comparison postprocess_altivec_template.c @ 63:650554bacd12 libpostproc

spelling

62:aac097949f15

  /*
    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. Quite a lot of load/stores
    can be removed by assuming proper alignement of
    src & stride :-(
  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  const int properStride = (stride % 16);

  /*
    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. Quite a lot of load/stores
    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]);

                                                    dclamp_P,
                                                    vec_cmpgt(q, (vector signed short)zero));
  const vector signed short dornotd = vec_sel((vector signed short)zero,
                                              dclampedfinal,
                                              vec_cmplt(absmE, vqp));
  /* add/substract to l4 and l5 */
  const vector signed short vb4minusd = vec_sub(vb4, dornotd);
  const vector signed short vb5plusd = vec_add(vb5, dornotd);
  /* finally, stores */
  const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
  const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);

  /*
    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. 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);

63:650554bacd12

  /*
    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. Quite a lot of load/stores
    can be removed by assuming proper alignment of
    src & stride :-(
  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  const int properStride = (stride % 16);

  /*
    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. Quite a lot of load/stores
    can be removed by assuming proper alignment of
    src & stride :-(
  */
  uint8_t *src2 = src;
  const vector signed int zero = vec_splat_s32(0);
  DECLARE_ALIGNED(16, short, qp[8]);

                                                    dclamp_P,
                                                    vec_cmpgt(q, (vector signed short)zero));
  const vector signed short dornotd = vec_sel((vector signed short)zero,
                                              dclampedfinal,
                                              vec_cmplt(absmE, vqp));
  /* add/subtract to l4 and l5 */
  const vector signed short vb4minusd = vec_sub(vb4, dornotd);
  const vector signed short vb5plusd = vec_add(vb5, dornotd);
  /* finally, stores */
  const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
  const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);

  /*
    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. Quite a lot of load/stores
    can be removed by assuming proper alignment of
    src & stride :-(
  */
  uint8_t *srcCopy = src;
  DECLARE_ALIGNED(16, uint8_t, dt[16]);
  const vector signed int zero = vec_splat_s32(0);