1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <emmintrin.h>
13 #include <assert.h>
14
15 #include "config/av1_rtcd.h"
16
17 #include "av1/common/convolve.h"
18 #include "aom_dsp/aom_dsp_common.h"
19 #include "aom_dsp/aom_filter.h"
20
av1_wiener_convolve_add_src_sse2(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const int16_t * filter_x,int x_step_q4,const int16_t * filter_y,int y_step_q4,int w,int h,const ConvolveParams * conv_params)21 void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride,
22 uint8_t *dst, ptrdiff_t dst_stride,
23 const int16_t *filter_x, int x_step_q4,
24 const int16_t *filter_y, int y_step_q4,
25 int w, int h,
26 const ConvolveParams *conv_params) {
27 const int bd = 8;
28 assert(x_step_q4 == 16 && y_step_q4 == 16);
29 assert(!(w & 7));
30 (void)x_step_q4;
31 (void)y_step_q4;
32
33 DECLARE_ALIGNED(16, uint16_t,
34 temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
35 int intermediate_height = h + SUBPEL_TAPS - 2;
36 memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
37 int i, j;
38 const int center_tap = ((SUBPEL_TAPS - 1) / 2);
39 const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;
40
41 const __m128i zero = _mm_setzero_si128();
42 // Add an offset to account for the "add_src" part of the convolve function.
43 const __m128i offset = _mm_insert_epi16(zero, 1 << FILTER_BITS, 3);
44
45 /* Horizontal filter */
46 {
47 const __m128i coeffs_x =
48 _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_x), offset);
49
50 // coeffs 0 1 0 1 2 3 2 3
51 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
52 // coeffs 4 5 4 5 6 7 6 7
53 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
54
55 // coeffs 0 1 0 1 0 1 0 1
56 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
57 // coeffs 2 3 2 3 2 3 2 3
58 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
59 // coeffs 4 5 4 5 4 5 4 5
60 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
61 // coeffs 6 7 6 7 6 7 6 7
62 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
63
64 const __m128i round_const = _mm_set1_epi32(
65 (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
66
67 for (i = 0; i < intermediate_height; ++i) {
68 for (j = 0; j < w; j += 8) {
69 const __m128i data =
70 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
71
72 // Filter even-index pixels
73 const __m128i src_0 = _mm_unpacklo_epi8(data, zero);
74 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
75 const __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero);
76 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
77 const __m128i src_4 = _mm_unpacklo_epi8(_mm_srli_si128(data, 4), zero);
78 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
79 const __m128i src_6 = _mm_unpacklo_epi8(_mm_srli_si128(data, 6), zero);
80 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
81
82 __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
83 _mm_add_epi32(res_2, res_6));
84 res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const),
85 conv_params->round_0);
86
87 // Filter odd-index pixels
88 const __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero);
89 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
90 const __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero);
91 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
92 const __m128i src_5 = _mm_unpacklo_epi8(_mm_srli_si128(data, 5), zero);
93 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
94 const __m128i src_7 = _mm_unpacklo_epi8(_mm_srli_si128(data, 7), zero);
95 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
96
97 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
98 _mm_add_epi32(res_3, res_7));
99 res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const),
100 conv_params->round_0);
101
102 // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
103 __m128i res = _mm_packs_epi32(res_even, res_odd);
104 res = _mm_min_epi16(
105 _mm_max_epi16(res, zero),
106 _mm_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1));
107 _mm_storeu_si128((__m128i *)&temp[i * MAX_SB_SIZE + j], res);
108 }
109 }
110 }
111
112 /* Vertical filter */
113 {
114 const __m128i coeffs_y =
115 _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_y), offset);
116
117 // coeffs 0 1 0 1 2 3 2 3
118 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
119 // coeffs 4 5 4 5 6 7 6 7
120 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
121
122 // coeffs 0 1 0 1 0 1 0 1
123 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
124 // coeffs 2 3 2 3 2 3 2 3
125 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
126 // coeffs 4 5 4 5 4 5 4 5
127 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
128 // coeffs 6 7 6 7 6 7 6 7
129 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
130
131 const __m128i round_const =
132 _mm_set1_epi32((1 << (conv_params->round_1 - 1)) -
133 (1 << (bd + conv_params->round_1 - 1)));
134
135 for (i = 0; i < h; ++i) {
136 for (j = 0; j < w; j += 8) {
137 // Filter even-index pixels
138 const uint16_t *data = &temp[i * MAX_SB_SIZE + j];
139 const __m128i src_0 =
140 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
141 *(__m128i *)(data + 1 * MAX_SB_SIZE));
142 const __m128i src_2 =
143 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
144 *(__m128i *)(data + 3 * MAX_SB_SIZE));
145 const __m128i src_4 =
146 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
147 *(__m128i *)(data + 5 * MAX_SB_SIZE));
148 const __m128i src_6 =
149 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
150 *(__m128i *)(data + 7 * MAX_SB_SIZE));
151
152 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
153 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
154 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
155 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
156
157 const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
158 _mm_add_epi32(res_4, res_6));
159
160 // Filter odd-index pixels
161 const __m128i src_1 =
162 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
163 *(__m128i *)(data + 1 * MAX_SB_SIZE));
164 const __m128i src_3 =
165 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
166 *(__m128i *)(data + 3 * MAX_SB_SIZE));
167 const __m128i src_5 =
168 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
169 *(__m128i *)(data + 5 * MAX_SB_SIZE));
170 const __m128i src_7 =
171 _mm_unpackhi_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
172 *(__m128i *)(data + 7 * MAX_SB_SIZE));
173
174 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
175 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
176 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
177 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
178
179 const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
180 _mm_add_epi32(res_5, res_7));
181
182 // Rearrange pixels back into the order 0 ... 7
183 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
184 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
185
186 const __m128i res_lo_round = _mm_srai_epi32(
187 _mm_add_epi32(res_lo, round_const), conv_params->round_1);
188 const __m128i res_hi_round = _mm_srai_epi32(
189 _mm_add_epi32(res_hi, round_const), conv_params->round_1);
190
191 const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
192 __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit);
193
194 __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
195 _mm_storel_epi64(p, res_8bit);
196 }
197 }
198 }
199 }
200