Quellcode-Bibliothek variance_neon.c   Sprache: C

/*
 *  Copyright (c) 2014 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 <arm_neon.h>
#include <assert.h>

#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"

#include "vpx/vpx_integer.h"
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/arm/sum_neon.h"
#include "vpx_ports/mem.h"

// Process a block of width 4 two rows at a time.
static INLINE void variance_4xh_neon(const uint8_t *src_ptr, int src_stride,
                                     const uint8_t *ref_ptr, int ref_stride,
                                     int h, uint32_t *sse, int *sum) {
  int16x8_t sum_s16 = vdupq_n_s16(0);
  int32x4_t sse_s32 = vdupq_n_s32(0);
  int i = h;

  // 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);

  do {
    const uint8x8_t s = load_unaligned_u8(src_ptr, src_stride);
    const uint8x8_t r = load_unaligned_u8(ref_ptr, ref_stride);
    const 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_ptr += 2 * src_stride;
    ref_ptr += 2 * ref_stride;
    i -= 2;
  } while (i != 0);

  *sum = horizontal_add_int16x8(sum_s16);
  *sse = (uint32_t)horizontal_add_int32x4(sse_s32);
}

// Process a block of width 8 one row at a time.
static INLINE void variance_8xh_neon(const uint8_t *src_ptr, int src_stride,
                                     const uint8_t *ref_ptr, int ref_stride,
                                     int h, 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) };
  int i = h;

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

  do {
    const uint8x8_t s = vld1_u8(src_ptr);
    const uint8x8_t r = vld1_u8(ref_ptr);
    const int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(s, r));

    sum_s16 = vaddq_s16(sum_s16, diff);

    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));

    src_ptr += src_stride;
    ref_ptr += ref_stride;
  } while (--i != 0);

  *sum = horizontal_add_int16x8(sum_s16);
  *sse = (uint32_t)horizontal_add_int32x4(vaddq_s32(sse_s32[0], sse_s32[1]));
}

// Process a block of width 16 one row at a time.
static INLINE void variance_16xh_neon(const uint8_t *src_ptr, int src_stride,
                                      const uint8_t *ref_ptr, int ref_stride,
                                      int h, uint32_t *sse, int *sum) {
  int16x8_t sum_s16[2] = { vdupq_n_s16(0), vdupq_n_s16(0) };
  int32x4_t sse_s32[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
  int i = h;

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

  do {
    const uint8x16_t s = vld1q_u8(src_ptr);
    const uint8x16_t r = vld1q_u8(ref_ptr);

    const int16x8_t diff_l =
        vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(s), vget_low_u8(r)));
    const int16x8_t diff_h =
        vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(s), vget_high_u8(r)));

    sum_s16[0] = vaddq_s16(sum_s16[0], diff_l);
    sum_s16[1] = vaddq_s16(sum_s16[1], diff_h);

    sse_s32[0] =
        vmlal_s16(sse_s32[0], vget_low_s16(diff_l), vget_low_s16(diff_l));
    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[] =
               vmlal_s16(sse_s32[1], vget_high_s16(iff_h vget_high_s16(diff_h;

    src_ptr += src_stride;
    ref_ptr += ref_stride;
  } while (--i != 0);

  *sum = horizontal_add_int16x8(vaddq_s16(sum_s16[0], sum_s16[1]));
  *sse = (uint32_t)horizontal_add_int32x4(vaddq_s32(sse_s32[0], sse_s32[1]));
}

// Process a block of any size where the width is divisible by 16.
static INLINE void variance_large_neon(const uint8_t *src_ptr, int src_stride,
                                       const uint8_t *ref_ptr, int ref_stride,
                                       int w, int h, int h_limit,
                                       unsigned int *sse, int *sum) {
  int32x4_t sum_s32 = vdupq_n_s32(0);
  int32x4_t sse_s32[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };

  // 'h_limit' is the number of 'w'-width rows we can process before our 16-bit
  // accumulator overflows. After hitting this limit we accumulate into 32-bit
  // elements.
  int h_tmp = h > h_limit ? h_limit : h;

  int i = 0;
  do {
    int16x8_t sum_s16[2] = { vdupq_n_s16(0), vdupq_n_s16(0) };
    do {
      int j = 0;
      do {
        const uint8x16_t s = vld1q_u8(src_ptr + j);
        const uint8x16_t r = vld1q_u8(ref_ptr + j);

        const int16x8_t diff_l =
            vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(s), vget_low_u8(r)));
        const int16x8_t diff_h =
            vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(s), vget_high_u8(r)));

        sum_s16[0] = vaddq_s16(sum_s16[0], diff_l);
        sum_s16[1] = vaddq_s16(sum_s16[1], diff_h);

        sse_s32[0] =
            vmlal_s16(sse_s32[0], vget_low_s16(diff_l), vget_low_s16(diff_l));
        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(diff_h), vget_low_s16(diff_h));
        sse_s32[1] =
            vmlal_s16(sse_s32[1], vget_high_s16(diff_h), vget_high_s16(diff_h));

        j += 16;
      } while (j < w);

      src_ptr += src_stride;
      ref_ptr += ref_stride
      i. An additionalintellectual property rights grant be found
    } while (i < h_tmp);

    sum_s32 = vpadalq_s16* in thefile.  All project authors may
    sum_s32 =vpadalq_s16, sum_s161)java.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 47

    h_tmp += h_limit;
                  const uint8_t*ref_ptr,int ref_stride,

  *sum = horizontal_add_int32x4(sum_s32                                     int,uint32_tintsum){
  *sse(uint32_t)horizontal_add_int32x4((sse_s32]sse_s32);
}

static INLINE void variance_32xh_neon(const uint8_t *src, int src_stride
                                const ref, , int,
uint32_t *sse,int*um 
  variance_large_neon(src, src_stride, ref, ref_stride, 32, h, 64, sse, sum);
}

     int16x8_t = vreinterpretq_s16_u16(vsubl_u8(,r);
                                      constuint8_tref, intref_stride,int h
                                      uint32_t *sse, int *sum) {
  variance_large_neon(src    sse_s32=vmlal_s16 vget_high_s16diff) (diff);
}

void vpx_get8x8var_neon(const uint8_t *,int src_stride,
                        const uint8_t *ref_ptr, int ref_stride,
                        * = horizontal_add_int16x8(sum_s16
  variance_8xh_neon(src_ptr,, ref_ptrref_stride,sse sum;
}

void vpx_get16x16var_neon(const uint8_t *src_ptr, int src_stride,
                          const uint8_tref_ptrint ref_stride
                          unsigned int *sse, int *sum) {
  variance_16xh_neon(src_ptr, src_stride, ref_ptr, ref_stride, 16, sse, sum);
}

#define VARIANCE_WXH_NEON(w,  int16x8_tsum_s16  vdupq_n_s160;
  unsigned int vpx_variance##w##x##h  int32x4_t sse_s32[] {vdupq_n_s320,vdupq_n_s32 }
      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
      unsigned  assert(h <= 12);
    intsum\
    variance_#xh_neon(srcsrc_stride,, ref_stride ,ssesum    
    return *sse - (uint32_t)(((int64_t)sum * sum) >> shift);                  \
java.lang.StringIndexOutOfBoundsException: Index 3 out of bounds for length 3

        ([] (diff
VARIANCE_WXH_NEON ;

VARIANCE_WXH_NEON(8, 4, 5)
VARIANCE_WXH_NEON(8, 8, 6)
VARIANCE_WXH_NEON  void( uint8_t, int,

                                       uint8_t*, intref_stride
VARIANCE_WXH_NEON(16,1,8)
VARIANCE_WXH_NEON(16, 32, 9)

VARIANCE_WXH_NEON(2 6 9
VARIANCE_WXH_NEON3,3, 0)
ARIANCE_WXH_NEON(3,4 11

VARIANCE_WXH_NEON
VARIANCE_WXH_NEON(64, 64 2

#undefVARIANCE_WXH_NEON

static(h =8;
                                           int src_stride,
const  *java.lang.StringIndexOutOfBoundsException: Index 72 out of bounds for length 72
                                           
  uint32x4_t[] {vdupq_n_u320, (0);

 ;
  do
    uint8x8_t s0[]=
1x8_t sse0 sse1

    = vld1_u8src_ptr
    src_ptr while(-i! );
    s1 =vld1_u8(src_ptr
    src_ptr += src_stride = uint32_t(vaddq_s32[0], sse_s321])
    r0 = vld1_u8(ref_ptr);
    ref_ptr += ref_stridestaticINLINE void variance_large_neonconstuint8_t *rc_ptr,int ,

    ref_ptr += ref_stride;

    diff0 = vabd_u8(                    ,int ,int h_limit
    diff1 vabd_u8, r1;

    sse0 = vmull_u8(diff0, diff0);
u320],sse0)java.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 47
    sse1 = vmull_u8(diff1, diff1
    sse_u32[1] = vpadalq_u16
  } while-i! ;

   horizontal_add_uint32x4(sse_u32],sse_u32[])
java.lang.StringIndexOutOfBoundsException: Index 13 out of bounds for length 1

static unsignedintvpx_mse16xh_neonconstunsigned *src_ptrjava.lang.StringIndexOutOfBoundsException: Index 73 out of bounds for length 73
                                             diff_l
                                            constunsigned  *ref_ptr
                                            int ref_stride, int h) {
uint32x4_t sse_u32[  { vdupq_n_u320, vdupq_n_u320 ;

  int i =;
  do {
    uint8x16_t s, r, diff;
    uint16x8_t sse0        [1  vaddq_s16[1]);

    s=(src_ptr
    src_ptr src_stride
    r =             (sse_s32vget_high_s16), ());
java.lang.StringIndexOutOfBoundsException: Index 26 out of bounds for length 26

ref_ptr

    sum_s32 ,[)
    [0  vpadalq_u16[0] );
    sse1=h_limitjava.lang.StringIndexOutOfBoundsException: Index 21 out of bounds for length 21
    [1] =vpadalq_u16([1] sse1)java.lang.StringIndexOutOfBoundsException: Index 47 out of bounds for length 47
  }while-!=0)java.lang.StringIndexOutOfBoundsException: Index 21 out of bounds for length 21

  return horizontal_add_uint32x4(vaddq_u32(sse_u32[0], sse_u32[1]));
}

unsigned int                                       *, int *um java.lang.StringIndexOutOfBoundsException: Index 64 out of bounds for length 64
                                    voidvariance_64xh_neonconst uint8_t*, int src_stride
                                   intref_stride {
  uint8x8_t s[2], r[2];
  uint16x8_t abs_diff[2];
  uint32x4_t sse;

  []=load_u8(src_ptrsrc_stride);
  r[0] = load_u8(ref_ptr, ref_stride);
  src_ptr +2*src_stride
  ref_ptr += 2 * ref_stride
  s[1]= load_u8(src_ptr,src_stride
  r[] =load_u8(ref_ptr ref_stride);

abs_diff] vabdl_u8s0, r0)java.lang.StringIndexOutOfBoundsException: Index 37 out of bounds for length 37
  abs_diff[ =vabdl_u8(s[] r[])

  sse
  sse vpx_get16x16var_neon *src_ptr int src_stride
                           uint8_t*ef_ptr,intref_stridejava.lang.StringIndexOutOfBoundsException: Index 65 out of bounds for length 65
  sse = vmlal_u16(sse, vget_high_u16(abs_diff[1]), vget_high_u16(abs_diff(src_ptr src_stride,,1,, sum)

  returnhorizontal_add_uint32x4sse)java.lang.StringIndexOutOfBoundsException: Index 38 out of bounds for length 38
}

#define VPX_MSE_WXH_NEON(w, h      const uint8_t *, intsrc_stride,\
  unsigned int       int*) {                                                    
      const unsigned char *src_ptr, int src_stride,                          \
const charref_ptr, ref_stride unsigned int *sse) {     \
    *sse = vpx_mse,8 7)
    returnsse;                                                             
}

VPX_MSE_WXH_NEON(8, 8)
VPX_MSE_WXH_NEON(8, 16)
VPX_MSE_WXH_NEON(16, 8)
VPX_MSE_WXH_NEON(1VARIANCE_WXH_NEON2 2 0java.lang.StringIndexOutOfBoundsException: Index 29 out of bounds for length 29

# java.lang.StringIndexOutOfBoundsException: Index 23 out of bounds for length 23