Module fuse 

Fusing chains of elementwise verbs into one blockwise pass.

A chain like +/ w * x runs one pass per verb: the product is written to memory in full and read back to be reduced. This pass finds maximal subtrees of elementwise primitives at compile time and replaces them with Expr::Fused, which evaluates the whole chain a block at a time — the block stays in cache, so the arrays at the leaves are read once and the result is written once.

The kernel is a postfix program over a small stack of block buffers. It covers only what it can compute exactly as the unfused pipeline would; everything else — a shape that needs broadcasting, a dtype the chain would narrow, an integer overflow — declines at run time and the original subtree, kept inside the node, evaluates instead. Fusion therefore cannot change a result or an error message.

A chain does not have to be written as one sentence. d =. {x} - m followed by +/ d * d names a value that nothing needs as an array, and the pass moves such a value into the sentences that read it — see inline_once for the rules that keep that sound.

Structs

FusedKernel

A fused elementwise chain and what is made of its values.

Summary

What a compiled kernel is made of.

Enums

Decline

Why a kernel handed its work back to the chain it came from.

Instr

One step of a kernel: postfix, so operands are already on the stack.

Yield

What one evaluation of a kernel produces.

Constants

BLOCK

Elements a block buffer holds.

Functions

decline_reason

Why this kernel would decline these inputs, or None if it would run.

fallback_count

Number of fallbacks since the process started.

inlined_names

The names the pass took out of the program: values it moved into the kernels that read them, so no sentence computes them as arrays any more.

is_fused

Does any sentence of this program run a fused kernel?

is_inlined

Did the pass move a named value into the sentences that read it?

pass

Optimise a compiled program’s sentences: move the values that are only named for the reader into the sentences that read them, then fuse every chain that is left.

summary

Describe a compiled kernel: what it computes, and with what.

unfused

The program as the plain evaluator would run it: the sentences it was compiled from, with every fused node replaced by the subtree it came from. The two must compute the same thing; tests hold them to it.