Module compute 

Hardware-agnostic compute target abstraction.

Defines the ComputeTarget trait: the contract that any GPU compilation and runtime target must satisfy (NVIDIA/PTX, AMD/HSACO, Apple/Metal, etc.).

Atlas is designed so that:

• Build time: kernel source files are compiled by a target-specific compiler into a target-specific binary format (PTX, SPIR-V, metallib).
• Runtime: the binary modules are loaded via GpuBackend::kernel() and executed via GpuBackend::launch().

This module covers the build-time contract. The runtime contract is defined by GpuBackend in spark-runtime.

Extending to new hardware

1. Create a HARDWARE.toml in kernels/<hw>/ with vendor = "<vendor>".
2. Implement ComputeTarget for the new vendor.
3. Write kernel source files in the vendor’s language (.cu, .metal, .cl).
4. Implement GpuBackend in spark-runtime for the vendor’s runtime API.

Structs

NvidiaTarget

NVIDIA CUDA compilation target: .cu → PTX via nvcc.

Enums

Vendor

Vendor identifier parsed from HARDWARE.toml.

Traits

ComputeTarget

Build-time compilation target contract.

Functions

target_for_vendor

Resolve the appropriate ComputeTarget from a HARDWARE.toml vendor field.