diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32f_index_max_32u.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32f_index_max_32u.h index 8ea27b9c2..56666d531 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32f_index_max_32u.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32f_index_max_32u.h @@ -59,6 +59,140 @@ #include #include +#ifdef LV_HAVE_AVX +#include + +static inline void volk_gnsssdr_32f_index_max_32u_a_avx(uint32_t* target, const float* src0, uint32_t num_points) +{ + if(num_points > 0) + { + uint32_t number = 0; + const uint32_t quarterPoints = num_points / 8; + + float* inputPtr = (float*)src0; + + __m256 indexIncrementValues = _mm256_set1_ps(8); + __m256 currentIndexes = _mm256_set_ps(-1,-2,-3,-4,-5,-6,-7,-8); + + float max = src0[0]; + float index = 0; + __m256 maxValues = _mm256_set1_ps(max); + __m256 maxValuesIndex = _mm256_setzero_ps(); + __m256 compareResults; + __m256 currentValues; + + __VOLK_ATTR_ALIGNED(32) float maxValuesBuffer[8]; + __VOLK_ATTR_ALIGNED(32) float maxIndexesBuffer[8]; + + for(;number < quarterPoints; number++) + { + + currentValues = _mm256_load_ps(inputPtr); inputPtr += 8; + currentIndexes = _mm256_add_ps(currentIndexes, indexIncrementValues); + + //compareResults = _mm_cmpgt_ps(maxValues, currentValues); + compareResults = _mm256_cmp_ps(maxValues, currentValues, 0x1e); + + maxValuesIndex = _mm256_blendv_ps(currentIndexes, maxValuesIndex, compareResults); + maxValues = _mm256_blendv_ps(currentValues, maxValues, compareResults); + } + + // Calculate the largest value from the remaining 8 points + _mm256_store_ps(maxValuesBuffer, maxValues); + _mm256_store_ps(maxIndexesBuffer, maxValuesIndex); + + for(number = 0; number < 8; number++) + { + if(maxValuesBuffer[number] > max) + { + index = maxIndexesBuffer[number]; + max = maxValuesBuffer[number]; + } + } + + number = quarterPoints * 8; + for(;number < num_points; number++) + { + if(src0[number] > max) + { + index = number; + max = src0[number]; + } + } + target[0] = (uint32_t)index; + } +} + +#endif /*LV_HAVE_AVX*/ + + +#ifdef LV_HAVE_AVX +#include + +static inline void volk_gnsssdr_32f_index_max_32u_u_avx(uint32_t* target, const float* src0, uint32_t num_points) +{ + if(num_points > 0) + { + uint32_t number = 0; + const uint32_t quarterPoints = num_points / 8; + + float* inputPtr = (float*)src0; + + __m256 indexIncrementValues = _mm256_set1_ps(8); + __m256 currentIndexes = _mm256_set_ps(-1,-2,-3,-4,-5,-6,-7,-8); + + float max = src0[0]; + float index = 0; + __m256 maxValues = _mm256_set1_ps(max); + __m256 maxValuesIndex = _mm256_setzero_ps(); + __m256 compareResults; + __m256 currentValues; + + __VOLK_ATTR_ALIGNED(32) float maxValuesBuffer[8]; + __VOLK_ATTR_ALIGNED(32) float maxIndexesBuffer[8]; + + for(;number < quarterPoints; number++) + { + + currentValues = _mm256_loadu_ps(inputPtr); inputPtr += 8; + currentIndexes = _mm256_add_ps(currentIndexes, indexIncrementValues); + + //compareResults = _mm_cmpgt_ps(maxValues, currentValues); + compareResults = _mm256_cmp_ps(maxValues, currentValues, 0x1e); + + maxValuesIndex = _mm256_blendv_ps(currentIndexes, maxValuesIndex, compareResults); + maxValues = _mm256_blendv_ps(currentValues, maxValues, compareResults); + } + + // Calculate the largest value from the remaining 8 points + _mm256_store_ps(maxValuesBuffer, maxValues); + _mm256_store_ps(maxIndexesBuffer, maxValuesIndex); + + for(number = 0; number < 8; number++) + { + if(maxValuesBuffer[number] > max) + { + index = maxIndexesBuffer[number]; + max = maxValuesBuffer[number]; + } + } + + number = quarterPoints * 8; + for(;number < num_points; number++) + { + if(src0[number] > max) + { + index = number; + max = src0[number]; + } + } + target[0] = (uint32_t)index; + } +} + +#endif /*LV_HAVE_AVX*/ + + #ifdef LV_HAVE_SSE4_1 #include @@ -125,6 +259,72 @@ static inline void volk_gnsssdr_32f_index_max_32u_a_sse4_1(uint32_t* target, con #endif /*LV_HAVE_SSE4_1*/ +#ifdef LV_HAVE_SSE4_1 +#include + +static inline void volk_gnsssdr_32f_index_max_32u_u_sse4_1(uint32_t* target, const float* src0, uint32_t num_points) +{ + if(num_points > 0) + { + uint32_t number = 0; + const uint32_t quarterPoints = num_points / 4; + + float* inputPtr = (float*)src0; + + __m128 indexIncrementValues = _mm_set1_ps(4); + __m128 currentIndexes = _mm_set_ps(-1,-2,-3,-4); + + float max = src0[0]; + float index = 0; + __m128 maxValues = _mm_set1_ps(max); + __m128 maxValuesIndex = _mm_setzero_ps(); + __m128 compareResults; + __m128 currentValues; + + __VOLK_ATTR_ALIGNED(16) float maxValuesBuffer[4]; + __VOLK_ATTR_ALIGNED(16) float maxIndexesBuffer[4]; + + for(;number < quarterPoints; number++) + { + + currentValues = _mm_loadu_ps(inputPtr); inputPtr += 4; + currentIndexes = _mm_add_ps(currentIndexes, indexIncrementValues); + + compareResults = _mm_cmpgt_ps(maxValues, currentValues); + + maxValuesIndex = _mm_blendv_ps(currentIndexes, maxValuesIndex, compareResults); + maxValues = _mm_blendv_ps(currentValues, maxValues, compareResults); + } + + // Calculate the largest value from the remaining 4 points + _mm_store_ps(maxValuesBuffer, maxValues); + _mm_store_ps(maxIndexesBuffer, maxValuesIndex); + + for(number = 0; number < 4; number++) + { + if(maxValuesBuffer[number] > max) + { + index = maxIndexesBuffer[number]; + max = maxValuesBuffer[number]; + } + } + + number = quarterPoints * 4; + for(;number < num_points; number++) + { + if(src0[number] > max) + { + index = number; + max = src0[number]; + } + } + target[0] = (uint32_t)index; + } +} + +#endif /*LV_HAVE_SSE4_1*/ + + #ifdef LV_HAVE_SSE #include @@ -192,6 +392,73 @@ static inline void volk_gnsssdr_32f_index_max_32u_a_sse(uint32_t* target, const #endif /*LV_HAVE_SSE*/ +#ifdef LV_HAVE_SSE + +#include + +static inline void volk_gnsssdr_32f_index_max_32u_u_sse(uint32_t* target, const float* src0, uint32_t num_points) +{ + if(num_points > 0) + { + uint32_t number = 0; + const uint32_t quarterPoints = num_points / 4; + + float* inputPtr = (float*)src0; + + __m128 indexIncrementValues = _mm_set1_ps(4); + __m128 currentIndexes = _mm_set_ps(-1,-2,-3,-4); + + float max = src0[0]; + float index = 0; + __m128 maxValues = _mm_set1_ps(max); + __m128 maxValuesIndex = _mm_setzero_ps(); + __m128 compareResults; + __m128 currentValues; + + __VOLK_ATTR_ALIGNED(16) float maxValuesBuffer[4]; + __VOLK_ATTR_ALIGNED(16) float maxIndexesBuffer[4]; + + for(;number < quarterPoints; number++) + { + currentValues = _mm_loadu_ps(inputPtr); inputPtr += 4; + currentIndexes = _mm_add_ps(currentIndexes, indexIncrementValues); + + compareResults = _mm_cmpgt_ps(maxValues, currentValues); + + maxValuesIndex = _mm_or_ps(_mm_and_ps(compareResults, maxValuesIndex) , _mm_andnot_ps(compareResults, currentIndexes)); + + maxValues = _mm_or_ps(_mm_and_ps(compareResults, maxValues) , _mm_andnot_ps(compareResults, currentValues)); + } + + // Calculate the largest value from the remaining 4 points + _mm_store_ps(maxValuesBuffer, maxValues); + _mm_store_ps(maxIndexesBuffer, maxValuesIndex); + + for(number = 0; number < 4; number++) + { + if(maxValuesBuffer[number] > max) + { + index = maxIndexesBuffer[number]; + max = maxValuesBuffer[number]; + } + } + + number = quarterPoints * 4; + for(;number < num_points; number++) + { + if(src0[number] > max) + { + index = number; + max = src0[number]; + } + } + target[0] = (uint32_t)index; + } +} + +#endif /*LV_HAVE_SSE*/ + + #ifdef LV_HAVE_GENERIC static inline void volk_gnsssdr_32f_index_max_32u_generic(uint32_t* target, const float* src0, uint32_t num_points)