Folding Factor Constraints

This page documents the divisibility constraints for PE and SIMD folding parameters for each FINN hardware layer.

Overview

Folding parameters (PE and SIMD) control the degree of parallelism for each hardware layer. These parameters must satisfy specific divisibility constraints based on the layer’s dimensions.

Common parameters:

• PE: Processing Elements - controls output parallelism

• SIMD: Single Instruction Multiple Data - controls input parallelism

Constraint notation:

• % denotes the modulo operator (must divide evenly)

• MH, MW: Matrix height/width (for MVAU)

• k_h, k_w: Kernel height/width (for VVAU)

Constraint Table

Folding factor constraints

Layer	Parameters	Constraints
Addstreams (deprecated, use ElementwiseBinary)	PE	inp_channels % PE == 0
ChannelwiseOp (deprecated, use ElementwiseBinary)	PE	channels % PE == 0
ConvolutionInputGenerator	SIMD	inp_channels % SIMD == 0
DuplicateStreams	PE	channels % PE == 0
ElementwiseBinary	PE	last_dim % PE == 0
StreamingEltwise (deprecated, use ElementwiseBinary)	PE	inp_channels % PE == 0
FMPadding	SIMD	inp_channels % SIMD == 0
FMPadding_Pixel	SIMD	inp_channels % SIMD == 0
Globalaccpool	PE	channels % PE == 0
Labelselect	PE	num_labels % PE == 0
MatrixVectorActivation	PE & SIMD	MH % PE == 0 & MW % SIMD == 0
Pool	PE	inp_channels % PE == 0
Thresholding	PE	MH % PE == 0
Upsampler	SIMD	inp_channels % SIMD == 0
VectorVectorActivation	PE & SIMD	k_h * k_w % SIMD == 0 & channels % PE == 0

Critical Multi-Layer Constraints

Some layer combinations require matching parallelism to ensure functional correctness:

Warning

ConvolutionInputGenerator paired with VVAU or Pool:

The SIMD parameter of ConvolutionInputGenerator must match the PE parameter of the following layer (VVAU or Pool).

• ConvolutionInputGenerator SIMD == VVAU PE

• ConvolutionInputGenerator SIMD == Pool PE

Violating this constraint causes functional incorrectness, not just performance issues.

Usage

These constraints are enforced by the SetFolding transformation and must be satisfied when manually configuring folding parameters via the folding config JSON file.

See Also

• Network Preparation - Network preparation and folding configuration

• finn.transformation.fpgadataflow.set_folding.SetFolding - Automatic folding transformation