diff options
| -rw-r--r-- | volk/apps/volk_profile.cc | 3 | ||||
| -rw-r--r-- | volk/include/volk/volk_32f_s32f_convert_32i_a16.h | 36 |
2 files changed, 38 insertions, 1 deletions
diff --git a/volk/apps/volk_profile.cc b/volk/apps/volk_profile.cc index fd9507207..c091a6289 100644 --- a/volk/apps/volk_profile.cc +++ b/volk/apps/volk_profile.cc @@ -13,7 +13,7 @@ extern "C" { int main(int argc, char *argv[]) { std::vector<std::string> results; - + /* //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4_a16, 1e-4, 2046, 10000, &results); //VOLK_PROFILE(volk_16i_branch_4_state_8_a16, 1e-4, 2046, 10000, &results); VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f_a16, 1e-5, 32768.0, 204600, 10000, &results); @@ -52,6 +52,7 @@ int main(int argc, char *argv[]) { VOLK_PROFILE(volk_32fc_x2_multiply_32fc_a16, 1e-4, 0, 204600, 1000, &results); VOLK_PROFILE(volk_32f_s32f_convert_16i_a16, 1, 32768, 204600, 10000, &results); VOLK_PROFILE(volk_32f_s32f_convert_16i_u, 1, 32768, 204600, 10000, &results); + */ VOLK_PROFILE(volk_32f_s32f_convert_32i_a16, 1, 2<<31, 204600, 10000, &results); VOLK_PROFILE(volk_32f_s32f_convert_32i_u, 1, 2<<31, 204600, 10000, &results); VOLK_PROFILE(volk_32f_convert_64f_a16, 1e-4, 0, 204600, 10000, &results); diff --git a/volk/include/volk/volk_32f_s32f_convert_32i_a16.h b/volk/include/volk/volk_32f_s32f_convert_32i_a16.h index 2927d616c..3f5044313 100644 --- a/volk/include/volk/volk_32f_s32f_convert_32i_a16.h +++ b/volk/include/volk/volk_32f_s32f_convert_32i_a16.h @@ -5,6 +5,42 @@ #include <inttypes.h> #include <stdio.h> +#ifdef LV_HAVE_AVX +#include <immintrin.h> + /*! + \brief Multiplies each point in the input buffer by the scalar value, then converts the result into a 32 bit integer value + \param inputVector The floating point input data buffer + \param outputVector The 32 bit output data buffer + \param scalar The value multiplied against each point in the input buffer + \param num_points The number of data values to be converted + */ +static inline void volk_32f_s32f_convert_32i_a16_avx(int32_t* outputVector, const float* inputVector, const float scalar, unsigned int num_points){ + unsigned int number = 0; + + const unsigned int eighthPoints = num_points / 8; + + const float* inputVectorPtr = (const float*)inputVector; + int32_t* outputVectorPtr = outputVector; + __m256 vScalar = _mm256_set1_ps(scalar); + __m256 inputVal1; + __m256i intInputVal1; + + for(;number < eighthPoints; number++){ + inputVal1 = _mm256_load_ps(inputVectorPtr); inputVectorPtr += 8; + + intInputVal1 = _mm256_cvtps_epi32(_mm256_mul_ps(inputVal1, vScalar)); + + _mm256_store_si256((__m256i*)outputVectorPtr, intInputVal1); + outputVectorPtr += 8; + } + + number = eighthPoints * 8; + for(; number < num_points; number++){ + outputVector[number] = (int32_t)(inputVector[number] * scalar); + } +} +#endif /* LV_HAVE_AVX */ + #ifdef LV_HAVE_SSE2 #include <emmintrin.h> /*! |