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
27 #include "libswscale/swscale.h"
28 #include "libswscale/swscale_internal.h"
29 #include "libavutil/attributes.h"
30 #include "libavutil/cpu.h"
31 #include "yuv2rgb_altivec.h"
32 #include "libavutil/ppc/util_altivec.h"
35 #define vzero vec_splat_s32(0)
38 #define GET_LS(a,b,c,s) {\
39 vector signed short l2 = vec_ld(((b) << 1) + 16, s);\
40 ls = vec_perm(a, l2, c);\
44 #define GET_LS(a,b,c,s) {\
46 a = vec_vsx_ld(((b) << 1) + 16, s);\
50 #define yuv2planeX_8(d1, d2, l1, src, x, perm, filter) do {\
51 vector signed short ls;\
52 GET_LS(l1, x, perm, src);\
53 vector signed int i1 = vec_mule(filter, ls);\
54 vector signed int i2 = vec_mulo(filter, ls);\
55 vector signed int vf1, vf2;\
56 vf1 = vec_mergeh(i1, i2);\
57 vf2 = vec_mergel(i1, i2);\
58 d1 = vec_add(d1, vf1);\
59 d2 = vec_add(d2, vf2);\
63 #define LOAD_FILTER(vf,f) {\
64 vector unsigned char perm0 = vec_lvsl(joffset, f);\
65 vf = vec_ld(joffset, f);\
66 vf = vec_perm(vf, vf, perm0);\
68 #define LOAD_L1(ll1,s,p){\
69 p = vec_lvsl(xoffset, s);\
70 ll1 = vec_ld(xoffset, s);\
73 #define LOAD_FILTER(vf,f) {\
74 vf = vec_vsx_ld(joffset, f);\
76 #define LOAD_L1(ll1,s,p){\
77 ll1 = vec_vsx_ld(xoffset, s);\
81 static void yuv2planeX_16_altivec(const int16_t *filter, int filterSize,
82 const int16_t **src, uint8_t *dest,
83 const uint8_t *dither, int offset, int x)
86 LOCAL_ALIGNED(16, int, val, [16]);
87 vector signed int vo1, vo2, vo3, vo4;
88 vector unsigned short vs1, vs2;
89 vector unsigned char vf;
90 vector unsigned int altivec_vectorShiftInt19 =
91 vec_add(vec_splat_u32(10), vec_splat_u32(9));
93 for (i = 0; i < 16; i++)
94 val[i] = dither[(x + i + offset) & 7] << 12;
97 vo2 = vec_ld(16, val);
98 vo3 = vec_ld(32, val);
99 vo4 = vec_ld(48, val);
101 for (j = 0; j < filterSize; j++) {
102 unsigned int joffset=j<<1;
103 unsigned int xoffset=x<<1;
104 vector unsigned char perm;
105 vector signed short l1,vLumFilter;
106 LOAD_FILTER(vLumFilter,filter);
107 vLumFilter = vec_splat(vLumFilter, 0);
108 LOAD_L1(l1,src[j],perm);
109 yuv2planeX_8(vo1, vo2, l1, src[j], x, perm, vLumFilter);
110 yuv2planeX_8(vo3, vo4, l1, src[j], x + 8, perm, vLumFilter);
113 vo1 = vec_sra(vo1, altivec_vectorShiftInt19);
114 vo2 = vec_sra(vo2, altivec_vectorShiftInt19);
115 vo3 = vec_sra(vo3, altivec_vectorShiftInt19);
116 vo4 = vec_sra(vo4, altivec_vectorShiftInt19);
117 vs1 = vec_packsu(vo1, vo2);
118 vs2 = vec_packsu(vo3, vo4);
119 vf = vec_packsu(vs1, vs2);
124 static inline void yuv2planeX_u(const int16_t *filter, int filterSize,
125 const int16_t **src, uint8_t *dest, int dstW,
126 const uint8_t *dither, int offset, int x)
130 for (i = x; i < dstW; i++) {
131 int t = dither[(i + offset) & 7] << 12;
132 for (j = 0; j < filterSize; j++)
133 t += src[j][i] * filter[j];
134 dest[i] = av_clip_uint8(t >> 19);
138 static void yuv2planeX_altivec(const int16_t *filter, int filterSize,
139 const int16_t **src, uint8_t *dest, int dstW,
140 const uint8_t *dither, int offset)
142 int dst_u = -(uintptr_t)dest & 15;
145 yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0);
147 for (i = dst_u; i < dstW - 15; i += 16)
148 yuv2planeX_16_altivec(filter, filterSize, src, dest + i, dither,
151 yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i);
155 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
157 // The neat trick: We only care for half the elements,
158 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
159 // and we're going to use vec_mule, so we choose
160 // carefully how to "unpack" the elements into the even slots.
161 #define GET_VF4(a, vf, f) {\
162 vf = vec_ld(a<< 3, f);\
164 vf = vec_mergel(vf, (vector signed short)vzero);\
166 vf = vec_mergeh(vf, (vector signed short)vzero);\
168 #define FIRST_LOAD(sv, pos, s, per) {\
169 sv = vec_ld(pos, s);\
170 per = vec_lvsl(pos, s);\
172 #define UPDATE_PTR(s0, d0, s1, d1) {\
176 #define LOAD_SRCV(pos, a, s, per, v0, v1, vf) {\
177 v1 = vec_ld(pos + a + 16, s);\
178 vf = vec_perm(v0, v1, per);\
180 #define LOAD_SRCV8(pos, a, s, per, v0, v1, vf) {\
181 if ((((uintptr_t)s + pos) % 16) > 8) {\
182 v1 = vec_ld(pos + a + 16, s);\
184 vf = vec_perm(v0, src_v1, per);\
186 #define GET_VFD(a, b, f, vf0, vf1, per, vf, off) {\
187 vf1 = vec_ld((a * 2 * filterSize) + (b * 2) + 16 + off, f);\
188 vf = vec_perm(vf0, vf1, per);\
190 #else /* else of #if HAVE_BIGENDIAN */
191 #define GET_VF4(a, vf, f) {\
192 vf = (vector signed short)vec_vsx_ld(a << 3, f);\
193 vf = vec_mergeh(vf, (vector signed short)vzero);\
195 #define FIRST_LOAD(sv, pos, s, per) {}
196 #define UPDATE_PTR(s0, d0, s1, d1) {}
197 #define LOAD_SRCV(pos, a, s, per, v0, v1, vf) {\
198 vf = vec_vsx_ld(pos + a, s);\
200 #define LOAD_SRCV8(pos, a, s, per, v0, v1, vf) LOAD_SRCV(pos, a, s, per, v0, v1, vf)
201 #define GET_VFD(a, b, f, vf0, vf1, per, vf, off) {\
202 vf = vec_vsx_ld((a * 2 * filterSize) + (b * 2) + off, f);\
204 #endif /* end of #if HAVE_BIGENDIAN */
206 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
207 const uint8_t *src, const int16_t *filter,
208 const int32_t *filterPos, int filterSize)
211 LOCAL_ALIGNED(16, int, tempo, [4]);
213 if (filterSize % 4) {
214 for (i = 0; i < dstW; i++) {
216 register int srcPos = filterPos[i];
217 register int val = 0;
218 for (j = 0; j < filterSize; j++)
219 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
220 dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
223 switch (filterSize) {
225 for (i = 0; i < dstW; i++) {
226 register int srcPos = filterPos[i];
228 vector unsigned char src_vF = unaligned_load(srcPos, src);
229 vector signed short src_v, filter_v;
230 vector signed int val_vEven, val_s;
231 src_v = // vec_unpackh sign-extends...
232 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
233 // now put our elements in the even slots
234 src_v = vec_mergeh(src_v, (vector signed short)vzero);
235 GET_VF4(i, filter_v, filter);
236 val_vEven = vec_mule(src_v, filter_v);
237 val_s = vec_sums(val_vEven, vzero);
238 vec_st(val_s, 0, tempo);
239 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
243 for (i = 0; i < dstW; i++) {
244 register int srcPos = filterPos[i];
245 vector unsigned char src_vF, src_v0, src_v1;
246 vector unsigned char permS;
247 vector signed short src_v, filter_v;
248 vector signed int val_v, val_s;
249 FIRST_LOAD(src_v0, srcPos, src, permS);
250 LOAD_SRCV8(srcPos, 0, src, permS, src_v0, src_v1, src_vF);
251 src_v = // vec_unpackh sign-extends...
252 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
253 filter_v = vec_ld(i << 4, filter);
254 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
255 val_s = vec_sums(val_v, vzero);
256 vec_st(val_s, 0, tempo);
257 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
262 for (i = 0; i < dstW; i++) {
263 register int srcPos = filterPos[i];
265 vector unsigned char src_vF = unaligned_load(srcPos, src);
266 vector signed short src_vA = // vec_unpackh sign-extends...
267 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
268 vector signed short src_vB = // vec_unpackh sign-extends...
269 (vector signed short)(VEC_MERGEL((vector unsigned char)vzero, src_vF));
270 vector signed short filter_v0 = vec_ld(i << 5, filter);
271 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
273 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
274 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
276 vector signed int val_s = vec_sums(val_v, vzero);
278 VEC_ST(val_s, 0, tempo);
279 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
284 for (i = 0; i < dstW; i++) {
285 register int j, offset = i * 2 * filterSize;
286 register int srcPos = filterPos[i];
288 vector signed int val_s, val_v = (vector signed int)vzero;
289 vector signed short filter_v0R;
290 vector unsigned char permF, src_v0, permS;
291 FIRST_LOAD(filter_v0R, offset, filter, permF);
292 FIRST_LOAD(src_v0, srcPos, src, permS);
294 for (j = 0; j < filterSize - 15; j += 16) {
295 vector unsigned char src_v1, src_vF;
296 vector signed short filter_v1R, filter_v2R, filter_v0, filter_v1;
297 LOAD_SRCV(srcPos, j, src, permS, src_v0, src_v1, src_vF);
298 vector signed short src_vA = // vec_unpackh sign-extends...
299 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
300 vector signed short src_vB = // vec_unpackh sign-extends...
301 (vector signed short)(VEC_MERGEL((vector unsigned char)vzero, src_vF));
302 GET_VFD(i, j, filter, filter_v0R, filter_v1R, permF, filter_v0, 0);
303 GET_VFD(i, j, filter, filter_v1R, filter_v2R, permF, filter_v1, 16);
305 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
306 val_v = vec_msums(src_vB, filter_v1, val_acc);
307 UPDATE_PTR(filter_v2R, filter_v0R, src_v1, src_v0);
310 if (j < filterSize - 7) {
311 // loading src_v0 is useless, it's already done above
312 vector unsigned char src_v1, src_vF;
313 vector signed short src_v, filter_v1R, filter_v;
314 LOAD_SRCV8(srcPos, j, src, permS, src_v0, src_v1, src_vF);
315 src_v = // vec_unpackh sign-extends...
316 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
317 GET_VFD(i, j, filter, filter_v0R, filter_v1R, permF, filter_v, 0);
318 val_v = vec_msums(src_v, filter_v, val_v);
320 val_s = vec_sums(val_v, vzero);
322 VEC_ST(val_s, 0, tempo);
323 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
327 #endif /* HAVE_ALTIVEC */
329 av_cold void ff_sws_init_swscale_ppc(SwsContext *c)
332 enum AVPixelFormat dstFormat = c->dstFormat;
334 if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
337 if (c->srcBpc == 8 && c->dstBpc <= 14) {
338 c->hyScale = c->hcScale = hScale_altivec_real;
340 if (!is16BPS(dstFormat) && !isNBPS(dstFormat) &&
341 dstFormat != AV_PIX_FMT_NV12 && dstFormat != AV_PIX_FMT_NV21 &&
343 c->yuv2planeX = yuv2planeX_altivec;
346 /* The following list of supported dstFormat values should
347 * match what's found in the body of ff_yuv2packedX_altivec() */
348 if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->needAlpha) {
349 switch (c->dstFormat) {
350 case AV_PIX_FMT_ABGR:
351 c->yuv2packedX = ff_yuv2abgr_X_altivec;
353 case AV_PIX_FMT_BGRA:
354 c->yuv2packedX = ff_yuv2bgra_X_altivec;
356 case AV_PIX_FMT_ARGB:
357 c->yuv2packedX = ff_yuv2argb_X_altivec;
359 case AV_PIX_FMT_RGBA:
360 c->yuv2packedX = ff_yuv2rgba_X_altivec;
362 case AV_PIX_FMT_BGR24:
363 c->yuv2packedX = ff_yuv2bgr24_X_altivec;
365 case AV_PIX_FMT_RGB24:
366 c->yuv2packedX = ff_yuv2rgb24_X_altivec;
370 #endif /* HAVE_ALTIVEC */