ArrayFire slower than Eigen in cpu mode ?

[William Tambellini]

Good morning gentlemen

In order to evaluate ArrayFire, I am converting some Eigen codes to ArrayFire :

Eigen (w and n are float32 matrix):
    EigenRowVector maxCoeff = w.colwise().maxCoeff();
    EigenMatrix normTerms = (w.rowwise() - maxCoeff).unaryExpr(Exp()).colwise().sum().unaryExpr(Log()) + maxCoeff;
    n.noalias() = w.rowwise() - normTerms.row(0);

ArrayFire:
      n = w - af::tile( af::log(af::sum( af::exp(w - af::tile(af::max(w), w.dims(0), 1)) )) + af::max(w), w.dims(0), 1);  // I m not fan of the af::tile() thing but did nt find better.

When run in CPU mode (not opengl), AF is about 3 to 4 times slower than Eigen.
Am I doing anything wrong ?
Cheers
WT.

[Umar Arshad]

The CPU backend of ArrayFire was implemented as a fallback for the CUDA/OpenCL

backends. With that said, we are evaluating ways of improving the performance of

the CPU backend.

I am also curious why you do not want to use the OpenCL backend to perform CPU

operations. We have been discussing how we can improve the CPU performance

but it looks like it will just be an exercise to repeat the OpenCL implementations with

extra complexity in the binary. The reason this is not a problem in libraries like Eigen

is because they are header only library whereas ArrayFire is an actual library. In order

to support all architectures, we will need implementations for AVX, SSE, and Non-vectorized

versions in the library and switch to it at runtime which requires additional logic for each

function. This is not a problem with the OpenCL backend because it compiles the

kernels at runtime so it is optimized for that architecture.

Umar

[William Tambellini]

Hi Umar
Thanks very much.
Of course, I m going to try the Intel OpenCL backend in order to compare.
Cheers
William T.