fastmath.rst

CUDA Fast Math

As noted in :ref:`fast-math`, for certain classes of applications that utilize floating point, strict IEEE-754 conformance is not required. For this subset of applications, performance speedups may be possible.

The CUDA target implements :ref:`fast-math` behavior with two differences.

• First, the fastmath argument to the :func:`@jit decorator <numba.cuda.jit>` is limited to the values True and False. When True, the following optimizations are enabled:

  • Flushing of denormals to zero.
  • Use of a fast approximation to the square root function.
  • Use of a fast approximation to the division operation.
  • Contraction of multiply and add operations into single fused multiply-add operations.

  See the documentation for nvvmCompileProgram for more details of these optimizations.

• Secondly, calls to a subset of math module functions on float32 operands will be implemented using fast approximate implementations from the libdevice library.

  • :func:`math.cos`: Implemented using __nv_fast_cosf.
  • :func:`math.sin`: Implemented using __nv_fast_sinf.
  • :func:`math.tan`: Implemented using __nv_fast_tanf.
  • :func:`math.exp`: Implemented using __nv_fast_expf.
  • :func:`math.log2`: Implemented using __nv_fast_log2f.
  • :func:`math.log10`: Implemented using __nv_fast_log10f.
  • :func:`math.log`: Implemented using __nv_fast_logf.
  • :func:`math.pow`: Implemented using __nv_fast_powf.