CRISP: Concurrent Rendering and Compute Simulation Platform for GPUs

Abstract

Programmable graphics and compute shaders have blurred the lines between graphics processing and general-purpose computation. APIs such as Asynchronous Compute enable the concurrent execution of raster-based graphics shaders with more general, parallel computation, and emerging graphics and computation-heavy workloads in areas such as augmented and virtual reality can benefit from spatially sharing a GPU. Although concurrent execution of graphics rendering and compute kernels are widely used today, contemporary simulators primarily focus on either general compute kernels or rendering pipelines, lacking insight into the challenges and potential opportunities arising from concurrently executing both graphics and compute. To bridge this gap, we present CRISP, a validated cycle-level GPU simulator capable of running graphics rendering and compute kernels concurrently. CRISP extends Accel-Sim by incorporating a detailed programmable graphics pipeline to support Vulkan GLSL Shaders. We demonstrate that CRISP models the rendering pipeline to a high degree of accuracy and highlight the new research opportunities enabled by CRISP through case studies demonstrating the impact of mipmapping, L2 composition using advanced shading techniques, and how prior work on concurrent kernel execution behaves when mixing graphics workloads with compute.