Special-Function Approximation Arithmetic
Fast, reduced-precision hardware approximation of a transcendental function (reciprocal, reciprocal sqrt, sin, log2, exp2, ...).
Vendor-Neutral Definition
d ≈ f(a) for f in {1/x, 1/sqrt(x), sin, cos, log2, exp2, ...}; precision (ULP error bound) is hardware- and instruction-specific and is generally lower than a fully-rounded IEEE result.
⚠️ Cross-vendor mappings describe semantic relationships and are not guaranteed one-to-one compiler
translations.
equivalent with restrictions
Both provide fast, reduced-precision hardware approximations of transcendental functions.
| Property | NVIDIA PTX | AMDGPU |
|---|---|---|
| Instruction | rsqrt, sin, lg2 | v_rsq_f32, v_sin_f32 |
| ISA Layer | Virtual | Native |
| Data Types | f32 | f32 |
| Version / Target Introduced | PTX ISA 1.0, PTX ISA 1.0, PTX ISA 1.0 | gfx942/gfx1100, gfx942/gfx1100 |
Important Differences
- Approximation error bounds (max ULP) are hardware-generation-specific on both vendors and are not guaranteed to match; do not assume bit-identical results across vendors even when both are described as 'approx'.
Source Evidence
- Parallel Thread Execution ISA ↗ — NVIDIA Corporation
- User Guide for AMDGPU Backend ↗ — LLVM Project
Verification method: documentation