xref: /aosp_15_r20/external/pytorch/aten/src/ATen/native/cuda/UnaryGeometricCosKernel.cu (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #define TORCH_ASSERT_NO_OPERATORS
2 #include <ATen/AccumulateType.h>
3 #include <ATen/Dispatch.h>
4 #include <ATen/OpMathType.h>
5 #include <ATen/native/DispatchStub.h>
6 #include <ATen/native/TensorIterator.h>
7 #include <ATen/native/UnaryOps.h>
8 #include <ATen/native/cuda/JitLoops.cuh>
9 #include <ATen/native/cuda/Loops.cuh>
10 #include <ATen/native/cuda/Math.cuh>
11 #include <limits>
12 
13 namespace at::native {
14 
15 #if AT_USE_JITERATOR()
16 CONSTEXPR_EXCEPT_WIN_CUDA char cos_name[] = "cos_impl";
17 #endif // AT_USE_JITERATOR()
18 
cos_kernel_cuda(TensorIteratorBase & iter)19 void cos_kernel_cuda(TensorIteratorBase& iter) {
20   auto common_dtype = iter.common_dtype();
21   if (at::isComplexType(common_dtype)) {
22 #if AT_USE_JITERATOR()
23     static const auto cos_string = jiterator_stringify(
24         template <typename T> T cos_impl(T a) { return std::cos(a); });
25     AT_DISPATCH_COMPLEX_TYPES_AND(
26         kComplexHalf, common_dtype, "cos_name", [&]() {
27           jitted_gpu_kernel<
28               /*name=*/cos_name,
29               /*return_dtype=*/scalar_t,
30               /*common_dtype=*/scalar_t,
31               /*arity=*/1>(iter, cos_string);
32         });
33 #else
34     AT_DISPATCH_COMPLEX_TYPES_AND(
35         kComplexHalf, common_dtype, "cos_name", [&]() {
36           gpu_kernel(iter, [] GPU_LAMBDA(scalar_t a) -> scalar_t {
37             using opmath_t = at::opmath_type<scalar_t>;
38             return ::cos(static_cast<opmath_t>(a));
39           });
40         });
41 #endif
42   } else {
43     AT_DISPATCH_FLOATING_TYPES_AND2(
44         ScalarType::Half,
45         ScalarType::BFloat16,
46         common_dtype,
47         "cos_cuda",
48         [&]() {
49           gpu_kernel(
50               iter, [] GPU_LAMBDA(scalar_t a) -> scalar_t { return ::cos(a); });
51         });
52   }
53 }
54 
55 REGISTER_DISPATCH(cos_stub, &cos_kernel_cuda);
56 
57 } // namespace at::native
58