Equivalent precision considerations

The SPU instruction set does not fully implement the IEEE 754 single-precision floating-point standard (default rounding mode is round to zero , denormals are treated as zero, and there are no infinities or NaNs).

Therefore, floating-point results on an SPE may differ slightly from floating-point results using the PPE's PowerPC instruction set. In addition, all estimation intrinsics (for example, ceiling, floor, reciprocal estimate, reciprocal square root estimate, exponent estimate, and log estimate) do not have equivalent accuracy on the SPU and PPE PowerPC instruction sets.

However, the instructions in the PPE's Vector/SIMD Multimedia Extension have a graphics rounding mode that allows programs written with Vector/SIMD Multimedia Extension instructions to produce floating-point results that are equivalent in precision to those written in the SPU instruction set. In this Vector/SIMD Multimedia Extension mode, as in the SPU environment, the default rounding mode is round to zero, denormals are treated as zero, and there are no infinities or NaNs.

Details on the graphics rounding mode can be found in Cell Broadband Engine, Programming Handbook.