2.5.4. Input Feature Tensor In-Memory Format
Input features are stored in FP16 format. FP16 format has 1 sign bit, 10 mantissa bits, and 5 exponent bits. The input features are converted by the IP hardware to its native format using a round to nearest, ties to even (RNE) rounding rule.
Feature elements are packed into CVEC-sized chunks in the channel dimension from low to high. The final CVEC chunk, at a given (d,h,w), is padded with zeros. The CVEC chunks are stored in NCDHW format. The order is as follows: batch, channel, depth, height, width, and CVEC, where CVEC is the fastest changing index and batch the slowest.
The following figure shows a sample memory layout for a 1×3×1×2×2 input tensor to a CVEC=2 architecture: