create_acl_quant_matmul Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/ACLUtils.h lines 217–245

  template <typename ACLQuantMatmulT>
  std::shared_ptr<ACLQuantMatmulT> create_acl_quant_matmul(
      const ACLQuantMatmulCacheKey& key) {
    std::optional<float*> bias_ptr;
    if (bias_.has_value()) {
      bias_ptr = (float*)bias_.value().get_data_handle();
    }
    auto acl_gemm = std::make_shared<ACLQuantMatmulT>(
        k_,
        n_,
        weight_scale_,
        weight_zero_point_,
        (int8_t*)weight_.get()->get_data_handle(),
        bias_ptr,
        key);

    // validate
    auto status = acl_gemm->validate();
    if (status.error_code() != arm_compute::ErrorCode::OK) {
      TORCH_WARN(
          "Arm Compute Library's Quantized Matmul Validation Failed: " +
          status.error_description());
      return nullptr;
    }

    // configure
    acl_gemm->configure();
    return acl_gemm;
  }