#include "torch_npu/csrc/aten/CustomFunctions.h"
#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
at::Tensor npu_dtype_cast_impl_op_api(const at::Tensor& self, at::ScalarType dtype)
{
if (self.dtype() == dtype) {
return self.clone();
}
at::Tensor result = npu_preparation::apply_tensor_without_format(self.sizes(),
self.options().dtype(dtype));
EXEC_NPU_CMD(aclnnCast, self, dtype, result);
return result;
}
}
at::Tensor npu_dtype_cast(const at::Tensor& self, at::ScalarType dtype)
{
DO_COMPATIBILITY(aclnnCast, acl_op::npu_dtype_cast(self, dtype));
return npu_dtype_cast_impl_op_api(self, dtype);
}
at::Tensor _npu_dtype_cast(const at::Tensor& self, at::ScalarType dtype)
{
DO_COMPATIBILITY(aclnnCast, acl_op::npu_dtype_cast(self, dtype));
return npu_dtype_cast_impl_op_api(self, dtype);
}
at::Tensor npu_dtype_cast_backward(const at::Tensor& grad, at::ScalarType dtype)
{
grad.requires_grad_();
return at_npu::native::custom_ops::npu_dtype_cast(grad, dtype);
}
at::Tensor _npu_dtype_cast_backward(const at::Tensor& grad, at::ScalarType dtype)
{
grad.requires_grad_();
return at_npu::native::custom_ops::npu_dtype_cast(grad, dtype);
}
}