xref: /aosp_15_r20/external/pytorch/aten/src/ATen/native/cuda/ActivationLeakyReluKernel.cu (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #define TORCH_ASSERT_NO_OPERATORS
2 #define _USE_MATH_DEFINES
3 
4 #include <ATen/native/Activation.h>
5 
6 #include <cmath>
7 
8 #include <thrust/tuple.h>
9 
10 #include <ATen/AccumulateType.h>
11 #include <ATen/Dispatch.h>
12 #include <ATen/core/TensorBase.h>
13 #include <c10/core/Scalar.h>
14 #include <c10/cuda/CUDAMathCompat.h>
15 #include <ATen/cuda/ApplyGridUtils.cuh>
16 #include <ATen/cuda/detail/OffsetCalculator.cuh>
17 #include <ATen/native/cuda/Loops.cuh>
18 
19 namespace at::native {
20 namespace {
21 
leaky_relu_kernel(TensorIteratorBase & iter,const Scalar & negval_)22 void leaky_relu_kernel(TensorIteratorBase& iter, const Scalar& negval_) {
23   AT_DISPATCH_FLOATING_TYPES_AND2(
24       at::ScalarType::Half,
25       at::ScalarType::BFloat16,
26       iter.dtype(),
27       "leaky_relu_cuda",
28       [&]() {
29         using opmath_t = at::opmath_type<scalar_t>;
30         auto negval = negval_.to<opmath_t>();
31         gpu_kernel(iter, [negval] GPU_LAMBDA(scalar_t a) -> scalar_t {
32           opmath_t aop = static_cast<opmath_t>(a);
33           return aop > opmath_t(0) ? aop : aop * negval;
34         });
35       });
36 }
37 
leaky_relu_backward_kernel(TensorIteratorBase & iter,const Scalar & negval_)38 void leaky_relu_backward_kernel(
39     TensorIteratorBase& iter,
40     const Scalar& negval_) {
41   AT_DISPATCH_FLOATING_TYPES_AND2(
42       at::ScalarType::Half,
43       at::ScalarType::BFloat16,
44       iter.dtype(),
45       "leaky_relu_backward_cuda",
46       [&]() {
47         using opmath_t = at::opmath_type<scalar_t>;
48         auto negval = negval_.to<opmath_t>();
49         gpu_kernel(
50             iter, [negval] GPU_LAMBDA(scalar_t a, scalar_t b) -> scalar_t {
51               opmath_t aop = static_cast<opmath_t>(a);
52               opmath_t bop = static_cast<opmath_t>(b);
53               return aop > opmath_t(0) ? bop : bop * negval;
54             });
55       });
56 }
57 } // namespace
58 
59 REGISTER_DISPATCH(leaky_relu_stub, &leaky_relu_kernel);
60 REGISTER_DISPATCH(leaky_relu_backward_stub, &leaky_relu_backward_kernel);
61 
62 } // namespace at::native
63