Reduction on two dimensional arrays

1

Hi,

I would like to know if the reduction kernel could be adapted to work on arrays whose ioshape is not three-dimensional (e.g. nupts, nvars, neles) but rather bidimensional (e.g. nvars, neles) and still perform a reduction over neles to get a nvars shape as a resulting array.
If so could you point out to the major changes that should be set for indexing the matrix appropriately?

Best

2

The overall logic should be simple to adapt (basically just removing the inner loop which does the accumulation over the first dimension). I would suggest starting with the OpenMP backend as its reduction code is the simplest.

Regards, Freddie.

3

Hi @fdw,

thanks for the hint, I changed the inner loop as you suggested to:

if (i < ncolb)
{
    ixdtype_t idx = SOA_IX(i, blockIdx.y, gridDim.y);
    r = data[idx];
    acc += r;
}

but when I test it with a sum of all elements I get this as results

    test = np.ones((5, 49055)).astype(np.float32)
    test= self._be.matrix(test, initval= test, tags={'align'})
    red = self._be.kernel('sum', test)
    self._be.run_kernels([red], wait=True)

with a result which is

[49054. 49054. 49054. 49054. 49055.]

Is there anything else that should be changed?

Best

4

As per my suggestion, start with the OpenMP backend first for testing.

Regards, Freddie.

5

(basically just removing the inner loop which does the accumulation over the first dimension)

In the openMP kernel is it

                #pragma omp simd
                for (ixdtype_t _xj = 0; _xj < SOA_SZ; _xj++)
                {
                 ...
                }

?

Best