tile_target.h File Reference

The complete target-specific contract of the tile algebra. More...

#include <optional>
#include "sdfg/data_flow/library_node.h"
#include "sdfg/structured_control_flow/structured_loop.h"
#include "sdfg/tiles/bank_conflicts.h"
#include "sdfg/tiles/tile.h"
#include "sdfg/types/type.h"

Go to the source code of this file.

Classes
class	sdfg::tiles::TileTarget

Namespaces
namespace	sdfg
namespace	sdfg::tiles

Detailed Description

The complete target-specific contract of the tile algebra.

Everything else in the tiles API is target-neutral: Layout, Tile, LocalityPlan, PackedBuffer, TiledCopy, and the Level / Space lattices. A backend (CUDA, ROCm, OpenMP, Tenstorrent, ...) makes its schedules and memory legible to that neutral core by implementing a handful of per-target facts:

1. how its loop schedules cooperate -> classify (a schedule -> AxisSchedule),
2. which memory tier backs each cooperation level -> space (a Level -> Space),
3. whether a schedule can drive a cooperative staging copy -> supports_cooperative_staging,
4. which concrete buffer realizes a memory tier -> storage_type (a Space -> StorageType),
5. its SIMD lane width for bank-conflict avoidance -> lane_width,
6. the banks of a banked memory tier (optional) -> bank_model.

(has_scratchpad is derived from space, not implemented.) These are the only places the tile algebra needs to know a target exists. A TileTarget replaces the hard-coded "CUDA_Offload" / "ROCM_Offload" / NV_Shared / gpu_warp_size checks that would otherwise scatter through AxisSchedule::classify and LocalStorage. New targets (including externally linked ones) plug in by implementing this interface and registering it; no edit to the tiles core is required.