Vector Class Discussion

inline __m128 vgather4(const float *base, const __m128i &indices)
{
__m128 res = _mm_load_ss(base+_mm_cvtsi128_si32(indices));
res = _mm_insert_ps(res,_mm_load_ss(base+_mm_extract_epi32(indices,1)),_MM_MK_INSERTPS_NDX(0,1,0));
res = _mm_insert_ps(res,_mm_load_ss(base+_mm_extract_epi32(indices,2)),_MM_MK_INSERTPS_NDX(0,2,0));
res = _mm_insert_ps(res,_mm_load_ss(base+_mm_extract_epi32(indices,3)),_MM_MK_INSERTPS_NDX(0,3,0));
return res;
}

inline __m256 vgather8(const float *base, const Vec8i &indices)
{
const __m128 low = vgather4(base,indices.get_low()),
high = vgather4(base,indices.get_high());
return _mm256_insertf128_ps(_mm256_castps128_ps256(low),high,1);
}

This looks like a lot of overhead (it compiles to 18 assembly instructions to basically load 8 values from memory so if one does only
one or two computations after loading a vector, there will be little performance benefit.
However there are many usage cases where data can be re-used many times for computations once loaded in a vector register so there is potential for significant speedup with the gather() instruction (for example I was able to get a speedup of a factor of three for a 3D grid tri-cubic interpolation code for fluid simulation advection using the above gather() instruction on top of <vectorclass.h>)

I wonder however if there may be a faster method for emulating gather() with pre-AVX2 means ?
(under AVX1 one could use an implementation basd on _mm256_blend_ps/_mm256_broadcast_ss but I am not sure if this would be faster though)

Usage example: simultaneously loading 8 pixels at 8 arbitrary coordinate pairs (ix,iy) from a 2D grid:

Vec8i ix,iy,grid_width;
float *p_grid_base;
...
ix = min(max(ix,0),grid_width); ... // calculate & clamp coords etc.
Vec8i idx = ix + iy * grid_width; // calc indices
Vec8f pixel_value = vgather8( p_grid_base, idx ); // load 8 pixels in parallel