Monte Carlo rendering produces photo-realistic images but requires computationally intensive sampling. Recent methods resort to CNN-based denoising or super-resolution networks. To improve rendering efficiency further, we investigate a joint denoising and super-resolution (JDSR) method based on Transformer that explicitly models non-local dependencies, a vital capability for high-quality image reconstruction. However, we observe that existing Transformers mainly rely on dense self-attention, exhibiting a critical limitation when processing noisy inputs: their attention maps frequently concentrate on irrelevant regions while aggregating excessive redundant features. To alleviate the problem, we propose a High-resolution Auxiliary Features-Guided Sparse Transformer (ASFormer), which exploits auxiliary features as guiding information to improve self-attention performance. Specifically, we introduce a dual-branch sparse self-attention mechanism (DB-SSA), consisting of a sparse self-attention (SSA) branch and a dense (DSA) one. The core is to leverage noise-free and informative auxiliary features to suppress irrelevant information using SSA, while ensuring the necessary information flows through DSA. Second, we propose Residual Multi-Scale Feature Extraction block (RMEB) to refine the representation of high-resolution auxiliary features, enhancing their effectiveness as guidance for reconstruction. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods while achieving faster running times.