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 <tmmintrin.h>
13 #include <assert.h>
14
15 #include "config/aom_dsp_rtcd.h"
16
17 #include "aom_dsp/aom_dsp_common.h"
18 #include "aom_dsp/aom_filter.h"
19 #include "aom_dsp/x86/convolve_sse2.h"
20 #include "av1/common/convolve.h"
21
av1_highbd_convolve_2d_sr_ssse3(const uint16_t * src,int src_stride,uint16_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_q4,const int subpel_y_q4,ConvolveParams * conv_params,int bd)22 void av1_highbd_convolve_2d_sr_ssse3(
23 const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
24 int h, const InterpFilterParams *filter_params_x,
25 const InterpFilterParams *filter_params_y, const int subpel_x_q4,
26 const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
27 DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
28 int im_h = h + filter_params_y->taps - 1;
29 int im_stride = 8;
30 int i, j;
31 const int fo_vert = filter_params_y->taps / 2 - 1;
32 const int fo_horiz = filter_params_x->taps / 2 - 1;
33 const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
34
35 // Check that, even with 12-bit input, the intermediate values will fit
36 // into an unsigned 16-bit intermediate array.
37 assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
38 __m128i coeffs_x[4], coeffs_y[4], s[16];
39
40 const __m128i round_const_x = _mm_set1_epi32(
41 ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
42 const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
43
44 const __m128i round_const_y =
45 _mm_set1_epi32(((1 << conv_params->round_1) >> 1) -
46 (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
47 const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
48
49 const int bits =
50 FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
51 const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
52 const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
53 const __m128i clip_pixel =
54 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
55 const __m128i zero = _mm_setzero_si128();
56
57 prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x);
58 prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
59
60 for (j = 0; j < w; j += 8) {
61 /* Horizontal filter */
62 {
63 for (i = 0; i < im_h; i += 1) {
64 const __m128i row00 =
65 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
66 const __m128i row01 =
67 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
68
69 // even pixels
70 s[0] = _mm_alignr_epi8(row01, row00, 0);
71 s[1] = _mm_alignr_epi8(row01, row00, 4);
72 s[2] = _mm_alignr_epi8(row01, row00, 8);
73 s[3] = _mm_alignr_epi8(row01, row00, 12);
74
75 __m128i res_even = convolve(s, coeffs_x);
76 res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
77 round_shift_x);
78
79 // odd pixels
80 s[0] = _mm_alignr_epi8(row01, row00, 2);
81 s[1] = _mm_alignr_epi8(row01, row00, 6);
82 s[2] = _mm_alignr_epi8(row01, row00, 10);
83 s[3] = _mm_alignr_epi8(row01, row00, 14);
84
85 __m128i res_odd = convolve(s, coeffs_x);
86 res_odd =
87 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x), round_shift_x);
88
89 __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
90 __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
91 __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
92
93 _mm_store_si128((__m128i *)&im_block[i * im_stride], res);
94 }
95 }
96 /* Vertical filter */
97 {
98 __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride));
99 __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride));
100 __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride));
101 __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride));
102 __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride));
103 __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride));
104 __m128i s6 = _mm_loadu_si128((__m128i *)(im_block + 6 * im_stride));
105
106 s[0] = _mm_unpacklo_epi16(s0, s1);
107 s[1] = _mm_unpacklo_epi16(s2, s3);
108 s[2] = _mm_unpacklo_epi16(s4, s5);
109
110 s[4] = _mm_unpackhi_epi16(s0, s1);
111 s[5] = _mm_unpackhi_epi16(s2, s3);
112 s[6] = _mm_unpackhi_epi16(s4, s5);
113
114 s[0 + 8] = _mm_unpacklo_epi16(s1, s2);
115 s[1 + 8] = _mm_unpacklo_epi16(s3, s4);
116 s[2 + 8] = _mm_unpacklo_epi16(s5, s6);
117
118 s[4 + 8] = _mm_unpackhi_epi16(s1, s2);
119 s[5 + 8] = _mm_unpackhi_epi16(s3, s4);
120 s[6 + 8] = _mm_unpackhi_epi16(s5, s6);
121
122 for (i = 0; i < h; i += 2) {
123 const int16_t *data = &im_block[i * im_stride];
124
125 __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));
126 __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * im_stride));
127
128 s[3] = _mm_unpacklo_epi16(s6, s7);
129 s[7] = _mm_unpackhi_epi16(s6, s7);
130
131 s[3 + 8] = _mm_unpacklo_epi16(s7, s8);
132 s[7 + 8] = _mm_unpackhi_epi16(s7, s8);
133
134 const __m128i res_a0 = convolve(s, coeffs_y);
135 __m128i res_a_round0 =
136 _mm_sra_epi32(_mm_add_epi32(res_a0, round_const_y), round_shift_y);
137 res_a_round0 = _mm_sra_epi32(
138 _mm_add_epi32(res_a_round0, round_const_bits), round_shift_bits);
139
140 const __m128i res_a1 = convolve(s + 8, coeffs_y);
141 __m128i res_a_round1 =
142 _mm_sra_epi32(_mm_add_epi32(res_a1, round_const_y), round_shift_y);
143 res_a_round1 = _mm_sra_epi32(
144 _mm_add_epi32(res_a_round1, round_const_bits), round_shift_bits);
145
146 if (w - j > 4) {
147 const __m128i res_b0 = convolve(s + 4, coeffs_y);
148 __m128i res_b_round0 = _mm_sra_epi32(
149 _mm_add_epi32(res_b0, round_const_y), round_shift_y);
150 res_b_round0 = _mm_sra_epi32(
151 _mm_add_epi32(res_b_round0, round_const_bits), round_shift_bits);
152
153 const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
154 __m128i res_b_round1 = _mm_sra_epi32(
155 _mm_add_epi32(res_b1, round_const_y), round_shift_y);
156 res_b_round1 = _mm_sra_epi32(
157 _mm_add_epi32(res_b_round1, round_const_bits), round_shift_bits);
158
159 __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
160 res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
161 res_16bit0 = _mm_max_epi16(res_16bit0, zero);
162
163 __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
164 res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
165 res_16bit1 = _mm_max_epi16(res_16bit1, zero);
166
167 _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
168 _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
169 res_16bit1);
170 } else if (w == 4) {
171 res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
172 res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
173 res_a_round0 = _mm_max_epi16(res_a_round0, zero);
174
175 res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
176 res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
177 res_a_round1 = _mm_max_epi16(res_a_round1, zero);
178
179 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
180 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
181 res_a_round1);
182 } else {
183 res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
184 res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
185 res_a_round0 = _mm_max_epi16(res_a_round0, zero);
186
187 res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
188 res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
189 res_a_round1 = _mm_max_epi16(res_a_round1, zero);
190
191 *((uint32_t *)(&dst[i * dst_stride + j])) =
192 _mm_cvtsi128_si32(res_a_round0);
193
194 *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) =
195 _mm_cvtsi128_si32(res_a_round1);
196 }
197 s[0] = s[1];
198 s[1] = s[2];
199 s[2] = s[3];
200
201 s[4] = s[5];
202 s[5] = s[6];
203 s[6] = s[7];
204
205 s[0 + 8] = s[1 + 8];
206 s[1 + 8] = s[2 + 8];
207 s[2 + 8] = s[3 + 8];
208
209 s[4 + 8] = s[5 + 8];
210 s[5 + 8] = s[6 + 8];
211 s[6 + 8] = s[7 + 8];
212
213 s6 = s8;
214 }
215 }
216 }
217 }
218