SqueezCL: Squeezing OpenCL Kernels for Approximate Computing on Contemporary GPUs

Approximate computing provides an opportunity for exploiting application characteristics to improve performance of computing systems. However, such opportunity must be balanced against generality of methods and quality guarantees that the system designer can provide to the application developer. Improved parallel processing in graphics processing units (GPUs) provides one such means for data-level parallel applications. We propose SqueezCL a software method to reduce the hardware resources used by an OpenCL kernel. SqueezCL transforms an exact OpenCL kernel to an approximate OpenCL kernel by squeezing dimensions of its data elements. The core of SqueezCL leverages bitwidth reduction to shrink the hardware resources. Selectively reducing the precision and size of data elements generates approximate kernels that can be executed faster at a cost to quality loss. Exploiting this opportunity is particularly important for GPU accelerators that are inherently subject to memory resource constraints. We evaluate SqueezCL on a diverse set of data-level parallel OpenCL benchmarks from the AMD APP SDK v2.9. Experimental result on the AMD Radeon HD 5870 shows that SqueezCL yields on average 1.1× higher performance with less than 10% quality loss without requiring any changes to the underlying GPU hardware.