Workspace and memory

Overview

The per-Queue workspace arena, leases and BumpAllocator sizing.

Functions

workspace_bytes()

template<typename Fn>

size_t batchlas::workspace_bytes ( Fn && layout )

inline

Bytes a workspace layout needs, obtained by replaying it against a sizing pool.

An algorithm describes its workspace once, in a *_layout function, and both its *_buffer_size entry point and its implementation go through it, so they cannot drift:

size_t foo_buffer_size(Queue& ctx, ...) {
    return workspace_bytes([&](BumpAllocator& p) { return foo_layout<B, T>(ctx, p, ...); });
}

Parameters

layout

callable taking a BumpAllocator&; its result is discarded

Returns

BumpAllocator::required_bytes() of the sizing pool after layout ran

Precondition

layout is pure with respect to the workspace: it may read the caller's views and build views over what it allocates, but never reads or writes workspace memory, never launches a kernel, and asks nested size queries about the caller's views only.

See also

Workspace

Definition at line 192 of file mempool.hh.

Files
file	include/batchlas/util/mempool.hh
	BumpAllocator, the linear sub-allocator over a workspace span, and workspace_bytes().
file	include/batchlas/util/workspace.hh
	WorkspaceLease: a scoped borrow of scratch memory from a Queue's workspace arena.

Classes
struct	batchlas::BumpAllocator
	Linear (bump) sub-allocator over a caller-supplied workspace buffer. More...
class	batchlas::WorkspaceLease
	A borrow of scratch memory from a Queue's workspace arena, returned to the arena when the handle goes out of scope. More...