Numerical behavior

A mathematical expression does not fully specify floating-point behavior. Read each operation's ordered algorithm: input conversion, intermediate precision, rounding boundaries, offsets and saturation can all affect the result.

Result semantics versus instructions

Quant specifies a nearest-even binary32 product followed by selected integer rounding, zero-point addition and INT8 saturation. An eligible implementation may produce that result through bit manipulation and FP16 instructions instead. It need not allocate an FP32 tensor. See Quant implementation selection.

Dequant specifies (float(input) - float(zero_point)) * scale with staged FP32 semantics. Replacing this with a differently rounded expression or an arbitrary fused multiply-add is not automatically equivalent.

Special values

Do not assume one universal NaN policy. Cast preserves its exact bit-rounding rule; Quant maps NaN to its zero point; Argmax selects the first NaN. Signed-zero and payload behavior are operation-specific. The operation pages also specify infinity, subnormal and overflow behavior.

Floating-point environment

Operations that require a particular rounding/flush/trap environment check the calling thread at execution, before output writes. Preparing a plan on one thread does not establish another thread's environment. Kernels do not silently repair caller FP state. Explicit Quant integer-rounding parameters do not change the required rounding of its intermediate binary32 product.

What an implementation explanation proves

A fused path can avoid a tensor-sized intermediate. A tiled path can avoid reading padding. These are structural properties, not universal latency guarantees. Measured latency depends on workload, compiler, device and memory state; use the qualification workflow before making comparisons.