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
26 #include "libswscale/swscale.h"
27 #include "libswscale/swscale_internal.h"
28 #include "libavutil/cpu.h"
29 #include "yuv2rgb_altivec.h"
31 #define vzero vec_splat_s32(0)
34 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW)
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;
44 for (i = 0 ; (i < dstW) &&
45 (((uintptr_t)dest + i) % 16) ; i++) {
47 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
49 perm1 = vec_lvsl(i << 2, val);
50 v1 = vec_ld(i << 2, val);
51 for ( ; i < (dstW - 15); i+=16) {
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);
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);
72 } else { // dest is properly aligned, great
73 for (i = 0; i < (dstW - 15); i+=16) {
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);
89 for ( ; i < dstW ; i++) {
91 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
96 yuv2yuvX_altivec_real(SwsContext *c,
97 const int16_t *lumFilter, const int16_t **lumSrc,
98 int lumFilterSize, const int16_t *chrFilter,
99 const int16_t **chrUSrc, const int16_t **chrVSrc,
100 int chrFilterSize, const int16_t **alpSrc,
101 uint8_t *dest[4], int dstW, int chrDstW)
103 uint8_t *yDest = dest[0], *uDest = dest[1], *vDest = dest[2];
104 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
107 DECLARE_ALIGNED(16, int, val)[dstW];
109 for (i = 0; i < (dstW -7); i+=4) {
110 vec_st(vini, i << 2, val);
112 for (; i < dstW; i++) {
116 for (j = 0; j < lumFilterSize; j++) {
117 vector signed short l1, vLumFilter = vec_ld(j << 1, lumFilter);
118 vector unsigned char perm, perm0 = vec_lvsl(j << 1, lumFilter);
119 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
120 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
122 perm = vec_lvsl(0, lumSrc[j]);
123 l1 = vec_ld(0, lumSrc[j]);
125 for (i = 0; i < (dstW - 7); i+=8) {
127 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
129 vector signed int v1 = vec_ld(offset, val);
130 vector signed int v2 = vec_ld(offset + 16, val);
132 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
134 vector signed int i1 = vec_mule(vLumFilter, ls);
135 vector signed int i2 = vec_mulo(vLumFilter, ls);
137 vector signed int vf1 = vec_mergeh(i1, i2);
138 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
140 vector signed int vo1 = vec_add(v1, vf1);
141 vector signed int vo2 = vec_add(v2, vf2);
143 vec_st(vo1, offset, val);
144 vec_st(vo2, offset + 16, val);
148 for ( ; i < dstW; i++) {
149 val[i] += lumSrc[j][i] * lumFilter[j];
152 altivec_packIntArrayToCharArray(val, yDest, dstW);
155 DECLARE_ALIGNED(16, int, u)[chrDstW];
156 DECLARE_ALIGNED(16, int, v)[chrDstW];
158 for (i = 0; i < (chrDstW -7); i+=4) {
159 vec_st(vini, i << 2, u);
160 vec_st(vini, i << 2, v);
162 for (; i < chrDstW; i++) {
167 for (j = 0; j < chrFilterSize; j++) {
168 vector signed short l1, l1_V, vChrFilter = vec_ld(j << 1, chrFilter);
169 vector unsigned char perm, perm0 = vec_lvsl(j << 1, chrFilter);
170 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
171 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
173 perm = vec_lvsl(0, chrUSrc[j]);
174 l1 = vec_ld(0, chrUSrc[j]);
175 l1_V = vec_ld(0, chrVSrc[j]);
177 for (i = 0; i < (chrDstW - 7); i+=8) {
179 vector signed short l2 = vec_ld((i << 1) + 16, chrUSrc[j]);
180 vector signed short l2_V = vec_ld((i << 1) + 16, chrVSrc[j]);
182 vector signed int v1 = vec_ld(offset, u);
183 vector signed int v2 = vec_ld(offset + 16, u);
184 vector signed int v1_V = vec_ld(offset, v);
185 vector signed int v2_V = vec_ld(offset + 16, v);
187 vector signed short ls = vec_perm(l1, l2, perm); // chrUSrc[j][i] ... chrUSrc[j][i+7]
188 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrVSrc[j][i] ... chrVSrc[j][i]
190 vector signed int i1 = vec_mule(vChrFilter, ls);
191 vector signed int i2 = vec_mulo(vChrFilter, ls);
192 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
193 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
195 vector signed int vf1 = vec_mergeh(i1, i2);
196 vector signed int vf2 = vec_mergel(i1, i2); // chrUSrc[j][i] * chrFilter[j] ... chrUSrc[j][i+7] * chrFilter[j]
197 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
198 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrVSrc[j][i] * chrFilter[j] ... chrVSrc[j][i+7] * chrFilter[j]
200 vector signed int vo1 = vec_add(v1, vf1);
201 vector signed int vo2 = vec_add(v2, vf2);
202 vector signed int vo1_V = vec_add(v1_V, vf1_V);
203 vector signed int vo2_V = vec_add(v2_V, vf2_V);
205 vec_st(vo1, offset, u);
206 vec_st(vo2, offset + 16, u);
207 vec_st(vo1_V, offset, v);
208 vec_st(vo2_V, offset + 16, v);
213 for ( ; i < chrDstW; i++) {
214 u[i] += chrUSrc[j][i] * chrFilter[j];
215 v[i] += chrVSrc[j][i] * chrFilter[j];
218 altivec_packIntArrayToCharArray(u, uDest, chrDstW);
219 altivec_packIntArrayToCharArray(v, vDest, chrDstW);
223 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
224 const uint8_t *src, const int16_t *filter,
225 const int16_t *filterPos, int filterSize)
228 DECLARE_ALIGNED(16, int, tempo)[4];
230 if (filterSize % 4) {
231 for (i=0; i<dstW; i++) {
233 register int srcPos = filterPos[i];
234 register int val = 0;
235 for (j=0; j<filterSize; j++) {
236 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
238 dst[i] = FFMIN(val>>7, (1<<15)-1);
242 switch (filterSize) {
245 for (i=0; i<dstW; i++) {
246 register int srcPos = filterPos[i];
248 vector unsigned char src_v0 = vec_ld(srcPos, src);
249 vector unsigned char src_v1, src_vF;
250 vector signed short src_v, filter_v;
251 vector signed int val_vEven, val_s;
252 if ((((int)src + srcPos)% 16) > 12) {
253 src_v1 = vec_ld(srcPos + 16, src);
255 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
257 src_v = // vec_unpackh sign-extends...
258 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
259 // now put our elements in the even slots
260 src_v = vec_mergeh(src_v, (vector signed short)vzero);
262 filter_v = vec_ld(i << 3, filter);
263 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
265 // The neat trick: We only care for half the elements,
266 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
267 // and we're going to use vec_mule, so we choose
268 // carefully how to "unpack" the elements into the even slots.
270 filter_v = vec_mergel(filter_v, (vector signed short)vzero);
272 filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
274 val_vEven = vec_mule(src_v, filter_v);
275 val_s = vec_sums(val_vEven, vzero);
276 vec_st(val_s, 0, tempo);
277 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
284 for (i=0; i<dstW; i++) {
285 register int srcPos = filterPos[i];
287 vector unsigned char src_v0 = vec_ld(srcPos, src);
288 vector unsigned char src_v1, src_vF;
289 vector signed short src_v, filter_v;
290 vector signed int val_v, val_s;
291 if ((((int)src + srcPos)% 16) > 8) {
292 src_v1 = vec_ld(srcPos + 16, src);
294 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
296 src_v = // vec_unpackh sign-extends...
297 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
298 filter_v = vec_ld(i << 4, filter);
299 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
301 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
302 val_s = vec_sums(val_v, vzero);
303 vec_st(val_s, 0, tempo);
304 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
311 for (i=0; i<dstW; i++) {
312 register int srcPos = filterPos[i];
314 vector unsigned char src_v0 = vec_ld(srcPos, src);
315 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
316 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
318 vector signed short src_vA = // vec_unpackh sign-extends...
319 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
320 vector signed short src_vB = // vec_unpackh sign-extends...
321 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
323 vector signed short filter_v0 = vec_ld(i << 5, filter);
324 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
325 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
327 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
328 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
330 vector signed int val_s = vec_sums(val_v, vzero);
332 vec_st(val_s, 0, tempo);
333 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
340 for (i=0; i<dstW; i++) {
342 register int srcPos = filterPos[i];
344 vector signed int val_s, val_v = (vector signed int)vzero;
345 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
346 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
348 vector unsigned char src_v0 = vec_ld(srcPos, src);
349 vector unsigned char permS = vec_lvsl(srcPos, src);
351 for (j = 0 ; j < filterSize - 15; j += 16) {
352 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
353 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
355 vector signed short src_vA = // vec_unpackh sign-extends...
356 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
357 vector signed short src_vB = // vec_unpackh sign-extends...
358 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
360 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
361 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
362 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
363 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
365 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
366 val_v = vec_msums(src_vB, filter_v1, val_acc);
368 filter_v0R = filter_v2R;
372 if (j < filterSize-7) {
373 // loading src_v0 is useless, it's already done above
374 //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
375 vector unsigned char src_v1, src_vF;
376 vector signed short src_v, filter_v1R, filter_v;
377 if ((((int)src + srcPos)% 16) > 8) {
378 src_v1 = vec_ld(srcPos + j + 16, src);
380 src_vF = vec_perm(src_v0, src_v1, permS);
382 src_v = // vec_unpackh sign-extends...
383 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
384 // loading filter_v0R is useless, it's already done above
385 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
386 filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
387 filter_v = vec_perm(filter_v0R, filter_v1R, permF);
389 val_v = vec_msums(src_v, filter_v, val_v);
392 val_s = vec_sums(val_v, vzero);
394 vec_st(val_s, 0, tempo);
395 dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
402 void ff_sws_init_swScale_altivec(SwsContext *c)
404 enum PixelFormat dstFormat = c->dstFormat;
406 if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
409 if (c->scalingBpp == 8) {
410 c->hScale = hScale_altivec_real;
412 if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
413 dstFormat != PIX_FMT_NV12 && dstFormat != PIX_FMT_NV21 &&
415 c->yuv2yuvX = yuv2yuvX_altivec_real;
418 /* The following list of supported dstFormat values should
419 * match what's found in the body of ff_yuv2packedX_altivec() */
420 if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
421 switch (c->dstFormat) {
422 case PIX_FMT_ABGR: c->yuv2packedX = ff_yuv2abgr_X_altivec; break;
423 case PIX_FMT_BGRA: c->yuv2packedX = ff_yuv2bgra_X_altivec; break;
424 case PIX_FMT_ARGB: c->yuv2packedX = ff_yuv2argb_X_altivec; break;
425 case PIX_FMT_RGBA: c->yuv2packedX = ff_yuv2rgba_X_altivec; break;
426 case PIX_FMT_BGR24: c->yuv2packedX = ff_yuv2bgr24_X_altivec; break;
427 case PIX_FMT_RGB24: c->yuv2packedX = ff_yuv2rgb24_X_altivec; break;