#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "torch_npu/csrc/core/npu/NpuVariables.h"
#include "torch_npu/csrc/custom_dtype/Init.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
using namespace c10_npu;
constexpr int ATTRS_DIM = 2;
constexpr int TENSORS_DIM = 4;
constexpr int OFFSET_H_INDEX = 2;
constexpr int OFFSET_W_INDEX = 3;
std::tuple<at::Tensor, at::Tensor> npu_deformable_conv2d_out(const at::Tensor &input, const at::Tensor &weight,
const at::Tensor &offset,
const c10::optional<at::Tensor> &bias,
at::IntArrayRef kernel_size, at::IntArrayRef stride,
at::IntArrayRef padding, at::IntArrayRef dilation,
int64_t groups, int64_t deformable_groups, bool modulated)
{
TORCH_CHECK(input.dim() >= TENSORS_DIM, "input has to be more than 4D, but got Tensor of dimension ",
input.dim(), OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(weight.dim() >= TENSORS_DIM, "weight has to be more than 4D, but got Tensor of dimension ",
weight.dim(), OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(offset.dim() >= TENSORS_DIM, "offset has to be more than 4D, but got Tensor of dimension ",
offset.dim(), OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(kernel_size.size() >= ATTRS_DIM, "kernel_size has to contain more than 2 elements, but got ",
kernel_size.size(), OPS_ERROR(ErrCode::PARAM));
int64_t n = input.size(0);
int64_t ci = input.size(1);
int64_t co = weight.size(0);
int64_t ho = offset.size(OFFSET_H_INDEX);
int64_t wo = offset.size(OFFSET_W_INDEX);
int64_t kh = kernel_size[0];
int64_t kw = kernel_size[1];
c10::SmallVector<int64_t, SIZE> deformable_offset_size = {n, ci, ho * kh, wo * kw};
c10::SmallVector<int64_t, SIZE> conv_output_size = {n, co, ho, wo};
auto deformable_offset = npu_preparation::apply_tensor_without_format(deformable_offset_size,
input.options().dtype(input.dtype()));
auto conv_output = npu_preparation::apply_tensor_without_format(conv_output_size,
input.options().dtype(input.dtype()));
EXEC_NPU_CMD(aclnnDeformableConv2d, input, weight, offset, bias, kernel_size, stride, padding, dilation,
groups, deformable_groups, modulated, conv_output, deformable_offset);
return std::make_tuple(conv_output, deformable_offset);
}
std::tuple<at::Tensor, at::Tensor> npu_deformable_conv2d(const at::Tensor &input, const at::Tensor &weight,
const at::Tensor &offset,
const c10::optional<at::Tensor> &bias,
at::IntArrayRef kernel_size, at::IntArrayRef stride,
at::IntArrayRef padding, at::IntArrayRef dilation,
int64_t groups, int64_t deformable_groups, bool modulated)
{
DO_COMPATIBILITY(aclnnDeformableConv2d,
acl_op::npu_deformable_conv2d(input, weight, offset, bias, kernel_size, stride, padding, dilation,
groups, deformable_groups, modulated));
if (c10_npu::IsAclnnOnly()) {
return npu_deformable_conv2d_out(input, weight, offset, bias, kernel_size, stride, padding, dilation,
groups, deformable_groups, modulated);
}
return acl_op::npu_deformable_conv2d(input, weight, offset, bias, kernel_size, stride, padding, dilation,
groups, deformable_groups, modulated);
}
}