nvrtc_warp_fusion.h File Reference

Runtime NVRTC codegen for the warp-register fusion harness (cuSZp family). More...

#include "backend/types.h"
#include <cstddef>
#include <cstdint>
#include <string>
#include <vector>

Go to the source code of this file.

Classes
struct	fz::fused::WarpFusionSpec

Namespaces
namespace	fz

Functions
std::string	fz::fused::generateWarpFusionSource (const WarpFusionSpec &spec)
	The CUDA source the codegen emits for spec (exposed for tests/inspection).
size_t	fz::fused::launchNvrtcWarpFused (const WarpFusionSpec &spec, const float *d_in, size_t n_ab, const uint8_t *pred_params, size_t params_bytes, uint8_t *d_out, MemoryPool *pool, fz::stream_t stream)

Detailed Description

Runtime NVRTC codegen for the warp-register fusion harness (cuSZp family).

The compile-time path (fused_block.cu) instantiates the rate/pack kernels for a fixed <ElemsPerLane, Predictor> chosen in C++ and dispatches on the predictor name. This path does the same at RUNTIME: a WarpFusionSpec (predictor policy type name + ElemsPerLane) is turned into source that wraps fused_rate_body/fused_pack_body, compiled+cached via the shared NVRTC JIT, and launched with the host-side CUB exclusive-scan of per-block costs in between. Adding a predictor becomes writing its policy in warp_fusion.cuh and having its stage declare the op — no new template instantiation, no per-shape launcher, no string dispatch in the registry.

Function Documentation

launchNvrtcWarpFused()

size_t fz::fused::launchNvrtcWarpFused	(	const WarpFusionSpec &	spec,
		const float *	d_in,
		size_t	n_ab,
		const uint8_t *	pred_params,
		size_t	params_bytes,
		uint8_t *	d_out,
		MemoryPool *	pool,
		fz::stream_t	stream
	)

Generic warp-register fused compress — the entry the registry runner uses with no per-predictor shape. NVRTC-composes spec (predictor + ElemsPerLane), uploads the predictor params blob (pred_params/params_bytes; a leading float inv2eb at offset 0, already resolved by the caller), runs the rate kernel, the CUB exclusive-scan of per-block costs, and the pack kernel, then reads back the archive length. n_ab is the padded block-covering element count; block_size = 32 * elems_per_lane. Byte-identical to the pre-instantiated launchers. Returns the archive length, or 0 if n_ab == 0.