This paper describes multigrain parallel processing on OSCAR (Optimally SCheduled Advanced multiprocessoR) chip multiprocessor architecture. OSCAR compiler cooperative chip multiprocessor architecture aims at development of scalable, high effective performance and cost effective chip multiprocessor with ease of use by compiler supports. OSCAR chip multiprocessor architecture integrates simple single issue processors having distributed shared data memory for optimal use of data locality over different loops and fine grain data transfer and synchronization, local data memory for private data recognized by compiler, and compiler controllable data transfer unit for overlapping data transfer to hide data transfer overhead. This OSCAR chip multiprocessor and OSCAR multigrain parallelizing compiler have been developed simultaneously. Performance of multigrain parallel processing on OSCAR chip multiprocessor architecture is evaluated using SPEC fp 2000/95 benchmark suite. When microSPARC like single issue core is used, OSCAR chip multiprocessor architecture gives us 2.36 times speedup in fpppp, 2.64 times in su2cor, 2.88 times in turb3d, 2.98 times in hydro2d, 3.84 times in tomcatv, 3.84 times in mgrid and 3.97 times in swim respectively for four processors against single processor.

A power-aware compiler controllable chip multiprocessor (CMP) is presented and its performance and power consumption are evaluated with the optimally scheduled advanced multiprocessor (OSCAR) parallelizing compiler. The CMP is equipped with power control registers that change clock frequency and power supply voltage to functional units including processor cores, memories, and an interconnection network. The OSCAR compiler carries out coarse-grain task parallelization of programs and reduces power consumption using architectural power control support and the compiler's power saving scheme. The performance evaluation shows that MPEG-2 encoding on the proposed CMP with four CPUs results in 82.6% power reduction in real-time execution mode with a deadline constraint on its sequential execution time. Furthermore, MP3 encoding on a heterogeneous CMP with four CPUs and four accelerators results in 53.9% power reduction at 21.1-fold speed-up in performance against its sequential execution in the fastest execution mode.

Utilization of local memory from real-time embedded systems to high performance systems with multi-core processors has become an important factor for satisfying hard deadline constraints. However, challenges lie in the area of efficiently managing the memory hierarchy, such as decomposing large data into small blocks to fit onto local memory and transferring blocks for reuse and replacement. To address this issue, this paper presents a compiler optimization method that automatically manage local memory of multi-core processors. The method selects and maps multi-dimensional data onto software specified memory blocks called Adjustable Blocks. These blocks are hierarchically divisible with varying sizes defined by the features of the input application. Moreover, the method introduces mapping structures called Template Arrays to maintain the indices of the decomposed multi-dimensional data. The proposed work is implemented on the OSCAR automatic parallelizing compiler and evaluations were performed on the Renesas RP2 8-core processor. Experimental results from NAS Parallel Benchmark, SPEC benchmark, and multimedia applications show the effectiveness of the method, obtaining maximum speed-ups of 20.44 with 8 cores utilizing local memory from single core sequential versions that use off-chip memory.

The advancement of multicore technology has made hundreds or even thousands of cores processor on a single chip possible. However, on a larger scale multicore, a hardware-based cache coherency mechanism becomes overwhelmingly complicated, hot, and expensive. Therefore, we propose a software coherence scheme managed by a parallelizing compiler for shared-memory multicore systems without a hardware cache coherence mechanism. Our proposed method is simple and efficient. It is built into OSCAR automatic parallelizing compiler. The OSCAR compiler parallelizes the coarse grain task, analyzes stale data and line sharing in the program, then solves those problems by simple program restructuring and data synchronization. Using our proposed method, we compiled 10 benchmark programs from SPEC2000, SPEC2006, NAS Parallel Benchmark (NPB), and MediaBench II. The compiled binaries then are run on Renesas RP2, an 8 cores SH-4A processor, and a custom 8-core Altera Nios II system on Altera Arria 10 FPGA. The cache coherence hardware on the RP2 processor is only available for up to 4 cores. The RP2's cache coherence hardware can also be turned off for non-coherence cache mode. The Nios II multicore system does not have any hardware cache coherence mechanism; therefore, running a parallel program is difficult without any compiler support. The proposed method performed as good as or better than the hardware cache coherence scheme while still provided the correct result as the hardware coherence mechanism. This method allows a massive array of shared memory CPU cores in an HPC setting or a simple non-coherent multicore embedded CPU to be easily programmed. For example, on the RP2 processor, the proposed software-controlled non-coherent-cache (NCC) method gave us 2.6 times speedup for SPEC 2000 “equake” with 4 cores against sequential execution while got only 2.5 times speedup for 4 cores MESI hardware coherent control. Also, the software coherence control gave us 4.4 times speedup for 8 cores with no hardware coherence mechanism available.

We built a 12.4 mm12.4 mm, 45-nm CMOS, chip that integrates eight 648-MHz general purpose cores, two matrix processor (MX-2) cores, four flexible engine (FE) cores and media IP (VPU5) to establish heterogeneous multi-core chip architecture. The general purpose core had its IPC (instructions per cycle) performance enhanced by adding 32-bit instructions to the existing 16-bit fixed-length instruction set and executing up to two 32-bit instructions per cycle. Considering these five-to-seven years of embedded LSI and increasing trend of access-master within LSI, we predict that the memory usage of single core will not exceed 32-bit physical area (i.e. 4 GB), but chip-total memory usage will exceed 4 GB. Based on this prediction, the physical address was expanded from 32-bit to 40-bit. The fabricated chip was tested and a parallel operation of eight general purpose cores and four FE cores and eight data transfer units (DTU) is obtained on AAC (Advanced Audio Coding) encode processing.

Emerging byte-addressable non-volatile memory devices attract much attention. A non-volatile main memory (NVMM) built on them enables larger memory size and lower power consumption than a traditional DRAM main memory. To fully utilize an NVMM, both software and hardware must be cooperatively optimized. Simultaneously, even focusing on a memory module, its micro architecture is still being developed though real non-volatile memory modules, such as Intel Optane DC persistent memory (DCPMM), have been on the market. Looking at existing NVMM evaluation environments, software simulators can evaluate various micro architectures with their long simulation time. Emulators can evaluate the whole system fast with less flexibility in their configuration than simulators. Thus, an NVMM emulator that can realize flexible and fast system evaluation still has an important role to explore the optimal system. In this paper, we introduce an NVMM emulator for embedded systems and explore a direction of optimization techniques for NVMMs by using it. It is implemented on an SoC-FPGA board employing three NVMM behaviour models: coarse-grain, fine-grain and DCPMM-based. The coarse and fine models enable NVMM performance evaluations based on extensions of traditional DRAM behaviour. The DCPMM-based model emulates the behaviour of a real DCPMM. Whole evaluation environment is also provided including Linux kernel modifications and several runtime functions. We first validate the developed emulator with an existing NVMM emulator, a cycle-accurate NVMM simulator and a real DCPMM. Then, the program behavior differences among three models are evaluated with SPEC CPU programs. As a result, the fine-grain model reveals the program execution time is affected by the frequency of NVMM memory requests rather than the cache hit ratio. Comparing with the fine-grain model and the coarse-grain model under the condition of the former's longer total write latency than the latter's, the former shows lower execution time for four of fourteen programs than the latter because of the bank-level parallelism and the row-buffer access locality exploited by the former model.