Auto folding for Reshape Op possibly broken

[fpjentzsch]

Reshape is a "pe_op" and doesn't have a "NumChannels" attribute. This means that during auto folding in SetFolding() its maximum PE will be determined by the last dimension of its normal output shape.

However, a valid folding requires PE to divide the last axis of both, the input and output shape.

Example (kws model) where I would expect this to fail:

[Layer with PE=1] -> 1x10x49x1 -> Reshape -> 1x490 -> DWC -> [Layer with SIMD=10]
Auto folding could set PE of Reshape > 1, which would be illegal.

Such a simple (no-op) operator as Reshape should probably always infer its stream width (PE) from the surrounding nodes, but it is unclear how this mechanism is best implemented.

[iksnagreb]

Hm, I see two options right now: Either we restrict valid PEs to common divisors of both, the input and output side, or we add a second folding parameter SIMD(?) and let Reshape actually act as a proper DWC (+ Shape adaptation). Both options would probably require to introduce a special case handling of Reshape within the auto-folding.

For inferring folding setting from surrounding nodes, we don't really have a mechanism right now?

[fpjentzsch]

Or maybe we can let the DWC take over the role of Reshape completely, as mentioned here: #163 (comment)

The DWC essentially infers folding/stream width from surrounding nodes, as it is always automatically inserted after folding as been applied.