#include <c10/core/impl/alloc_cpu.h>
#include <c10/core/Allocator.h>

#include <torch/csrc/Device.h>
#include <torch/extension.h>

#include <ATen/native/cpu/Loops.h>
#include <ATen/native/DispatchStub.h>
#include <ATen/EmptyTensor.h>


static uint64_t add_counter = 0;
static uint64_t last_saved_value = 0;

// basic dummy add function
at::Tensor custom_add_Tensor(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha) {
  add_counter += 1;
  // Since this custom device is just for testing, not bothering to implement kernels.
  return at::empty(self.sizes(), self.options());
}

// A dummy allocator for our custom device, that secretly uses the CPU
struct DummyCustomAllocator final : at::Allocator {
  DummyCustomAllocator() = default;
  at::DataPtr allocate(size_t nbytes) const override {
    void* data = c10::alloc_cpu(nbytes);
    return {data, data, &ReportAndDelete, at::Device(at::DeviceType::PrivateUse1, 0)};
  }

  static void ReportAndDelete(void* ptr) {
    if (!ptr) {
      return;
    }
    c10::free_cpu(ptr);
  }

  at::DeleterFnPtr raw_deleter() const override {
    return &ReportAndDelete;
  }
};

// Register our dummy allocator
static DummyCustomAllocator global_custom_alloc;
REGISTER_ALLOCATOR(c10::DeviceType::PrivateUse1, &global_custom_alloc);

// basic dummy empty function, so we can directly construct tensors on the custom device
// This dummy test device will just use the CPU allocator, and ignores pinned memory.
at::Tensor custom_empty_memory_format(at::IntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory, c10::optional<at::MemoryFormat> memory_format) {
  constexpr c10::DispatchKeySet private_use_ks(c10::DispatchKey::PrivateUse1);
  return at::detail::empty_generic(size, &global_custom_alloc, private_use_ks, c10::dtype_or_default(dtype), memory_format);
}
at::Tensor custom_empty_symint(c10::IntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory, c10::optional<at::MemoryFormat> memory_format) {
  constexpr c10::DispatchKeySet private_use_ks(c10::DispatchKey::PrivateUse1);
  return at::detail::empty_generic(size, &global_custom_alloc, private_use_ks, c10::dtype_or_default(dtype), memory_format);
}

at::Tensor & custom_fill__scalar(at::Tensor & self, const at::Scalar & value) {
  // Not bothering to implement.
  return self;
}

// basic dummy copy_() function, so we can copy from the custom device to/from CPU
at::Tensor custom__copy_from(const at::Tensor& self, const at::Tensor& dst, bool non_blocking) {
  TORCH_CHECK(self.is_cpu() || self.device().type() == c10::DeviceType::PrivateUse1, "Dummy test only allows copy from cpu -> dummy device.");
  TORCH_CHECK(dst.is_cpu() || dst.device().type() == c10::DeviceType::PrivateUse1, "Dummy test only allows copy from cpu -> dummy device.");

  // Some dummy asserts for the basic use case: inputs are the same size / dtype, all contiguous.
  TORCH_CHECK(self.sizes() == dst.sizes());
  TORCH_CHECK(self.scalar_type() == dst.scalar_type());
  TORCH_CHECK(self.is_contiguous() && dst.is_contiguous());

  std::memcpy(dst.storage().data_ptr().get(), self.storage().data_ptr().get(), self.storage().nbytes());
  return dst;
}


// This macro does the heavy lifting.
// With TORCH_LIBRARY_IMPL, you can register custom kernels for your backend.
// For open registration, we're registering all of our kernels to the PrivateUse1 dispatch key.
// Later in this file, we map a custom device to the PrivateUse1 device type,
// which allows user code that puts a tensor on your custom_device to eventually get plumbed
// into the kernels registered here.
//
// This macro registers your kernels to the PyTorch Dispatcher.
// More details on the dispatcher can be found at http://blog.ezyang.com/2020/09/lets-talk-about-the-pytorch-dispatcher/.
TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
  m.impl("add.Tensor", &custom_add_Tensor);
  m.impl("empty.memory_format", &custom_empty_symint);
  m.impl("fill_.Scalar", &custom_fill__scalar);
  m.impl("_copy_from", &custom__copy_from);
}

// This basic implementation doesn't bother dealing with different device indices
// (e.g. custom_device:0 vs. custom_device:1).
// We could do that by letting the user pass in a device index in our exposed device function.
// Note that if you do that, you'll also need to register a device guard to core.
// See `c10/core/impl/DeviceGuardImplInterface.h:C10_REGISTER_GUARD_IMPL`.
c10::Device get_custom_device() {
  return c10::Device(c10::DeviceType::PrivateUse1, 0);
}

bool custom_add_called() {
  bool called = false;
  if (add_counter > last_saved_value) {
    called = true;
    last_saved_value = add_counter;
  }
  return called;
}

// Here, we're exposing a custom device object that corresponds to our custom backend.
// We do this using pybind: exposing an "extension_name.custom_device()" function in python,
// that's implemented in C++.
// The implementation in this file maps directly to the `PrivateUse1` device type.
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("custom_device", &get_custom_device, "get custom device object");
    m.def("custom_add_called", &custom_add_called, "check if our custom add function was called");
}