1/*
2 *  Copyright 2012 The LibYuv 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
11#include "libyuv/basic_types.h"
12#include "libyuv/row.h"
13
14#ifdef __cplusplus
15namespace libyuv {
16extern "C" {
17#endif
18
19#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
20
21uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
22  uint32 sse;
23  asm volatile (  // NOLINT
24    "pxor      %%xmm0,%%xmm0                   \n"
25    "pxor      %%xmm5,%%xmm5                   \n"
26    LABELALIGN
27  "1:                                          \n"
28    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
29    "lea       " MEMLEA(0x10, 0) ",%0          \n"
30    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
31    "lea       " MEMLEA(0x10, 1) ",%1          \n"
32    "sub       $0x10,%2                        \n"
33    "movdqa    %%xmm1,%%xmm3                   \n"
34    "psubusb   %%xmm2,%%xmm1                   \n"
35    "psubusb   %%xmm3,%%xmm2                   \n"
36    "por       %%xmm2,%%xmm1                   \n"
37    "movdqa    %%xmm1,%%xmm2                   \n"
38    "punpcklbw %%xmm5,%%xmm1                   \n"
39    "punpckhbw %%xmm5,%%xmm2                   \n"
40    "pmaddwd   %%xmm1,%%xmm1                   \n"
41    "pmaddwd   %%xmm2,%%xmm2                   \n"
42    "paddd     %%xmm1,%%xmm0                   \n"
43    "paddd     %%xmm2,%%xmm0                   \n"
44    "jg        1b                              \n"
45
46    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
47    "paddd     %%xmm1,%%xmm0                   \n"
48    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
49    "paddd     %%xmm1,%%xmm0                   \n"
50    "movd      %%xmm0,%3                       \n"
51
52  : "+r"(src_a),      // %0
53    "+r"(src_b),      // %1
54    "+r"(count),      // %2
55    "=g"(sse)         // %3
56  :
57  : "memory", "cc"
58#if defined(__SSE2__)
59    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
60#endif
61  );  // NOLINT
62  return sse;
63}
64
65#endif  // defined(__x86_64__) || defined(__i386__)
66
67#if !defined(LIBYUV_DISABLE_X86) && \
68    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
69#define HAS_HASHDJB2_SSE41
70static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
71static uvec32 kHashMul0 = {
72  0x0c3525e1,  // 33 ^ 15
73  0xa3476dc1,  // 33 ^ 14
74  0x3b4039a1,  // 33 ^ 13
75  0x4f5f0981,  // 33 ^ 12
76};
77static uvec32 kHashMul1 = {
78  0x30f35d61,  // 33 ^ 11
79  0x855cb541,  // 33 ^ 10
80  0x040a9121,  // 33 ^ 9
81  0x747c7101,  // 33 ^ 8
82};
83static uvec32 kHashMul2 = {
84  0xec41d4e1,  // 33 ^ 7
85  0x4cfa3cc1,  // 33 ^ 6
86  0x025528a1,  // 33 ^ 5
87  0x00121881,  // 33 ^ 4
88};
89static uvec32 kHashMul3 = {
90  0x00008c61,  // 33 ^ 3
91  0x00000441,  // 33 ^ 2
92  0x00000021,  // 33 ^ 1
93  0x00000001,  // 33 ^ 0
94};
95
96uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
97  uint32 hash;
98  asm volatile (  // NOLINT
99    "movd      %2,%%xmm0                       \n"
100    "pxor      %%xmm7,%%xmm7                   \n"
101    "movdqa    %4,%%xmm6                       \n"
102    LABELALIGN
103  "1:                                          \n"
104    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
105    "lea       " MEMLEA(0x10, 0) ",%0          \n"
106    "pmulld    %%xmm6,%%xmm0                   \n"
107    "movdqa    %5,%%xmm5                       \n"
108    "movdqa    %%xmm1,%%xmm2                   \n"
109    "punpcklbw %%xmm7,%%xmm2                   \n"
110    "movdqa    %%xmm2,%%xmm3                   \n"
111    "punpcklwd %%xmm7,%%xmm3                   \n"
112    "pmulld    %%xmm5,%%xmm3                   \n"
113    "movdqa    %6,%%xmm5                       \n"
114    "movdqa    %%xmm2,%%xmm4                   \n"
115    "punpckhwd %%xmm7,%%xmm4                   \n"
116    "pmulld    %%xmm5,%%xmm4                   \n"
117    "movdqa    %7,%%xmm5                       \n"
118    "punpckhbw %%xmm7,%%xmm1                   \n"
119    "movdqa    %%xmm1,%%xmm2                   \n"
120    "punpcklwd %%xmm7,%%xmm2                   \n"
121    "pmulld    %%xmm5,%%xmm2                   \n"
122    "movdqa    %8,%%xmm5                       \n"
123    "punpckhwd %%xmm7,%%xmm1                   \n"
124    "pmulld    %%xmm5,%%xmm1                   \n"
125    "paddd     %%xmm4,%%xmm3                   \n"
126    "paddd     %%xmm2,%%xmm1                   \n"
127    "sub       $0x10,%1                        \n"
128    "paddd     %%xmm3,%%xmm1                   \n"
129    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
130    "paddd     %%xmm2,%%xmm1                   \n"
131    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
132    "paddd     %%xmm2,%%xmm1                   \n"
133    "paddd     %%xmm1,%%xmm0                   \n"
134    "jg        1b                              \n"
135    "movd      %%xmm0,%3                       \n"
136  : "+r"(src),        // %0
137    "+r"(count),      // %1
138    "+rm"(seed),      // %2
139    "=g"(hash)        // %3
140  : "m"(kHash16x33),  // %4
141    "m"(kHashMul0),   // %5
142    "m"(kHashMul1),   // %6
143    "m"(kHashMul2),   // %7
144    "m"(kHashMul3)    // %8
145  : "memory", "cc"
146#if defined(__SSE2__)
147    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
148#endif
149  );  // NOLINT
150  return hash;
151}
152#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
153
154#ifdef __cplusplus
155}  // extern "C"
156}  // namespace libyuv
157#endif
158
159