#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/OpUtils.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
at::Tensor npu_linear(
const at::Tensor &input,
const at::Tensor &weight,
const c10::optional<at::Tensor> &bias)
{
const at::Tensor &bias_opt = bias.value_or(at::Tensor());
const at::Tensor &weight_t = weight.t();
auto output_size = {input.size(0), weight.size(0)};
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, input.options());
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (bias_opt.defined()) {
const at::Scalar beta = 1;
const at::Scalar alpha = 1;
DO_COMPATIBILITY(aclnnAddmm, acl_op::addmm(bias_opt, input, weight_t, beta, alpha));
EXEC_NPU_CMD(aclnnAddmm, bias_opt, input, weight_t, beta, alpha, result, cube_math_type);
return result;
}
DO_COMPATIBILITY(aclnnMm, acl_op::mm(input, weight_t));
EXEC_NPU_CMD(aclnnMm, input, weight_t, result, cube_math_type);
return result;
}
std::tuple<at::Tensor, at::Tensor> npu_linear_backward(
const at::Tensor &grad,
const at::Tensor &input,
const at::Tensor &weight)
{
DO_COMPATIBILITY(aclnnMm, acl_op::npu_linear_backward(grad, input, weight));
at::Tensor input_grad = npu_preparation::apply_tensor_without_format(input.sizes(), grad.options());
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
EXEC_NPU_CMD(aclnnMm, grad, weight, input_grad, cube_math_type);
const at::Tensor &grad_t = grad.t();
at::Tensor weight_grad = npu_preparation::apply_tensor_without_format(weight.sizes(), grad.options());
EXEC_NPU_CMD(aclnnMm, grad_t, input, weight_grad, cube_math_type);
return std::tie(input_grad, weight_grad);
}
}