vEven, vzero);
263 vec_st(val_s, 0, tempo);
264 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
271 for(i=0; i<dstW; i++) {
272 register int srcPos = filterPos[i];
274 vector unsigned char src_v0 = vec_ld(srcPos, src);
275 vector unsigned char src_v1;
276 if ((((int)src + srcPos)% 16) > 8) {
277 src_v1 = vec_ld(srcPos + 16, src);
279 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
281 vector signed short src_v = // vec_unpackh sign-extends...
282 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
283 vector signed short filter_v = vec_ld(i << 4, filter);
284 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
286 vector signed int val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
287 vector signed int val_s = vec_sums(val_v, vzero);
288 vec_st(val_s, 0, tempo);
289 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
296 for(i=0; i<dstW; i++) {
297 register int srcPos = filterPos[i];
299 vector unsigned char src_v0 = vec_ld(srcPos, src);
300 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
301 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
303 vector signed short src_vA = // vec_unpackh sign-extends...
304 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
305 vector signed short src_vB = // vec_unpackh sign-extends...
306 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
308 vector signed short filter_v0 = vec_ld(i << 5, filter);
309 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
310 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
312 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
313 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
315 vector signed int val_s = vec_sums(val_v, vzero);
317 vec_st(val_s, 0, tempo);
318 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
325 for(i=0; i<dstW; i++) {
327 register int srcPos = filterPos[i];
329 vector signed int val_v = (vector signed int)vzero;
330 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
331 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
333 vector unsigned char src_v0 = vec_ld(srcPos, src);
334 vector unsigned char permS = vec_lvsl(srcPos, src);
336 for (j = 0 ; j < filterSize - 15; j += 16) {
337 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
338 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
340 vector signed short src_vA = // vec_unpackh sign-extends...
341 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
342 vector signed short src_vB = // vec_unpackh sign-extends...
343 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
345 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
346 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
347 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
348 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
350 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
351 val_v = vec_msums(src_vB, filter_v1, val_acc);
353 filter_v0R = filter_v2R;
357 if (j < (filterSize-7)) {
358 // loading src_v0 is useless, it's already done above
359 //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
360 vector unsigned char src_v1;
361 if ((((int)src + srcPos)% 16) > 8) {
362 src_v1 = vec_ld(srcPos + j + 16, src);
364 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
366 vector signed short src_v = // vec_unpackh sign-extends...
367 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
368 // loading filter_v0R is useless, it's already done above
369 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
370 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
371 vector signed short filter_v = vec_perm(filter_v0R, filter_v1R, permF);
373 val_v = vec_msums(src_v, filter_v, val_v);
376 vector signed int val_s = vec_sums(val_v, vzero);
378 vec_st(val_s, 0, tempo);
379 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
386 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
387 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
388 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
389 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
390 uint8_t *ysrc = src[0];
391 uint8_t *usrc = src[1];
392 uint8_t *vsrc = src[2];
393 const int width = c->srcW;
394 const int height = srcSliceH;
395 const int lumStride = srcStride[0];
396 const int chromStride = srcStride[1];
397 const int dstStride = dstStride_a[0];
398 const vector unsigned char yperm = vec_lvsl(0, ysrc);
399 const int vertLumPerChroma = 2;
400 register unsigned int y;
403 yv12toyuy2( ysrc, usrc, vsrc, dst,c->srcW,srcSliceH, lumStride, chromStride, dstStride);
409 1) dst is 16 bytes-aligned
410 2) dstStride is a multiple of 16
411 3) width is a multiple of 16
412 4) lum&chrom stride are multiple of 8
415 for(y=0; y<height; y++)
418 for (i = 0; i < width - 31; i+= 32) {
419 const unsigned int j = i >> 1;
420 vector unsigned char v_yA = vec_ld(i, ysrc);
421 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
422 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
423 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
424 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
425 vector unsigned char v_uA = vec_ld(j, usrc);
426 vector unsigned char v_uB = vec_ld(j + 16, usrc);
427 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
428 vector unsigned char v_vA = vec_ld(j, vsrc);
429 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
430 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
431 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
432 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
433 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
434 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
435 vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b);
436 vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b);
437 vec_st(v_yuy2_0, (i << 1), dst);
438 vec_st(v_yuy2_1, (i << 1) + 16, dst);
439 vec_st(v_yuy2_2, (i << 1) + 32, dst);
440 vec_st(v_yuy2_3, (i << 1) + 48, dst);
443 const unsigned int j = i >> 1;
444 vector unsigned char v_y1 = vec_ld(i, ysrc);
445 vector unsigned char v_u = vec_ld(j, usrc);
446 vector unsigned char v_v = vec_ld(j, vsrc);
447 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
448 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
449 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
450 vec_st(v_yuy2_0, (i << 1), dst);
451 vec_st(v_yuy2_1, (i << 1) + 16, dst);
453 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
465 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
466 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
467 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
468 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
469 uint8_t *ysrc = src[0];
470 uint8_t *usrc = src[1];
471 uint8_t *vsrc = src[2];
472 const int width = c->srcW;
473 const int height = srcSliceH;
474 const int lumStride = srcStride[0];
475 const int chromStride = srcStride[1];
476 const int dstStride = dstStride_a[0];
477 const int vertLumPerChroma = 2;
478 const vector unsigned char yperm = vec_lvsl(0, ysrc);
479 register unsigned int y;
482 yv12touyvy( ysrc, usrc, vsrc, dst,c->srcW,srcSliceH, lumStride, chromStride, dstStride);
488 1) dst is 16 bytes-aligned
489 2) dstStride is a multiple of 16
490 3) width is a multiple of 16
491 4) lum&chrom stride are multiple of 8
494 for(y=0; y<height; y++)
497 for (i = 0; i < width - 31; i+= 32) {
498 const unsigned int j = i >> 1;
499 vector unsigned char v_yA = vec_ld(i, ysrc);
500 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
501 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
502 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
503 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
504 vector unsigned char v_uA = vec_ld(j, usrc);
505 vector unsigned char v_uB = vec_ld(j + 16, usrc);
506 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
507 vector unsigned char v_vA = vec_ld(j, vsrc);
508 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
509 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
510 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
511 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
512 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
513 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
514 vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2);
515 vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2);
516 vec_st(v_uyvy_0, (i << 1), dst);
517 vec_st(v_uyvy_1, (i << 1) + 16, dst);
518 vec_st(v_uyvy_2, (i << 1) + 32, dst);
519 vec_st(v_uyvy_3, (i << 1) + 48, dst);
522 const unsigned int j = i >> 1;
523 vector unsigned char v_y1 = vec_ld(i, ysrc);
524 vector unsigned char v_u = vec_ld(j, usrc);
525 vector unsigned char v_v = vec_ld(j, vsrc);
526 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
527 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
528 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
529 vec_st(v_uyvy_0, (i << 1), dst);
530 vec_st(v_uyvy_1, (i << 1) + 16, dst);
532 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )