Blaze is not generating vectorized code on Arm

[Adel Ahmadyan]

Hi,

I'm using Blaze for small matrix multiplication (Multiplying 6x3 X 3x6 matrices of floats). However, the performance is lower than expected. I suspect the reason is the generated code is not Neon-ized.

Here is an example:

blaze::DynamicMatrix<blaze::StaticMatrix<float, 6, 6>> dst; blaze::DynamicVector<blaze::StaticMatrix<float, 6,3>> a, b;

// initialization, resizing and some logic

for (...) { dst(i,j) += a[i] * trans(b[j]); }

• Target is 64-bit ARM (iPhone, ...)
• When I check for alignment, all matrices report to be aligned (dst.isAligned(), a.isAligned(), b.isAligned()).
• When I look at the assembly code, that line calls addAssign function, and inside it, there are unrolled ARM code, but no neon.
• blaze is linked against accelerate, which includes blas and lapack.

Any help or suggestion is appreciated. Thanks Adel

[Klaus Iglberger]

Hi Adel!

Thanks for creating this issue. This is not a bug, but a missing feature: Blaze currently does not support vectorisation on the ARM architecture. Issue #49 is devoted to add support for ARM platforms.

In order to make use of Accelerate for small matrices, you can manually reduce the BLAS thresholds in the <blaze/config/Thresholds.h> configuration file or define the according symbol before the <blaze/Blaze.h> include. The following code snippet demonstrates this for the threshold for the row-major dense matrix multiplication:

#define BLAZE_DMATDMATMULT_THRESHOLD 0UL
#include <blaze/Blaze.h>

Please set all necessary thresholds to 0 in order to unconditionally use BLAS kernels.

Best regards,

Klaus!

[Klaus Iglberger]

Duplicate of #49.