1 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
2 #include <ATen/core/Tensor.h>
3 #include <ATen/Config.h>
4 #include <ATen/native/Activation.h>
5
6 #ifndef AT_PER_OPERATOR_HEADERS
7 #include <ATen/NativeFunctions.h>
8 #else
9 #include <ATen/ops/gelu_native.h>
10 #include <ATen/ops/gelu_backward_native.h>
11 #endif
12
13 #if !AT_MKLDNN_ENABLED()
14
15 namespace at { namespace native {
16
mkldnn_gelu(const Tensor & input,c10::string_view approximate)17 Tensor mkldnn_gelu(const Tensor& input, c10::string_view approximate) {
18 TORCH_CHECK(false, "mkldnn_gelu: ATen not compiled with MKLDNN support");
19 }
20
mkldnn_gelu_backward(const Tensor & grad_output,const Tensor & input,c10::string_view approximate)21 Tensor mkldnn_gelu_backward(const Tensor& grad_output, const Tensor& input, c10::string_view approximate) {
22 TORCH_CHECK(false, "mkldnn_gelu_backward: ATen not compiled with MKLDNN support");
23 }
24
25 }}
26
27 #else // AT_MKLDNN_ENABLED
28
29 #include <ATen/native/mkldnn/MKLDNNCommon.h>
30 #include <ATen/native/mkldnn/Utils.h>
31
32 namespace at { namespace native {
33
mkldnn_gelu(const Tensor & input,c10::string_view approximate)34 Tensor mkldnn_gelu(const Tensor& input, c10::string_view approximate) {
35 if (input.scalar_type() == ScalarType::BFloat16) {
36 TORCH_CHECK(mkldnn_bf16_device_check(),
37 "mkldnn_gelu: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq");
38 }
39 TORCH_CHECK(get_gelutype_enum(approximate) == GeluType::None,
40 "mkldnn_gelu: fast, approximate gelu is not supported");
41 const ideep::tensor& x = itensor_from_tensor(input);
42 ideep::tensor y;
43 ideep::eltwise_forward::compute(
44 x, y, ideep::algorithm::eltwise_gelu_erf, ideep::prop_kind::forward_training, /*alpha*/ 0.0);
45 return new_with_itensor_mkldnn(std::move(y), optTypeMetaToScalarType(input.options().dtype_opt()),
46 input.options().device_opt());
47 }
48
mkldnn_gelu_backward(const Tensor & grad_output,const Tensor & input,c10::string_view approximate)49 Tensor mkldnn_gelu_backward(const Tensor& grad_output, const Tensor& input, c10::string_view approximate) {
50 TORCH_CHECK(get_gelutype_enum(approximate) == GeluType::None,
51 "mkldnn_gelu_backward: fast, approximate gelu is not supported");
52 const ideep::tensor& x = itensor_from_tensor(input);
53 ideep::tensor grady = itensor_from_tensor(grad_output);
54 ideep::tensor gradx;
55 ideep::eltwise_backward::compute(x, grady, gradx,
56 ideep::algorithm::eltwise_gelu_erf, /*alpha*/ 0.0);
57 return new_with_itensor_mkldnn(std::move(gradx),
58 optTypeMetaToScalarType(grad_output.options().dtype_opt()),
59 grad_output.options().device_opt());
60 }
61
62 }}
63
64 #endif // AT_MKLDNN_ENABLED
65