Layers Supported

Current MAESTRO supports the following layer operations in a mapping.

CONV2D (CONV)

MAESTRO supports CONV2D operations (CONV layers in most CNNs)

Depth-wise convolution (DWCONV)

MAESTRO supports depth-wise convolutions. To specify depth-wise separable convolutions, please use CONV2D for point-wise convolution part.

Transposed Convolution (TRCONV)

MAESTRO supports transposed convolution that doubles the resolution of a featuremap (2X upconv)

Fully-connected (FC)

FC layers are indirectly supported via CONV2D operation. To model FC, please follow the following rule regarding layer dimension sizes: Y = R = 1, X = S, c = number of input features, k = number of output features

Matrix Multiplication (GEMM)

For GEMM layer with [M,K] x [K, N] matrix-matrix multiplication:

Directly supported via GEMM operations by specifying M, N, K.

Or indirectly supported via CONV2D operation. To specify GEMM, please follow the following rule regarding layer dimension sizes: Assuming standard M, N, K convention of GEMMs, C = K (in GEMM), Y = M (in GEMM), K = N (in GEMM), X = R = S = 1

Note

For more information on the Mapping, please see Mapping Description.