| content |
Contents |
Implementation
The usual layout convention is that a simple array is read out vertically. That is, a single word line, which runs horizontally, causes a row of bit cells to put their data on bit lines, which run vertically. Sense amps, which convert low-swing read bitlines into full-swing logic levels, are usually at the bottom (by convention). Larger register files are then sometimes constructed by tiling mirrored and rotated simple arrays.
Register files have one word line per entry per port, one bit line per bit of width per read port, and two bit lines per bit of width per write port. Each bit cell also has a Vdd and Vss. Therefore, the wire pitch area increases as the square of the number of ports, and the transistor area increases linearly. At some point, it may be smaller and/or faster to have multiple redundant register files, with smaller numbers of read ports, than a single register file with all the read ports. The MIPS R8000's integer unit, for example, had a 9 read 4 write port 32 entry 64-bit register file implemented in a 0.35-µm process, which could be seen when looking at the chip from arm's length.
Decoder
- The decoder is often broken into predecoder and decoder proper.
- The decoder is a series of AND gates that drive word lines.
- There is one decoder per read or write port. If the array has four read and two write ports, for example, it has 6 word lines per bit cell in the array, and six AND gates per row in the decoder. Note that the decoder has to be pitch matched to the array, which forces those AND gates to be wide and short
Array
The basic scheme for a bit cell:
- State is stored in pair of inverters
- Data is read out by nmos transistor to a bit line.
- Data is written by shorting one side or the other to ground through a two-nmos stack.
- So: read ports take one transistor per bit cell, write ports take four!
Many optimizations are possible:
- Sharing lines between cells, for example, Vdd and Vss.
- Read bit lines are often precharged to something between Vdd and Vss.
- Read bit lines often swing only a fraction of the way to Vdd or Vss. A sense amplifier converts this small-swing signal into a full logic level. Small swing signals are faster because the bit line has little drive but a great deal of parasitic capacitance.
- Write bit lines may be braided, so that they couple equally to the nearby read bitlines. Because write bitlines are full swing, they can cause significant disturbances on read bitlines.
- If Vdd is a horizontal line, it can be switched off, by yet another decoder, if any of the write ports are writing that line during that cycle. This optimization increases the speed of the write.
Microarchitecture
Most register files make no special provision to prevent multiple write ports from writing the same entry simultaneously. Instead, the instruction scheduling hardware ensures that only one instruction in any particular cycle writes a particular entry. If multiple instructions targeting the same register are issued, all but one have their write enables turned off.
The crossed inverters take some finite time to settle after a write operation, during which a read operation will either take longer or return garbage. It is common to have bypass multiplexors that bypass written data to the read ports when a simultaneous read and write to the same entry is commanded. These bypass multiplexors are often just part of a larger bypass network that forwards results that have not yet been committed between functional units.
The register file is usually pitch matched to the datapath that it serves. Pitch matching avoids having the many busses passing over the datapath turn corners, which would use a lot of area. But since every unit must have the same bit pitch, every unit in the datapath ends up with the bit pitch forced by the widest unit, which can waste area in the other units. Register files, because they have two wires per bit per write port, and because all the bit lines must contact the silicon at every bit cell, can often set the pitch of a datapath.
Area can sometimes be saved, on machines with multiple units in a datapath, by having two datapaths side-by-side, each of which has smaller bit pitch than a single datapath would have. This case usually forces multiple copies of a register file, one for each datapath.
The DEC Alpha EV-6, for instance, had two copies of the integer register file, and took an extra cycle to propagate data between the two. The issue logic tried to reduce the number of operations forwarding data between the two. The MIPS R8000 floating-point unit had two copies of the floating-point register file, each with four write and four read ports, and wrote both copies at the same time.
Processors that do register renaming can arrange for each functional unit to write to a subset of the physical register file. This arrangement can eliminate the need for multiple write ports per bit cell, for a large savings in area. The resulting register file, effectively a stack of register files with single write ports, then benefits from replication and subsetting the read ports. At the limit, this technique would place a stack of 1-write, 2-read regfiles at the inputs to each functional unit. Since regfiles with a small number of ports are often dominated by transistor area, it is best not to push this technique to this limit, but it is useful all the same.
The SPARC ISA defines register windows, in which the 5-bit architectural names of the registers actually point into a window on a much larger register file, with hundreds of entries. Implementing multiported register files with hundreds of entries requires a lot of area. The register window slides by 16 registers when moved, so that each architectural register name can refer to only a small number of registers in the larger array, e.g. architectural register r20 can only refer to physical registers #20, #36, #52, #68, #84, #100, #116, if there are just seven windows in the physical file.
To save area, some SPARC implementations implement a 32-entry register file, in which each cell has seven "bits". Only one is read and writeable through the external ports, but the contents of the bits can be rotated. A rotation accomplishes in a single cycle a movement of the register window. Because most of the wires accomplishing the state movement are local, tremendous bandwidth is possible with little power.
This same technique is used in the R10000 register renaming mapping file, which stores a 6-bit virtual register number for each of the physical registers. In the renaming file, the renaming state is checkpointed whenever a branch is taken, so that when a branch is detected to be mispredicted, the old renaming state can be recovered in a single cycle. (See Register renaming.)
External links
- Register File Design Considerations in Dynamically Scheduled Processors - Farkas, Jouppi, Chow - 1995 (obsolete?)
- Register file design considerations in dynamically scheduled processors - Farkas, Jouppi, Chow - 1995
See also
