comparison ppc/h264_template_altivec.c @ 5603:861eb234e6ba libavcodec

remove alignment correction of the destination pointers in luma_16x6 interpolations, since they are always 16-bytes aligned in practice. Add asserts to ease narrowing down potential image corructions on exotic plateforms Based on a patch by Mauricio Alvarez % lokifo A gmail P com % Original thread: Date: Jun 26, 2007 1:07 PM Subject: Re: [FFmpeg-devel] [PATCH] h264 luma interpolation 8x8 for altivec

5602:3b21f3268707

 *
 * You should have received a copy of the GNU Lesser 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
 */

/* this code assume that stride % 16 == 0 */
void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) {
  POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1);
    DECLARE_ALIGNED_16(signed int, ABCD[4]) =

  const vec_u8_t permP3 = vec_lvsl(+3, src);
  const vec_s16_t v5ss = vec_splat_s16(5);
  const vec_u16_t v5us = vec_splat_u16(5);
  const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
  const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
  const vec_u8_t dstperm = vec_lvsr(0, dst);
  const vec_u8_t neg1 = (const vec_u8_t) vec_splat_s8(-1);
  const vec_u8_t dstmask = vec_perm(zero_u8v, neg1, dstperm);

  vec_u8_t srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;

  register int align = ((((unsigned long)src) - 2) % 16);

                      srcM1A, srcM1B, srcM2A, srcM2B,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
                      pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
                      psumA, psumB, sumA, sumB;

  vec_u8_t sum, dst1, dst2, vdst, fsum, rsum, fdst1, fdst2;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);

  for (i = 0 ; i < 16 ; i ++) {
    vec_u8_t srcR1 = vec_ld(-2, src);

    sumA = vec_sra(psumA, v5us);
    sumB = vec_sra(psumB, v5us);

    sum = vec_packsu(sumA, sumB);

    dst1 = vec_ld(0, dst);
    dst2 = vec_ld(16, dst);
    vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));

    OP_U8_ALTIVEC(fsum, sum, vdst);

    rsum = vec_perm(fsum, fsum, dstperm);
    fdst1 = vec_sel(dst1, rsum, dstmask);
    fdst2 = vec_sel(rsum, dst2, dstmask);

    vec_st(fdst1, 0, dst);
    vec_st(fdst2, 16, dst);

    src += srcStride;
    dst += dstStride;
  }
POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);

  const vec_u8_t perm = vec_lvsl(0, src);
  const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
  const vec_u16_t v5us = vec_splat_u16(5);
  const vec_s16_t v5ss = vec_splat_s16(5);
  const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
  const vec_u8_t dstperm = vec_lvsr(0, dst);
  const vec_u8_t neg1 = (const vec_u8_t)vec_splat_s8(-1);
  const vec_u8_t dstmask = vec_perm(zero_u8v, neg1, dstperm);

  uint8_t *srcbis = src - (srcStride * 2);

  const vec_u8_t srcM2a = vec_ld(0, srcbis);
  const vec_u8_t srcM2b = vec_ld(16, srcbis);

  vec_s16_t pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
                      psumA, psumB, sumA, sumB,
                      srcP3ssA, srcP3ssB,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;

  vec_u8_t sum, dst1, dst2, vdst, fsum, rsum, fdst1, fdst2, srcP3a, srcP3b, srcP3;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);

  for (i = 0 ; i < 16 ; i++) {
    srcP3a = vec_ld(0, srcbis += srcStride);

    sumA = vec_sra(psumA, v5us);
    sumB = vec_sra(psumB, v5us);

    sum = vec_packsu(sumA, sumB);

    dst1 = vec_ld(0, dst);
    dst2 = vec_ld(16, dst);
    vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));

    OP_U8_ALTIVEC(fsum, sum, vdst);

    rsum = vec_perm(fsum, fsum, dstperm);
    fdst1 = vec_sel(dst1, rsum, dstmask);
    fdst2 = vec_sel(rsum, dst2, dstmask);

    vec_st(fdst1, 0, dst);
    vec_st(fdst2, 16, dst);

    dst += dstStride;
  }
  POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
}

  const vec_s32_t v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
  const vec_u32_t v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));

  register int align = ((((unsigned long)src) - 2) % 16);

  const vec_u8_t neg1 = (const vec_u8_t) vec_splat_s8(-1);

  vec_s16_t srcP0A, srcP0B, srcP1A, srcP1B,
                      srcP2A, srcP2B, srcP3A, srcP3B,
                      srcM1A, srcM1B, srcM2A, srcM2B,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
                      pp1A, pp1B, pp2A, pp2B, psumA, psumB;

  const vec_u8_t dstperm = vec_lvsr(0, dst);

  const vec_u8_t dstmask = vec_perm(zero_u8v, neg1, dstperm);

  const vec_u8_t mperm = (const vec_u8_t)
    AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
        0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F);
  int16_t *tmpbis = tmp;

  vec_s32_t pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
                    pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
                    pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
                    ssumAe, ssumAo, ssumBe, ssumBo;
  vec_u8_t fsum, sumv, sum, dst1, dst2, vdst, rsum, fdst1, fdst2;
  vec_s16_t ssume, ssumo;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
  src -= (2 * srcStride);
  for (i = 0 ; i < 21 ; i ++) {

    ssumo = vec_packs(ssumAo, ssumBo);

    sumv = vec_packsu(ssume, ssumo);
    sum = vec_perm(sumv, sumv, mperm);

    dst1 = vec_ld(0, dst);
    dst2 = vec_ld(16, dst);
    vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));

    OP_U8_ALTIVEC(fsum, sum, vdst);

    rsum = vec_perm(fsum, fsum, dstperm);
    fdst1 = vec_sel(dst1, rsum, dstmask);
    fdst2 = vec_sel(rsum, dst2, dstmask);

    vec_st(fdst1, 0, dst);
    vec_st(fdst2, 16, dst);

    dst += dstStride;
  }
  POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
}

5603:861eb234e6ba

 *
 * You should have received a copy of the GNU Lesser 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
 */

//#define DEBUG_ALIGNMENT
#ifdef DEBUG_ALIGNMENT
#define ASSERT_ALIGNED(ptr) assert(((unsigned long)ptr&0x0000000F));
#else
#define ASSERT_ALIGNED(ptr) ;
#endif

/* this code assume that stride % 16 == 0 */
void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) {
  POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1);
    DECLARE_ALIGNED_16(signed int, ABCD[4]) =

  const vec_u8_t permP3 = vec_lvsl(+3, src);
  const vec_s16_t v5ss = vec_splat_s16(5);
  const vec_u16_t v5us = vec_splat_u16(5);
  const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
  const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));

  vec_u8_t srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;

  register int align = ((((unsigned long)src) - 2) % 16);

                      srcM1A, srcM1B, srcM2A, srcM2B,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
                      pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
                      psumA, psumB, sumA, sumB;

  vec_u8_t sum, vdst, fsum;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);

  for (i = 0 ; i < 16 ; i ++) {
    vec_u8_t srcR1 = vec_ld(-2, src);

    sumA = vec_sra(psumA, v5us);
    sumB = vec_sra(psumB, v5us);

    sum = vec_packsu(sumA, sumB);

    ASSERT_ALIGNED(dst);
    vdst = vec_ld(0, dst);

    OP_U8_ALTIVEC(fsum, sum, vdst);

    vec_st(fsum, 0, dst);

    src += srcStride;
    dst += dstStride;
  }
POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);

  const vec_u8_t perm = vec_lvsl(0, src);
  const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
  const vec_u16_t v5us = vec_splat_u16(5);
  const vec_s16_t v5ss = vec_splat_s16(5);
  const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));

  uint8_t *srcbis = src - (srcStride * 2);

  const vec_u8_t srcM2a = vec_ld(0, srcbis);
  const vec_u8_t srcM2b = vec_ld(16, srcbis);

  vec_s16_t pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
                      psumA, psumB, sumA, sumB,
                      srcP3ssA, srcP3ssB,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;

  vec_u8_t sum, vdst, fsum, srcP3a, srcP3b, srcP3;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);

  for (i = 0 ; i < 16 ; i++) {
    srcP3a = vec_ld(0, srcbis += srcStride);

    sumA = vec_sra(psumA, v5us);
    sumB = vec_sra(psumB, v5us);

    sum = vec_packsu(sumA, sumB);

    ASSERT_ALIGNED(dst);
    vdst = vec_ld(0, dst);

    OP_U8_ALTIVEC(fsum, sum, vdst);

    vec_st(fsum, 0, dst);

    dst += dstStride;
  }
  POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
}

  const vec_s32_t v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
  const vec_u32_t v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));

  register int align = ((((unsigned long)src) - 2) % 16);

  vec_s16_t srcP0A, srcP0B, srcP1A, srcP1B,
                      srcP2A, srcP2B, srcP3A, srcP3B,
                      srcM1A, srcM1B, srcM2A, srcM2B,
                      sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
                      pp1A, pp1B, pp2A, pp2B, psumA, psumB;

  const vec_u8_t mperm = (const vec_u8_t)
    AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
        0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F);
  int16_t *tmpbis = tmp;

  vec_s32_t pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
                    pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
                    pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
                    ssumAe, ssumAo, ssumBe, ssumBo;
  vec_u8_t fsum, sumv, sum, vdst;
  vec_s16_t ssume, ssumo;

  POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
  src -= (2 * srcStride);
  for (i = 0 ; i < 21 ; i ++) {

    ssumo = vec_packs(ssumAo, ssumBo);

    sumv = vec_packsu(ssume, ssumo);
    sum = vec_perm(sumv, sumv, mperm);

    ASSERT_ALIGNED(dst);
    vdst = vec_ld(0, dst);

    OP_U8_ALTIVEC(fsum, sum, vdst);

    vec_st(fsum, 0, dst);

    dst += dstStride;
  }
  POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
}