We show in this work that memory intensive computations can result in severe performance problems due to off-chip memory access and CPU-GPU context switch overheads in a wide range of deep learning models. For this problem, current just-in-time (JIT) kernel fusion and code generation techniques have limitations, such as rough fusion plan exploration strategies and limited code generation ability. We propose FusionStitching, a deep learning compiler capable of fusing memory intensive operators, with varied data dependencies and non-homogeneous parallelism, into large GPU kernels to reduce global memory access and context switch overhead automatically. FusionStitching widens the range of operation combinations that fusion can target beyond previous JIT works by introducing data reuse of intermediate values. It explores large fusion spaces to decide optimal fusion plans with considerations of memory access costs, kernel calls and resource usage constraints. FusionStitching tunes the optimal stitching scheme with a domain-specific cost model efficiently. Experimental results show that FusionStitching can reach up to 2.21x speedup compared to state-of-the-art, with 1.45x on average. Besides these experimental results, we integrated our approach into a compiler product and deployed it onto a production cluster for AI workloads with thousands of GPUs. The system has been in operation for more than 4 months and saves 7,000 GPU hours on average for approximately 30,000 tasks per month.