Emit form

emit_form picks how a product, a dot, a sum or a norm is written in the generated code. It is an argument of build_bundle.

bundle = hawk.build_bundle(kernel, emit_form="auto")   # the default

Value	What it writes
"auto"	each site by size: unrolled when small, a loop when large
"static"	every site unrolled, element by element
"rolled"	every site as a loop over the element index

Any other value raises HawkError.

The size rule

Under "auto" a site is unrolled when its fold length and its result element count are both at most 16, and written as a loop above that. The fold is the length that is summed over: the inner size of a matrix-vector or matrix-matrix product, or the length of the vector in a dot, sum or norm.

What does not change

The three forms give the same bits. They differ only in compile time and in how many values a lane keeps live. A loop keeps one element live where the unrolled text keeps a whole fold, so a long fold or a wide result does not spill.

"static" is the text hawk wrote before the option existed. A wide product can fail to compile in that form. "rolled" is the fast-compile mode: the compiler does no per-element work, and a fold or result of any size builds.

The spill report

For a CUDA target the sidecar’s ptxas rows carry two more fields beside the register and local-frame figures: working_set_B, the working set the kernel is expected to need, and spills. spills is true when registers were spilled or the local frame is more than 1.25 times the working set. A kernel flagged this way also warns when it is built; if it does, try emit_form="rolled".