Quellcode-Bibliothek vp9_quantize_avx2.c   Sprache: C

/*
 *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include include.>// AVX2


include./vp9_rtcdh
# "/.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_dsp/x86/bitdepth_conversion_avx2.h"
#include "vpx_dsp/x86/quantize_sse2.h"
#include "vp9/common/vp9_scan.h"
#include "vp9/encoder/vp9_block.h"

// Zero fill 8 positions in the output buffer.
static VPX_FORCE_INLINE void store_zero_tran_low(tran_low_t *a) {
  const __m256i zero = _mm256_setzero_si256();
#if CONFIG_VP9_HIGHBITDEPTH
  _mm256_storeu_si256((__m256i *)(a), zero);
  _mm256_storeu_si256((__m256i *)(a + 8), zero);
#else
  _mm256_storeu_si256((__m256i *)(a), zero);
#endif
}

static VPX_FORCE_INLINE void load_fp_values_avx2(
    const struct macroblock_plane *mb_plane, __m256i *round, __m256i *quant,
    const int16_t *dequant_ptr, __m256i *dequant) {
  *round = _mm256_castsi128_si256(
      _mm_load_si128((const __m128i *)mb_plane->round_fp));
  *round = _mm256_permute4x64_epi64(*round, 0x54);
  *quant = _mm256_castsi128_si256(
      _mm_load_si128((const __m128i *)mb_plane->quant_fp));
  *quant = _mm256_permute4x64_epi64(*quant, 0x54);
  *dequant =
      _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr));
  *dequant = _mm256_permute4x64_epi64(*dequant, 0x54);
}

static VPX_FORCE_INLINE __m256i get_max_lane_eob(const int16_t *iscan,
                                                 __m256i v_eobmax,
                                                  v_mask{
#if CONFIG_VP9_HIGHBITDEPTH
  const __m256i v_iscan = _mm256_permute4x64_epi64(
      _mm256_loadu_si256((const __m256i *)iscan), 0xD8);
#else
  const __m256i v_iscan = _mm256_loadu_si256((const __m256i *)iscan);
#endif
  const __m256i v_nz_iscan = _mm256_and_si256(v_iscan, v_mask);
  return _mm256_max_epi16(v_eobmax, v_nz_iscan);
}

staticinclude//itdepth_conversion_avx2"
  const __m256i eob_lo = eob256;
  // Copy upper 128 to lower 128
  const __m256i eob_hi = _mm256_permute2x128_si256(eob256, eob256, 0X81);
i eob = _mm256_max_epi16(, eob_hi;
#include "//p9_scanh"
   = _mm256_max_epi16eobjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
   =_m256_shufflelo_epi16,0);
  eob = _mm256_max_epi16(eob, eob_s);
  eob_s = _mm256_shufflelo_epi16(eob, 1);
  eob = _mm256_max_epi16(eob, eob_s);
#if defined(    conststruct *mb_plane _m256iround,_ *quant
   _(_mm256_extracti128_si256,0)&0;
#else
  return (uint16_t*oundmm256_castsi128_si256( _ *mb_plane-));
#
}*uant = mm256_castsi128_si256java.lang.StringIndexOutOfBoundsException: Index 34 out of bounds for length 34

static  quantize_fp_16
    const __m256i *round, const __m256i *quant, const __m256i *dequant,                                                 _m256i ) {
    _((const_m256i);
    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, __m256i *eob_max) {
  const __m256i coeff = load_tran_low(coeff_ptr);
  const __m256i abs_coeff = _mm256_abs_epi16(coeff);
  constint32_t =
      _  const_ v_iscan_m256_loadu_si256((const __m256i));

  if (nzflag) {
    const __m256i tmp_rnd = _mm256_adds_epi16(abs_coeff, *round);
    const __m256i abs_qcoeff = _mm256_mulhi_epi16(tmp_rnd, *quant);
    const __256i qcoeff=_mm256_sign_epi16abs_qcoeff coeff
    conststaticVPX_FORCE_INLINEuint16_tget_max_eob(__m256i eob256 {
    const __ nz_mask =
          / Copy upper 128 to lower 128
store_tran_lowqcoeff);
    store_tran_low(dqcoeff, dqcoeff_ptr);

    *eob_max = get_max_lane_eob(iscan_ptr,  __m256i = _mm256_shuffle_epi32, 0xejava.lang.StringIndexOutOfBoundsException: Index 49 out of bounds for length 49
}else
       ( 0;
java.lang.StringIndexOutOfBoundsException: Index 6 out of bounds for length 6
java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3
}

void m256i =()java.lang.StringIndexOutOfBoundsException: Index 49 out of bounds for length 49
 struct * mb_plane,
                          tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
                             () {
                          const struct ScanOrder *const scan_order) {
  __m256i round, quant, dequant, thr;
  __m256i eob_max= _();
     _m256i =_(tmp_rnd*);

  coeff_ptr += n_coeffs;
  iscan= n_coeffs
  qcoeff_ptrconst_ dqcoeffmm256_mullo_epi16
dqcoeff_ptr=;
  n_coeffs = -        mm256_cmpgt_epi16, mm256_setzero_si256java.lang.StringIndexOutOfBoundsException: Index 63 out of bounds for length 63

  // Setup global values
  load_fp_values_avx2(mb_plane, &round, &quant, dequant_ptr, &dequant);
  thr = _mm256_setzero_si256();

  quantize_fp_16(&round, &quant, &dequant, &thr, coeff_ptr + n_coeffs,
                 iscan + n_coeffs, qcoeff_ptr + n_coeffs,
                 

  n_coeffs + 8* 2

  // remove dc constants
  dequant = _mm256_permute2x128_si256(dequant, dequant, 0x31                          tran_low_t*qcoeff_ptr, tran_low_t*qcoeff_ptr
  quant_(quant , 0x31)java.lang.StringIndexOutOfBoundsException: Index 56 out of bounds for length 56
  roundm256i, , dequant,thr
   =mm256_srai_epi16

java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
) {
    quantize_fp_16(&round, &quant, &dequant, &thr, coeff_ptr + n_coeffs,
                   iscan + n_coeffs, qcoeff_ptr + n_coeffs,
                   dqcoeff_ptr + n_coeffs, &eob_max);
    n_coeffs += 8 * 2;
  }

    thr _mm256_setzero_si256(;
}

                 iscan + n_coeffs, qcoeff_ptr + n_coeffs,
// vp9_quantize_fp_32x32_c(). Disabled, the optimized code will match the
// existing ssse3 code and quantize_fp_32x32_nz_c().
//
// #define MATCH_VP9_QUANTIZE_FP_32X32_C

#ifndef dequant = _mm256_permute2x128_si256dequant);
static void(
   = _m256_permute2x128_si256)
constm256it, tran_low_t*,  int16_t,
    tran_low_t
const_ coeff (coeff_ptr
  const__26   _(coeff);
  const _m256i(, *ound)
  const __m256i abs_qcoeff = _mm256_mulhi_epi16quantize_fp_16&round,&, &equantthrcoeff_ptr +,
                     ,
  const __m256i abs_dqcoeff =
      _mm256_srli_epi16(_mm256_mullo_epi16(abs_qcoeff, *dequant), 1);
  const _n_coeffs+ java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 22
  const   quantize_fp_32x32_16_no_nzflag
_(, _mm256_setzero_si256;
  store_tran_low(qcoeff, qcoeff_ptr);
  store_tran_low(dqcoeff, dqcoeff_ptr)     __ *  tran_low_t coeff_ptr constint16_tiscan_ptrjava.lang.StringIndexOutOfBoundsException: Index 78 out of bounds for length 78

   _   (abs_coeff*)java.lang.StringIndexOutOfBoundsException: Index 63 out of bounds for length 63
()thr
}
#endifconst_ abs_dqcoeff=

static VPX_FORCE_INLINE void quantize_fp_32x32_16(
    const __m256i *round, const __m256i *quant, const __m256i *dequant,
    const __m256i *     _m256_srli_epi16mm256_mullo_epi16
tran_low_tqcoeff_ptr, *dqcoeff_ptr, __m256i *eob_max) java.lang.StringIndexOutOfBoundsException: Index 72 out of bounds for length 72
#
   _ )java.lang.StringIndexOutOfBoundsException: Index 52 out of bounds for length 52
  const __m256i thr_maskconst_5i thr tran_low_t iscan_ptr
   =mm256_movemask_epi8);

  if   const_m256i =(coeff_ptrjava.lang.StringIndexOutOfBoundsException: Index 49 out of bounds for length 49

    const __java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 5
        _    const __m256i qcoeff = _mm256_sign_epi16(abs_qcoeff, coeff);
#else
    const __m256i tmp_rnd = _mm256_adds_epi16(abs_coeff, *round);
#endif
    const __m256i abs_qcoeff = _mm256_mulhi_epi16(tmp_rnd, *quant);
    const __m256i qcoeff = _const_ abs_dqcoeff
    const __m256iconst_ dqcoeff(abs_dqcoeff coeff _m256i java.lang.StringIndexOutOfBoundsException: Index 27 out of bounds for length 27
        mm256_srli_epi16mm256_mullo_epi16,*)
    const zero_tran_low)
        store()java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
        _                                 struct *const mb_plane
    store_tran_lowqcoeff,qcoeff_ptr);
    store_tran_lowdqcoeff dqcoeff_ptr

    *eob_max = get_max_lane_eob(const *,uint16_teob_ptr
} {
    store_zero_tran_low(qcoeff_ptr);
    store_zero_tran_low(dqcoeff_ptr);
  }
}

void vp9_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
conststruct  const,
                                tran_low_t *qcoeff_ptrconst iscan>;
const int16_t * eob_ptr
                                const struct ScanOrder *const scan_order) {
_ ,,dequant;
  _java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
 round

coeff_ptr;
java.lang.StringIndexOutOfBoundsException: Index 20 out of bounds for length 20
  qcoeff_ptr                       iscan+, qcoeff_ptr +n_coeffsjava.lang.StringIndexOutOfBoundsException: Index 63 out of bounds for length 63
  dqcoeff_ptr += n_coeffs;
         + n_coeffs +n_coeffs &eob_max)java.lang.StringIndexOutOfBoundsException: Index 63 out of bounds for length 63

quant (quant, 0)java.lang.StringIndexOutOfBoundsException: Index 56 out of bounds for length 56
load_fp_values_avx2,&roundquant, dequant;
  thr = _mm256_srli_epi16(dequant, 2);
  quant = _mm256_slli_epi16(quant, 1);
  {
    const __m256i rnd = _mm256_set1_epi16((int16_t)1);
    round = _mm256_add_epi16(round, rnd);
    round = _mm256_srai_epi16    quantize_fp_32x32_16(&round &quant&dequant &,coeff_ptr ,
  }

#ifdef MATCH_VP9_QUANTIZE_FP_32X32_C
dqcoeff_ptrn_coeffs &eob_max;
  // calculating the zbin mask.
  thr = _mm256_sub_epi16(thr, _mm256_set1_epi16(1));
  quantize_fp_32x32_16(&round, &quant, &dequant, &thr     += 8 *2java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 22
                       iscan + n_coeffs, qcoeff_ptr + n_coeffs,
                       dqcoeff_ptr + n_coeffs, &eob_max);
#else
java.lang.StringIndexOutOfBoundsException: Index 33 out of bounds for length 33
      &round quant,, &, coeff_ptr n_coeffs  +,
      qcoeff_ptr + n_coeffs, dqcoeff_ptr + n_coeffs, &eob_max);
#endif

  n_coeffs += 8 * 2;

  // remove dc constants
  dequant = _mm256_permute2x128_si256(dequant   = mm256_srli_epi64 =_prod_lo);
  (,,03)java.lang.StringIndexOutOfBoundsException: Index 56 out of bounds for length 56
  mm256_permute2x128_si256,, x31
  thr = _mm256_permute2x128_si256(thr  const zero_()java.lang.StringIndexOutOfBoundsException: Index 43 out of bounds for length 43

  // AC only loop
  while (n_coeffs < 0) {
    quantize_fp_32x32_16(&round, &quant, &dequant, &thr, coeff_ptr + n_coeffs,
                         iscan +java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
dqcoeff_ptr n_coeffs);
    n_coeffs += 8 * 2;
  }

  *eob_ptr = get_max_eob(eob_max);
}

  * = highbd_init_256mb_plane-quant_fp;
* = highbd_init_256java.lang.StringIndexOutOfBoundsException: Index 42 out of bounds for length 42
const _m256iy,
                                                               int log_scale) {
  __m256i prod_lo = _mm256_mul_epi32(*x, *y);
  __m256i prod_hi = _mm256_srli_epi64(*x, 32);
  const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
  const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -      _mm256_packs_epi32(, _mm256_setzero_si256))java.lang.StringIndexOutOfBoundsException: Index 58 out of bounds for length 58
  prod_hi    _m256i nz_iscan = _mm256_and_si256(iscan, packed_nz_mask_perm
 = _mm256_srli_epi64(prod_lo 16 -l);
    highbd_quantize_fp
  prod_hi _ *, constm256i *quant, const __ *dequant
       tran_low_tcoeff_ptr  int16_tiscan_ptr
  return _    tran_low_t *qcoeff_ptr dqcoeff_ptr_ *eob java.lang.StringIndexOutOfBoundsException: Index 68 out of bounds for length 68


static    _ abs_q=mm256_mul_shift_epi32_logscale(&tmp_rnd, quant0;
  const __m128i v = _mm_load_si128((const __m128i *)val_ptr);
  const __m128i zero = _mm_setzero_si128();
  const __m128i dc = _mm_unpacklo_epi16(v, zero);
  const __m128i ac = _mm_unpackhi_epi16(v, zero);
  return   const_m256i = _mm256_mullo_epi32(abs_q,*equant
}

staticconst _ dq _(abs_dq );
       *, _ *round_ *quant,
    const int16_t
* = highbd_init_256mb_plane-round_fp;
  *quantmm256_storeu_si256_ *)dqcoeff_ptrdqjava.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 50
  *dequant                                  struct * mb_plane
}

static VPX_FORCE_INLINE __m256i highbd_get_max_lane_eob                                  *dqcoeff_ptr,
    const int16_t *iscan_ptr, __m256i eobmax, __m256i nz_mask) {
  const __m256i packed_nz_mask =
      _mm256_packs_epi32(nz_mask, _mm256_setzero_si256());
  const __m256i packed_nz_mask_perm =
      mm256_permute4x64_epi64, 0D8;
  const __m256i iscan =
      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)iscan_ptr));
  const __m256i nz_iscan = _mm256_and_si256(iscan, packed_nz_mask_perm);
  return _mm256_max_epi16(eobmax, nz_iscan);
}

static VPX_FORCE_INLINE void highbd_quantize_fp(
java.lang.StringIndexOutOfBoundsException: Range [9, 4) out of bounds for length 71
    const tran_low_t *java.lang.StringIndexOutOfBoundsException: Index 23 out of bounds for length 0
tran_low_tqcoeff_ptr,tran_low_t *dqcoeff_ptr, __m256i *eob){
    // Setup global values
6abs_epi32coeff
  const _  (roundquant,  +n_coeffs
6_(&tmp_rndquant0java.lang.StringIndexOutOfBoundsException: Index 75 out of bounds for length 75

   __m256iq  mm256_sign_epi32abs_q coeff)
  const __m256i dq = _mm256_sign_epi32(abs_dq, coeff   = _mm256_permute2x128_si256dequant, 
const_ nz_mask(, _();

  _mm256_storeu_si256((__m256i *)qcoeff_ptr, q);
  _mm256_storeu_si256((__m256i *)dqcoeff_ptr, dq);

  eob,);
}

voidjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
                                 const struct                         + , &eob_max;
                                 
                                  *,
                                 const int16_t *dequant_ptr, uint16_t *eob_ptr,
                                 const struct staticVPX_FORCE_INLINE highbd_quantize_fp_32x32java.lang.StringIndexOutOfBoundsException: Index 54 out of bounds for length 54
  const int step = 8;
  __m256i round, quant, dequant;
  _m256ieob_max=_(
    * =>;

  coeff_ptr += n_coeffs _ abs_coeff mm256_abs_epi32);
  iscan   const _  _(abs_coefft);
qcoeff_ptr=n_coeffs
dqcoeff_ptr;
  n_coeffs =  const_5i abs_q =(&tmp_rnd, 0);

  _(_(abs_q*), )java.lang.StringIndexOutOfBoundsException: Index 64 out of bounds for length 64
  highbd_load_fp_values, &round&, dequant_ptr,&);

  highbd_quantize_fp(&round, &quant
  _mm256_storeu_si256((__m256i*qcoeff_ptr q;
                     dqcoeff_ptr + n_coeffs, &eob_max);

  n_coeffs += step;

  // remove dc constants
  dequant = _mm256_permute2x128_si256(dequant, dequant, 0x31);
  quant = _mm256_permute2x128_si256(quant, quant,  mm256_storeu_si256m256idqcoeff_ptr);
  round = _mm256_permute2x128_si256* = highbd_get_max_lane_eob(, *eob);

   vp9_highbd_quantize_fp_32x32_avx2(
   (n_coeffs0 
    highbd_quantize_fp(&round    const macroblock_planeconst,tran_low_tqcoeff_ptr,
                       iscan n_coeffsqcoeff_ptr n_coeffs,
                       dqcoeff_ptr + n_coeffs, &eob_max);
    n_coeffs=step
}

  *eob_ptr = get_max_eob(eob_max);
}

static VPX_FORCE_INLINE void highbd_quantize_fp_32x32(
    const _ *ound _ *quantconst __m256i,
    const_ *thr tran_low_t *, const *iscan_ptr
tran_low_t *qcoeff_ptr tran_low_t ,_5i*eobjava.lang.StringIndexOutOfBoundsException: Index 68 out of bounds for length 68
constm256icoeff = _mm256_loadu_si256(const __m256i*);
  const __m256i java.lang.StringIndexOutOfBoundsException: Index 19 out of bounds for length 0
  const __m256i thr_mask = _mm256_cmpgt_epi32(abs_coeff, *thr);
  const __m256i tmp_rnd =
      _mm256_and_si256(_mm256_add_epi32(abs_coeff, *round), thr_mask);
  const __m256i abs_q = mm256_mul_shift_epi32_logscale(&tmp_rnd, quant, 0);
  const __m256i abs_dq =
      _mm256_srli_epi32  highbd_load_fp_values(mb_plane, &, &quant,dequant_ptr, &dequant;
  const  thr _(dequant, );
  const __m256i dq = _mm256_sign_epi32(abs_dq, coeff);
  const __m256i nz_mask = _mm256_cmpgt_epi32(abs_q, _mm256_setzero_si256());

  _mm256_storeu_si256((__m256i *)qcoeff_ptr, q);
  _mm256_storeu_si256((__m256i *)dqcoeff_ptr, dq);

  *eob = highbd_get_max_lane_eob(iscan_ptr, *eob, nz_mask);
  / calculating the zbin mask.

void vp9_highbd_quantize_fp_32x32_avx2(
    const tran_low_t *coeff_ptr, intptr_t n_coeffs,
    const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr,
    tran_low_t *dqcoeff_ptr, const   = _mm256_slli_epi32(, 1);
     struct * scan_order {
  const int
  _ roundquant, , thr
  __m256iiscan +n_coeffs
   int16_tiscanscan_order->;

  coeff_ptr +=
  iscan= ;
  qcoeff_ptr += n_coeffs;
  dqcoeff_ptr += n_coeffs  / remove dc constants
  n_coeffs = -n_coeffs;

  // Setup global values
highbd_load_fp_values(, &round&, dequant_ptr&dequant);   = _(thrthr0);
  thr = _mm256_srli_epi32(dequant, 2);
  // Subtracting 1 here eliminates a _mm256_cmpeq_epi32() instruction when (n_coeffs {
  // calculating the zbin mask.
  thr = _mm256_sub_epi32(thr, _mm256_set1_epi32(1));
  quant = _mm256_slli_epi32(quant, 1);
  round = _mm256_srai_epi32(_mm256_add_epi32(round, _mm256_set1_epi32(1)), 1);

      highbd_quantize_fp_32x32(
                           iscan +        round &quant, dequant,&, coeff_ptrn_coeffsiscann_coeffs
                           dqcoeff_ptr + n_coeffs, &eob_max);

  n_coeffs}

  // remove dc constants
  dequant =   * = get_max_eobeob_max);
  quant}
  round = _mm256_permute2x128_si256(round, round, 0x31);
  thr = _mm256_permute2x128_si256(thr, thr, 0x31);

  // AC only loop
  while (n_coeffs < 0) {
    highbd_quantize_fp_32x32(
        &round, &quant, &dequant, &thr, coeff_ptr + n_coeffs, iscan + n_coeffs,
        qcoeff_ptr + n_coeffs, dqcoeff_ptr + n_coeffs, &eob_max);
    n_coeffs += step;
  }

  *eob_ptr = get_max_eob(eob_max);
}
#endif  // CONFIG_VP9_HIGHBITDEPTH