1 /*
2  *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3  *
4  *  Usee 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 <immintrin.h>  // AVX2
12 
13 #include "./vp9_rtcd.h"
14 #include "vpx/vpx_integer.h"
15 
vp9_block_error_avx2(const int16_t * coeff,const int16_t * dqcoeff,intptr_t block_size,int64_t * ssz)16 int64_t vp9_block_error_avx2(const int16_t *coeff,
17                              const int16_t *dqcoeff,
18                              intptr_t block_size,
19                              int64_t *ssz) {
20   __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
21   __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
22   __m256i sse_reg_64hi, ssz_reg_64hi;
23   __m128i sse_reg128, ssz_reg128;
24   int64_t sse;
25   int i;
26   const __m256i zero_reg = _mm256_set1_epi16(0);
27 
28   // init sse and ssz registerd to zero
29   sse_reg = _mm256_set1_epi16(0);
30   ssz_reg = _mm256_set1_epi16(0);
31 
32   for (i = 0 ; i < block_size ; i+= 16) {
33     // load 32 bytes from coeff and dqcoeff
34     coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
35     dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
36     // dqcoeff - coeff
37     dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
38     // madd (dqcoeff - coeff)
39     dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
40     // madd coeff
41     coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
42     // expand each double word of madd (dqcoeff - coeff) to quad word
43     exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
44     exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
45     // expand each double word of madd (coeff) to quad word
46     exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
47     exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
48     // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
49     sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
50     ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
51     sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
52     ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
53   }
54   // save the higher 64 bit of each 128 bit lane
55   sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
56   ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
57   // add the higher 64 bit to the low 64 bit
58   sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
59   ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
60 
61   // add each 64 bit from each of the 128 bit lane of the 256 bit
62   sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
63                              _mm256_extractf128_si256(sse_reg, 1));
64 
65   ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
66                              _mm256_extractf128_si256(ssz_reg, 1));
67 
68   // store the results
69   _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
70 
71   _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
72   return sse;
73 }
74