1 /*
2  * Copyright (C) 2013 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #ifndef ANDROID_AUDIO_RESAMPLER_FIR_OPS_H
18 #define ANDROID_AUDIO_RESAMPLER_FIR_OPS_H
19 
20 namespace android {
21 
22 #if defined(__arm__) && !defined(__thumb__)
23 #define USE_INLINE_ASSEMBLY (true)
24 #else
25 #define USE_INLINE_ASSEMBLY (false)
26 #endif
27 
28 #if defined(__aarch64__) || defined(__ARM_NEON__)
29 #ifndef USE_NEON
30 #define USE_NEON (true)
31 #endif
32 #else
33 #define USE_NEON (false)
34 #endif
35 #if USE_NEON
36 #include <arm_neon.h>
37 #endif
38 
39 #if defined(__SSSE3__)  // Should be supported in x86 ABI for both 32 & 64-bit.
40 #define USE_SSE (true)
41 #include <tmmintrin.h>
42 #else
43 #define USE_SSE (false)
44 #endif
45 
46 template<typename T, typename U>
47 struct is_same
48 {
49     static const bool value = false;
50 };
51 
52 template<typename T>
53 struct is_same<T, T>  // partial specialization
54 {
55     static const bool value = true;
56 };
57 
58 static inline
59 int32_t mulRL(int left, int32_t in, uint32_t vRL)
60 {
61 #if USE_INLINE_ASSEMBLY
62     int32_t out;
63     if (left) {
64         asm( "smultb %[out], %[in], %[vRL] \n"
65              : [out]"=r"(out)
66              : [in]"%r"(in), [vRL]"r"(vRL)
67              : );
68     } else {
69         asm( "smultt %[out], %[in], %[vRL] \n"
70              : [out]"=r"(out)
71              : [in]"%r"(in), [vRL]"r"(vRL)
72              : );
73     }
74     return out;
75 #else
76     int16_t v = left ? static_cast<int16_t>(vRL) : static_cast<int16_t>(vRL>>16);
77     return static_cast<int32_t>((static_cast<int64_t>(in) * v) >> 16);
78 #endif
79 }
80 
81 static inline
82 int32_t mulAdd(int16_t in, int16_t v, int32_t a)
83 {
84 #if USE_INLINE_ASSEMBLY
85     int32_t out;
86     asm( "smlabb %[out], %[v], %[in], %[a] \n"
87          : [out]"=r"(out)
88          : [in]"%r"(in), [v]"r"(v), [a]"r"(a)
89          : );
90     return out;
91 #else
92     return a + v * in;
93 #endif
94 }
95 
96 static inline
97 int32_t mulAdd(int16_t in, int32_t v, int32_t a)
98 {
99 #if USE_INLINE_ASSEMBLY
100     int32_t out;
101     asm( "smlawb %[out], %[v], %[in], %[a] \n"
102          : [out]"=r"(out)
103          : [in]"%r"(in), [v]"r"(v), [a]"r"(a)
104          : );
105     return out;
106 #else
107     return a + static_cast<int32_t>((static_cast<int64_t>(v) * in) >> 16);
108 #endif
109 }
110 
111 static inline
112 int32_t mulAdd(int32_t in, int32_t v, int32_t a)
113 {
114 #if USE_INLINE_ASSEMBLY
115     int32_t out;
116     asm( "smmla %[out], %[v], %[in], %[a] \n"
117          : [out]"=r"(out)
118          : [in]"%r"(in), [v]"r"(v), [a]"r"(a)
119          : );
120     return out;
121 #else
122     return a + static_cast<int32_t>((static_cast<int64_t>(v) * in) >> 32);
123 #endif
124 }
125 
126 static inline
127 int32_t mulAddRL(int left, uint32_t inRL, int16_t v, int32_t a)
128 {
129 #if 0 // USE_INLINE_ASSEMBLY Seems to fail with Clang b/34110890
130     int32_t out;
131     if (left) {
132         asm( "smlabb %[out], %[v], %[inRL], %[a] \n"
133              : [out]"=r"(out)
134              : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
135              : );
136     } else {
137         asm( "smlabt %[out], %[v], %[inRL], %[a] \n"
138              : [out]"=r"(out)
139              : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
140              : );
141     }
142     return out;
143 #else
144     int16_t s = left ? static_cast<int16_t>(inRL) : static_cast<int16_t>(inRL>>16);
145     return a + v * s;
146 #endif
147 }
148 
149 static inline
150 int32_t mulAddRL(int left, uint32_t inRL, int32_t v, int32_t a)
151 {
152 #if 0 // USE_INLINE_ASSEMBLY Seems to fail with Clang b/34110890
153     int32_t out;
154     if (left) {
155         asm( "smlawb %[out], %[v], %[inRL], %[a] \n"
156              : [out]"=r"(out)
157              : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
158              : );
159     } else {
160         asm( "smlawt %[out], %[v], %[inRL], %[a] \n"
161              : [out]"=r"(out)
162              : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
163              : );
164     }
165     return out;
166 #else
167     int16_t s = left ? static_cast<int16_t>(inRL) : static_cast<int16_t>(inRL>>16);
168     return a + static_cast<int32_t>((static_cast<int64_t>(v) * s) >> 16);
169 #endif
170 }
171 
172 } // namespace android
173 
174 #endif /*ANDROID_AUDIO_RESAMPLER_FIR_OPS_H*/
175