CUDA Fast Math

As noted in Fastmath, 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 Fastmath behavior with two differences.

• First, the fastmath argument to the @jit decorator 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.

  • math.cos(): Implemented using __nv_fast_cosf.

  • math.sin(): Implemented using __nv_fast_sinf.

  • math.tan(): Implemented using __nv_fast_tanf.

  • math.exp(): Implemented using __nv_fast_expf.

  • math.log2(): Implemented using __nv_fast_log2f.

  • math.log10(): Implemented using __nv_fast_log10f.

  • math.log(): Implemented using __nv_fast_logf.

  • math.pow(): Implemented using __nv_fast_powf.