Line data Source code
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 <assert.h>
13 : #include <emmintrin.h>
14 : #include <stdio.h>
15 :
16 : #include "aom_dsp/x86/synonyms.h"
17 :
18 : #include "./aom_dsp_rtcd.h"
19 :
20 0 : static uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src,
21 : int stride) {
22 0 : const __m128i v_val_0_w =
23 : _mm_loadl_epi64((const __m128i *)(src + 0 * stride));
24 0 : const __m128i v_val_1_w =
25 0 : _mm_loadl_epi64((const __m128i *)(src + 1 * stride));
26 0 : const __m128i v_val_2_w =
27 0 : _mm_loadl_epi64((const __m128i *)(src + 2 * stride));
28 0 : const __m128i v_val_3_w =
29 0 : _mm_loadl_epi64((const __m128i *)(src + 3 * stride));
30 :
31 0 : const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
32 0 : const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
33 0 : const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
34 0 : const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
35 :
36 0 : const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
37 0 : const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
38 0 : const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
39 :
40 0 : const __m128i v_sum_d =
41 0 : _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));
42 :
43 0 : return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
44 : }
45 :
46 : #ifdef __GNUC__
47 : // This prevents GCC/Clang from inlining this function into
48 : // aom_sum_squares_2d_i16_sse2, which in turn saves some stack
49 : // maintenance instructions in the common case of 4x4.
50 : __attribute__((noinline))
51 : #endif
52 : static uint64_t
53 0 : aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
54 : int height) {
55 : int r, c;
56 :
57 0 : const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
58 0 : __m128i v_acc_q = _mm_setzero_si128();
59 :
60 0 : for (r = 0; r < height; r += 8) {
61 0 : __m128i v_acc_d = _mm_setzero_si128();
62 :
63 0 : for (c = 0; c < width; c += 8) {
64 0 : const int16_t *b = src + c;
65 :
66 0 : const __m128i v_val_0_w =
67 : _mm_load_si128((const __m128i *)(b + 0 * stride));
68 0 : const __m128i v_val_1_w =
69 0 : _mm_load_si128((const __m128i *)(b + 1 * stride));
70 0 : const __m128i v_val_2_w =
71 0 : _mm_load_si128((const __m128i *)(b + 2 * stride));
72 0 : const __m128i v_val_3_w =
73 0 : _mm_load_si128((const __m128i *)(b + 3 * stride));
74 0 : const __m128i v_val_4_w =
75 0 : _mm_load_si128((const __m128i *)(b + 4 * stride));
76 0 : const __m128i v_val_5_w =
77 0 : _mm_load_si128((const __m128i *)(b + 5 * stride));
78 0 : const __m128i v_val_6_w =
79 0 : _mm_load_si128((const __m128i *)(b + 6 * stride));
80 0 : const __m128i v_val_7_w =
81 0 : _mm_load_si128((const __m128i *)(b + 7 * stride));
82 :
83 0 : const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
84 0 : const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
85 0 : const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
86 0 : const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
87 0 : const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
88 0 : const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
89 0 : const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
90 0 : const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
91 :
92 0 : const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
93 0 : const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
94 0 : const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
95 0 : const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
96 :
97 0 : const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
98 0 : const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
99 :
100 0 : v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
101 0 : v_acc_d = _mm_add_epi32(v_acc_d, v_sum_4567_d);
102 : }
103 :
104 0 : v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
105 0 : v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));
106 :
107 0 : src += 8 * stride;
108 : }
109 :
110 0 : v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
111 :
112 : #if ARCH_X86_64
113 0 : return (uint64_t)_mm_cvtsi128_si64(v_acc_q);
114 : #else
115 : {
116 : uint64_t tmp;
117 : _mm_storel_epi64((__m128i *)&tmp, v_acc_q);
118 : return tmp;
119 : }
120 : #endif
121 : }
122 :
123 0 : uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width,
124 : int height) {
125 : // 4 elements per row only requires half an XMM register, so this
126 : // must be a special case, but also note that over 75% of all calls
127 : // are with size == 4, so it is also the common case.
128 0 : if (LIKELY(width == 4 && height == 4)) {
129 0 : return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
130 0 : } else if (LIKELY(width % 8 == 0 && height % 8 == 0)) {
131 : // Generic case
132 0 : return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
133 : } else {
134 0 : return aom_sum_squares_2d_i16_c(src, stride, width, height);
135 : }
136 : }
137 :
138 : //////////////////////////////////////////////////////////////////////////////
139 : // 1D version
140 : //////////////////////////////////////////////////////////////////////////////
141 :
142 0 : static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
143 0 : const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
144 0 : __m128i v_acc0_q = _mm_setzero_si128();
145 0 : __m128i v_acc1_q = _mm_setzero_si128();
146 :
147 0 : const int16_t *const end = src + n;
148 :
149 0 : assert(n % 64 == 0);
150 :
151 0 : while (src < end) {
152 0 : const __m128i v_val_0_w = xx_load_128(src);
153 0 : const __m128i v_val_1_w = xx_load_128(src + 8);
154 0 : const __m128i v_val_2_w = xx_load_128(src + 16);
155 0 : const __m128i v_val_3_w = xx_load_128(src + 24);
156 0 : const __m128i v_val_4_w = xx_load_128(src + 32);
157 0 : const __m128i v_val_5_w = xx_load_128(src + 40);
158 0 : const __m128i v_val_6_w = xx_load_128(src + 48);
159 0 : const __m128i v_val_7_w = xx_load_128(src + 56);
160 :
161 0 : const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
162 0 : const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
163 0 : const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
164 0 : const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
165 0 : const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
166 0 : const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
167 0 : const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
168 0 : const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
169 :
170 0 : const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
171 0 : const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
172 0 : const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
173 0 : const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
174 :
175 0 : const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
176 0 : const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
177 :
178 0 : const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);
179 :
180 0 : v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
181 0 : v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));
182 :
183 0 : src += 64;
184 : }
185 :
186 0 : v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
187 0 : v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
188 :
189 : #if ARCH_X86_64
190 0 : return (uint64_t)_mm_cvtsi128_si64(v_acc0_q);
191 : #else
192 : {
193 : uint64_t tmp;
194 : _mm_storel_epi64((__m128i *)&tmp, v_acc0_q);
195 : return tmp;
196 : }
197 : #endif
198 : }
199 :
200 0 : uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
201 0 : if (n % 64 == 0) {
202 0 : return aom_sum_squares_i16_64n_sse2(src, n);
203 0 : } else if (n > 64) {
204 0 : int k = n & ~(64 - 1);
205 0 : return aom_sum_squares_i16_64n_sse2(src, k) +
206 0 : aom_sum_squares_i16_c(src + k, n - k);
207 : } else {
208 0 : return aom_sum_squares_i16_c(src, n);
209 : }
210 : }
|
