diff --git a/doc/index.md b/doc/index.md index 8712cb3..eb53141 100644 --- a/doc/index.md +++ b/doc/index.md @@ -1,8 +1,8 @@ -# RISC-V Extension Interface +# CORE-V X-Interface **Draft Specification Version 0.1** -The RISC-V Extension Interface (REI) defines a unified framework for RISC-V CPU cores to implement ISA extensions in external functional units and to share accelerator- and co-processor structures among multiple cores. +The CORE-V X-Interface (core-v-xif) for RISC-V extension defines a unified framework for RISC-V CPU cores to implement ISA extensions in external functional units and to share accelerator- and co-processor structures among multiple cores. ## Background The different RISC-V cores originally designed and used at IIS (CV32E40P, CVA6, Ibex, Snitch) all feature various, and configurable extensions over the baseline RV32I/RV64I ISA. @@ -11,24 +11,24 @@ The integration of these various instructions into the pipeline varies between d Some ISA extensions (Xpulp, B, M) are deeply integrated into the core pipeline which complicates verification, debugging, and reduces portability to other designs. In contrast, some designs implement the F and D extensions outside of the core pipeline (CV32E40P, Snitch) and use a variety of interfaces and interconnects. -The goal of the REI is to reduce the overall divergence between the different cores and provide a generalized infrastructure suitable to implement custom co-processors and ISA extensions. +The goal of the core-v-xif is to reduce the overall divergence between the different cores and provide a generalized infrastructure suitable to implement custom co-processors and ISA extensions. ## Overview A top-level specification of the operating principle is given [here](operating-principle.md). -The REI aims to provide an entirely accelerator-agnostic instruction offloading mechanism. -The core concept of the REI is to move as much logic required to implement an ISA extension outside of the CPU cores implementing a base instruction set. -No extension-specific changes should be necessary to a CPU core architecture supporting the REI in order to implement new custom extensions. +The core-v-xif aims to provide an entirely accelerator-agnostic instruction offloading mechanism. +The core concept of the core-v-xif is to move as much logic required to implement an ISA extension outside of the CPU cores implementing a base instruction set. +No extension-specific changes should be necessary to a CPU core architecture supporting the core-v-xif in order to implement new custom extensions. -The REI comprises the following core components +The core-v-xif comprises the following core components - CPU core: In order to communicate with the accelerator adapter, the offloading core exposes instruction data not implemented by the supported base ISA and potential operands to the accelerator adapter. - The [accelerator interconnect module](accelerator-interconnect.md) implements a hierarchical interconnect beween the offloading core and core-private or shared accelerator units. - The [accelerator adapter module](accelerator-adapter.md) implements communication with the offloading core, accelerator units and accelerator predecoders. - The [accelerator predecoder module](accelerator-predecoder.md) implements limited decoding of instruction-specific metadata necessary for correct operation (source operands, memory operation, writeback status). One predecoder module is instantiated per connected accelerator. - Accelerator Subsystem: Extension-specific wrapper module, implementing full decoding of the offloaded instructions, execution and maintains architectural state (external register file, CSRs). -The REI defines the following interfaces: +The core-v-xif defines the following interfaces: ### Instruction Offloading Interfaces @@ -49,20 +49,20 @@ The REI defines the following interfaces: A simplified schematic of the overall architecture is given below. -![REI Architecture](img/rei-architecture.svg) +![core-v-xif Architecture](img/rei-architecture.svg) ## Properties ### Accelerator-Agnostic Instruction Offloading -The REI enables decoupled development of accelerators and CPU cores through a mechanism facilitating accelerator-agnostic instruction offloading. +The core-v-xif enables decoupled development of accelerators and CPU cores through a mechanism facilitating accelerator-agnostic instruction offloading. ### Dual-Writeback Instructions -The REI optionally supports implementation of custom ISA extensions mandating dual register writebacks. +The core-v-xif optionally supports implementation of custom ISA extensions mandating dual register writebacks. In order to accomodate that need we provision the possibility to reserve multiple destination registers for a single offloaded instruction. For even destination registers other than `X0`, `Xn` and `Xn+1` are reserved for writeback upon offloading a dual-writeback instruction, where `Xn` denotes the destination register addresss extracted from `instr_data[11:7]`. ### Ternary Operations -The REI optionally supports ISA extensions implementing instructions which use three source operands. +The core-v-xif optionally supports ISA extensions implementing instructions which use three source operands. Ternary instructions must be encoded in the R4-type instruction format defined by the RISC-V specification. ### Hierarchical Interconnect @@ -78,7 +78,7 @@ The order in which offload requests are issued is determined by validity of sour The offloading core may provide internal structures to facilitate multiple instructions to be issued independently, resulting in a pseudo multi-issue pipeline. ### Memory Operations -The REI defines two distinct modes of memory access for external accelerator units. +The core-v-xif defines two distinct modes of memory access for external accelerator units. - The internal mode for accessing memories for extension accelerators is through the offloading core's load/store unit. For executing offloaded memory operations, control over the offloading core's load/store unit is handed to the accelerator unit. Offloaded load and store operations happen directly through the core's memory interface. @@ -86,11 +86,11 @@ The REI defines two distinct modes of memory access for external accelerator uni For this purpose, external mode memory operations are defined. ## Interface Subset Naming Convention -The naming scheme to describe the subset of optional features included in a hardware implementation of the REI implementation comprises the following components. +The naming scheme to describe the subset of optional features included in a hardware implementation of the core-v-xif implementation comprises the following components. | Component | Description | | --------- | ----------- | -| REIv[X] | REI Version [X] of the specification draft | +| core-v-xifv[X] | core-v-xif Version [X] of the specification draft | | XLEN | Base ISA bit width | | T | Support for ternary operations | | D | Support for dual-writeback instructions | @@ -98,5 +98,5 @@ The naming scheme to describe the subset of optional features included in a hard | E | Support for 'external mode' memory operations | The specification version indicator is separated from the rest of the description string by an underscore (`_`). -For example, a hardware implementation of the spec version 0.1 (this version) based on a 32 bit core supporting ternary operations and internal mode memory operations would be described with `REIv0.1_32TM`. +For example, a hardware implementation of the spec version 0.1 (this version) based on a 32 bit core supporting ternary operations and internal mode memory operations would be described with `core-v-xifv0.1_32TM`.