Convolution is one of the most computationally intensive operations that must be performed for machine learning model inference. A traditional approach to computing convolutions is known as the Im2Col + BLAS method. This article proposes SConv: a direct-convolution algorithm based on an MLIR/LLVM code-generation toolchain that can be integrated into machine-learning compilers. This algorithm introduces: (a) Convolution Slicing Analysis (CSA)—a convolution-specific 3D cache-blocking analysis pass that focuses on tile reuse over the cache hierarchy; (b) Convolution Slicing Optimization—a code-generation pass that uses CSA to generate a tiled direct-convolution macro-kernel; and (c) Vector-based Packing—an architecture-specific optimized input-tensor packing solution based on vector-register shift instructions for convolutions with unitary stride. Experiments conducted on 393 convolutions from full ONNX-MLIR machine learning models indicate that the elimination of the Im2Col transformation and the use of fast packing routines result in a total packing time reduction, on full model inference, of 2.3×–4.0× on Intel x86 and 3.3×–5.9× on IBM POWER10. The speed-up over an Im2Col + BLAS method based on current BLAS implementations for end-to-end machine-learning model inference is in the range of 11%–27% for Intel x86 and 11%–34% for IBM POWER10 architectures. The total convolution speedup for model inference is 13%–28% on Intel x86 and 23%–39% on IBM POWER10. SConv also outperforms BLAS GEMM, when computing pointwise convolutions in more than 82% of the 219 tested instances.

Convolution is one of the most computationally intensive machine learning model operations, usually solved by the known Im2Col + BLAS method. This work proposes a novel convolution-algorithm to improve upon Im2Col + BLAS by introducing (a) CSA: a convolution specific 3D cache-blocking analysis that focuses on tile reuse over the cache hierarchy, (b) CSO: a macro-kernel that follows CSA to compute the convolution by tiling it, (c) a specialized microkernel that seeks to achieve peak hardware performance, and (d) packing routines for the input tensor and filters to bridge the gap between tiling and micro-kernel. Our approach speeds up end-to-end machine learning model inference by up to 26% and 21% for x86 and POWER10 architectures, respectively.

Although code generation for Convolution Neural Network (CNN) models has been extensively studied, performing efficient data slicing and parallelization for highly-constrained Multicore Neural Processor Units (NPUs) is still a challenging problem. Given the size of convolutions' in-put/output tensors and the small footprint of NPU on-chip memories, minimizing memory transactions while maximizing parallelism and MAC utilization are central to any effective solution. This paper proposes a TensorFlow XLA/LLVM compiler optimization pass for Multicore NPUs, called Tensor Slicing Optimization (TSO), which: (a) maximizes convolution parallelism and memory usage across NPU cores; and (b) reduces data transfers between host and NPU on-chip memories by using DRAM memory burst time estimates to guide tensor slicing. To evaluate the proposed approach, a set of experiments was performed using the NeuroMorphic Processor (NMP), a multicore NPU containing 32 RISC-V cores extended with novel CNN instructions. Experimental results show that TSO is capable of identifying the best tensor slicing that minimizes execution time for a set of CNN models. Speed-ups of up to 21.7% result when comparing the TSO burst-based technique to a no-burst data slicing approach.

Although heterogeneous computing has enabled impressive program speed-ups, knowledge about the architecture of the target device is still critical to reap full hardware benefits. Programming such architectures is complex and is usually done by means of specialized languages (e.g. CUDA, OpenCL). The cost of moving and keeping host/device data coherent may easily eliminate any performance gains achieved by acceleration. Although this problem has been extensively studied for multicore architectures and was recently tackled in discrete GPUs through CUDA8, no generic solution exists for integrated CPU/GPUs architectures like those found in mobile devices (e.g. ARM Mali). This paper proposes Data Coherence Analysis (DCA), a set of two data-flow analyses that determine how variables are used by host/device at each program point. It also introduces Data Coherence Optimization (DCO), a code optimization technique that uses DCA information to: (a) allocate OpenCL shared buffers between host and devices; and (b) insert appropriate OpenCL function calls into program points so as to minimize the number of data coherence operations. DCO was implemented in AClang LLVM (www.aclang.org) a compiler capable of translating OpenMP 4.X annotated loops to OpenCL kernels, thus hiding the complexity of directly programming in OpenCL. Experimental results using DCA and DCO in AClang to compile programs from the Parboil, Polybench and Rodinia benchmarks reveal performance speed-ups of up to 5.25x on an Exynos 8890 Octacore CPU with ARM Mali-T880 MP12 GPU and up to 2.03x on a 2.4 GHz dual-core Intel Core i5 processor equipped with an Intel Iris GPU unit.