1 /*
2  *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include <arm_neon.h>
12 #include <assert.h>
13 
14 #include "./vp9_rtcd.h"
15 
vp9_block_error_fp_neon(const int16_t * coeff,const int16_t * dqcoeff,int block_size)16 int64_t vp9_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff,
17                                 int block_size) {
18   int64x2_t error = vdupq_n_s64(0);
19 
20   assert(block_size >= 8);
21   assert((block_size % 8) == 0);
22 
23   do {
24     const int16x8_t c = vld1q_s16(coeff);
25     const int16x8_t d = vld1q_s16(dqcoeff);
26     const int16x8_t diff = vsubq_s16(c, d);
27     const int16x4_t diff_lo = vget_low_s16(diff);
28     const int16x4_t diff_hi = vget_high_s16(diff);
29     // diff is 15-bits, the squares 30, so we can store 2 in 31-bits before
30     // accumulating them in 64-bits.
31     const int32x4_t err0 = vmull_s16(diff_lo, diff_lo);
32     const int32x4_t err1 = vmlal_s16(err0, diff_hi, diff_hi);
33     const int64x2_t err2 = vaddl_s32(vget_low_s32(err1), vget_high_s32(err1));
34     error = vaddq_s64(error, err2);
35     coeff += 8;
36     dqcoeff += 8;
37     block_size -= 8;
38   } while (block_size != 0);
39 
40   return vgetq_lane_s64(error, 0) + vgetq_lane_s64(error, 1);
41 }
42