如何在不影响性能的情况下抽象SIMD代码以处理不同的数据类型

Question

I have been writing code for performing matrix operations. Initially it was just for x86 and now am porting it to different architectures. Also, I want it to support different data types other than float.

Consider the following code for adding to float arrays

void add(float *a, float *b, float *dst, int len)
{
        int k = 0;
        for(; k + 8 <  len; k += 8,a +=  8, b += 8, dst+= 8){
            __m256 x = _mm256_load_ps(a);
            __m256 y = _mm256_load_ps(b);
            __m256 z = _mm256_add_ps(x, y);
            _mm256_store_ps(dst, z);
        }
}

Here is what I have thought of to improve the code to support several platforms and data types.

1. For different the data types, I was going to change the function to template function

2. For simd instructions I thought of having macros that rename all architecture specific intrinsic functions to generic simd instructions such as SIMD_ADD . Problem is that different data-types require different intrinsic functions and the return type of the intrinsic is dependent on the data type too.

3. Also if I were to write a subtract function, I would end up copying most of the code just to replace SIMD_ADD macro for SIMD_SUB macro. Is their a neat way such that the I dont have to repeat the same code for all element wise operations such multiplication, division and subtraction ?

How would one tackle points 2 & 3 without abstracting to the extent of affecting the code's performance ?

Answer 1

I ended up having template classes for simd instructions with a specialization for each datatype. Unfortunately the compiler will not inline it automatically therefore you must use the compiler specific attributes to force it to inline