1 /*
2 * Copyright (c) 2017 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 #include <immintrin.h> // AVX512
11 #include "./vpx_dsp_rtcd.h"
12 #include "vpx/vpx_integer.h"
13
vpx_sad64x64x4d_avx512(const uint8_t * src_ptr,int src_stride,const uint8_t * const ref_array[4],int ref_stride,uint32_t res[4])14 void vpx_sad64x64x4d_avx512(const uint8_t *src_ptr, int src_stride,
15 const uint8_t *const ref_array[4], int ref_stride,
16 uint32_t res[4]) {
17 __m512i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
18 __m512i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
19 __m512i sum_mlow, sum_mhigh;
20 int i;
21 const uint8_t *ref0, *ref1, *ref2, *ref3;
22
23 ref0 = ref_array[0];
24 ref1 = ref_array[1];
25 ref2 = ref_array[2];
26 ref3 = ref_array[3];
27 sum_ref0 = _mm512_set1_epi16(0);
28 sum_ref1 = _mm512_set1_epi16(0);
29 sum_ref2 = _mm512_set1_epi16(0);
30 sum_ref3 = _mm512_set1_epi16(0);
31 for (i = 0; i < 64; i++) {
32 // load src and all ref[]
33 src_reg = _mm512_loadu_si512((const __m512i *)src_ptr);
34 ref0_reg = _mm512_loadu_si512((const __m512i *)ref0);
35 ref1_reg = _mm512_loadu_si512((const __m512i *)ref1);
36 ref2_reg = _mm512_loadu_si512((const __m512i *)ref2);
37 ref3_reg = _mm512_loadu_si512((const __m512i *)ref3);
38 // sum of the absolute differences between every ref[] to src
39 ref0_reg = _mm512_sad_epu8(ref0_reg, src_reg);
40 ref1_reg = _mm512_sad_epu8(ref1_reg, src_reg);
41 ref2_reg = _mm512_sad_epu8(ref2_reg, src_reg);
42 ref3_reg = _mm512_sad_epu8(ref3_reg, src_reg);
43 // sum every ref[]
44 sum_ref0 = _mm512_add_epi32(sum_ref0, ref0_reg);
45 sum_ref1 = _mm512_add_epi32(sum_ref1, ref1_reg);
46 sum_ref2 = _mm512_add_epi32(sum_ref2, ref2_reg);
47 sum_ref3 = _mm512_add_epi32(sum_ref3, ref3_reg);
48
49 src_ptr += src_stride;
50 ref0 += ref_stride;
51 ref1 += ref_stride;
52 ref2 += ref_stride;
53 ref3 += ref_stride;
54 }
55 {
56 __m256i sum256;
57 __m128i sum128;
58 // in sum_ref[] the result is saved in the first 4 bytes
59 // the other 4 bytes are zeroed.
60 // sum_ref1 and sum_ref3 are shifted left by 4 bytes
61 sum_ref1 = _mm512_bslli_epi128(sum_ref1, 4);
62 sum_ref3 = _mm512_bslli_epi128(sum_ref3, 4);
63
64 // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
65 sum_ref0 = _mm512_or_si512(sum_ref0, sum_ref1);
66 sum_ref2 = _mm512_or_si512(sum_ref2, sum_ref3);
67
68 // merge every 64 bit from each sum_ref[]
69 sum_mlow = _mm512_unpacklo_epi64(sum_ref0, sum_ref2);
70 sum_mhigh = _mm512_unpackhi_epi64(sum_ref0, sum_ref2);
71
72 // add the low 64 bit to the high 64 bit
73 sum_mlow = _mm512_add_epi32(sum_mlow, sum_mhigh);
74
75 // add the low 128 bit to the high 128 bit
76 sum256 = _mm256_add_epi32(_mm512_castsi512_si256(sum_mlow),
77 _mm512_extracti32x8_epi32(sum_mlow, 1));
78 sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum256),
79 _mm256_extractf128_si256(sum256, 1));
80
81 _mm_storeu_si128((__m128i *)(res), sum128);
82 }
83 }
84