Image Test Libraries for the on-line self-test of functional units in GPUs running CNNs

The widespread use of artificial intelligence (AI)-based systems brings with it several questions about the deployment of such systems in safety-critical contexts. Several industry standards exist, such as ISO26262 for automotive, that require detecting hardware faults during the mission of the device. Similarly, new standards are being released concerning the functional safety of AI systems (e.g., ISO/IEC CD TR 5469). Hardware solutions have been proposed for the in-field testing of the hardware executing AI applications, but when used in conjunction with complex applications such as Convolutional Neural Networks (CNNs) in image processing tasks, they may increase the hardware cost and affect the application performances. In this thesis, a methodology to develop high-quality test images, to be interleaved with the normal inference process of the CNN application is proposed. An Image Test Library (ITL) that targets GPU single-precision floating-point multipliers is developed with the aim of performing an on-line test of said functional units. The proposed approach does not require changing the actual CNN (thus incurring in very costly memory operations) since it is able to exploit the actual CNN structure. In particular, the ITL is built to exploit the convolution operation between an input image and a series of filters, which consists of multiply-and-add operations, to pass test patterns generated beforehand to multipliers. Since the fundamental objective is to keep the CNN structure, and thus also the weights, unchanged, the only elements that can be manipulated are the input images. For this reason, test patterns for a multiplier must be generated with methods exploiting Automatic Test Pattern Generation (ATPG) techniques, putting the already trained network weights as constraints. The generated test patterns are placed into the right spots of the input image (or images), where it is guaranteed that a certain multiplier will multiply them by the weight used as constraint. The main issue that arises at this point is ensuring the correct placement of test patterns into the ITL, which depends on the scheduling algorithm of the GPU and the convolution algorithm. In particular, the thesis work consisted of analyzing existing implementations of convolution algorithms such as GEMM (General Matrix Multiplication), analyzing the GPU scheduling policy and developing an algorithm that exploits this knowledge to correctly predict, given an input pixel-weight pair, which multiplier will perform that multiplication. The experiments that have been performed on the first layer of a ResNet-20 CNN and a DenseNet-121 CNN, show that a 6/8-image ITL is able to achieve about 95\% of stuck-at test coverage on the single-precision floating-point multipliers in a GPU. The obtained ITL requires a very low test application time and has a very low memory footprint, needing space only to store the test images and the golden test responses.