CPU target detection

Status: current · line numbers checked 2026-10-08.

Configure decides whether CPU device code is compiled (BATCHLAS_HAS_CPU_TARGET) and drops what depends on it. Logic: cmake/BatchLASDetectSYCL.cmake. Options: cmake/BatchLASOptions.cmake.

A GPU-only DPC++ may have no CPU backend. Without a CPU entry in -fsycl-targets, code that submits kernels to a CPU device (the NETLIB SYCL helpers, Queue("cpu") work that is not host LAPACK) fails to compile or finds no kernel at run time.

Two stages

Stage 1: detect_sycl_cpu_target() (cmake/BatchLASDetectSYCL.cmake:490) may append a CPU target to BATCHLAS_SYCL_TARGETS:

BATCHLAS_CPU_TARGET	Effect
none	No CPU target; OFF.
native_cpu or spir64_x86_64	Appended; ON.
auto (default, cmake/BatchLASOptions.cmake:165)	sycl-ls lists [opencl:cpu] or [host:cpu]: append spir64_x86_64. Lists [native_cpu:cpu]: append native_cpu. Otherwise, or if sycl-ls is missing or fails: OFF.
Anything else	Configure WARNING, then treated as auto.

Stage 2: the final gate (cmake/BatchLASDetectSYCL.cmake:643-666) runs after -- Using SYCL targets: ... is printed. It decides from the compile flags, not from device detection.

1. An explicit native_cpu or spir64_x86_64 forces ON.
2. If no entry of BATCHLAS_SYCL_TARGETS matches cpu, spir64 or native_cpu, the result is OFF, whatever sycl-ls said. Configure then prints:

   -- CPU device detected by sycl-ls, but no CPU target in fsycl-targets
   -- CPU kernels will not be compiled - disabling CPU-dependent tests/benchmarks

3. BATCHLAS_CPU_TARGET=none forces OFF (and drops the CPU target itself).

The result is exported as BATCHLAS_HAS_CPU_TARGET (cmake/backend_config.h.in, #cmakedefine01) and to Python as compiled_features()["has_cpu_target"].

What BATCHLAS_HAS_CPU_TARGET=OFF turns off

• minibench_cli_tests is not built. It is the only member of CPU_SYCL_DEPENDENT_TESTS (tests/CMakeLists.txt:137-139).
• Every NETLIB instantiation of every typed suite is compiled out. backend_types, backend_types_filtered and backend_types_complex (tests/test_utils.hh) add the NETLIB block only when BATCHLAS_HAS_HOST_BACKEND && BATCHLAS_HAS_CPU_TARGET. Any NETLIB configuration still reached is skipped in BatchLASTest::SetUp().

Note: Other test binaries still build (for example ormqr_cta_tests, syev_cta_tests, sytrd_cta_tests), because their GPU kernels compile and host LAPACK needs no SYCL kernel. Their NETLIB cases are absent. Benchmarks have no base and CPU-dependent split.

Traps

• Typed-test indices shift. Without the NETLIB block, SteqrTest/0 is float/CUDA, not float/NETLIB. tests/known-failures.txt therefore pins each entry with type_param=, and applies those entries only to the dev-tests preset.
• dev-gpu and dev-gpu-tests set BATCHLAS_CPU_TARGET=none. This saves about 30% per compile and device link. Their coverage is reduced, and ctest still reports green.
• native_cpu has sub-group size 1. A CPU target does not make CTA kernels runnable. See Running the BatchLAS tests.
• Stage 2 matches by substring. Any target name containing cpu counts.