numba_vulkan.probes

Behaviour of a device that its reported features do not reveal.

Two kinds of float behaviour make libclc and kernels go wrong on some drivers, and both are found out by running small kernels on each device:

• Vulkan lets a driver compute Fma as a multiplication followed by an addition, rounding twice. libclc assumes a fused operation; without it, its sin, cos and tan lose all accuracy beyond small arguments. llvmpipe does not fuse, in either precision. Kernels for such a device are compiled with a software version (numba_vulkan.legalize.emulate_fma).

• llvmpipe (Mesa 26.1) gets Trunc and RoundEven of double values wrong in its vectorised code (values that differ between invocations); the scalar path is correct. Kernels for such a device compute both from Floor (numba_vulkan.legalize.emulate_rounding64).

The probe therefore gives each invocation its own operands. Its results depend only on the driver, so they are kept in probes.json in the cache directory (~/.cache/numba-vulkan), per device and driver version, unless NUMBA_VULKAN_CACHE=0.

The environment variables NUMBA_VULKAN_SOFT_FMA and NUMBA_VULKAN_SOFT_ROUNDING override the probe: 1 applies the workaround on every device, 0 on none.

Attributes

FMA_ENV_VAR
ROUNDING_ENV_VAR

Functions

workarounds(device)

Which float64 workarounds kernels for a device need.

Module Contents

numba_vulkan.probes.FMA_ENV_VAR = 'NUMBA_VULKAN_SOFT_FMA'

numba_vulkan.probes.ROUNDING_ENV_VAR = 'NUMBA_VULKAN_SOFT_ROUNDING'

numba_vulkan.probes.workarounds(device)

Which float64 workarounds kernels for a device need.

Parameters:

device (numba_vulkan.runtime.Device) – The device, which must not yet use any of them.

Returns:

soft_fma and soft_rounding, as for numba_vulkan.narrowing.Mode.

Return type:

dict