Naive dot product performance

[ffevotte]

The naive dot product implementation currently uses "explicit" SIMD instructions:

AccurateArithmetic.jl/src/accumulators/dot.jl

Line 7 in 1fd8e58

function add!(acc::DotAcc, x, y)

This should prevent LLVM from transforming those into vfmadd instructions, which is a shame since those would probably be both more accurate and more efficient. We should

• either change those to non-explicit ("fuseable") SIMD instructions, relying on LLVM to turn those into vfmadds, or
• explicitly use an fma here.

I wonder if this is what explains the differences between OpenBLAS.ddot and AccurateArithmetic.dot_naive observed in the paper (fig. 4)