This Xbar connects the largest components in the design; the cores
and the L2 banks. We already have a full buffer on the core side.
However, the valid path going to the L2 comes back as a ready path.
Putting a flow Q also on the outputs of the l1tol2 cuts this path
in half at no cost to IPC.
Fundamental new features:
* Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces.
* Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile.
* Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile.
* Defined RocketCoreParams: All the parameters that can be varied per-core.
* Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes.
* Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created.
* Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little.
* Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support.
Additional changes that got rolled in along the way:
* rocket: Fix critical path through BTB for I$ index bits > pgIdxBits
* coreplex: tiles connected via :=*
* groundtest: updated to use TileParams
* tilelink: cache cork requirements are relaxed to allow more cacheless masters
We want to keep the banks split in the outer SoC if there is an L3.
Furthermore, each channel might go to different memory subsystems,
like DDR/HMC/Zero, from rocketchip.
* [rocket] Refactor Tile into cake pattern with traits
* [rocket] cacheDataBits &etc in HasCoreParameters
* [rocket] pass TLEdgeOut implicitly rather than relying on val edge in HasCoreParameters
* [rocket] frontend and icache now diplomatic
* [rocket] file name capitalization
* [rocket] re-add hook for inserting externally-defined Cores
* [rocket] add FPUCoreIO
* [groundtest] move TL1 Config instances to where they are used
* [unittest] remove legacy unit tests
* [groundtest] remove legacy device tests
* Configs: use a uniform syntax without Match exceptions
The old style of specifying Configs used total functions. The only way to
indicate that a key was not matched was to throw an exception. Not only was
this a performance concern, but it also caused confusing error messages
whenever you had a match failure from a lookup within a lookup. The
exception could get handled by an outer-lookup that then reported the wrong
key as missing.
This bug is ancient. I don't understand how it never mattered before.
Anyway, in processors with a custom CacheBlockBytes, this value is wrong!
The symptom is that TL1 components end up missing high address bits.
This causes, for example, a system to jump to 0 instead of RAM.
I don't understand how this very serious bug did not cause problems before.
1. Bundles be created after base class Module constructor runs
2. Bundles must be created before Module(...) runs
Solution: pass a bundle constructor to the cake base class
Require the constructor to take a parameter so people don't use it by
accident; they should get a type error.
Consistently name all the cake arguments with an _io, _coreplex, _outer,
so that they don't shadow the base class variables you should be using.
A lot of utility code was just being imported willy-nilly from one
package to another. This moves the common code into util to make things
more sensible. The code moved were
* The AsyncQueue and AsyncDecoupledCrossing from junctions.
* All of the code in rocket's util.scala
* The BlackBox asynchronous reset registers from uncore.tilelink2
* The implicit definitions from uncore.util
This would allow, for instance, putting the coreplex on a separate clock
domain and crossing the IOs over through asynchronous queues.
The ExampleMultiClockTop* classes are removed since they no longer fit
into the class hierarchy.
We have a handful of TileLink-related helper objects
(wrappers, unwrappers, width adapters, and enqueuers). Previously, using
them could be error-prone, because you had to make sure the implicit
parameters they took in had the same TLId as the TileLinkIO bundles
passed in as inputs. This is rather silly, we should just use the
parameters in the bundle.
A chip's power-up sequence, or awake-from-sleep sequence, may wish to
set the reset PC based upon dynamic properties, e.g., the settings of
external pins. Support this by passing the reset vector to the Coreplex.
ExampleTop simply hard-wires the reset vector, as was the case before.
Additionally, allow MTVEC to *not* be reset. In most cases, including
riscv-tests, pk, and bbl, overriding MTVEC is one of the first things
that the boot sequence does. So the reset value is superfluous.
Unfortunately, I had to touch a lot of code, which weren't quite possible to split up into multiple commits.
This commit gets rid of the "extra" infrastructure to add periphery devices into Top.
They fit in the same part of the address space as DRAM would be, and
are coherent (because they are not cacheable).
They are currently limited to single cores without DRAM. We intend
to lift both restrictions, probably when we add support for
heterogeneous tiles.
This usually shouldn't be used in Tiles that are meant to be P&R'd once
and multiply instantiated, as their RTL would no longer be homogeneous.
However, it is useful for conditionalizing RTL generation for
heterogeneous tiles.
