| 1 | /* |
| 2 | * AltiVec-enhanced yuv2yuvX |
| 3 | * |
| 4 | * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org> |
| 5 | * based on the equivalent C code in swscale.c |
| 6 | * |
| 7 | * This file is part of FFmpeg. |
| 8 | * |
| 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. |
| 13 | * |
| 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. |
| 18 | * |
| 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 |
| 22 | */ |
| 23 | |
| 24 | #include <inttypes.h> |
| 25 | |
| 26 | #include "config.h" |
| 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 | |
| 33 | #if HAVE_ALTIVEC |
| 34 | #define vzero vec_splat_s32(0) |
| 35 | |
| 36 | #define yuv2planeX_8(d1, d2, l1, src, x, perm, filter) do { \ |
| 37 | vector signed short l2 = vec_ld(((x) << 1) + 16, src); \ |
| 38 | vector signed short ls = vec_perm(l1, l2, perm); \ |
| 39 | vector signed int i1 = vec_mule(filter, ls); \ |
| 40 | vector signed int i2 = vec_mulo(filter, ls); \ |
| 41 | vector signed int vf1 = vec_mergeh(i1, i2); \ |
| 42 | vector signed int vf2 = vec_mergel(i1, i2); \ |
| 43 | d1 = vec_add(d1, vf1); \ |
| 44 | d2 = vec_add(d2, vf2); \ |
| 45 | l1 = l2; \ |
| 46 | } while (0) |
| 47 | |
| 48 | static void yuv2planeX_16_altivec(const int16_t *filter, int filterSize, |
| 49 | const int16_t **src, uint8_t *dest, |
| 50 | const uint8_t *dither, int offset, int x) |
| 51 | { |
| 52 | register int i, j; |
| 53 | DECLARE_ALIGNED(16, int, val)[16]; |
| 54 | vector signed int vo1, vo2, vo3, vo4; |
| 55 | vector unsigned short vs1, vs2; |
| 56 | vector unsigned char vf; |
| 57 | vector unsigned int altivec_vectorShiftInt19 = |
| 58 | vec_add(vec_splat_u32(10), vec_splat_u32(9)); |
| 59 | |
| 60 | for (i = 0; i < 16; i++) |
| 61 | val[i] = dither[(x + i + offset) & 7] << 12; |
| 62 | |
| 63 | vo1 = vec_ld(0, val); |
| 64 | vo2 = vec_ld(16, val); |
| 65 | vo3 = vec_ld(32, val); |
| 66 | vo4 = vec_ld(48, val); |
| 67 | |
| 68 | for (j = 0; j < filterSize; j++) { |
| 69 | vector signed short l1, vLumFilter = vec_ld(j << 1, filter); |
| 70 | vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter); |
| 71 | vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0); |
| 72 | vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter |
| 73 | |
| 74 | perm = vec_lvsl(x << 1, src[j]); |
| 75 | l1 = vec_ld(x << 1, src[j]); |
| 76 | |
| 77 | yuv2planeX_8(vo1, vo2, l1, src[j], x, perm, vLumFilter); |
| 78 | yuv2planeX_8(vo3, vo4, l1, src[j], x + 8, perm, vLumFilter); |
| 79 | } |
| 80 | |
| 81 | vo1 = vec_sra(vo1, altivec_vectorShiftInt19); |
| 82 | vo2 = vec_sra(vo2, altivec_vectorShiftInt19); |
| 83 | vo3 = vec_sra(vo3, altivec_vectorShiftInt19); |
| 84 | vo4 = vec_sra(vo4, altivec_vectorShiftInt19); |
| 85 | vs1 = vec_packsu(vo1, vo2); |
| 86 | vs2 = vec_packsu(vo3, vo4); |
| 87 | vf = vec_packsu(vs1, vs2); |
| 88 | vec_st(vf, 0, dest); |
| 89 | } |
| 90 | |
| 91 | static inline void yuv2planeX_u(const int16_t *filter, int filterSize, |
| 92 | const int16_t **src, uint8_t *dest, int dstW, |
| 93 | const uint8_t *dither, int offset, int x) |
| 94 | { |
| 95 | int i, j; |
| 96 | |
| 97 | for (i = x; i < dstW; i++) { |
| 98 | int t = dither[(i + offset) & 7] << 12; |
| 99 | for (j = 0; j < filterSize; j++) |
| 100 | t += src[j][i] * filter[j]; |
| 101 | dest[i] = av_clip_uint8(t >> 19); |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | static void yuv2planeX_altivec(const int16_t *filter, int filterSize, |
| 106 | const int16_t **src, uint8_t *dest, int dstW, |
| 107 | const uint8_t *dither, int offset) |
| 108 | { |
| 109 | int dst_u = -(uintptr_t)dest & 15; |
| 110 | int i; |
| 111 | |
| 112 | yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0); |
| 113 | |
| 114 | for (i = dst_u; i < dstW - 15; i += 16) |
| 115 | yuv2planeX_16_altivec(filter, filterSize, src, dest + i, dither, |
| 116 | offset, i); |
| 117 | |
| 118 | yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i); |
| 119 | } |
| 120 | |
| 121 | static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW, |
| 122 | const uint8_t *src, const int16_t *filter, |
| 123 | const int32_t *filterPos, int filterSize) |
| 124 | { |
| 125 | register int i; |
| 126 | DECLARE_ALIGNED(16, int, tempo)[4]; |
| 127 | |
| 128 | if (filterSize % 4) { |
| 129 | for (i = 0; i < dstW; i++) { |
| 130 | register int j; |
| 131 | register int srcPos = filterPos[i]; |
| 132 | register int val = 0; |
| 133 | for (j = 0; j < filterSize; j++) |
| 134 | val += ((int)src[srcPos + j]) * filter[filterSize * i + j]; |
| 135 | dst[i] = FFMIN(val >> 7, (1 << 15) - 1); |
| 136 | } |
| 137 | } else |
| 138 | switch (filterSize) { |
| 139 | case 4: |
| 140 | for (i = 0; i < dstW; i++) { |
| 141 | register int srcPos = filterPos[i]; |
| 142 | |
| 143 | vector unsigned char src_v0 = vec_ld(srcPos, src); |
| 144 | vector unsigned char src_v1, src_vF; |
| 145 | vector signed short src_v, filter_v; |
| 146 | vector signed int val_vEven, val_s; |
| 147 | if ((((uintptr_t)src + srcPos) % 16) > 12) { |
| 148 | src_v1 = vec_ld(srcPos + 16, src); |
| 149 | } |
| 150 | src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
| 151 | |
| 152 | src_v = // vec_unpackh sign-extends... |
| 153 | (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
| 154 | // now put our elements in the even slots |
| 155 | src_v = vec_mergeh(src_v, (vector signed short)vzero); |
| 156 | |
| 157 | filter_v = vec_ld(i << 3, filter); |
| 158 | // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2). |
| 159 | |
| 160 | // The neat trick: We only care for half the elements, |
| 161 | // high or low depending on (i<<3)%16 (it's 0 or 8 here), |
| 162 | // and we're going to use vec_mule, so we choose |
| 163 | // carefully how to "unpack" the elements into the even slots. |
| 164 | if ((i << 3) % 16) |
| 165 | filter_v = vec_mergel(filter_v, (vector signed short)vzero); |
| 166 | else |
| 167 | filter_v = vec_mergeh(filter_v, (vector signed short)vzero); |
| 168 | |
| 169 | val_vEven = vec_mule(src_v, filter_v); |
| 170 | val_s = vec_sums(val_vEven, vzero); |
| 171 | vec_st(val_s, 0, tempo); |
| 172 | dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1); |
| 173 | } |
| 174 | break; |
| 175 | |
| 176 | case 8: |
| 177 | for (i = 0; i < dstW; i++) { |
| 178 | register int srcPos = filterPos[i]; |
| 179 | |
| 180 | vector unsigned char src_v0 = vec_ld(srcPos, src); |
| 181 | vector unsigned char src_v1, src_vF; |
| 182 | vector signed short src_v, filter_v; |
| 183 | vector signed int val_v, val_s; |
| 184 | if ((((uintptr_t)src + srcPos) % 16) > 8) { |
| 185 | src_v1 = vec_ld(srcPos + 16, src); |
| 186 | } |
| 187 | src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
| 188 | |
| 189 | src_v = // vec_unpackh sign-extends... |
| 190 | (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
| 191 | filter_v = vec_ld(i << 4, filter); |
| 192 | // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2) |
| 193 | |
| 194 | val_v = vec_msums(src_v, filter_v, (vector signed int)vzero); |
| 195 | val_s = vec_sums(val_v, vzero); |
| 196 | vec_st(val_s, 0, tempo); |
| 197 | dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1); |
| 198 | } |
| 199 | break; |
| 200 | |
| 201 | case 16: |
| 202 | for (i = 0; i < dstW; i++) { |
| 203 | register int srcPos = filterPos[i]; |
| 204 | |
| 205 | vector unsigned char src_v0 = vec_ld(srcPos, src); |
| 206 | vector unsigned char src_v1 = vec_ld(srcPos + 16, src); |
| 207 | vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
| 208 | |
| 209 | vector signed short src_vA = // vec_unpackh sign-extends... |
| 210 | (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
| 211 | vector signed short src_vB = // vec_unpackh sign-extends... |
| 212 | (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF)); |
| 213 | |
| 214 | vector signed short filter_v0 = vec_ld(i << 5, filter); |
| 215 | vector signed short filter_v1 = vec_ld((i << 5) + 16, filter); |
| 216 | // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2) |
| 217 | |
| 218 | vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero); |
| 219 | vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc); |
| 220 | |
| 221 | vector signed int val_s = vec_sums(val_v, vzero); |
| 222 | |
| 223 | vec_st(val_s, 0, tempo); |
| 224 | dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1); |
| 225 | } |
| 226 | break; |
| 227 | |
| 228 | default: |
| 229 | for (i = 0; i < dstW; i++) { |
| 230 | register int j; |
| 231 | register int srcPos = filterPos[i]; |
| 232 | |
| 233 | vector signed int val_s, val_v = (vector signed int)vzero; |
| 234 | vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter); |
| 235 | vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter); |
| 236 | |
| 237 | vector unsigned char src_v0 = vec_ld(srcPos, src); |
| 238 | vector unsigned char permS = vec_lvsl(srcPos, src); |
| 239 | |
| 240 | for (j = 0; j < filterSize - 15; j += 16) { |
| 241 | vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src); |
| 242 | vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS); |
| 243 | |
| 244 | vector signed short src_vA = // vec_unpackh sign-extends... |
| 245 | (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
| 246 | vector signed short src_vB = // vec_unpackh sign-extends... |
| 247 | (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF)); |
| 248 | |
| 249 | vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter); |
| 250 | vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter); |
| 251 | vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF); |
| 252 | vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF); |
| 253 | |
| 254 | vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v); |
| 255 | val_v = vec_msums(src_vB, filter_v1, val_acc); |
| 256 | |
| 257 | filter_v0R = filter_v2R; |
| 258 | src_v0 = src_v1; |
| 259 | } |
| 260 | |
| 261 | if (j < filterSize - 7) { |
| 262 | // loading src_v0 is useless, it's already done above |
| 263 | // vector unsigned char src_v0 = vec_ld(srcPos + j, src); |
| 264 | vector unsigned char src_v1, src_vF; |
| 265 | vector signed short src_v, filter_v1R, filter_v; |
| 266 | if ((((uintptr_t)src + srcPos) % 16) > 8) { |
| 267 | src_v1 = vec_ld(srcPos + j + 16, src); |
| 268 | } |
| 269 | src_vF = vec_perm(src_v0, src_v1, permS); |
| 270 | |
| 271 | src_v = // vec_unpackh sign-extends... |
| 272 | (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
| 273 | // loading filter_v0R is useless, it's already done above |
| 274 | // vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter); |
| 275 | filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter); |
| 276 | filter_v = vec_perm(filter_v0R, filter_v1R, permF); |
| 277 | |
| 278 | val_v = vec_msums(src_v, filter_v, val_v); |
| 279 | } |
| 280 | |
| 281 | val_s = vec_sums(val_v, vzero); |
| 282 | |
| 283 | vec_st(val_s, 0, tempo); |
| 284 | dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1); |
| 285 | } |
| 286 | } |
| 287 | } |
| 288 | #endif /* HAVE_ALTIVEC */ |
| 289 | |
| 290 | av_cold void ff_sws_init_swscale_ppc(SwsContext *c) |
| 291 | { |
| 292 | #if HAVE_ALTIVEC |
| 293 | enum AVPixelFormat dstFormat = c->dstFormat; |
| 294 | |
| 295 | if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)) |
| 296 | return; |
| 297 | |
| 298 | if (c->srcBpc == 8 && c->dstBpc <= 14) { |
| 299 | c->hyScale = c->hcScale = hScale_altivec_real; |
| 300 | } |
| 301 | if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) && |
| 302 | dstFormat != AV_PIX_FMT_NV12 && dstFormat != AV_PIX_FMT_NV21 && |
| 303 | !c->alpPixBuf) { |
| 304 | c->yuv2planeX = yuv2planeX_altivec; |
| 305 | } |
| 306 | |
| 307 | /* The following list of supported dstFormat values should |
| 308 | * match what's found in the body of ff_yuv2packedX_altivec() */ |
| 309 | if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) { |
| 310 | switch (c->dstFormat) { |
| 311 | case AV_PIX_FMT_ABGR: |
| 312 | c->yuv2packedX = ff_yuv2abgr_X_altivec; |
| 313 | break; |
| 314 | case AV_PIX_FMT_BGRA: |
| 315 | c->yuv2packedX = ff_yuv2bgra_X_altivec; |
| 316 | break; |
| 317 | case AV_PIX_FMT_ARGB: |
| 318 | c->yuv2packedX = ff_yuv2argb_X_altivec; |
| 319 | break; |
| 320 | case AV_PIX_FMT_RGBA: |
| 321 | c->yuv2packedX = ff_yuv2rgba_X_altivec; |
| 322 | break; |
| 323 | case AV_PIX_FMT_BGR24: |
| 324 | c->yuv2packedX = ff_yuv2bgr24_X_altivec; |
| 325 | break; |
| 326 | case AV_PIX_FMT_RGB24: |
| 327 | c->yuv2packedX = ff_yuv2rgb24_X_altivec; |
| 328 | break; |
| 329 | } |
| 330 | } |
| 331 | #endif /* HAVE_ALTIVEC */ |
| 332 | } |