1*ec779b8eSAndroid Build Coastguard Worker /* 2*ec779b8eSAndroid Build Coastguard Worker * Copyright (C) 2013 The Android Open Source Project 3*ec779b8eSAndroid Build Coastguard Worker * 4*ec779b8eSAndroid Build Coastguard Worker * Licensed under the Apache License, Version 2.0 (the "License"); 5*ec779b8eSAndroid Build Coastguard Worker * you may not use this file except in compliance with the License. 6*ec779b8eSAndroid Build Coastguard Worker * You may obtain a copy of the License at 7*ec779b8eSAndroid Build Coastguard Worker * 8*ec779b8eSAndroid Build Coastguard Worker * http://www.apache.org/licenses/LICENSE-2.0 9*ec779b8eSAndroid Build Coastguard Worker * 10*ec779b8eSAndroid Build Coastguard Worker * Unless required by applicable law or agreed to in writing, software 11*ec779b8eSAndroid Build Coastguard Worker * distributed under the License is distributed on an "AS IS" BASIS, 12*ec779b8eSAndroid Build Coastguard Worker * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13*ec779b8eSAndroid Build Coastguard Worker * See the License for the specific language governing permissions and 14*ec779b8eSAndroid Build Coastguard Worker * limitations under the License. 15*ec779b8eSAndroid Build Coastguard Worker */ 16*ec779b8eSAndroid Build Coastguard Worker 17*ec779b8eSAndroid Build Coastguard Worker #ifndef ANDROID_AUDIO_RESAMPLER_FIR_OPS_H 18*ec779b8eSAndroid Build Coastguard Worker #define ANDROID_AUDIO_RESAMPLER_FIR_OPS_H 19*ec779b8eSAndroid Build Coastguard Worker 20*ec779b8eSAndroid Build Coastguard Worker namespace android { 21*ec779b8eSAndroid Build Coastguard Worker 22*ec779b8eSAndroid Build Coastguard Worker #if defined(__arm__) && !defined(__thumb__) 23*ec779b8eSAndroid Build Coastguard Worker #define USE_INLINE_ASSEMBLY (true) 24*ec779b8eSAndroid Build Coastguard Worker #else 25*ec779b8eSAndroid Build Coastguard Worker #define USE_INLINE_ASSEMBLY (false) 26*ec779b8eSAndroid Build Coastguard Worker #endif 27*ec779b8eSAndroid Build Coastguard Worker 28*ec779b8eSAndroid Build Coastguard Worker #if defined(__aarch64__) || defined(__ARM_NEON__) 29*ec779b8eSAndroid Build Coastguard Worker #ifndef USE_NEON 30*ec779b8eSAndroid Build Coastguard Worker #define USE_NEON (true) 31*ec779b8eSAndroid Build Coastguard Worker #endif 32*ec779b8eSAndroid Build Coastguard Worker #else 33*ec779b8eSAndroid Build Coastguard Worker #define USE_NEON (false) 34*ec779b8eSAndroid Build Coastguard Worker #endif 35*ec779b8eSAndroid Build Coastguard Worker #if USE_NEON 36*ec779b8eSAndroid Build Coastguard Worker #include <arm_neon.h> 37*ec779b8eSAndroid Build Coastguard Worker #endif 38*ec779b8eSAndroid Build Coastguard Worker 39*ec779b8eSAndroid Build Coastguard Worker #if defined(__AVX2__) // Should be supported in x86 ABI for both 32 & 64-bit. 40*ec779b8eSAndroid Build Coastguard Worker #define USE_AVX2 (true) // Inference AVX2/FMA Intrinsics 41*ec779b8eSAndroid Build Coastguard Worker #define USE_SSE (true) 42*ec779b8eSAndroid Build Coastguard Worker #include <immintrin.h> 43*ec779b8eSAndroid Build Coastguard Worker #elif defined(__SSSE3__) // Should be supported in x86 ABI for both 32 & 64-bit. 44*ec779b8eSAndroid Build Coastguard Worker #define USE_SSE (true) // Inference SSE Intrinsics 45*ec779b8eSAndroid Build Coastguard Worker #define USE_AVX2 (false) 46*ec779b8eSAndroid Build Coastguard Worker #include <tmmintrin.h> 47*ec779b8eSAndroid Build Coastguard Worker #else 48*ec779b8eSAndroid Build Coastguard Worker #define USE_SSE (false) 49*ec779b8eSAndroid Build Coastguard Worker #define USE_AVX2(false) 50*ec779b8eSAndroid Build Coastguard Worker #endif 51*ec779b8eSAndroid Build Coastguard Worker 52*ec779b8eSAndroid Build Coastguard Worker 53*ec779b8eSAndroid Build Coastguard Worker template<typename T, typename U> 54*ec779b8eSAndroid Build Coastguard Worker struct is_same 55*ec779b8eSAndroid Build Coastguard Worker { 56*ec779b8eSAndroid Build Coastguard Worker static const bool value = false; 57*ec779b8eSAndroid Build Coastguard Worker }; 58*ec779b8eSAndroid Build Coastguard Worker 59*ec779b8eSAndroid Build Coastguard Worker template<typename T> 60*ec779b8eSAndroid Build Coastguard Worker struct is_same<T, T> // partial specialization 61*ec779b8eSAndroid Build Coastguard Worker { 62*ec779b8eSAndroid Build Coastguard Worker static const bool value = true; 63*ec779b8eSAndroid Build Coastguard Worker }; 64*ec779b8eSAndroid Build Coastguard Worker 65*ec779b8eSAndroid Build Coastguard Worker static inline 66*ec779b8eSAndroid Build Coastguard Worker int32_t mulRL(int left, int32_t in, uint32_t vRL) 67*ec779b8eSAndroid Build Coastguard Worker { 68*ec779b8eSAndroid Build Coastguard Worker #if USE_INLINE_ASSEMBLY 69*ec779b8eSAndroid Build Coastguard Worker int32_t out; 70*ec779b8eSAndroid Build Coastguard Worker if (left) { 71*ec779b8eSAndroid Build Coastguard Worker asm( "smultb %[out], %[in], %[vRL] \n" 72*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 73*ec779b8eSAndroid Build Coastguard Worker : [in]"%r"(in), [vRL]"r"(vRL) 74*ec779b8eSAndroid Build Coastguard Worker : ); 75*ec779b8eSAndroid Build Coastguard Worker } else { 76*ec779b8eSAndroid Build Coastguard Worker asm( "smultt %[out], %[in], %[vRL] \n" 77*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 78*ec779b8eSAndroid Build Coastguard Worker : [in]"%r"(in), [vRL]"r"(vRL) 79*ec779b8eSAndroid Build Coastguard Worker : ); 80*ec779b8eSAndroid Build Coastguard Worker } 81*ec779b8eSAndroid Build Coastguard Worker return out; 82*ec779b8eSAndroid Build Coastguard Worker #else 83*ec779b8eSAndroid Build Coastguard Worker int16_t v = left ? static_cast<int16_t>(vRL) : static_cast<int16_t>(vRL>>16); 84*ec779b8eSAndroid Build Coastguard Worker return static_cast<int32_t>((static_cast<int64_t>(in) * v) >> 16); 85*ec779b8eSAndroid Build Coastguard Worker #endif 86*ec779b8eSAndroid Build Coastguard Worker } 87*ec779b8eSAndroid Build Coastguard Worker 88*ec779b8eSAndroid Build Coastguard Worker static inline 89*ec779b8eSAndroid Build Coastguard Worker int32_t mulAdd(int16_t in, int16_t v, int32_t a) 90*ec779b8eSAndroid Build Coastguard Worker { 91*ec779b8eSAndroid Build Coastguard Worker #if USE_INLINE_ASSEMBLY 92*ec779b8eSAndroid Build Coastguard Worker int32_t out; 93*ec779b8eSAndroid Build Coastguard Worker asm( "smlabb %[out], %[v], %[in], %[a] \n" 94*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 95*ec779b8eSAndroid Build Coastguard Worker : [in]"%r"(in), [v]"r"(v), [a]"r"(a) 96*ec779b8eSAndroid Build Coastguard Worker : ); 97*ec779b8eSAndroid Build Coastguard Worker return out; 98*ec779b8eSAndroid Build Coastguard Worker #else 99*ec779b8eSAndroid Build Coastguard Worker return a + v * in; 100*ec779b8eSAndroid Build Coastguard Worker #endif 101*ec779b8eSAndroid Build Coastguard Worker } 102*ec779b8eSAndroid Build Coastguard Worker 103*ec779b8eSAndroid Build Coastguard Worker static inline 104*ec779b8eSAndroid Build Coastguard Worker int32_t mulAdd(int16_t in, int32_t v, int32_t a) 105*ec779b8eSAndroid Build Coastguard Worker { 106*ec779b8eSAndroid Build Coastguard Worker #if USE_INLINE_ASSEMBLY 107*ec779b8eSAndroid Build Coastguard Worker int32_t out; 108*ec779b8eSAndroid Build Coastguard Worker asm( "smlawb %[out], %[v], %[in], %[a] \n" 109*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 110*ec779b8eSAndroid Build Coastguard Worker : [in]"%r"(in), [v]"r"(v), [a]"r"(a) 111*ec779b8eSAndroid Build Coastguard Worker : ); 112*ec779b8eSAndroid Build Coastguard Worker return out; 113*ec779b8eSAndroid Build Coastguard Worker #else 114*ec779b8eSAndroid Build Coastguard Worker return a + static_cast<int32_t>((static_cast<int64_t>(v) * in) >> 16); 115*ec779b8eSAndroid Build Coastguard Worker #endif 116*ec779b8eSAndroid Build Coastguard Worker } 117*ec779b8eSAndroid Build Coastguard Worker 118*ec779b8eSAndroid Build Coastguard Worker static inline 119*ec779b8eSAndroid Build Coastguard Worker int32_t mulAdd(int32_t in, int32_t v, int32_t a) 120*ec779b8eSAndroid Build Coastguard Worker { 121*ec779b8eSAndroid Build Coastguard Worker #if USE_INLINE_ASSEMBLY 122*ec779b8eSAndroid Build Coastguard Worker int32_t out; 123*ec779b8eSAndroid Build Coastguard Worker asm( "smmla %[out], %[v], %[in], %[a] \n" 124*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 125*ec779b8eSAndroid Build Coastguard Worker : [in]"%r"(in), [v]"r"(v), [a]"r"(a) 126*ec779b8eSAndroid Build Coastguard Worker : ); 127*ec779b8eSAndroid Build Coastguard Worker return out; 128*ec779b8eSAndroid Build Coastguard Worker #else 129*ec779b8eSAndroid Build Coastguard Worker return a + static_cast<int32_t>((static_cast<int64_t>(v) * in) >> 32); 130*ec779b8eSAndroid Build Coastguard Worker #endif 131*ec779b8eSAndroid Build Coastguard Worker } 132*ec779b8eSAndroid Build Coastguard Worker 133*ec779b8eSAndroid Build Coastguard Worker static inline 134*ec779b8eSAndroid Build Coastguard Worker int32_t mulAddRL(int left, uint32_t inRL, int16_t v, int32_t a) 135*ec779b8eSAndroid Build Coastguard Worker { 136*ec779b8eSAndroid Build Coastguard Worker #if 0 // USE_INLINE_ASSEMBLY Seems to fail with Clang b/34110890 137*ec779b8eSAndroid Build Coastguard Worker int32_t out; 138*ec779b8eSAndroid Build Coastguard Worker if (left) { 139*ec779b8eSAndroid Build Coastguard Worker asm( "smlabb %[out], %[v], %[inRL], %[a] \n" 140*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 141*ec779b8eSAndroid Build Coastguard Worker : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a) 142*ec779b8eSAndroid Build Coastguard Worker : ); 143*ec779b8eSAndroid Build Coastguard Worker } else { 144*ec779b8eSAndroid Build Coastguard Worker asm( "smlabt %[out], %[v], %[inRL], %[a] \n" 145*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 146*ec779b8eSAndroid Build Coastguard Worker : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a) 147*ec779b8eSAndroid Build Coastguard Worker : ); 148*ec779b8eSAndroid Build Coastguard Worker } 149*ec779b8eSAndroid Build Coastguard Worker return out; 150*ec779b8eSAndroid Build Coastguard Worker #else 151*ec779b8eSAndroid Build Coastguard Worker int16_t s = left ? static_cast<int16_t>(inRL) : static_cast<int16_t>(inRL>>16); 152*ec779b8eSAndroid Build Coastguard Worker return a + v * s; 153*ec779b8eSAndroid Build Coastguard Worker #endif 154*ec779b8eSAndroid Build Coastguard Worker } 155*ec779b8eSAndroid Build Coastguard Worker 156*ec779b8eSAndroid Build Coastguard Worker static inline 157*ec779b8eSAndroid Build Coastguard Worker int32_t mulAddRL(int left, uint32_t inRL, int32_t v, int32_t a) 158*ec779b8eSAndroid Build Coastguard Worker { 159*ec779b8eSAndroid Build Coastguard Worker #if 0 // USE_INLINE_ASSEMBLY Seems to fail with Clang b/34110890 160*ec779b8eSAndroid Build Coastguard Worker int32_t out; 161*ec779b8eSAndroid Build Coastguard Worker if (left) { 162*ec779b8eSAndroid Build Coastguard Worker asm( "smlawb %[out], %[v], %[inRL], %[a] \n" 163*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 164*ec779b8eSAndroid Build Coastguard Worker : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a) 165*ec779b8eSAndroid Build Coastguard Worker : ); 166*ec779b8eSAndroid Build Coastguard Worker } else { 167*ec779b8eSAndroid Build Coastguard Worker asm( "smlawt %[out], %[v], %[inRL], %[a] \n" 168*ec779b8eSAndroid Build Coastguard Worker : [out]"=r"(out) 169*ec779b8eSAndroid Build Coastguard Worker : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a) 170*ec779b8eSAndroid Build Coastguard Worker : ); 171*ec779b8eSAndroid Build Coastguard Worker } 172*ec779b8eSAndroid Build Coastguard Worker return out; 173*ec779b8eSAndroid Build Coastguard Worker #else 174*ec779b8eSAndroid Build Coastguard Worker int16_t s = left ? static_cast<int16_t>(inRL) : static_cast<int16_t>(inRL>>16); 175*ec779b8eSAndroid Build Coastguard Worker return a + static_cast<int32_t>((static_cast<int64_t>(v) * s) >> 16); 176*ec779b8eSAndroid Build Coastguard Worker #endif 177*ec779b8eSAndroid Build Coastguard Worker } 178*ec779b8eSAndroid Build Coastguard Worker 179*ec779b8eSAndroid Build Coastguard Worker } // namespace android 180*ec779b8eSAndroid Build Coastguard Worker 181*ec779b8eSAndroid Build Coastguard Worker #endif /*ANDROID_AUDIO_RESAMPLER_FIR_OPS_H*/ 182