search

stanford-ppl / spatial-lang

Spatial: "Specify Parameterized Accelerators Through Inordinately Abstract Language"
MIT License
100 stars 12 forks source link

Partial Mem Reduce #242

Closed mattfel1 closed 6 years ago

mattfel1 commented 6 years ago

Is there already an issue for this? It would be nice to be able to do something like

MemReduce( (0::16, 0::16), accum) (N by 1){i => 
...
}{_+_}

In cases where we try to save area by reusing hardware to do convolutions and the layer dimensions are different, we could use the same large memory but only reduce over a portion of it since we know this portion is all we will care to store into DRAM.

shadjis commented 6 years ago

This would be helpful for me too

mattfel1 commented 6 years ago

Implemented. See the BlockReduce1D and BlockReduce2D unit tests. We can figure out how to make it prettier later if we want.

For now, whatever range you accumulate into for the accumulator is the same range it reads from in the result tile:

image

If the Partial tile has different dimensions, the read addresses will wrap around in hw and you will probably get XXX data in scala. I think we mainly just need to accumulate starting at the origin in most cases so this should be fine.