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
24 #include "libavutil/attributes.h"
25 #include "libavutil/mem_internal.h"
27 static void FUNC(yuv2planeX_8_16)(const int16_t *filter, int filterSize,
28 const int16_t **src, uint8_t *dest,
29 const uint8_t *dither, int offset, int x)
32 LOCAL_ALIGNED(16, int, val, [16]);
33 vector signed int vo1, vo2, vo3, vo4;
34 vector unsigned short vs1, vs2;
35 vector unsigned char vf;
36 vector unsigned int altivec_vectorShiftInt19 =
37 vec_add(vec_splat_u32(10), vec_splat_u32(9));
39 for (i = 0; i < 16; i++)
40 val[i] = dither[(x + i + offset) & 7] << 12;
43 vo2 = vec_ld(16, val);
44 vo3 = vec_ld(32, val);
45 vo4 = vec_ld(48, val);
47 for (j = 0; j < filterSize; j++) {
48 unsigned int joffset=j<<1;
49 unsigned int xoffset=x<<1;
50 vector unsigned char av_unused perm;
51 vector signed short l1,vLumFilter;
52 LOAD_FILTER(vLumFilter,filter);
53 vLumFilter = vec_splat(vLumFilter, 0);
54 LOAD_L1(l1,src[j],perm);
55 yuv2planeX_8(vo1, vo2, l1, src[j], x, perm, vLumFilter);
56 yuv2planeX_8(vo3, vo4, l1, src[j], x + 8, perm, vLumFilter);
59 vo1 = vec_sra(vo1, altivec_vectorShiftInt19);
60 vo2 = vec_sra(vo2, altivec_vectorShiftInt19);
61 vo3 = vec_sra(vo3, altivec_vectorShiftInt19);
62 vo4 = vec_sra(vo4, altivec_vectorShiftInt19);
63 vs1 = vec_packsu(vo1, vo2);
64 vs2 = vec_packsu(vo3, vo4);
65 vf = vec_packsu(vs1, vs2);
70 static inline void yuv2planeX_u(const int16_t *filter, int filterSize,
71 const int16_t **src, uint8_t *dest, int dstW,
72 const uint8_t *dither, int offset, int x)
76 for (i = x; i < dstW; i++) {
77 int t = dither[(i + offset) & 7] << 12;
78 for (j = 0; j < filterSize; j++)
79 t += src[j][i] * filter[j];
80 dest[i] = av_clip_uint8(t >> 19);
84 static void FUNC(yuv2planeX)(const int16_t *filter, int filterSize,
85 const int16_t **src, uint8_t *dest, int dstW,
86 const uint8_t *dither, int offset)
88 int dst_u = -(uintptr_t)dest & 15;
91 yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0);
93 for (i = dst_u; i < dstW - 15; i += 16)
94 FUNC(yuv2planeX_8_16)(filter, filterSize, src, dest + i, dither,
97 yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i);
100 static void FUNC(hScale_real)(SwsContext *c, int16_t *dst, int dstW,
101 const uint8_t *src, const int16_t *filter,
102 const int32_t *filterPos, int filterSize)
105 LOCAL_ALIGNED(16, int, tempo, [4]);
107 if (filterSize % 4) {
108 for (i = 0; i < dstW; i++) {
110 register int srcPos = filterPos[i];
111 register int val = 0;
112 for (j = 0; j < filterSize; j++)
113 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
114 dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
117 switch (filterSize) {
119 for (i = 0; i < dstW; i++) {
120 register int srcPos = filterPos[i];
122 vector unsigned char src_vF = unaligned_load(srcPos, src);
123 vector signed short src_v, filter_v;
124 vector signed int val_vEven, val_s;
125 src_v = // vec_unpackh sign-extends...
126 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
127 // now put our elements in the even slots
128 src_v = vec_mergeh(src_v, (vector signed short)vzero);
129 GET_VF4(i, filter_v, filter);
130 val_vEven = vec_mule(src_v, filter_v);
131 val_s = vec_sums(val_vEven, vzero);
132 vec_st(val_s, 0, tempo);
133 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
137 for (i = 0; i < dstW; i++) {
138 register int srcPos = filterPos[i];
139 vector unsigned char src_vF, av_unused src_v0, av_unused src_v1;
140 vector unsigned char av_unused permS;
141 vector signed short src_v, filter_v;
142 vector signed int val_v, val_s;
143 FIRST_LOAD(src_v0, srcPos, src, permS);
144 LOAD_SRCV8(srcPos, 0, src, permS, src_v0, src_v1, src_vF);
145 src_v = // vec_unpackh sign-extends...
146 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
147 filter_v = vec_ld(i << 4, filter);
148 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
149 val_s = vec_sums(val_v, vzero);
150 vec_st(val_s, 0, tempo);
151 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
156 for (i = 0; i < dstW; i++) {
157 register int srcPos = filterPos[i];
159 vector unsigned char src_vF = unaligned_load(srcPos, src);
160 vector signed short src_vA = // vec_unpackh sign-extends...
161 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
162 vector signed short src_vB = // vec_unpackh sign-extends...
163 (vector signed short)(VEC_MERGEL((vector unsigned char)vzero, src_vF));
164 vector signed short filter_v0 = vec_ld(i << 5, filter);
165 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
167 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
168 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
170 vector signed int val_s = vec_sums(val_v, vzero);
172 VEC_ST(val_s, 0, tempo);
173 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
178 for (i = 0; i < dstW; i++) {
179 register int j, av_unused offset = i * 2 * filterSize;
180 register int srcPos = filterPos[i];
182 vector signed int val_s, val_v = (vector signed int)vzero;
183 vector signed short av_unused filter_v0R;
184 vector unsigned char av_unused permF, av_unused src_v0, av_unused permS;
185 FIRST_LOAD(filter_v0R, offset, filter, permF);
186 FIRST_LOAD(src_v0, srcPos, src, permS);
188 for (j = 0; j < filterSize - 15; j += 16) {
189 vector unsigned char av_unused src_v1, src_vF;
190 vector signed short av_unused filter_v1R, av_unused filter_v2R,
191 filter_v0, filter_v1, src_vA, src_vB;
192 vector signed int val_acc;
193 LOAD_SRCV(srcPos, j, src, permS, src_v0, src_v1, src_vF);
194 src_vA = // vec_unpackh sign-extends...
195 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
196 src_vB = // vec_unpackh sign-extends...
197 (vector signed short)(VEC_MERGEL((vector unsigned char)vzero, src_vF));
198 GET_VFD(i, j, filter, filter_v0R, filter_v1R, permF, filter_v0, 0);
199 GET_VFD(i, j, filter, filter_v1R, filter_v2R, permF, filter_v1, 16);
201 val_acc = vec_msums(src_vA, filter_v0, val_v);
202 val_v = vec_msums(src_vB, filter_v1, val_acc);
203 UPDATE_PTR(filter_v2R, filter_v0R, src_v1, src_v0);
206 if (j < filterSize - 7) {
207 // loading src_v0 is useless, it's already done above
208 vector unsigned char av_unused src_v1, src_vF;
209 vector signed short src_v, av_unused filter_v1R, filter_v;
210 LOAD_SRCV8(srcPos, j, src, permS, src_v0, src_v1, src_vF);
211 src_v = // vec_unpackh sign-extends...
212 (vector signed short)(VEC_MERGEH((vector unsigned char)vzero, src_vF));
213 GET_VFD(i, j, filter, filter_v0R, filter_v1R, permF, filter_v, 0);
214 val_v = vec_msums(src_v, filter_v, val_v);
216 val_s = vec_sums(val_v, vzero);
218 VEC_ST(val_s, 0, tempo);
219 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);