Codeplay Reference Silicon (RefSi) or simply Codeplay Silicon is the name of a computer platform developed by Codeplay to be used as a complete reference target for running the Acoran compute stack on a RISC-V-based accelerator. It can be seen as a System-On-Chip which includes components typically found in hardware accelerator platforms, such as a CPU, DRAM, accelerator cores, fast on-chip (TCM) memory, a DMA controller, a command processor and a bus to link the components together.
Since RefSi is a virtual platform, many of its components are simulated. Components that are not simulated are used as-is from a x86 machine (e.g. CPU, system RAM) we call the ‘host’, in reference to the “offload execution model” where the “Host CPU” offloads the key computations to an accelerator.
The platform is meant as an example for how Acoran maps to hardware with similar components and features. It is used to demonstrate how such hardware features can be integrated within Acoran components such as ComputeCpp, oneAPI Construction Kit and libraries such as SYCL-BLAS and SYCL-DNN.
In this section we will describe the first version RefSi Accelerator architecture (RefSi M1), its components and how these components interact. A related but different reference architecture is RefSi G1, which is more CPU-like and does not feature a DMA controller or command processor.
The RefSi M1 platform was designed with an ‘offloading’ execution model in mind, which is exposed through standards like SYCL and OpenCL. In this model, the role of the CPU is to run the ‘host’ part of compute applications. Typically, this is logic that is not easily executed in parallel or that would not benefit from highly parallel execution. The CPU is also responsible for scheduling and dispatching work (in the form of compute kernels) to the accelerator. In some scenarios, the CPU is also used to compile compute kernels down to machine code that can be executed by accelerator cores.
Unlike many components in the RefSi platform, the CPU is not simulated. The x86 CPU that executes the Acoran for RISC-V software (the ‘system CPU’) is used instead. There is no hard dependency on the x86 architecture however, and an ARM or RISC-V CPU could equally well be used for this platform.
The details of this CPU are dependant on the host system where the RefSi simulator runs.
This is the memory used by the host CPU. While not technically part of the RefSi SoC, it is listed here to contrast with the DRAM that is part of the SoC. Due to RefSi being a virtual platform, system RAM is used to simulate every kind of memory that is exposed by the platform. However, accelerator cores cannot access system RAM directly and must use DRAM instead. While the DMA controller could technically be used by the cores to transfer data from system RAM, cores are not expected to access any system pointers.
Size: dependent on the host system.
RefSi accelerators have their own dedicated RAM, separate from the system RAM that is used by the CPU. DRAM can be accessed by the CPU, accelerator cores, the DMA controller and the command processor.
This is the memory used to allocate global buffers when using SYCL or OpenCL. DRAM can be pinned in the CPU’s address space and used to transfer data between the CPU and accelerator cores.
DRAM is also used to store command buffers that can be executed by the command processor as well as kernel executables (e.g. ELFs) prior to being loaded in an accelerator core’s TCIM.
The DRAM size is set to 4 GB.
Since their purpose is to execute compute kernels, accelerator cores are at the heart of the RefSi platform which provides two 64-bit RISC-V cores, each with 4 hardware threads (called ‘harts’). These 8 independent (i.e. no lockstep execution) execution units provide some amount of parallelism while keeping the system simple and easy to understand. Cores support the RVV extensible vector extension to increase parallelism and demonstrate how Acoran can be mapped to vector hardware.
Following typical accelerator architecture, the cores in the RefSi M1 platform feature fast on-chip memory (tightly-coupled memory, or TCM) that can be used to improve kernel performance by using tiling and double-buffering strategies. Each core has a dedicated amount of TCM and DMA transfers to and from TCM are supported.
While these RISC-V cores are capable of executing a general-purpose operating system or even RTOS, for simplicity’s sake they are only used to execute kernels using a ‘bare metal’ approach. Following the ‘offloading’ style of execution model exposed by OpenCL and SYCL, the cores are idle until the command processor wakes them up with work and as soon as kernel execution is completed.
The ISA for RISC-V cores is set to RV64GCVZbc, the VLEN to 512 and ELEN to 64.
Each RISC-V accelerator core in the platform is associated with a chunk of fast on-chip TCM (tightly-coupled memory) called. TCDM is meant for the storage of data that is frequently accessed or requires low latency.
Size: 4 MB TCDM per core (data)
The DMA controller is responsible for managing data transfers between different kinds of memory, such as system RAM, DRAM and accelerator cores’ TCM. While platform components such as the CPU and accelerator cores can access memory directly, performing data transfers using the DMA controller reduces transfer overheads since only a minimal number of instructions are needed to start a transfer. Transfers are also asynchronous to other components (CPU, CMP, cores), which increases parallelism within the platform.
The DMA controller includes a single FIFO queue of transfer requests, and two DMA channels for performing DMA transfers. This means that up to two concurrent data transfers are possible.
DMA transfers can be done between DRAM and TCM in both directions, between system RAM and DRAM also in both directions, as well as within DRAM, so long as the source and destination memory regions do not overlap. 1D, 2D and 3D transfers are supported by the DMA controller.
The DMA controller can be accessed through the bus (e.g. by the CPU or accelerator cores) or through the command processor using the same interface.
Accelerator cores can access the DMA controller through a set of DMA registers that are mapped into the cores’ address space. These registers allow data transfers to be configured, started and waited for completion using regular load and store instructions from the RISC-V ISA. Each hart has its own set of DMA registers, allowing for transparent and independent access to the DMA controller from OpenCL and SYCL kernels.
The command processor (CMP) can be seen as a scheduler for the RefSi M1 platform. Its purpose is to manage kernel execution as well as DMA transfers. It is a simple programmable processor that can execute high-level commands such as ‘Execute Kernel Slice’ or ‘Wait For Pending DMA Transfers To Complete’.
The CMP includes a FIFO queue of command requests, which are made up of a start address and size for a command buffer. Command requests are executed in-order, and commands specified within a command buffer are also executed in-order. Once all commands in the command buffer have finished executing, the CMP signals the CPU that a command request has been completed through the bus. The FIFO queue size can contain a maximum of 16 requests. Attempting to submit a request when the queue is full results the sender of the request to stall until an entry is available in the queue.
In the first version of Codeplay Reference Silicon (M1) there is no plan to include synchronization between different commands within a command buffer (e.g. to implement compute graphs). However, this could be done in the future to allow for overlapping kernel execution and DMA transfers.
The bus connects all components in the platform together, except for DRAM. The CPU can use it to enqueue a command request onto the CMP FIFO as well as be notified when the command processor has finished executing a request.
In the current revision of this design, the specific mechanism used by the bus to communicate between components has not been explicitely specified.
Dedicated matrix multiplication unit