GPU kernel optimization is essential for highperformance machine learning systems, but it remains laborintensive. Triton simplifies GPU kernel development with a Python-based interface, yet autotuning still requires searching a large configuration space, which is costly. We present TTX, a Triton tuning framework with an XGBoost-based performance predictor. By using input shapes, tuning parameters, and IR-level features, TTX accurately estimates kernel latency and quickly selects promising candidates for further search. Experiments show that TTX achieves around $10 \%$ MAPE on operators such as matrix multiplication, batched matrix multiplication, and convolution across multiple GPUs. It reaches about $80 \%$ of the best performance with the top-1 candidate and over $95 \%$ with the top- 50 candidates.

Performance modeling, a pivotal domain in program cost analysis, currently relies on manually crafted models constrained by various program and hardware limitations, especially in the intricate landscape of GPGPU. Meanwhile, Large Language Models (LLMs) have demonstrated their effectiveness in addressing diverse programming challenges. Our work establishes a connection between LLMs and performance modeling, employing the LLM as a performance estimator. Through experimental exploration with carefully designed large-scale OpenCL datasets, we highlight the potential capability as well as the main difficulties of using LLMs in handling performance modeling tasks for OpenCL device source programs. As the first study for this line of work, our LLM-based performance model achieves a mean absolute percentage error of 24.25% for a large-scale generated validation set. On a set of publicly available OpenCL programs, our model achieves a mean absolute percentage error of 46.1%.

Widely used Deep Learning (DL) frameworks, such as TensorFlow, PyTorch, and MXNet, heavily rely on the NVIDIA cuDNN for performance. However, using cuDNN does not always give the best performance. One reason is that it is hard to handle every case of versatile DNN models and GPU architectures with a library that has a fixed implementation. Another reason is that cuDNN lacks kernel fusion functionality that gives a lot of chances to improve performance. In this paper, we propose a DL optimization framework for versatile GPU workloads, called DeepCuts. It considers both kernel implementation parameters and GPU architectures. It analyzes the DL workload, groups multiple DL operations into a single GPU kernel, and generates optimized GPU kernels considering kernel implementation parameters and GPU architecture parameters. The evaluation result with various DL workloads for inference and training indicates that DeepCuts outperforms cuDNN/cuBLAS-based implementations and the state-of-the-art DL optimization frameworks, such as TVM, TensorFlow XLA, and TensorRT.

Tuning the kernel work-group size for GPUs is a challenging problem. In this paper, using the performance counters provided by GPUs, we characterize a large body of OpenCL kernels to identify the performance factors that affect the choice of a good work-group size. Based on the characterization, we realize that the most influential performance factors with regard to the work-group size include occupancy, coalesced global memory accesses, cache contention, and variation in the amount of workload in the kernel. By tackling the performance factors one by one, we propose auto-tuning techniques that selects the best work-group size and shape for GPU kernels. We show the effectiveness of our auto-tuner by evaluating it with a set of 54 OpenCL kernels on three different NVIDIA GPUs and one AMD GPU. On average, the auto-tuner needs to spend no more than 8 percent of the time required by an exhaustive search to find an optimal work-group size. The execution time of the selected sub-optimal work-group size is at most 1.14x slower than that of the optimal work-group size found by the exhaustive search, on average.

Heterogeneous systems are widening their user-base, and heterogeneous computing is becoming popular in supercomputing. Among others, OpenCL and CUDA are the most popular programming models for heterogeneous systems. Although OpenCL inherited many features from CUDA and they have almost the same platform model, they are not compatible with each other. In this paper, we present similarities and differences between them and propose an automatic translation framework for both OpenCL to CUDA and CUDA to OpenCL. We describe features that make it difficult to translate from one to the other and provide our solution. We show that our translator achieves comparable performance between the original and target applications in both directions. Since each programming model separately has a wide user-base and large code-base, our translation framework is useful to extend the code-base for each programming model and unifies the efforts to develop applications for heterogeneous systems.

To exploit the abundant computational power of the world's fastest supercomputers, an even workload distribution to the typically heterogeneous compute devices is necessary. While relatively accurate performance models exist for conventional CPUs, accurate performance estimation models for modern GPUs do not exist. This paper presents two accurate models for modern GPUs: a sampling-based linear model, and a model based on machine-learning (ML) techniques which improves the accuracy of the linear model and is applicable to modern GPUs with and without caches. We first construct the sampling-based linear model to predict the runtime of an arbitrary OpenCL kernel. Based on an analysis of NVIDIA GPUs' scheduling policies we determine the earliest sampling points that allow an accurate estimation. The linear model cannot capture well the significant effects that memory coalescing or caching as implemented in modern GPUs have on performance. We therefore propose a model based on ML techniques that takes several compiler-generated statistics about the kernel as well as the GPU's hardware performance counters as additional inputs to obtain a more accurate runtime performance estimation for modern GPUs. We demonstrate the effectiveness and broad applicability of the model by applying it to three different NVIDIA GPU architectures and one AMD GPU architecture. On an extensive set of OpenCL benchmarks, on average, the proposed model estimates the runtime performance with less than 7 percent error for a second-generation GTX 280 with no on-chip caches and less than 5 percent for the Fermi-based GTX 580 with hardware caches. On the Kepler-based GTX 680, the linear model has an error of less than 10 percent. On an AMD GPU architecture, Radeon HD 6970, the model estimates with 8 percent of error rates. The proposed technique outperforms existing models by a factor of 5 to 6 in terms of accuracy.

In this paper, we present the design and implementation of an Open Computing Language (OpenCL) framework that targets heterogeneous accelerator multicore architectures with local memory. The architecture consists of a general-purpose processor core and multiple accelerator cores that typically do not have any cache. Each accelerator core, instead, has a small internal local memory. Our OpenCL runtime is based on software-managed caches and coherence protocols that guarantee OpenCL memory consistency to overcome the limited size of the local memory. To boost performance, the runtime relies on three source-code transformation techniques, work-item coalescing, web-based variable expansion and preload-poststore buffering, performed by our OpenCL C source-to-source translator. Work-item coalescing is a procedure to serialize multiple SPMD-like tasks that execute concurrently in the presence of barriers and to sequentially run them on a single accelerator core. It requires the web-based variable expansion technique to allocate local memory for private variables. Preload-poststore buffering is a buffering technique that eliminates the overhead of software cache accesses. Together with work-item coalescing, it has a synergistic effect on boosting performance. We show the effectiveness of our OpenCL framework, evaluating its performance with a system that consists of two Cell BE processors. The experimental result shows that our approach is promising.