diff options
| author | Josh Blum <josh@joshknows.com> | 2010-11-23 13:50:37 -0800 |
|---|---|---|
| committer | Josh Blum <josh@joshknows.com> | 2010-11-23 13:50:37 -0800 |
| commit | 30ce5acedd3e0dc6fc97d7597781a0a4828812f2 (patch) | |
| tree | 28d010c0938121f587e1820849c34ee900e7f74b /host/lib/transport | |
| parent | 8ce75a3ca7a51f4bdee87d78a610a0f2519473ae (diff) | |
| parent | 13ae4786e091d5581baf31c9967dca822ef15e39 (diff) | |
| download | uhd-30ce5acedd3e0dc6fc97d7597781a0a4828812f2.tar.gz uhd-30ce5acedd3e0dc6fc97d7597781a0a4828812f2.tar.bz2 uhd-30ce5acedd3e0dc6fc97d7597781a0a4828812f2.zip | |
Merge branch 'usrp_e100' into next
Conflicts:
images/Makefile
Diffstat (limited to 'host/lib/transport')
| -rw-r--r-- | host/lib/transport/CMakeLists.txt | 6 | ||||
| -rw-r--r-- | host/lib/transport/convert_types_impl.hpp | 48 |
2 files changed, 54 insertions, 0 deletions
diff --git a/host/lib/transport/CMakeLists.txt b/host/lib/transport/CMakeLists.txt index b9ec7a6ad..b95d46381 100644 --- a/host/lib/transport/CMakeLists.txt +++ b/host/lib/transport/CMakeLists.txt @@ -58,6 +58,12 @@ IF(HAVE_EMMINTRIN_H) ADD_DEFINITIONS(-DHAVE_EMMINTRIN_H) ENDIF(HAVE_EMMINTRIN_H) +INCLUDE(CheckIncludeFileCXX) +CHECK_INCLUDE_FILE_CXX(arm_neon.h HAVE_ARM_NEON_H) + +IF(HAVE_ARM_NEON_H) + ADD_DEFINITIONS(-DHAVE_ARM_NEON_H) +ENDIF(HAVE_ARM_NEON_H) ######################################################################## # Setup defines for interface address discovery ######################################################################## diff --git a/host/lib/transport/convert_types_impl.hpp b/host/lib/transport/convert_types_impl.hpp index 90618dec6..48ff99725 100644 --- a/host/lib/transport/convert_types_impl.hpp +++ b/host/lib/transport/convert_types_impl.hpp @@ -32,6 +32,14 @@ #include <emmintrin.h> #endif +#ifdef HAVE_ARM_NEON_H + #define USE_ARM_NEON_H +#endif + +#if defined(USE_ARM_NEON_H) + #include <arm_neon.h> +#endif + /*********************************************************************** * Typedefs **********************************************************************/ @@ -135,6 +143,26 @@ static UHD_INLINE void fc32_to_item32_nswap( } } +#elif defined(USE_ARM_NEON_H) +static UHD_INLINE void fc32_to_item32_nswap( + const fc32_t *input, item32_t *output, size_t nsamps) +{ + size_t i; + + float32x4_t Q0 = vdupq_n_f32(shorts_per_float); + for (i=0; i < (nsamps & ~0x03); i+=2) { + float32x4_t Q1 = vld1q_f32(reinterpret_cast<const float *>(&input[i])); + float32x4_t Q2 = vmulq_f32(Q1, Q0); + int32x4_t Q3 = vcvtq_s32_f32(Q2); + int16x4_t D8 = vmovn_s32(Q3); + int16x4_t D9 = vrev32_s16(D8); + vst1_s16((reinterpret_cast<int16_t *>(&output[i])), D9); + } + + for (; i < nsamps; i++) + output[i] = fc32_to_item32(input[i]); +} + #else static UHD_INLINE void fc32_to_item32_nswap( const fc32_t *input, item32_t *output, size_t nsamps @@ -238,6 +266,26 @@ static UHD_INLINE void item32_to_fc32_nswap( } } +#elif defined(USE_ARM_NEON_H) +static UHD_INLINE void item32_to_fc32_nswap( + const item32_t *input, fc32_t *output, size_t nsamps) +{ + size_t i; + + float32x4_t Q1 = vdupq_n_f32(floats_per_short); + for (i=0; i < (nsamps & ~0x03); i+=2) { + int16x4_t D0 = vld1_s16(reinterpret_cast<const int16_t *>(&input[i])); + int16x4_t D1 = vrev32_s16(D0); + int32x4_t Q2 = vmovl_s16(D1); + float32x4_t Q3 = vcvtq_f32_s32(Q2); + float32x4_t Q4 = vmulq_f32(Q3, Q1); + vst1q_f32((reinterpret_cast<float *>(&output[i])), Q4); + } + + for (; i < nsamps; i++) + output[i] = item32_to_fc32(input[i]); +} + #else static UHD_INLINE void item32_to_fc32_nswap( const item32_t *input, fc32_t *output, size_t nsamps |