Edit Question - KnowledgeHub

Title Body

The Intel Advanced Vector Extensions (AVX) offers no dot product in the 256-bit version (YMM register) for double precision floating point variables . The "Why?" question have been very briefly treated in another forum ( here ) and on Stack Overflow ( here ). But the question I am facing is how to replace this missing instruction with other AVX instructions in an efficient way? The dot product in 256-bit version exists for single precision floating point variables ( reference here ): __m256 _mm256_dp_ps(__m256 m1, __m256 m2, const int mask); The idea is to find an efficient equivalent for this missing instruction: __m256d _mm256_dp_pd(__m256d m1, __m256d m2, const int mask); To be more specific, the code I would like to transform from __m128 (four floats) to __m256d (4 doubles) use the following instructions: __m128 val0 = ...; // Four float values __m128 val1 = ...; // __m128 val2 = ...; // __m128 val3 = ...; // __m128 val4 = ...; // __m128 res = _mm_or_ps( _mm_dp_ps(val1, val0, 0xF1), _mm_or_ps( _mm_dp_ps(val2, val0, 0xF2), _mm_or_ps( _mm_dp_ps(val3, val0, 0xF4), _mm_dp_ps(val4, val0, 0xF8) ))); The result of this code is a _m128 vector of four floats containing the results of the dot products between val1 and val0 , val2 and val0 , val3 and val0 , val4 and val0<

Tags (comma-separated) Cancel