git.sesse.net Git - ffmpeg/blob - libswscale/ppc/yuv2rgb_altivec.c

   1 /*
   2  * AltiVec acceleration for colorspace conversion
   3  *
   4  * copyright (C) 2004 Marc Hoffman <marc.hoffman@analog.com>
   5  *
   6  * This file is part of Libav.
   7  *
   8  * Libav is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * Libav is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with Libav; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 /*
  24  * Convert I420 YV12 to RGB in various formats,
  25  * it rejects images that are not in 420 formats,
  26  * it rejects images that don't have widths of multiples of 16,
  27  * it rejects images that don't have heights of multiples of 2.
  28  * Reject defers to C simulation code.
  29  *
  30  * Lots of optimizations to be done here.
  31  *
  32  * 1. Need to fix saturation code. I just couldn't get it to fly with packs
  33  * and adds, so we currently use max/min to clip.
  34  *
  35  * 2. The inefficient use of chroma loading needs a bit of brushing up.
  36  *
  37  * 3. Analysis of pipeline stalls needs to be done. Use shark to identify
  38  * pipeline stalls.
  39  *
  40  *
  41  * MODIFIED to calculate coeffs from currently selected color space.
  42  * MODIFIED core to be a macro where you specify the output format.
  43  * ADDED UYVY conversion which is never called due to some thing in swscale.
  44  * CORRECTED algorithim selection to be strict on input formats.
  45  * ADDED runtime detection of AltiVec.
  46  *
  47  * ADDED altivec_yuv2packedX vertical scl + RGB converter
  48  *
  49  * March 27,2004
  50  * PERFORMANCE ANALYSIS
  51  *
  52  * The C version uses 25% of the processor or ~250Mips for D1 video rawvideo
  53  * used as test.
  54  * The AltiVec version uses 10% of the processor or ~100Mips for D1 video
  55  * same sequence.
  56  *
  57  * 720 * 480 * 30  ~10MPS
  58  *
  59  * so we have roughly 10 clocks per pixel. This is too high, something has
  60  * to be wrong.
  61  *
  62  * OPTIMIZED clip codes to utilize vec_max and vec_packs removing the
  63  * need for vec_min.
  64  *
  65  * OPTIMIZED DST OUTPUT cache/DMA controls. We are pretty much guaranteed to
  66  * have the input video frame, it was just decompressed so it probably resides
  67  * in L1 caches. However, we are creating the output video stream. This needs
  68  * to use the DSTST instruction to optimize for the cache. We couple this with
  69  * the fact that we are not going to be visiting the input buffer again so we
  70  * mark it Least Recently Used. This shaves 25% of the processor cycles off.
  71  *
  72  * Now memcpy is the largest mips consumer in the system, probably due
  73  * to the inefficient X11 stuff.
  74  *
  75  * GL libraries seem to be very slow on this machine 1.33Ghz PB running
  76  * Jaguar, this is not the case for my 1Ghz PB.  I thought it might be
  77  * a versioning issue, however I have libGL.1.2.dylib for both
  78  * machines. (We need to figure this out now.)
  79  *
  80  * GL2 libraries work now with patch for RGB32.
  81  *
  82  * NOTE: quartz vo driver ARGB32_to_RGB24 consumes 30% of the processor.
  83  *
  84  * Integrated luma prescaling adjustment for saturation/contrast/brightness
  85  * adjustment.
  86  */
  87
  88 #include <stdio.h>
  89 #include <stdlib.h>
  90 #include <string.h>
  91 #include <inttypes.h>
  92 #include <assert.h>
  93
  94 #include "config.h"
  95 #include "libswscale/rgb2rgb.h"
  96 #include "libswscale/swscale.h"
  97 #include "libswscale/swscale_internal.h"
  98 #include "libavutil/attributes.h"
  99 #include "libavutil/cpu.h"
 100 #include "yuv2rgb_altivec.h"
 101
 102 #undef PROFILE_THE_BEAST
 103 #undef INC_SCALING
 104
 105 typedef unsigned char ubyte;
 106 typedef signed char   sbyte;
 107
 108 /* RGB interleaver, 16 planar pels 8-bit samples per channel in
 109  * homogeneous vector registers x0,x1,x2 are interleaved with the
 110  * following technique:
 111  *
 112  *    o0 = vec_mergeh(x0, x1);
 113  *    o1 = vec_perm(o0, x2, perm_rgb_0);
 114  *    o2 = vec_perm(o0, x2, perm_rgb_1);
 115  *    o3 = vec_mergel(x0, x1);
 116  *    o4 = vec_perm(o3, o2, perm_rgb_2);
 117  *    o5 = vec_perm(o3, o2, perm_rgb_3);
 118  *
 119  * perm_rgb_0:   o0(RG).h v1(B) --> o1*
 120  *            0   1  2   3   4
 121  *           rgbr|gbrg|brgb|rgbr
 122  *           0010 0100 1001 0010
 123  *           0102 3145 2673 894A
 124  *
 125  * perm_rgb_1:   o0(RG).h v1(B) --> o2
 126  *            0   1  2   3   4
 127  *           gbrg|brgb|bbbb|bbbb
 128  *           0100 1001 1111 1111
 129  *           B5CD 6EF7 89AB CDEF
 130  *
 131  * perm_rgb_2:   o3(RG).l o2(rgbB.l) --> o4*
 132  *            0   1  2   3   4
 133  *           gbrg|brgb|rgbr|gbrg
 134  *           1111 1111 0010 0100
 135  *           89AB CDEF 0182 3945
 136  *
 137  * perm_rgb_2:   o3(RG).l o2(rgbB.l) ---> o5*
 138  *            0   1  2   3   4
 139  *           brgb|rgbr|gbrg|brgb
 140  *           1001 0010 0100 1001
 141  *           a67b 89cA BdCD eEFf
 142  *
 143  */
 144 static const vector unsigned char
 145     perm_rgb_0 = { 0x00, 0x01, 0x10, 0x02, 0x03, 0x11, 0x04, 0x05,
 146                    0x12, 0x06, 0x07, 0x13, 0x08, 0x09, 0x14, 0x0a },
 147     perm_rgb_1 = { 0x0b, 0x15, 0x0c, 0x0d, 0x16, 0x0e, 0x0f, 0x17,
 148                    0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f },
 149     perm_rgb_2 = { 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
 150                    0x00, 0x01, 0x18, 0x02, 0x03, 0x19, 0x04, 0x05 },
 151     perm_rgb_3 = { 0x1a, 0x06, 0x07, 0x1b, 0x08, 0x09, 0x1c, 0x0a,
 152                    0x0b, 0x1d, 0x0c, 0x0d, 0x1e, 0x0e, 0x0f, 0x1f };
 153
 154 #define vec_merge3(x2, x1, x0, y0, y1, y2)     \
 155     do {                                       \
 156         __typeof__(x0) o0, o2, o3;             \
 157         o0 = vec_mergeh(x0, x1);               \
 158         y0 = vec_perm(o0, x2, perm_rgb_0);     \
 159         o2 = vec_perm(o0, x2, perm_rgb_1);     \
 160         o3 = vec_mergel(x0, x1);               \
 161         y1 = vec_perm(o3, o2, perm_rgb_2);     \
 162         y2 = vec_perm(o3, o2, perm_rgb_3);     \
 163     } while (0)
 164
 165 #define vec_mstbgr24(x0, x1, x2, ptr)          \
 166     do {                                       \
 167         __typeof__(x0) _0, _1, _2;             \
 168         vec_merge3(x0, x1, x2, _0, _1, _2);    \
 169         vec_st(_0, 0, ptr++);                  \
 170         vec_st(_1, 0, ptr++);                  \
 171         vec_st(_2, 0, ptr++);                  \
 172     } while (0)
 173
 174 #define vec_mstrgb24(x0, x1, x2, ptr)          \
 175     do {                                       \
 176         __typeof__(x0) _0, _1, _2;             \
 177         vec_merge3(x2, x1, x0, _0, _1, _2);    \
 178         vec_st(_0, 0, ptr++);                  \
 179         vec_st(_1, 0, ptr++);                  \
 180         vec_st(_2, 0, ptr++);                  \
 181     } while (0)
 182
 183 /* pack the pixels in rgb0 format
 184  * msb R
 185  * lsb 0
 186  */
 187 #define vec_mstrgb32(T, x0, x1, x2, x3, ptr)                            \
 188     do {                                                                \
 189         T _0, _1, _2, _3;                                               \
 190         _0 = vec_mergeh(x0, x1);                                        \
 191         _1 = vec_mergeh(x2, x3);                                        \
 192         _2 = (T) vec_mergeh((vector unsigned short) _0,                 \
 193                             (vector unsigned short) _1);                \
 194         _3 = (T) vec_mergel((vector unsigned short) _0,                 \
 195                             (vector unsigned short) _1);                \
 196         vec_st(_2, 0 * 16, (T *) ptr);                                  \
 197         vec_st(_3, 1 * 16, (T *) ptr);                                  \
 198         _0 = vec_mergel(x0, x1);                                        \
 199         _1 = vec_mergel(x2, x3);                                        \
 200         _2 = (T) vec_mergeh((vector unsigned short) _0,                 \
 201                             (vector unsigned short) _1);                \
 202         _3 = (T) vec_mergel((vector unsigned short) _0,                 \
 203                             (vector unsigned short) _1);                \
 204         vec_st(_2, 2 * 16, (T *) ptr);                                  \
 205         vec_st(_3, 3 * 16, (T *) ptr);                                  \
 206         ptr += 4;                                                       \
 207     } while (0)
 208
 209 /*
 210  * 1     0       1.4021   | | Y |
 211  * 1    -0.3441 -0.7142   |x| Cb|
 212  * 1     1.7718  0        | | Cr|
 213  *
 214  *
 215  * Y:      [-128 127]
 216  * Cb/Cr : [-128 127]
 217  *
 218  * typical YUV conversion works on Y: 0-255 this version has been
 219  * optimized for JPEG decoding.
 220  */
 221
 222 #define vec_unh(x)                                                      \
 223     (vector signed short)                                               \
 224         vec_perm(x, (__typeof__(x)) { 0 },                              \
 225                  ((vector unsigned char) {                              \
 226                      0x10, 0x00, 0x10, 0x01, 0x10, 0x02, 0x10, 0x03,    \
 227                      0x10, 0x04, 0x10, 0x05, 0x10, 0x06, 0x10, 0x07 }))
 228
 229 #define vec_unl(x)                                                      \
 230     (vector signed short)                                               \
 231         vec_perm(x, (__typeof__(x)) { 0 },                              \
 232                  ((vector unsigned char) {                              \
 233                      0x10, 0x08, 0x10, 0x09, 0x10, 0x0A, 0x10, 0x0B,    \
 234                      0x10, 0x0C, 0x10, 0x0D, 0x10, 0x0E, 0x10, 0x0F }))
 235
 236 #define vec_clip_s16(x)                                                 \
 237     vec_max(vec_min(x, ((vector signed short) {                         \
 238                     235, 235, 235, 235, 235, 235, 235, 235 })),         \
 239             ((vector signed short) { 16, 16, 16, 16, 16, 16, 16, 16 }))
 240
 241 #define vec_packclp(x, y)                                               \
 242     (vector unsigned char)                                              \
 243         vec_packs((vector unsigned short)                               \
 244                       vec_max(x, ((vector signed short) { 0 })),        \
 245                   (vector unsigned short)                               \
 246                       vec_max(y, ((vector signed short) { 0 })))
 247
 248 //#define out_pixels(a, b, c, ptr) vec_mstrgb32(__typeof__(a), ((__typeof__(a)) { 255 }), a, a, a, ptr)
 249
 250 static inline void cvtyuvtoRGB(SwsContext *c, vector signed short Y,
 251                                vector signed short U, vector signed short V,
 252                                vector signed short *R, vector signed short *G,
 253                                vector signed short *B)
 254 {
 255     vector signed short vx, ux, uvx;
 256
 257     Y = vec_mradds(Y, c->CY, c->OY);
 258     U = vec_sub(U, (vector signed short)
 259                        vec_splat((vector signed short) { 128 }, 0));
 260     V = vec_sub(V, (vector signed short)
 261                        vec_splat((vector signed short) { 128 }, 0));
 262
 263     // ux  = (CBU * (u << c->CSHIFT) + 0x4000) >> 15;
 264     ux = vec_sl(U, c->CSHIFT);
 265     *B = vec_mradds(ux, c->CBU, Y);
 266
 267     // vx  = (CRV * (v << c->CSHIFT) + 0x4000) >> 15;
 268     vx = vec_sl(V, c->CSHIFT);
 269     *R = vec_mradds(vx, c->CRV, Y);
 270
 271     // uvx = ((CGU * u) + (CGV * v)) >> 15;
 272     uvx = vec_mradds(U, c->CGU, Y);
 273     *G  = vec_mradds(V, c->CGV, uvx);
 274 }
 275
 276 /*
 277  * ------------------------------------------------------------------------------
 278  * CS converters
 279  * ------------------------------------------------------------------------------
 280  */
 281
 282 #define DEFCSP420_CVT(name, out_pixels)                                       \
 283 static int altivec_ ## name(SwsContext *c, const unsigned char **in,          \
 284                             int *instrides, int srcSliceY, int srcSliceH,     \
 285                             unsigned char **oplanes, int *outstrides)         \
 286 {                                                                             \
 287     int w = c->srcW;                                                          \
 288     int h = srcSliceH;                                                        \
 289     int i, j;                                                                 \
 290     int instrides_scl[3];                                                     \
 291     vector unsigned char y0, y1;                                              \
 292                                                                               \
 293     vector signed char u, v;                                                  \
 294                                                                               \
 295     vector signed short Y0, Y1, Y2, Y3;                                       \
 296     vector signed short U, V;                                                 \
 297     vector signed short vx, ux, uvx;                                          \
 298     vector signed short vx0, ux0, uvx0;                                       \
 299     vector signed short vx1, ux1, uvx1;                                       \
 300     vector signed short R0, G0, B0;                                           \
 301     vector signed short R1, G1, B1;                                           \
 302     vector unsigned char R, G, B;                                             \
 303                                                                               \
 304     const vector unsigned char *y1ivP, *y2ivP, *uivP, *vivP;                  \
 305     vector unsigned char align_perm;                                          \
 306                                                                               \
 307     vector signed short lCY       = c->CY;                                    \
 308     vector signed short lOY       = c->OY;                                    \
 309     vector signed short lCRV      = c->CRV;                                   \
 310     vector signed short lCBU      = c->CBU;                                   \
 311     vector signed short lCGU      = c->CGU;                                   \
 312     vector signed short lCGV      = c->CGV;                                   \
 313     vector unsigned short lCSHIFT = c->CSHIFT;                                \
 314                                                                               \
 315     const ubyte *y1i = in[0];                                                 \
 316     const ubyte *y2i = in[0] + instrides[0];                                  \
 317     const ubyte *ui  = in[1];                                                 \
 318     const ubyte *vi  = in[2];                                                 \
 319                                                                               \
 320     vector unsigned char *oute =                                              \
 321         (vector unsigned char *)                                              \
 322             (oplanes[0] + srcSliceY * outstrides[0]);                         \
 323     vector unsigned char *outo =                                              \
 324         (vector unsigned char *)                                              \
 325             (oplanes[0] + srcSliceY * outstrides[0] + outstrides[0]);         \
 326                                                                               \
 327     /* loop moves y{1, 2}i by w */                                            \
 328     instrides_scl[0] = instrides[0] * 2 - w;                                  \
 329     /* loop moves ui by w / 2 */                                              \
 330     instrides_scl[1] = instrides[1] - w / 2;                                  \
 331     /* loop moves vi by w / 2 */                                              \
 332     instrides_scl[2] = instrides[2] - w / 2;                                  \
 333                                                                               \
 334     for (i = 0; i < h / 2; i++) {                                             \
 335         vec_dstst(outo, (0x02000002 | (((w * 3 + 32) / 32) << 16)), 0);       \
 336         vec_dstst(oute, (0x02000002 | (((w * 3 + 32) / 32) << 16)), 1);       \
 337                                                                               \
 338         for (j = 0; j < w / 16; j++) {                                        \
 339             y1ivP = (const vector unsigned char *) y1i;                       \
 340             y2ivP = (const vector unsigned char *) y2i;                       \
 341             uivP  = (const vector unsigned char *) ui;                        \
 342             vivP  = (const vector unsigned char *) vi;                        \
 343                                                                               \
 344             align_perm = vec_lvsl(0, y1i);                                    \
 345             y0 = (vector unsigned char)                                       \
 346                      vec_perm(y1ivP[0], y1ivP[1], align_perm);                \
 347                                                                               \
 348             align_perm = vec_lvsl(0, y2i);                                    \
 349             y1 = (vector unsigned char)                                       \
 350                      vec_perm(y2ivP[0], y2ivP[1], align_perm);                \
 351                                                                               \
 352             align_perm = vec_lvsl(0, ui);                                     \
 353             u = (vector signed char)                                          \
 354                     vec_perm(uivP[0], uivP[1], align_perm);                   \
 355                                                                               \
 356             align_perm = vec_lvsl(0, vi);                                     \
 357             v = (vector signed char)                                          \
 358                     vec_perm(vivP[0], vivP[1], align_perm);                   \
 359                                                                               \
 360             u = (vector signed char)                                          \
 361                     vec_sub(u,                                                \
 362                             (vector signed char)                              \
 363                                 vec_splat((vector signed char) { 128 }, 0));  \
 364             v = (vector signed char)                                          \
 365                     vec_sub(v,                                                \
 366                             (vector signed char)                              \
 367                                 vec_splat((vector signed char) { 128 }, 0));  \
 368                                                                               \
 369             U = vec_unpackh(u);                                               \
 370             V = vec_unpackh(v);                                               \
 371                                                                               \
 372             Y0 = vec_unh(y0);                                                 \
 373             Y1 = vec_unl(y0);                                                 \
 374             Y2 = vec_unh(y1);                                                 \
 375             Y3 = vec_unl(y1);                                                 \
 376                                                                               \
 377             Y0 = vec_mradds(Y0, lCY, lOY);                                    \
 378             Y1 = vec_mradds(Y1, lCY, lOY);                                    \
 379             Y2 = vec_mradds(Y2, lCY, lOY);                                    \
 380             Y3 = vec_mradds(Y3, lCY, lOY);                                    \
 381                                                                               \
 382             /* ux  = (CBU * (u << CSHIFT) + 0x4000) >> 15 */                  \
 383             ux  = vec_sl(U, lCSHIFT);                                         \
 384             ux  = vec_mradds(ux, lCBU, (vector signed short) { 0 });          \
 385             ux0 = vec_mergeh(ux, ux);                                         \
 386             ux1 = vec_mergel(ux, ux);                                         \
 387                                                                               \
 388             /* vx  = (CRV * (v << CSHIFT) + 0x4000) >> 15; */                 \
 389             vx  = vec_sl(V, lCSHIFT);                                         \
 390             vx  = vec_mradds(vx, lCRV, (vector signed short) { 0 });          \
 391             vx0 = vec_mergeh(vx, vx);                                         \
 392             vx1 = vec_mergel(vx, vx);                                         \
 393                                                                               \
 394             /* uvx = ((CGU * u) + (CGV * v)) >> 15 */                         \
 395             uvx  = vec_mradds(U, lCGU, (vector signed short) { 0 });          \
 396             uvx  = vec_mradds(V, lCGV, uvx);                                  \
 397             uvx0 = vec_mergeh(uvx, uvx);                                      \
 398             uvx1 = vec_mergel(uvx, uvx);                                      \
 399                                                                               \
 400             R0 = vec_add(Y0, vx0);                                            \
 401             G0 = vec_add(Y0, uvx0);                                           \
 402             B0 = vec_add(Y0, ux0);                                            \
 403             R1 = vec_add(Y1, vx1);                                            \
 404             G1 = vec_add(Y1, uvx1);                                           \
 405             B1 = vec_add(Y1, ux1);                                            \
 406                                                                               \
 407             R = vec_packclp(R0, R1);                                          \
 408             G = vec_packclp(G0, G1);                                          \
 409             B = vec_packclp(B0, B1);                                          \
 410                                                                               \
 411             out_pixels(R, G, B, oute);                                        \
 412                                                                               \
 413             R0 = vec_add(Y2, vx0);                                            \
 414             G0 = vec_add(Y2, uvx0);                                           \
 415             B0 = vec_add(Y2, ux0);                                            \
 416             R1 = vec_add(Y3, vx1);                                            \
 417             G1 = vec_add(Y3, uvx1);                                           \
 418             B1 = vec_add(Y3, ux1);                                            \
 419             R  = vec_packclp(R0, R1);                                         \
 420             G  = vec_packclp(G0, G1);                                         \
 421             B  = vec_packclp(B0, B1);                                         \
 422                                                                               \
 423                                                                               \
 424             out_pixels(R, G, B, outo);                                        \
 425                                                                               \
 426             y1i += 16;                                                        \
 427             y2i += 16;                                                        \
 428             ui  += 8;                                                         \
 429             vi  += 8;                                                         \
 430         }                                                                     \
 431                                                                               \
 432         outo += (outstrides[0]) >> 4;                                         \
 433         oute += (outstrides[0]) >> 4;                                         \
 434                                                                               \
 435         ui  += instrides_scl[1];                                              \
 436         vi  += instrides_scl[2];                                              \
 437         y1i += instrides_scl[0];                                              \
 438         y2i += instrides_scl[0];                                              \
 439     }                                                                         \
 440     return srcSliceH;                                                         \
 441 }
 442
 443 #define out_abgr(a, b, c, ptr)                                          \
 444     vec_mstrgb32(__typeof__(a), ((__typeof__(a)) { 255 }), c, b, a, ptr)
 445 #define out_bgra(a, b, c, ptr)                                          \
 446     vec_mstrgb32(__typeof__(a), c, b, a, ((__typeof__(a)) { 255 }), ptr)
 447 #define out_rgba(a, b, c, ptr)                                          \
 448     vec_mstrgb32(__typeof__(a), a, b, c, ((__typeof__(a)) { 255 }), ptr)
 449 #define out_argb(a, b, c, ptr)                                          \
 450     vec_mstrgb32(__typeof__(a), ((__typeof__(a)) { 255 }), a, b, c, ptr)
 451 #define out_rgb24(a, b, c, ptr) vec_mstrgb24(a, b, c, ptr)
 452 #define out_bgr24(a, b, c, ptr) vec_mstbgr24(a, b, c, ptr)
 453
 454 DEFCSP420_CVT(yuv2_abgr,  out_abgr)
 455 DEFCSP420_CVT(yuv2_bgra,  out_bgra)
 456 DEFCSP420_CVT(yuv2_rgba,  out_rgba)
 457 DEFCSP420_CVT(yuv2_argb,  out_argb)
 458 DEFCSP420_CVT(yuv2_rgb24, out_rgb24)
 459 DEFCSP420_CVT(yuv2_bgr24, out_bgr24)
 460
 461 // uyvy|uyvy|uyvy|uyvy
 462 // 0123 4567 89ab cdef
 463 static const vector unsigned char
 464     demux_u = { 0x10, 0x00, 0x10, 0x00,
 465                 0x10, 0x04, 0x10, 0x04,
 466                 0x10, 0x08, 0x10, 0x08,
 467                 0x10, 0x0c, 0x10, 0x0c },
 468     demux_v = { 0x10, 0x02, 0x10, 0x02,
 469                 0x10, 0x06, 0x10, 0x06,
 470                 0x10, 0x0A, 0x10, 0x0A,
 471                 0x10, 0x0E, 0x10, 0x0E },
 472     demux_y = { 0x10, 0x01, 0x10, 0x03,
 473                 0x10, 0x05, 0x10, 0x07,
 474                 0x10, 0x09, 0x10, 0x0B,
 475                 0x10, 0x0D, 0x10, 0x0F };
 476
 477 /*
 478  * this is so I can play live CCIR raw video
 479  */
 480 static int altivec_uyvy_rgb32(SwsContext *c, const unsigned char **in,
 481                               int *instrides, int srcSliceY, int srcSliceH,
 482                               unsigned char **oplanes, int *outstrides)
 483 {
 484     int w = c->srcW;
 485     int h = srcSliceH;
 486     int i, j;
 487     vector unsigned char uyvy;
 488     vector signed short Y, U, V;
 489     vector signed short R0, G0, B0, R1, G1, B1;
 490     vector unsigned char R, G, B;
 491     vector unsigned char *out;
 492     const ubyte *img;
 493
 494     img = in[0];
 495     out = (vector unsigned char *) (oplanes[0] + srcSliceY * outstrides[0]);
 496
 497     for (i = 0; i < h; i++)
 498         for (j = 0; j < w / 16; j++) {
 499             uyvy = vec_ld(0, img);
 500
 501             U = (vector signed short)
 502                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_u);
 503             V = (vector signed short)
 504                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_v);
 505             Y = (vector signed short)
 506                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_y);
 507
 508             cvtyuvtoRGB(c, Y, U, V, &R0, &G0, &B0);
 509
 510             uyvy = vec_ld(16, img);
 511
 512             U = (vector signed short)
 513                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_u);
 514             V = (vector signed short)
 515                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_v);
 516             Y = (vector signed short)
 517                     vec_perm(uyvy, (vector unsigned char) { 0 }, demux_y);
 518
 519             cvtyuvtoRGB(c, Y, U, V, &R1, &G1, &B1);
 520
 521             R = vec_packclp(R0, R1);
 522             G = vec_packclp(G0, G1);
 523             B = vec_packclp(B0, B1);
 524
 525             // vec_mstbgr24 (R,G,B, out);
 526             out_rgba(R, G, B, out);
 527
 528             img += 32;
 529         }
 530     return srcSliceH;
 531 }
 532
 533 /* Ok currently the acceleration routine only supports
 534  * inputs of widths a multiple of 16
 535  * and heights a multiple 2
 536  *
 537  * So we just fall back to the C codes for this.
 538  */
 539 av_cold SwsFunc ff_yuv2rgb_init_altivec(SwsContext *c)
 540 {
 541     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
 542         return NULL;
 543
 544     /*
 545      * and this seems not to matter too much I tried a bunch of
 546      * videos with abnormal widths and MPlayer crashes elsewhere.
 547      * mplayer -vo x11 -rawvideo on:w=350:h=240 raw-350x240.eyuv
 548      * boom with X11 bad match.
 549      *
 550      */
 551     if ((c->srcW & 0xf) != 0)
 552         return NULL;
 553
 554     switch (c->srcFormat) {
 555     case AV_PIX_FMT_YUV410P:
 556     case AV_PIX_FMT_YUV420P:
 557     /*case IMGFMT_CLPL:        ??? */
 558     case AV_PIX_FMT_GRAY8:
 559     case AV_PIX_FMT_NV12:
 560     case AV_PIX_FMT_NV21:
 561         if ((c->srcH & 0x1) != 0)
 562             return NULL;
 563
 564         switch (c->dstFormat) {
 565         case AV_PIX_FMT_RGB24:
 566             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space RGB24\n");
 567             return altivec_yuv2_rgb24;
 568         case AV_PIX_FMT_BGR24:
 569             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space BGR24\n");
 570             return altivec_yuv2_bgr24;
 571         case AV_PIX_FMT_ARGB:
 572             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space ARGB\n");
 573             return altivec_yuv2_argb;
 574         case AV_PIX_FMT_ABGR:
 575             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space ABGR\n");
 576             return altivec_yuv2_abgr;
 577         case AV_PIX_FMT_RGBA:
 578             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space RGBA\n");
 579             return altivec_yuv2_rgba;
 580         case AV_PIX_FMT_BGRA:
 581             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space BGRA\n");
 582             return altivec_yuv2_bgra;
 583         default: return NULL;
 584         }
 585         break;
 586
 587     case AV_PIX_FMT_UYVY422:
 588         switch (c->dstFormat) {
 589         case AV_PIX_FMT_BGR32:
 590             av_log(c, AV_LOG_WARNING, "ALTIVEC: Color Space UYVY -> RGB32\n");
 591             return altivec_uyvy_rgb32;
 592         default: return NULL;
 593         }
 594         break;
 595     }
 596     return NULL;
 597 }
 598
 599 av_cold void ff_yuv2rgb_init_tables_altivec(SwsContext *c,
 600                                             const int inv_table[4],
 601                                             int brightness,
 602                                             int contrast,
 603                                             int saturation)
 604 {
 605     union {
 606         DECLARE_ALIGNED(16, signed short, tmp)[8];
 607         vector signed short vec;
 608     } buf;
 609
 610     buf.tmp[0] = ((0xffffLL) * contrast >> 8) >> 9;                               // cy
 611     buf.tmp[1] = -256 * brightness;                                               // oy
 612     buf.tmp[2] =   (inv_table[0] >> 3) * (contrast >> 16) * (saturation >> 16);   // crv
 613     buf.tmp[3] =   (inv_table[1] >> 3) * (contrast >> 16) * (saturation >> 16);   // cbu
 614     buf.tmp[4] = -((inv_table[2] >> 1) * (contrast >> 16) * (saturation >> 16));  // cgu
 615     buf.tmp[5] = -((inv_table[3] >> 1) * (contrast >> 16) * (saturation >> 16));  // cgv
 616
 617     c->CSHIFT = (vector unsigned short) vec_splat_u16(2);
 618     c->CY     = vec_splat((vector signed short) buf.vec, 0);
 619     c->OY     = vec_splat((vector signed short) buf.vec, 1);
 620     c->CRV    = vec_splat((vector signed short) buf.vec, 2);
 621     c->CBU    = vec_splat((vector signed short) buf.vec, 3);
 622     c->CGU    = vec_splat((vector signed short) buf.vec, 4);
 623     c->CGV    = vec_splat((vector signed short) buf.vec, 5);
 624     return;
 625 }
 626
 627 static av_always_inline void ff_yuv2packedX_altivec(SwsContext *c,
 628                                                     const int16_t *lumFilter,
 629                                                     const int16_t **lumSrc,
 630                                                     int lumFilterSize,
 631                                                     const int16_t *chrFilter,
 632                                                     const int16_t **chrUSrc,
 633                                                     const int16_t **chrVSrc,
 634                                                     int chrFilterSize,
 635                                                     const int16_t **alpSrc,
 636                                                     uint8_t *dest,
 637                                                     int dstW, int dstY,
 638                                                     enum AVPixelFormat target)
 639 {
 640     int i, j;
 641     vector signed short X, X0, X1, Y0, U0, V0, Y1, U1, V1, U, V;
 642     vector signed short R0, G0, B0, R1, G1, B1;
 643
 644     vector unsigned char R, G, B;
 645     vector unsigned char *out, *nout;
 646
 647     vector signed short RND   = vec_splat_s16(1 << 3);
 648     vector unsigned short SCL = vec_splat_u16(4);
 649     DECLARE_ALIGNED(16, unsigned int, scratch)[16];
 650
 651     vector signed short *YCoeffs, *CCoeffs;
 652
 653     YCoeffs = c->vYCoeffsBank + dstY * lumFilterSize;
 654     CCoeffs = c->vCCoeffsBank + dstY * chrFilterSize;
 655
 656     out = (vector unsigned char *) dest;
 657
 658     for (i = 0; i < dstW; i += 16) {
 659         Y0 = RND;
 660         Y1 = RND;
 661         /* extract 16 coeffs from lumSrc */
 662         for (j = 0; j < lumFilterSize; j++) {
 663             X0 = vec_ld(0, &lumSrc[j][i]);
 664             X1 = vec_ld(16, &lumSrc[j][i]);
 665             Y0 = vec_mradds(X0, YCoeffs[j], Y0);
 666             Y1 = vec_mradds(X1, YCoeffs[j], Y1);
 667         }
 668
 669         U = RND;
 670         V = RND;
 671         /* extract 8 coeffs from U,V */
 672         for (j = 0; j < chrFilterSize; j++) {
 673             X = vec_ld(0, &chrUSrc[j][i / 2]);
 674             U = vec_mradds(X, CCoeffs[j], U);
 675             X = vec_ld(0, &chrVSrc[j][i / 2]);
 676             V = vec_mradds(X, CCoeffs[j], V);
 677         }
 678
 679         /* scale and clip signals */
 680         Y0 = vec_sra(Y0, SCL);
 681         Y1 = vec_sra(Y1, SCL);
 682         U  = vec_sra(U, SCL);
 683         V  = vec_sra(V, SCL);
 684
 685         Y0 = vec_clip_s16(Y0);
 686         Y1 = vec_clip_s16(Y1);
 687         U  = vec_clip_s16(U);
 688         V  = vec_clip_s16(V);
 689
 690         /* now we have
 691          * Y0 = y0 y1 y2 y3 y4 y5 y6 y7    Y1 = y8 y9 y10 y11 y12 y13 y14 y15
 692          * U  = u0 u1 u2 u3 u4 u5 u6 u7    V  = v0 v1 v2 v3 v4 v5 v6 v7
 693          *
 694          * Y0 = y0 y1 y2 y3 y4 y5 y6 y7    Y1 = y8 y9 y10 y11 y12 y13 y14 y15
 695          * U0 = u0 u0 u1 u1 u2 u2 u3 u3    U1 = u4 u4 u5 u5 u6 u6 u7 u7
 696          * V0 = v0 v0 v1 v1 v2 v2 v3 v3    V1 = v4 v4 v5 v5 v6 v6 v7 v7
 697          */
 698
 699         U0 = vec_mergeh(U, U);
 700         V0 = vec_mergeh(V, V);
 701
 702         U1 = vec_mergel(U, U);
 703         V1 = vec_mergel(V, V);
 704
 705         cvtyuvtoRGB(c, Y0, U0, V0, &R0, &G0, &B0);
 706         cvtyuvtoRGB(c, Y1, U1, V1, &R1, &G1, &B1);
 707
 708         R = vec_packclp(R0, R1);
 709         G = vec_packclp(G0, G1);
 710         B = vec_packclp(B0, B1);
 711
 712         switch (target) {
 713         case AV_PIX_FMT_ABGR:
 714             out_abgr(R, G, B, out);
 715             break;
 716         case AV_PIX_FMT_BGRA:
 717             out_bgra(R, G, B, out);
 718             break;
 719         case AV_PIX_FMT_RGBA:
 720             out_rgba(R, G, B, out);
 721             break;
 722         case AV_PIX_FMT_ARGB:
 723             out_argb(R, G, B, out);
 724             break;
 725         case AV_PIX_FMT_RGB24:
 726             out_rgb24(R, G, B, out);
 727             break;
 728         case AV_PIX_FMT_BGR24:
 729             out_bgr24(R, G, B, out);
 730             break;
 731         default:
 732         {
 733             /* If this is reached, the caller should have called yuv2packedXinC
 734              * instead. */
 735             static int printed_error_message;
 736             if (!printed_error_message) {
 737                 av_log(c, AV_LOG_ERROR,
 738                        "altivec_yuv2packedX doesn't support %s output\n",
 739                        sws_format_name(c->dstFormat));
 740                 printed_error_message = 1;
 741             }
 742             return;
 743         }
 744         }
 745     }
 746
 747     if (i < dstW) {
 748         i -= 16;
 749
 750         Y0 = RND;
 751         Y1 = RND;
 752         /* extract 16 coeffs from lumSrc */
 753         for (j = 0; j < lumFilterSize; j++) {
 754             X0 = vec_ld(0, &lumSrc[j][i]);
 755             X1 = vec_ld(16, &lumSrc[j][i]);
 756             Y0 = vec_mradds(X0, YCoeffs[j], Y0);
 757             Y1 = vec_mradds(X1, YCoeffs[j], Y1);
 758         }
 759
 760         U = RND;
 761         V = RND;
 762         /* extract 8 coeffs from U,V */
 763         for (j = 0; j < chrFilterSize; j++) {
 764             X = vec_ld(0, &chrUSrc[j][i / 2]);
 765             U = vec_mradds(X, CCoeffs[j], U);
 766             X = vec_ld(0, &chrVSrc[j][i / 2]);
 767             V = vec_mradds(X, CCoeffs[j], V);
 768         }
 769
 770         /* scale and clip signals */
 771         Y0 = vec_sra(Y0, SCL);
 772         Y1 = vec_sra(Y1, SCL);
 773         U  = vec_sra(U, SCL);
 774         V  = vec_sra(V, SCL);
 775
 776         Y0 = vec_clip_s16(Y0);
 777         Y1 = vec_clip_s16(Y1);
 778         U  = vec_clip_s16(U);
 779         V  = vec_clip_s16(V);
 780
 781         /* now we have
 782          * Y0 = y0 y1 y2 y3 y4 y5 y6 y7    Y1 = y8 y9 y10 y11 y12 y13 y14 y15
 783          * U  = u0 u1 u2 u3 u4 u5 u6 u7    V  = v0 v1 v2 v3 v4 v5 v6 v7
 784          *
 785          * Y0 = y0 y1 y2 y3 y4 y5 y6 y7    Y1 = y8 y9 y10 y11 y12 y13 y14 y15
 786          * U0 = u0 u0 u1 u1 u2 u2 u3 u3    U1 = u4 u4 u5 u5 u6 u6 u7 u7
 787          * V0 = v0 v0 v1 v1 v2 v2 v3 v3    V1 = v4 v4 v5 v5 v6 v6 v7 v7
 788          */
 789
 790         U0 = vec_mergeh(U, U);
 791         V0 = vec_mergeh(V, V);
 792
 793         U1 = vec_mergel(U, U);
 794         V1 = vec_mergel(V, V);
 795
 796         cvtyuvtoRGB(c, Y0, U0, V0, &R0, &G0, &B0);
 797         cvtyuvtoRGB(c, Y1, U1, V1, &R1, &G1, &B1);
 798
 799         R = vec_packclp(R0, R1);
 800         G = vec_packclp(G0, G1);
 801         B = vec_packclp(B0, B1);
 802
 803         nout = (vector unsigned char *) scratch;
 804         switch (target) {
 805         case AV_PIX_FMT_ABGR:
 806             out_abgr(R, G, B, nout);
 807             break;
 808         case AV_PIX_FMT_BGRA:
 809             out_bgra(R, G, B, nout);
 810             break;
 811         case AV_PIX_FMT_RGBA:
 812             out_rgba(R, G, B, nout);
 813             break;
 814         case AV_PIX_FMT_ARGB:
 815             out_argb(R, G, B, nout);
 816             break;
 817         case AV_PIX_FMT_RGB24:
 818             out_rgb24(R, G, B, nout);
 819             break;
 820         case AV_PIX_FMT_BGR24:
 821             out_bgr24(R, G, B, nout);
 822             break;
 823         default:
 824             /* Unreachable, I think. */
 825             av_log(c, AV_LOG_ERROR,
 826                    "altivec_yuv2packedX doesn't support %s output\n",
 827                    sws_format_name(c->dstFormat));
 828             return;
 829         }
 830
 831         memcpy(&((uint32_t *) dest)[i], scratch, (dstW - i) / 4);
 832     }
 833 }
 834
 835 #define YUV2PACKEDX_WRAPPER(suffix, pixfmt)                             \
 836 void ff_yuv2 ## suffix ## _X_altivec(SwsContext *c,                     \
 837                                      const int16_t *lumFilter,          \
 838                                      const int16_t **lumSrc,            \
 839                                      int lumFilterSize,                 \
 840                                      const int16_t *chrFilter,          \
 841                                      const int16_t **chrUSrc,           \
 842                                      const int16_t **chrVSrc,           \
 843                                      int chrFilterSize,                 \
 844                                      const int16_t **alpSrc,            \
 845                                      uint8_t *dest, int dstW, int dstY) \
 846 {                                                                       \
 847     ff_yuv2packedX_altivec(c, lumFilter, lumSrc, lumFilterSize,         \
 848                            chrFilter, chrUSrc, chrVSrc,                 \
 849                            chrFilterSize, alpSrc,                       \
 850                            dest, dstW, dstY, pixfmt);                   \
 851 }
 852
 853 YUV2PACKEDX_WRAPPER(abgr,  AV_PIX_FMT_ABGR);
 854 YUV2PACKEDX_WRAPPER(bgra,  AV_PIX_FMT_BGRA);
 855 YUV2PACKEDX_WRAPPER(argb,  AV_PIX_FMT_ARGB);
 856 YUV2PACKEDX_WRAPPER(rgba,  AV_PIX_FMT_RGBA);
 857 YUV2PACKEDX_WRAPPER(rgb24, AV_PIX_FMT_RGB24);
 858 YUV2PACKEDX_WRAPPER(bgr24, AV_PIX_FMT_BGR24);