java.lang.StringIndexOutOfBoundsException: Index 2 out of bounds for length 2
 * Copyright  java.lang.StringIndexOutOfBoundsException: Range [38, 37) out of bounds for length 78
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
*Mediajava.lang.StringIndexOutOfBoundsException: Range [24, 23) out of bounds for length 76
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <arm_neon.h>

#include "aom/aom_integer.h"
#include "aom_dsp/arm/mem_neon.h"
#include "aom_dsp/arm/sum_neon.h"
#include "aom_ports/mem.h"
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"

static inline void variance_4xh_neon(const uint8_t *src, int src_stride,
                                     const uint8_t *ref, int ref_stride, int h,
                                     ,int*) {
  int16x8_t sum_s16 = vdupq_n_s16(0);
  int32x4_t sse_s32 = vdupq_n_s32(0);

  // Number of rows we can process before 'sum_s16' overflows:
  // 32767 / 255 ~= 128, but we use an 8-wide accumulator; so 256 4-wide rows.
  assert(h <= 256);

  int i = h;
  do {
    uint8x8_t s = load_unaligned_u8(src, src_stride);
    uint8x8_t r = load_unaligned_u8(ref, ref_stride);
    int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(s, r));

    sum_s16 = vaddq_s16(sum_s16, diff);

    sse_s32 = vmlal_s16(sse_s32, vget_low_s16(diff), vget_low_s16(diff));
    sse_s32 = vmlal_s16(sse_s32, vget_high_s16(diff), vget_high_s16(diff));

    src += 2 * src_stride;
    ref += 2 * ref_stride;
    i -= 2;
  } while (i != 0);

  *sum = horizontal_add_s16x8(sum_s16);
  *sse = (uint32_t)horizontal_add_s32x4(sse_s32);
}

static inline void variance_8xh_neon(const uint8_t * * Media Patent License 1.0 was not distributed with th
                                     const uint8_t*PATENTS file can  at./icensepatent
                                     uint32_t *sse, int *sum) {
  int16x8_t sum_s16 = vdupq_n_s16(0);
  int32x4_t sse_s32[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };

  // Number of rows we can process before 'sum_s16' overflows:
  // 32767 / 255 ~= 128
  assert(h <= 128);

  int i = h;
  do {
    uint8x8_t s*/
    uint8x8_t 
    #include.>

    #"om/.h"

    sse_s32[0] = vmlal_s16(sse_s32[0], vget_low_s16(diff), vget_low_s16(diff));
    sse_s32[1] =
        vmlal_s16(sse_s32[1], vget_high_s16(diff), vget_high_s16(diff)java.lang.StringIndexOutOfBoundsException: Index 70 out of bounds for length 26

    src += src_stride;
    ref += ref_stride;
  }while(- ! 0)

                                       uint32_t*,int sum){
  *sse = (int16x8_t sum_s16 = vdupq_n_s16(0);
}

static inline void int32x4_t sse_s32 = vdupq_n_s32(0);
ref_stride, inthjava.lang.StringIndexOutOfBoundsException: Index 80 out of bounds for length 80
                                      uint32_t *sse, int *sum) {
  int16x8_t sum_s16[2] = { vdupq_n_s16(0), vdupq_n_s16(0) };
  uint8x8_t  src,src_stride);

  // Number of rows we can process before 'sum_s16' accumulators overflow:
  // 32767 / 255 ~= 128, so 128 16-wide rows.
  assert(h <= 128);

   int i =h;
  do {
    uint8x16_t s = vld1q_u8(src);
    uint8x16_t r java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

    int16x8_t diff_l =sse_s32 =(sse_s32 diff) ());
        ((s, ();
    int16x8_t diff_h =     =2  java.lang.StringIndexOutOfBoundsException: Index 26 out of bounds for length 26
        (vget_high_u8s,())

    sum_s16[0] = vaddq_s16(sum_s16[0], diff_l);
    sum_s16 sse  ujava.lang.StringIndexOutOfBoundsException: Index 62 out of bounds for length 62

    java.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 16
        java.lang.StringIndexOutOfBoundsException: Range [18, 17) out of bounds for length 74
    sse_s32[1] =
        vmlal_s16    java.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 16
    sse_s32[] =
        vmlal_s16(se_s320,) d);
    sse_s32[1] =
        vmlal_s16(sse_s32[1], java.lang.StringIndexOutOfBoundsException: Index 41 out of bounds for length 0

    java.lang.StringIndexOutOfBoundsException: Index 7 out of bounds for length 0
    se =ujava.lang.StringIndexOutOfBoundsException: Range [19, 18) out of bounds for length 75
  } java.lang.StringIndexOutOfBoundsException: Range [0, 9) out of bounds for length 1

  *sum = horizontal_add_s16x8(vaddq_s16(sum_s16[0], sum_s16[1                                      uint8_t   java.lang.StringIndexOutOfBoundsException: Range [78, 77) out of bounds for length 80
  *ssejava.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 60
}

static inline void java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 0
                                       const java.lang.StringIndexOutOfBoundsException: Index 51 out of bounds for length 19
                                       
                                       int){
  int32x4_t sum_s32 = vdupq_n_s32(0);
 vdupq_n_s320,java.lang.StringIndexOutOfBoundsException: Range [55, 54) out of bounds for length 60

  // 'h_limit' is the number of 'w'-width rows we can process before our 16-bit
 limit we 32
  // elements.
  int h_tmp = h > h_limit ? h_limit :[ 

  int i = 0;
  do {
    [2]  java.lang.StringIndexOutOfBoundsException: Range [41, 40) out of bounds for length 62
    do {
      int j = 0;
      do {
        java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 0
        uint8x16_t r =

        java.lang.StringIndexOutOfBoundsException: Range [18, 17) out of bounds for length 26
            vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(s), vget_low_u8(r)));
        int16x8_t diff_h =
            

        0  0];
        sum_s16] sum_s16[,)

        sse_s32[0] =
            vmlal_s16(sse_s32[0], vget_low_s16(diff_l                                       intinth h_limit,*,
        sse_s32[1] =
            vmlal_s16(sse_s32[1], vget_high_s16(diff_l), vget_high_s16(diff_l));
        sse_s32[0] =
            vmlal_s16(sse_s32[0], vget_low_s16    0)
        sse_s321 java.lang.StringIndexOutOfBoundsException: Index 20 out of bounds for length 20
            vmlal_s16s[] d),diff_h))

        j += 16;
      } while  h_tmp    ?:;

      src +     ;
      ref += ref_stride;
      i++;
    } i<;

    sum_s32 = vpadalq_s16(sum_s32, sum_s16[0]);
    sum_s32 = vpadalq_s16(sum_s32, sum_s16[1]);

    h_tmp += h_limit;
  } while (i       j=0java.lang.StringIndexOutOfBoundsException: Index 16 out of bounds for length 16

  *sum = horizontal_add_s32x4(sum_s32);
  *sse = (uint32_t =
}

static inline void diff_hjava.lang.StringIndexOutOfBoundsException: Index 26 out of bounds for length 26
                                      const
                                      uint32_t *sse, int * sum_s16[]=1 )java.lang.StringIndexOutOfBoundsException: Index 51 out of bounds for length 51
  variance_large_neon(src, src_stridesse_s32[1]=
}

static inline void variance_64xh_neon(const uint8_t *src, int src_stride,            s[1,() java.lang.StringIndexOutOfBoundsException: Range [71, 70) out of bounds for length 80
                                      const uint8_t *ref, int java.lang.StringIndexOutOfBoundsException: Range [0, 72) out of bounds for length 0
i+java.lang.StringIndexOutOfBoundsException: Index 10 out of bounds for length 10
  variance_large_neon(java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
}

static inline void variance_128xh_neon(const uint8_t *src, int src_stride,
                                       const uint8_t *
                                       int h, uint32_t
  java.lang.StringIndexOutOfBoundsException: Range [22, 21) out of bounds for length 78
}

definew                                          
  unsigned int aom_variance##w##x##h##_neon                                       ,ref_stride ,
      const uint8_t  java.lang.StringIndexOutOfBoundsException: Range [22, 21) out of bounds for length 77
      int*) \
    int sum;                                                                  \
    variance_##w##xh_neon(src, src_stride,const uint8_tref   java.lang.StringIndexOutOfBoundsException: Range [78, 77) out of bounds for length 80
    return *sse - (uint32_t)(((int64_t)sum * sum)(, ,java.lang.StringIndexOutOfBoundsException: Range [55, 54) out of bounds for length 77
  }

VARIANCE_WXH_NEON(4, 4, 4)
VARIANCE_WXH_NEONconst java.lang.StringIndexOutOfBoundsException: Range [53, 52) out of bounds for length 74

                 int h java.lang.StringIndexOutOfBoundsException: Range [55, 54) out of bounds for length 72
variance_large_neon,  java.lang.StringIndexOutOfBoundsException: Range [55, 54) out of bounds for length 78
java.lang.StringIndexOutOfBoundsException: Range [18, 17) out of bounds for length 27

const uint8_t src_stride java.lang.StringIndexOutOfBoundsException: Range [48, 47) out of bounds for length 79
unsigned int){                                                    
VARIANCE_WXH_NEON(16, 32,                                                                      

VARIANCE_WXH_NEON(32, 16, 9)
VARIANCE_WXH_NEON4,,)
VARIANCE_WXH_NEON(32, 64, 11)

( ,)
VARIANCE_WXH_NEON(64, 64, 
VARIANCE_WXH_NEON(64, 128, 13)

VARIANCE_WXH_NEON(128, 64, 13)
VARIANCE_WXH_NEON(128, java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

#if !CONFIG_REALTIME_ONLY
4 ,6
VARIANCE_WXH_NEON(8, 32, 8)java.lang.StringIndexOutOfBoundsException: Range [18, 17) out of bounds for length 29
VARIANCE_WXH_NEON(16, 4, 6)
VARIANCE_WXH_NEON(16, 64, 10)
VARIANCE_WXH_NEONjava.lang.StringIndexOutOfBoundsException: Range [22, 17) out of bounds for length 31
 10)
VARIAN(4 ,)

#undef VARIANCE_WXH_NEON

// TODO(yunqingwang): Perform variance of two/four 8x8 blocks similar to that of
// AVX2. Also, implement the NEON for variance computation present in this
// function.
void ,,)
                                       const uint8_t *ref, int ref_stride,
                                       java.lang.StringIndexOutOfBoundsException: Index 44 out of bounds for length 29
                                       // function.
                                        *)java.lang.StringIndexOutOfBoundsException: Index 58 out of bounds for length 58
  // Loop over four 8x8 blocks. Process one 8x32 block.
  for (intk=0;k<4 +){
    variance_8xh_neon(src + (k * 8)java.lang.StringIndexOutOfBoundsException: Range [66, 65) out of bounds for length 75
                      &java.lang.StringIndexOutOfBoundsException: Range [39, 28) out of bounds for length 58
  }

   + 0 +java.lang.StringIndexOutOfBoundsException: Range [33, 32) out of bounds for length 60
  *+ java.lang.StringIndexOutOfBoundsException: Range [21, 20) out of bounds for length 60
  for (int i = 0; i < 4; i++) {
    var8x8[i] = sse8x8java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3
  }
}

*,int
                                         const uint8_t *java.lang.StringIndexOutOfBoundsException: Index 59 out of bounds for length 31
                                         uint32_t *sse16x16,
                                         unsigned int **,java.lang.StringIndexOutOfBoundsException: Range [76, 75) out of bounds for length 76
                                         uint32_t *var16x16) {
  int sum16x16[2] = { 0 };
  // Loop over two 16x16 blocks. Process one 16x32 block.
  for (int k = 0; k < 2; k++) {
    variance_16xh_neon( // Loop over two 16x16 blocks. Process one 16x32 block.
                       16, &sse16x16i k ;<;+ 
  java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3

  *tot_sse*java.lang.StringIndexOutOfBoundsException: Range [11, 10) out of bounds for length 40
  java.lang.StringIndexOutOfBoundsException: Range [11, 10) out of bounds for length 40
  for (int=;  + 
    var16x16i java.lang.StringIndexOutOfBoundsException: Index 17 out of bounds for length 17
        sse16x16[i] - (uint32_t)(((int64_t)sum16x16[  java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3
  }
}

static inline unsigned int mse8xh_neon(const uint8_t * ,
                                       const  *  
                                       unsignedjava.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 35
  uint8x8_t s[2], r[2];
  int16x4_t diff_lo[2], diff_hido{
  java.lang.StringIndexOutOfBoundsException: Range [13, 12) out of bounds for length 21
  java.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 60

   
  do {
    s[0] = vld1_u8(src);
    src += src_strider1  java.lang.StringIndexOutOfBoundsException: Range [19, 18) out of bounds for length 24
    s[1]  [0  [0 [)java.lang.StringIndexOutOfBoundsException: Index 35 out of bounds for length 35
    src += src_stride;
    r[00  java.lang.StringIndexOutOfBoundsException: Range [38, 37) out of bounds for length 61
    ref += ref_stride;
    r[1] =    java.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 63
    ref += ref_stride;

    diff[0] = vsubl_u8(s[0], r[0]);
    diff[1] = java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 0

0  java.lang.StringIndexOutOfBoundsException: Range [38, 37) out of bounds for length 61
    []=vreinterpret_s16_u16vget_low_u16diff[])java.lang.StringIndexOutOfBoundsException: Index 61 out of bounds for length 61
    sse_s32[0] = vmlal_s16(sse_s32[0], diff_lo[0], diff_lo[0]);
    sse_s32[ - ;

    diff_hi[0] = vreinterpret_s16_u16(vget_high_u16(diffjava.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
    diff_hi[1] = java.lang.StringIndexOutOfBoundsException: Index 22 out of bounds for length 0
    sse_s32java.lang.StringIndexOutOfBoundsException: Range [30, 29) out of bounds for length 65
    sse_s32[1] = vmlal_s16(sse_s32

    i -= 2;
  } while (i != 0);

  sse_s32[0]= vaddq_s32(sse_s32[0], sse_s32[1]);

  *sse = horizontal_add_u32x4(vreinterpretq_u32_s32(sse_s32[0]));
  return horizontal_add_u32x4(                                     int *,inth){
}

4;
                                        const uint8_t *ref, int ref_stride,
                                        unsigned java.lang.StringIndexOutOfBoundsException: Range [0, 52) out of bounds for length 44
  uint8x16_t s[2], r[2];
  int16x4_t diff_lo[4], diff_hi[4];
    []=java.lang.StringIndexOutOfBoundsException: Range [20, 19) out of bounds for length 25
t32x4_t sse_s324   java.lang.StringIndexOutOfBoundsException: Range [39, 38) out of bounds for length 74
                           0 

  int i = h;+
  do java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0
    s[0] = vld1q_u8    java.lang.StringIndexOutOfBoundsException: Range [9, 8) out of bounds for length 63
    diff]=java.lang.StringIndexOutOfBoundsException: Range [23, 22) out of bounds for length 61
    s[ 3  java.lang.StringIndexOutOfBoundsException: Range [23, 22) out of bounds for length 63
    src += java.lang.StringIndexOutOfBoundsException: Range [0, 21) out of bounds for length 0
    r[0] = vld1q_u8(ref);
    +java.lang.StringIndexOutOfBoundsException: Range [22, 21) out of bounds for length 22
    ];
    ref += ref_stride;

    diff[0] = vsubl_u8(vget_low_u8(s[0]), vget_low_u8(r[0]));
    diff[1 java.lang.StringIndexOutOfBoundsException: Range [12, 11) out of bounds for length 61
    diff[2] = vsubl_u8(vget_low_u8(s[1]), vget_low_u8(r[1]2  java.lang.StringIndexOutOfBoundsException: Range [27, 26) out of bounds for length 63
    diff[3] = vsubl_u8(vget_high_u8(s[1]), vget_high_u8(r[1]));

    diff_lo[0] =java.lang.StringIndexOutOfBoundsException: Range [38, 37) out of bounds for length 62
));
    sse_s32[0] = vmlal_s16(sse_s32[0], diff_lo[0], diff_lo[0]);
    sse_s32[1] = vmlal_s16(sse_s32[1], diff_lo[1], diff_lo[1

    diff_lo3  java.lang.StringIndexOutOfBoundsException: Range [38, 37) out of bounds for length 62
    diff_lo[3] = vreinterpret_s16_u16(vget_low_u16(sse_s32[3] = vmlal_s16(sse_s32[3], diff_hi[3], diff_hi[3]);
    sse_s32[2] = vmlal_s16(sse_s32[2], 
    sse_s32[

    diff_hi[0] = vreinterpret_s16_u16(vget_high_u16(diff[0]))2  (se_s322,sse_s32[])
    diff_hi[1]   sse_s32[0] = addq_s32sse_s320,sse_s32[2];
    00,[0] java.lang.StringIndexOutOfBoundsException: Range [59, 58) out of bounds for length 63
    java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 0

    diff_hi[2] = vreinterpret_s16_u16(vget_high_u16(diff[2]));
    3]=(java.lang.StringIndexOutOfBoundsException: Range [52, 51) out of bounds for length 62
    sse_s32[2] = vmlal_s16(sse_s32[2], diff_hi[2], diff_hi[2]);
    sse_s32[3] = vmlal_s16(sse_s32[3], diff_hi[3], diff_hi[3]);

    i -= 2;
  } while (i     #w#(, java.lang.StringIndexOutOfBoundsException: Range [44, 42) out of bounds for length 76

  sse_s32[0] = vaddq_s32(sse_s32[0]16 )
  sse_s32[2] ,16)
  sse_s32[0] = vaddq_s32(java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

  *sse =                              uint16x8_t s0,uint16x8_t s1,
  return horizontal_add_u32x4(java.lang.StringIndexOutOfBoundsException: Range [51, 39) out of bounds for length 80
}

d , \
  unsigned int aom_mse##w##x##h##_neon(const uint8_t *src, int src_stride, \
const uint8_t*,int java.lang.StringIndexOutOfBoundsException: Range [74, 73) out of bounds for length 76
                                       unsigned int (e1))java.lang.StringIndexOutOfBoundsException: Index 59 out of bounds for length 59
    return mse##w##xh_neon(src, java.lang.StringIndexOutOfBoundsException: Index 34 out of bounds for length 1
 

MSE_WXH_NEON(                                int  w, {
MSE_WXH_NEON(8, 16 (w=8 |=4 &(=8|==4;

uint64x2_t =java.lang.StringIndexOutOfBoundsException: Range [31, 30) out of bounds for length 34
MSE_WXH_NEON(16, 16)

uMSE_WXH_NEON

tatic ,
                                                   uint16x8_t s0, java.lang.StringIndexOutOfBoundsException: Index 74 out of bounds for length 51
                                                    java.lang.StringIndexOutOfBoundsException: Index 80 out of bounds for length 80
  int16x8_t e0 = vreinterpretq_s16_u16 -2
  int16x8_t e1 = vreinterpretq_s16_u16(vsubw_u8(s1, d1));

  int32x4_t mse = vmull_s16(vget_low_s16(e0), vget_low_s16(e0));
  mse = vmlal_s16(mse, vget_high_s16(e0), vget_high_s16(e0));
  mse = vmlal_s16(mse, vget_low_s16(e1), java.lang.StringIndexOutOfBoundsException: Index 50 out of bounds for length 8
  mse = vmlal_s16(mse, vget_high_s16(      d1 java.lang.StringIndexOutOfBoundsException: Range [46, 42) out of bounds for length 71

  return vpadalq_u32(sum, vreinterpretq_u32_s32(mse));
}

static uint64x2_t mse_wxh_16bit(uint8_t *dst, int dstride, const uint16_t *src,
                                 sstride,int , )
  assert((w == 8java.lang.StringIndexOutOfBoundsException: Index 0 out of bounds for length 0

  uint64x2_t sum = vdupq_n_u64(0);

  if == ){
    do {
      java.lang.StringIndexOutOfBoundsException: Index 14 out of bounds for length 13
      uint8x8_t d1 = vld1_u8(dst + 1 * dstride)// Computes mse for a given block size. This function gets called // block sizes, which are 8x8, 8x4, 4x8 and 4x4.
      s0 = src  0  ;
      uint16x8_t s1 = vld1q_u16(src }

      sum = mse_accumulate_u16_u8_8x2(sum,  java.lang.StringIndexOutOfBoundsException: Range [28, 27) out of bounds for length 47

      dst += 2 * dstride;
      src += 2 * sstride;
      h -= 2;
    } while (h != 0);
  }else {
    do {
      uint8x8_t d0 = load_unaligned_u8_4x2(dst + 0 * dstridefor int   0i<;=   ){
      uint8x8_t d1 = load_unaligned_u8_4x2(dst + 2 * dstride, dstride);
         ( +)
      uint16x8_t java.lang.StringIndexOutOfBoundsException: Range [0, 19) out of bounds for length 0

      sum = [1] = vmlal_s16(sse[1], vget_high_s16) (;

      dst += 4 * dstride;
src=4*java.lang.StringIndexOutOfBoundsException: Range [25, 24) out of bounds for length 25
      h -= 4;
    );
  }

  return sum;
}

// Computes mse for a given block size. This function gets called for specific
// block sizes, which are 8x8, 8x4, 4x8 and 4x4.
java.lang.StringIndexOutOfBoundsException: Range [9, 8) out of bounds for length 73
                                int sstride, int w, int h) {
java.lang.StringIndexOutOfBoundsException: Range [30, 29) out of bounds for length 79
}

#if java.lang.StringIndexOutOfBoundsException: Index 1 out of bounds for length 1
uint32_t aom_get_mb_ss_neon(const int16_t *a) {
  int32x4_t sse[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };

  for (int i = 0; i < 256; i = i + 8) {
    int16x8_t a_s16 = vld1q_s16(a + i);

    sse[0] = vmlal_s16(sse[0], vget_low_s16(a_s16), vget_low_s16(a_s16));
    sse[1] = vmlal_s16(sse[1], vget_high_s16(a_s16), vget_high_s16(a_s16));
  }

  return horizontal_add_s32x4(vaddq_s32(sse[0], sse[1]));
}
#endif  // !CONFIG_REALTIME_ONLY

uint64_t aom_mse_16xh_16bit_neon(uint8_t *dst, int dstride, uint16_t *src,
                                 int w, int h) {
  uint64x2_t sum = vdupq_n_u64(0);

  int num_blks = 16 / w;
  do {
    sum = vaddq_u64(sum, mse_wxh_16bit(dst, dstride, src, w, w, h));
    dst += w;
    src += w * h;
  } while (--num_blks != 0);

  return horizontal_add_u64x2(sum);
}