#include "utils.hpp"
#include <ATen/ATen.h>
#include <torch/library.h>
#include <ATen/autocast_mode.h>
#include <ATen/core/dispatch/Dispatcher.h>
namespace vision {
namespace ops {
namespace {
std::tuple<at::Tensor, at::Tensor> roi_pool_autocast(
const at::Tensor& input,
const at::Tensor& rois,
double spatial_scale,
int64_t pooled_height,
int64_t pooled_width)
{
c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::AutocastPrivateUse1);
static auto roi_pool_handle = c10::Dispatcher::singleton()
.findSchemaOrThrow("torchvision::roi_pool", "")
.typed<std::tuple<at::Tensor, at::Tensor>(const at::Tensor &, const at::Tensor &, double, int64_t, int64_t)>();
auto result = roi_pool_handle.call(
vision::autocast::cached_cast_npu(input, at::kFloat),
vision::autocast::cached_cast_npu(rois, at::kFloat),
spatial_scale, pooled_height, pooled_width);
return std::make_tuple(
std::get<0>(result).to(input.scalar_type()),
std::get<1>(result).to(input.scalar_type()));
}
}
TORCH_LIBRARY_IMPL(torchvision, AutocastPrivateUse1, m)
{
m.impl(
TORCH_SELECTIVE_NAME("torchvision::roi_pool"),
TORCH_FN(roi_pool_autocast));
}
}
}