1 /*
2  *  Copyright (c) 2014 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 <emmintrin.h>
12 #include <stdio.h>
13 
14 #include "vp9/common/vp9_common.h"
15 
vp9_highbd_block_error_sse2(tran_low_t * coeff,tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bps)16 int64_t vp9_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
17                                     intptr_t block_size, int64_t *ssz,
18                                     int bps) {
19   int i, j, test;
20   uint32_t temp[4];
21   __m128i max, min, cmp0, cmp1, cmp2, cmp3;
22   int64_t error = 0, sqcoeff = 0;
23   const int shift = 2 * (bps - 8);
24   const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
25 
26   for (i = 0; i < block_size; i+=8) {
27     // Load the data into xmm registers
28     __m128i mm_coeff = _mm_load_si128((__m128i*) (coeff + i));
29     __m128i mm_coeff2 = _mm_load_si128((__m128i*) (coeff + i + 4));
30     __m128i mm_dqcoeff = _mm_load_si128((__m128i*) (dqcoeff + i));
31     __m128i mm_dqcoeff2 = _mm_load_si128((__m128i*) (dqcoeff + i + 4));
32     // Check if any values require more than 15 bit
33     max = _mm_set1_epi32(0x3fff);
34     min = _mm_set1_epi32(0xffffc000);
35     cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
36             _mm_cmplt_epi32(mm_coeff, min));
37     cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
38             _mm_cmplt_epi32(mm_coeff2, min));
39     cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
40             _mm_cmplt_epi32(mm_dqcoeff, min));
41     cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
42             _mm_cmplt_epi32(mm_dqcoeff2, min));
43     test = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(cmp0, cmp1),
44             _mm_or_si128(cmp2, cmp3)));
45 
46     if (!test) {
47       __m128i mm_diff, error_sse2, sqcoeff_sse2;;
48       mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
49       mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
50       mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
51       error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
52       sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
53       _mm_storeu_si128((__m128i*)temp, error_sse2);
54       error = error + temp[0] + temp[1] + temp[2] + temp[3];
55       _mm_storeu_si128((__m128i*)temp, sqcoeff_sse2);
56       sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
57     } else {
58       for (j = 0; j < 8; j++) {
59         const int64_t diff = coeff[i + j] - dqcoeff[i + j];
60         error +=  diff * diff;
61         sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
62       }
63     }
64   }
65   assert(error >= 0 && sqcoeff >= 0);
66   error = (error + rounding) >> shift;
67   sqcoeff = (sqcoeff + rounding) >> shift;
68 
69   *ssz = sqcoeff;
70   return error;
71 }
72