#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/OpAdapter.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/OpUtils.h"
namespace op_api {
const int FOUR_DIM = 4;
const int AXIS_THREE = 3;
const int AXIS_TWO = 2;
using npu_preparation = at_npu::native::OpPreparation;
namespace {
at::Tensor _linspace_from_neg_one(const at::Tensor& grad, int64_t num_steps, bool align_corners)
{
if (num_steps <= 1) {
return at::tensor(0, grad.options());
}
auto range = at::linspace(-1, 1, num_steps, grad.options());
if (!align_corners && num_steps != 0) {
range = range * (num_steps - 1) / num_steps;
}
return range;
}
}
at::Tensor affine_grid_generator_backward(
const at::Tensor& grad,
at::IntArrayRef size,
bool align_corners)
{
TORCH_CHECK(size.size() == FOUR_DIM, "AffineGridGeneratorBackward needs 4d (spatial) input."
+ OPS_ERROR(ErrCode::PARAM));
c10::SmallVector<int64_t, SIZE> output_size = {size[0], 3, 2};
at::Tensor result = npu_preparation::apply_tensor_with_format(grad, output_size, ACL_FORMAT_ND);
c10::SmallVector<int64_t, SIZE> assist_size = {size[0], size[2], size[3], 3};
at::Tensor assist = npu_preparation::apply_tensor_without_format(grad, assist_size);
assist.select(-1, 0).copy_(_linspace_from_neg_one(grad, size[AXIS_THREE], align_corners));
assist.select(-1, 1).copy_(_linspace_from_neg_one(grad, size[AXIS_TWO], align_corners).unsqueeze_(-1));
assist.select(-1, AXIS_TWO).fill_(1);
AT_ASSERT(grad.sizes() == at::IntArrayRef({size[0], size[2], size[3], 2}), OPS_ERROR(ErrCode::VALUE));
auto reassist = assist.view({size[0], size[2] * size[3], 3}).transpose(1, 2);
auto grad_view = grad.view({size[0], size[2] * size[3], 2});
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
EXEC_NPU_CMD(aclnnBatchMatMul, reassist, grad_view, result, cube_math_type);
auto fresult = result.transpose(1, 2);
return fresult;
}
}