1 /*
2 * Copyright (c) 2018, 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/x86/convolve_sse2.h"
18
av1_highbd_convolve_y_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)19 void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride,
20 uint16_t *dst, int dst_stride, int w, int h,
21 const InterpFilterParams *filter_params_x,
22 const InterpFilterParams *filter_params_y,
23 const int subpel_x_q4,
24 const int subpel_y_q4,
25 ConvolveParams *conv_params, int bd) {
26 int i, j;
27 const int fo_vert = filter_params_y->taps / 2 - 1;
28 const uint16_t *const src_ptr = src - fo_vert * src_stride;
29 (void)filter_params_x;
30 (void)subpel_x_q4;
31 (void)conv_params;
32
33 assert(conv_params->round_0 <= FILTER_BITS);
34 assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
35 ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
36
37 __m128i s[16], coeffs_y[4];
38
39 const int bits = FILTER_BITS;
40
41 const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
42 const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
43 const __m128i clip_pixel =
44 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
45 const __m128i zero = _mm_setzero_si128();
46
47 prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
48
49 for (j = 0; j < w; j += 8) {
50 const uint16_t *data = &src_ptr[j];
51 /* Vertical filter */
52 {
53 __m128i s0 = _mm_loadu_si128((__m128i *)(data + 0 * src_stride));
54 __m128i s1 = _mm_loadu_si128((__m128i *)(data + 1 * src_stride));
55 __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_stride));
56 __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_stride));
57 __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_stride));
58 __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_stride));
59 __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_stride));
60
61 s[0] = _mm_unpacklo_epi16(s0, s1);
62 s[1] = _mm_unpacklo_epi16(s2, s3);
63 s[2] = _mm_unpacklo_epi16(s4, s5);
64
65 s[4] = _mm_unpackhi_epi16(s0, s1);
66 s[5] = _mm_unpackhi_epi16(s2, s3);
67 s[6] = _mm_unpackhi_epi16(s4, s5);
68
69 s[0 + 8] = _mm_unpacklo_epi16(s1, s2);
70 s[1 + 8] = _mm_unpacklo_epi16(s3, s4);
71 s[2 + 8] = _mm_unpacklo_epi16(s5, s6);
72
73 s[4 + 8] = _mm_unpackhi_epi16(s1, s2);
74 s[5 + 8] = _mm_unpackhi_epi16(s3, s4);
75 s[6 + 8] = _mm_unpackhi_epi16(s5, s6);
76
77 for (i = 0; i < h; i += 2) {
78 data = &src_ptr[i * src_stride + j];
79
80 __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * src_stride));
81 __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * src_stride));
82
83 s[3] = _mm_unpacklo_epi16(s6, s7);
84 s[7] = _mm_unpackhi_epi16(s6, s7);
85
86 s[3 + 8] = _mm_unpacklo_epi16(s7, s8);
87 s[7 + 8] = _mm_unpackhi_epi16(s7, s8);
88
89 const __m128i res_a0 = convolve(s, coeffs_y);
90 __m128i res_a_round0 = _mm_sra_epi32(
91 _mm_add_epi32(res_a0, round_const_bits), round_shift_bits);
92
93 const __m128i res_a1 = convolve(s + 8, coeffs_y);
94 __m128i res_a_round1 = _mm_sra_epi32(
95 _mm_add_epi32(res_a1, round_const_bits), round_shift_bits);
96
97 if (w - j > 4) {
98 const __m128i res_b0 = convolve(s + 4, coeffs_y);
99 __m128i res_b_round0 = _mm_sra_epi32(
100 _mm_add_epi32(res_b0, round_const_bits), round_shift_bits);
101
102 const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
103 __m128i res_b_round1 = _mm_sra_epi32(
104 _mm_add_epi32(res_b1, round_const_bits), round_shift_bits);
105
106 __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
107 res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
108 res_16bit0 = _mm_max_epi16(res_16bit0, zero);
109
110 __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
111 res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
112 res_16bit1 = _mm_max_epi16(res_16bit1, zero);
113
114 _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
115 _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
116 res_16bit1);
117 } else if (w == 4) {
118 res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
119 res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
120 res_a_round0 = _mm_max_epi16(res_a_round0, zero);
121
122 res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
123 res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
124 res_a_round1 = _mm_max_epi16(res_a_round1, zero);
125
126 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
127 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
128 res_a_round1);
129 } else {
130 res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
131 res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
132 res_a_round0 = _mm_max_epi16(res_a_round0, zero);
133
134 res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
135 res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
136 res_a_round1 = _mm_max_epi16(res_a_round1, zero);
137
138 *((uint32_t *)(&dst[i * dst_stride + j])) =
139 _mm_cvtsi128_si32(res_a_round0);
140
141 *((uint32_t *)(&dst[i * dst_stride + j + dst_stride])) =
142 _mm_cvtsi128_si32(res_a_round1);
143 }
144
145 s[0] = s[1];
146 s[1] = s[2];
147 s[2] = s[3];
148
149 s[4] = s[5];
150 s[5] = s[6];
151 s[6] = s[7];
152
153 s[0 + 8] = s[1 + 8];
154 s[1 + 8] = s[2 + 8];
155 s[2 + 8] = s[3 + 8];
156
157 s[4 + 8] = s[5 + 8];
158 s[5 + 8] = s[6 + 8];
159 s[6 + 8] = s[7 + 8];
160
161 s6 = s8;
162 }
163 }
164 }
165 }
166
av1_highbd_convolve_x_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)167 void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride,
168 uint16_t *dst, int dst_stride, int w, int h,
169 const InterpFilterParams *filter_params_x,
170 const InterpFilterParams *filter_params_y,
171 const int subpel_x_q4,
172 const int subpel_y_q4,
173 ConvolveParams *conv_params, int bd) {
174 int i, j;
175 const int fo_horiz = filter_params_x->taps / 2 - 1;
176 const uint16_t *const src_ptr = src - fo_horiz;
177 (void)subpel_y_q4;
178 (void)filter_params_y;
179
180 // Check that, even with 12-bit input, the intermediate values will fit
181 // into an unsigned 16-bit intermediate array.
182 assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
183
184 __m128i s[4], coeffs_x[4];
185
186 const __m128i round_const_x =
187 _mm_set1_epi32(((1 << conv_params->round_0) >> 1));
188 const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
189
190 const int bits = FILTER_BITS - conv_params->round_0;
191
192 const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
193 const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
194 const __m128i clip_pixel =
195 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
196 const __m128i zero = _mm_setzero_si128();
197
198 prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x);
199
200 for (j = 0; j < w; j += 8) {
201 /* Horizontal filter */
202 {
203 for (i = 0; i < h; i += 1) {
204 const __m128i row00 =
205 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
206 const __m128i row01 =
207 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
208
209 // even pixels
210 s[0] = _mm_alignr_epi8(row01, row00, 0);
211 s[1] = _mm_alignr_epi8(row01, row00, 4);
212 s[2] = _mm_alignr_epi8(row01, row00, 8);
213 s[3] = _mm_alignr_epi8(row01, row00, 12);
214
215 __m128i res_even = convolve(s, coeffs_x);
216 res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
217 round_shift_x);
218
219 // odd pixels
220 s[0] = _mm_alignr_epi8(row01, row00, 2);
221 s[1] = _mm_alignr_epi8(row01, row00, 6);
222 s[2] = _mm_alignr_epi8(row01, row00, 10);
223 s[3] = _mm_alignr_epi8(row01, row00, 14);
224
225 __m128i res_odd = convolve(s, coeffs_x);
226 res_odd =
227 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x), round_shift_x);
228
229 res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_bits),
230 round_shift_bits);
231 res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_bits),
232 round_shift_bits);
233
234 __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
235 __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
236 __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
237
238 res = _mm_min_epi16(res, clip_pixel);
239 res = _mm_max_epi16(res, zero);
240
241 if (w - j > 4) {
242 _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
243 } else if (w == 4) {
244 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res);
245 } else {
246 *((uint32_t *)(&dst[i * dst_stride + j])) = _mm_cvtsi128_si32(res);
247 }
248 }
249 }
250 }
251 }
252