xref: /aosp_15_r20/external/pytorch/c10/util/Unroll.h (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #pragma once
2 #include <c10/macros/Macros.h>
3 #include <type_traits>
4 
5 // Utility to guarantee complete unrolling of a loop where the bounds are known
6 // at compile time. Various pragmas achieve similar effects, but are not as
7 // portable across compilers.
8 
9 // Example: c10::ForcedUnroll<4>{}(f); is equivalent to f(0); f(1); f(2); f(3);
10 
11 namespace c10 {
12 
13 template <int n>
14 struct ForcedUnroll {
15   template <typename Func, typename... Args>
operatorForcedUnroll16   C10_ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
17     ForcedUnroll<n - 1>{}(f, args...);
18     f(std::integral_constant<int, n - 1>{}, args...);
19   }
20 };
21 
22 template <>
23 struct ForcedUnroll<1> {
24   template <typename Func, typename... Args>
25   C10_ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
26     f(std::integral_constant<int, 0>{}, args...);
27   }
28 };
29 
30 } // namespace c10
31