comparison postprocess_altivec_template.c @ 41:8dc2a966afa7 libpostproc

Many steps to avutilized this mess: DECLARE_ALIGNED

40:7bd9f7dcef91

 * You should have received a copy of the GNU General Public License
 * along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */


#ifdef CONFIG_DARWIN
#define AVV(x...) (x)
#else
#define AVV(x...) {x}

    this code makes no assumption on src or stride.
    One could remove the recomputation of the perm
    vector by assuming (stride % 16) == 0, unfortunately
    this is not always true.
  */
  short __attribute__ ((aligned(16))) data[8];
  int numEq;
  uint8_t *src2 = src;
  vector signed short v_dcOffset;
  vector signed short v2QP;
  vector unsigned short v4QP;

  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  const int properStride = (stride % 16);
  const int srcAlign = ((unsigned long)src2 % 16);
  short __attribute__ ((aligned(16))) qp[8];
  qp[0] = c->QP;
  vector signed short vqp = vec_ld(0, qp);
  vqp = vec_splat(vqp, 0);

  src2 += stride*3;

    can be removed by assuming proper alignement of
    src & stride :-(
  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  short __attribute__ ((aligned(16))) qp[8];
  qp[0] = 8*c->QP;
  vector signed short vqp = vec_ld(0, qp);
  vqp = vec_splat(vqp, 0);

#define LOAD_LINE(i)                                                    \

    this is not always true. Quite a lot of load/stores
    can be removed by assuming proper alignement of
    src & stride :-(
  */
  uint8_t *srcCopy = src;
  uint8_t __attribute__((aligned(16))) dt[16];
  const vector signed int zero = vec_splat_s32(0);
  vector unsigned char v_dt;
  dt[0] = deringThreshold;
  v_dt = vec_splat(vec_ld(0, dt), 0);

      return;

    v_avg = vec_avg(v_min, v_max);
  }

  signed int __attribute__((aligned(16))) S[8];
  {
    const vector unsigned short mask1 = (vector unsigned short)
      AVV(0x0001, 0x0002, 0x0004, 0x0008,
          0x0010, 0x0020, 0x0040, 0x0080);
    const vector unsigned short mask2 = (vector unsigned short)

  }

  /* I'm not sure the following is actually faster
     than straight, unvectorized C code :-( */

  int __attribute__((aligned(16))) tQP2[4];
  tQP2[0]= c->QP/2 + 1;
  vector signed int vQP2 = vec_ld(0, tQP2);
  vQP2 = vec_splat(vQP2, 0);
  const vector signed int vsint32_8 = vec_splat_s32(8);
  const vector unsigned int vuint32_4 = vec_splat_u32(4);

41:8dc2a966afa7

 * You should have received a copy of the GNU General Public License
 * along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <avutil.h>

#ifdef CONFIG_DARWIN
#define AVV(x...) (x)
#else
#define AVV(x...) {x}

    this code makes no assumption on src or stride.
    One could remove the recomputation of the perm
    vector by assuming (stride % 16) == 0, unfortunately
    this is not always true.
  */
  DECLARE_ALIGNED(16, short, data[8]);
  int numEq;
  uint8_t *src2 = src;
  vector signed short v_dcOffset;
  vector signed short v2QP;
  vector unsigned short v4QP;

  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  const int properStride = (stride % 16);
  const int srcAlign = ((unsigned long)src2 % 16);
  DECLARE_ALIGNED(16, short, qp[8]);
  qp[0] = c->QP;
  vector signed short vqp = vec_ld(0, qp);
  vqp = vec_splat(vqp, 0);

  src2 += stride*3;

    can be removed by assuming proper alignement of
    src & stride :-(
  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  DECLARE_ALIGNED(16, short, qp[8]);
  qp[0] = 8*c->QP;
  vector signed short vqp = vec_ld(0, qp);
  vqp = vec_splat(vqp, 0);

#define LOAD_LINE(i)                                                    \

    this is not always true. Quite a lot of load/stores
    can be removed by assuming proper alignement of
    src & stride :-(
  */
  uint8_t *srcCopy = src;
  DECLARE_ALIGNED(16, uint8_t, dt[16]);
  const vector signed int zero = vec_splat_s32(0);
  vector unsigned char v_dt;
  dt[0] = deringThreshold;
  v_dt = vec_splat(vec_ld(0, dt), 0);

      return;

    v_avg = vec_avg(v_min, v_max);
  }

  DECLARE_ALIGNED(16, signed int, S[8]);
  {
    const vector unsigned short mask1 = (vector unsigned short)
      AVV(0x0001, 0x0002, 0x0004, 0x0008,
          0x0010, 0x0020, 0x0040, 0x0080);
    const vector unsigned short mask2 = (vector unsigned short)

  }

  /* I'm not sure the following is actually faster
     than straight, unvectorized C code :-( */

  DECLARE_ALIGNED(16, int, tQP2[4]);
  tQP2[0]= c->QP/2 + 1;
  vector signed int vQP2 = vec_ld(0, tQP2);
  vQP2 = vec_splat(vQP2, 0);
  const vector signed int vsint32_8 = vec_splat_s32(8);
  const vector unsigned int vuint32_4 = vec_splat_u32(4);