Standardize math functions in SYCL kernels: std::/::f → sycl::, __f → sycl::native::

[Copilot]

SYCL kernels were inconsistently mixing sycl::, std::, ::, and sycl::native:: math functions. The rule: high-precision calls (expf, std::exp, ::expf) → sycl::exp; CUDA fast intrinsics (__expf) → sycl::native::exp.

Changes

• src/comm/XPUMathCompat.h: Replace ::expf/::tanhf/::exp/::tanh with sycl::exp/sycl::tanh

• 58 SYCL kernel files: Replace all std:: math functions with sycl:: equivalents in device code:

  • Transcendentals: exp, log, log1p, log2, log10, expm1, exp2, sqrt, pow
  • Trig: sin, cos, tan, sinh, cosh, tanh, asin, acos, atan, atan2
  • Misc: erf, erfc, lgamma, fabs, fmax, fmin, hypot, floor, ceil, round, trunc, fmod

• Complex type handling: Added specializations that preserve std:: for complex number math where SYCL has no equivalent (e.g., std::log1p(complex), complex sqrt). Affects SqrtFunctor, ExpFunctor in UnaryKernels.cpp; LogspaceFunctor in RangeFactoriesKernel.cpp; Rsqrt<c10::complex<T>> in ForeachUnaryKernels.cpp.

• sycl::native::: No existing __expf-style intrinsic calls found — no changes required for this category.

Kept as std:: (intentional):

• Host-side code (computed before sycl_kernel_submit, e.g., AdaptiveAveragePooling, AttentionKernels scale factor)
• std::modf / std::frexp — SYCL variants require multi_ptr, different API
• compat_pow in SharedReduceOps.h — dispatched with complex accumulator types

// Before
return x_acc * std::tanh(std::log1p(std::exp(x_acc)));

// After
return x_acc * sycl::tanh(sycl::log1p(sycl::exp(x_acc)));

// XPUMathCompat.h — Before
__MATH_FUNCTIONS_DECL__ float exp(float x) { return ::expf(x); }

// After
__MATH_FUNCTIONS_DECL__ float exp(float x) { return sycl::exp(x); }

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.