x86 FP emulation: Flush denormals to zero for vcvtneps2bf16
Floating-point emulation for vcvtneps2bf16 now flushes denormal inputs to zero, aligning behavior and improving predictability.
The emulation code for the vcvtneps2bf16 instruction, used in truncsfbf-5.c and truncsfbf-6.c, has been updated to flush denormal floating-point inputs to zero. This change ensures consistent behavior by treating small, subnormal numbers as zero before conversion, and includes new test cases to verify this behavior.
In Details
The vcvtneps2bf16 instruction emulation in truncsfbf-5.c and truncsfbf-6.c has been modified. The CALC macro now explicitly flushes subnormal inputs to zero, aligning with certain floating-point behavior specifications and improving the predictability of conversions to BFloat16. Tests are added to cover denormal inputs.
- vcvtneps2bf16
- An x86 AVX512 instruction that converts packed single-precision floating-point numbers to BFloat16 format.
- denormal
- A floating-point number that is too small to be represented with the full precision of the format, resulting in a loss of significance. Flushing denormals to zero is a common behavior for simplifying calculations or avoiding performance penalties.
- BFloat16
- A 16-bit floating-point format with a similar exponent range to IEEE 754 single-precision (32-bit) but reduced precision, often used in machine learning.