Lines Matching refs:packing

1 # The packing stage in gemmlowp
6 of computations described there: packing, kernel, unpacking.
8 This page goes into more details about the first stage: packing.
12 contract that the packing stage must honor.
15 [low-precision.md](low-precision.md) as the packing stage also has to compute
20 As explained in [design.md](design.md), the primary purpose of packing is to
39 This suggests the following approach to implementing packing: take the exact
125 ## The innermost loop of the packing stage, PackRun, and PackingRegisterBlock
132 that it expects; the packing code has to honor it. So there's an asymmetry here:
133 the kernel is the master, the packing is the slave. That's why the packing code
134 is templatized in the KernelSideFormat. At larger scales, the packing is
162 // one may provide optimized packing code paths.
164 // The packing of a block proceeds in two steps:
176 ## Other things that the packing stage has to do
178 Besides storing matrix entries in a suitable order, the packing stages also has
181 First, packing has to compute the vectors of sums of entries along the depth
186 the packing stage we have to requantize inputs accordingly. See
190 ## Specialized packing paths for specific formats on specific CPU architectures
200 in which case a given path is to be used in place of the generic packing code.
204 specialized packing paths used on that CPU architecture are consistent with it.