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"
33 #define vzero vec_splat_s32(0)
35 #define yuv2planeX_8(d1, d2, l1, src, x, perm, filter) do { \
36 vector signed short l2 = vec_ld(((x) << 1) + 16, src); \
37 vector signed short ls = vec_perm(l1, l2, perm); \
38 vector signed int i1 = vec_mule(filter, ls); \
39 vector signed int i2 = vec_mulo(filter, ls); \
40 vector signed int vf1 = vec_mergeh(i1, i2); \
41 vector signed int vf2 = vec_mergel(i1, i2); \
42 d1 = vec_add(d1, vf1); \
43 d2 = vec_add(d2, vf2); \
47 static void yuv2planeX_16_altivec(const int16_t *filter, int filterSize,
48 const int16_t **src, uint8_t *dest,
49 const uint8_t *dither, int offset, int x)
52 DECLARE_ALIGNED(16, int, val)[16];
53 vector signed int vo1, vo2, vo3, vo4;
54 vector unsigned short vs1, vs2;
55 vector unsigned char vf;
56 vector unsigned int altivec_vectorShiftInt19 =
57 vec_add(vec_splat_u32(10), vec_splat_u32(9));
59 for (i = 0; i < 16; i++)
60 val[i] = dither[(x + i + offset) & 7] << 12;
63 vo2 = vec_ld(16, val);
64 vo3 = vec_ld(32, val);
65 vo4 = vec_ld(48, val);
67 for (j = 0; j < filterSize; j++) {
68 vector signed short l1, vLumFilter = vec_ld(j << 1, filter);
69 vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter);
70 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
71 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
73 perm = vec_lvsl(x << 1, src[j]);
74 l1 = vec_ld(x << 1, src[j]);
76 yuv2planeX_8(vo1, vo2, l1, src[j], x, perm, vLumFilter);
77 yuv2planeX_8(vo3, vo4, l1, src[j], x + 8, perm, vLumFilter);
80 vo1 = vec_sra(vo1, altivec_vectorShiftInt19);
81 vo2 = vec_sra(vo2, altivec_vectorShiftInt19);
82 vo3 = vec_sra(vo3, altivec_vectorShiftInt19);
83 vo4 = vec_sra(vo4, altivec_vectorShiftInt19);
84 vs1 = vec_packsu(vo1, vo2);
85 vs2 = vec_packsu(vo3, vo4);
86 vf = vec_packsu(vs1, vs2);
90 static inline void yuv2planeX_u(const int16_t *filter, int filterSize,
91 const int16_t **src, uint8_t *dest, int dstW,
92 const uint8_t *dither, int offset, int x)
96 for (i = x; i < dstW; i++) {
97 int t = dither[(i + offset) & 7] << 12;
98 for (j = 0; j < filterSize; j++)
99 t += src[j][i] * filter[j];
100 dest[i] = av_clip_uint8(t >> 19);
104 static void yuv2planeX_altivec(const int16_t *filter, int filterSize,
105 const int16_t **src, uint8_t *dest, int dstW,
106 const uint8_t *dither, int offset)
108 int dst_u = -(uintptr_t)dest & 15;
111 yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0);
113 for (i = dst_u; i < dstW - 15; i += 16)
114 yuv2planeX_16_altivec(filter, filterSize, src, dest + i, dither,
117 yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i);
120 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
121 const uint8_t *src, const int16_t *filter,
122 const int32_t *filterPos, int filterSize)
125 DECLARE_ALIGNED(16, int, tempo)[4];
127 if (filterSize % 4) {
128 for (i = 0; i < dstW; i++) {
130 register int srcPos = filterPos[i];
131 register int val = 0;
132 for (j = 0; j < filterSize; j++)
133 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
134 dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
137 switch (filterSize) {
139 for (i = 0; i < dstW; i++) {
140 register int srcPos = filterPos[i];
142 vector unsigned char src_v0 = vec_ld(srcPos, src);
143 vector unsigned char src_v1, src_vF;
144 vector signed short src_v, filter_v;
145 vector signed int val_vEven, val_s;
146 if ((((uintptr_t)src + srcPos) % 16) > 12) {
147 src_v1 = vec_ld(srcPos + 16, src);
149 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
151 src_v = // vec_unpackh sign-extends...
152 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
153 // now put our elements in the even slots
154 src_v = vec_mergeh(src_v, (vector signed short)vzero);
156 filter_v = vec_ld(i << 3, filter);
157 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
159 // The neat trick: We only care for half the elements,
160 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
161 // and we're going to use vec_mule, so we choose
162 // carefully how to "unpack" the elements into the even slots.
164 filter_v = vec_mergel(filter_v, (vector signed short)vzero);
166 filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
168 val_vEven = vec_mule(src_v, filter_v);
169 val_s = vec_sums(val_vEven, vzero);
170 vec_st(val_s, 0, tempo);
171 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
176 for (i = 0; i < dstW; i++) {
177 register int srcPos = filterPos[i];
179 vector unsigned char src_v0 = vec_ld(srcPos, src);
180 vector unsigned char src_v1, src_vF;
181 vector signed short src_v, filter_v;
182 vector signed int val_v, val_s;
183 if ((((uintptr_t)src + srcPos) % 16) > 8) {
184 src_v1 = vec_ld(srcPos + 16, src);
186 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
188 src_v = // vec_unpackh sign-extends...
189 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
190 filter_v = vec_ld(i << 4, filter);
191 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
193 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
194 val_s = vec_sums(val_v, vzero);
195 vec_st(val_s, 0, tempo);
196 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
201 for (i = 0; i < dstW; i++) {
202 register int srcPos = filterPos[i];
204 vector unsigned char src_v0 = vec_ld(srcPos, src);
205 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
206 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
208 vector signed short src_vA = // vec_unpackh sign-extends...
209 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
210 vector signed short src_vB = // vec_unpackh sign-extends...
211 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
213 vector signed short filter_v0 = vec_ld(i << 5, filter);
214 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
215 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
217 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
218 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
220 vector signed int val_s = vec_sums(val_v, vzero);
222 vec_st(val_s, 0, tempo);
223 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
228 for (i = 0; i < dstW; i++) {
230 register int srcPos = filterPos[i];
232 vector signed int val_s, val_v = (vector signed int)vzero;
233 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
234 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
236 vector unsigned char src_v0 = vec_ld(srcPos, src);
237 vector unsigned char permS = vec_lvsl(srcPos, src);
239 for (j = 0; j < filterSize - 15; j += 16) {
240 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
241 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
243 vector signed short src_vA = // vec_unpackh sign-extends...
244 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
245 vector signed short src_vB = // vec_unpackh sign-extends...
246 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
248 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
249 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
250 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
251 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
253 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
254 val_v = vec_msums(src_vB, filter_v1, val_acc);
256 filter_v0R = filter_v2R;
260 if (j < filterSize - 7) {
261 // loading src_v0 is useless, it's already done above
262 // vector unsigned char src_v0 = vec_ld(srcPos + j, src);
263 vector unsigned char src_v1, src_vF;
264 vector signed short src_v, filter_v1R, filter_v;
265 if ((((uintptr_t)src + srcPos) % 16) > 8) {
266 src_v1 = vec_ld(srcPos + j + 16, src);
268 src_vF = vec_perm(src_v0, src_v1, permS);
270 src_v = // vec_unpackh sign-extends...
271 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
272 // loading filter_v0R is useless, it's already done above
273 // vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
274 filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
275 filter_v = vec_perm(filter_v0R, filter_v1R, permF);
277 val_v = vec_msums(src_v, filter_v, val_v);
280 val_s = vec_sums(val_v, vzero);
282 vec_st(val_s, 0, tempo);
283 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
288 av_cold void ff_sws_init_swScale_altivec(SwsContext *c)
290 enum AVPixelFormat dstFormat = c->dstFormat;
292 if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
295 if (c->srcBpc == 8 && c->dstBpc <= 14) {
296 c->hyScale = c->hcScale = hScale_altivec_real;
298 if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
299 dstFormat != AV_PIX_FMT_NV12 && dstFormat != AV_PIX_FMT_NV21 &&
301 c->yuv2planeX = yuv2planeX_altivec;
304 /* The following list of supported dstFormat values should
305 * match what's found in the body of ff_yuv2packedX_altivec() */
306 if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
307 switch (c->dstFormat) {
308 case AV_PIX_FMT_ABGR:
309 c->yuv2packedX = ff_yuv2abgr_X_altivec;
311 case AV_PIX_FMT_BGRA:
312 c->yuv2packedX = ff_yuv2bgra_X_altivec;
314 case AV_PIX_FMT_ARGB:
315 c->yuv2packedX = ff_yuv2argb_X_altivec;
317 case AV_PIX_FMT_RGBA:
318 c->yuv2packedX = ff_yuv2rgba_X_altivec;
320 case AV_PIX_FMT_BGR24:
321 c->yuv2packedX = ff_yuv2bgr24_X_altivec;
323 case AV_PIX_FMT_RGB24:
324 c->yuv2packedX = ff_yuv2rgb24_X_altivec;