github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-23 13:27:29 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

Documented split and join.

Browse source
This commit is contained in:
cdkersey 2014-09-09 03:19:05 -04:00
parent 56aaff1f87
commit dde43648fd
1 changed files with 12 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

13
doc/harp_iset.tex
View file

@ -132,7 +132,8 @@ The bit fields in the instruction encodings depend heavily on this quality.
30 "skep" 1REG 31 "reti" NONE 32 "tlbrm" 1REG
33 "itof" 2REG 34 "ftoi" 2REG 35 "fadd" 3REG
36 "fsub" 3REG 37 "fmul" 3REG 38 "fdiv" 3REG
39 "fneg" 2REG
39 "fneg" 2REG 3a "wspawn" 2REG 3b "split" NONE
3c "join" NONE
\end{verbatim}
\subsection{Word Encoding}
@ -346,6 +347,8 @@ format, which can be fixed point or floating point.
\texttt{jalis} \%link, \%n, \textsc{\#RelDest}&Jump and link immediate, spawning N active lanes.\\
\texttt{jalrs} \%link, \%n, \%dest&Jump and link indirect, spawning N active lanes.\\
\texttt{jmprt} \%addr&Jump indirect, terminating execution on all but a single lane.\\
\texttt{split}&Control flow diverge.\\
\texttt{join}&Control flow reconverge.\\
\end{tabular}
\end{center}
@ -412,6 +415,14 @@ The current response to this is to trap to the operating system (interrupt numbe
The \texttt{clone}, \texttt{jalis}, \texttt{jalrs}, and \texttt{jmprt} instructions form the basis of SIMD context control in the HARP instruction set.
Context is created using \texttt{clone}, the waiting threads are spawned using \texttt{jalrs} or \texttt{jalis}, ``jump-and-link immediate/register and spawn'', and finally the parallel section returns using \texttt{jmprt}, ``jump register and terminate'', best thought of as ``return and terminate.''
There are times when a control flow operation will need to be predicated, going one direction on some lanes and the other direction on other lanes.
For this, the HARP instruction set provides the \texttt{split} and \texttt{join} instructions.
When a predicated \texttt{split} is first encountered, only the lanes for which the \texttt{split}'s predicate are true are allowed to continue.
The other lanes are masked out until the corresponding \texttt{join} is encountered.
The first time \texttt{join} is reached, control flow returns to the instruction following the corresponding \texttt{split} with the set of masked-out lanes complemented.
The second time the same \texttt{join} is reached, control flow falls through and the original lane mask is restored.
A hardware stack is maintained to keep track of nested \texttt{split}s.
\section{Default I/O Devices}
The emulator currently only supports a single I/O device, simple console I/O.
Writing to the address \texttt{0x800...0} (an address with its MSB set and all other bits cleared) causes text to be written to the display.