FP matrix multiplication
- raw README.md
- raw benchmark.4th
- raw matmul.4th
- raw package.4th
- raw runbenchs
- raw test.4th
This package provides floating-point matrix multiplication. It runs on Forth-94 and Forth-2012 systems with the floating-point wordset.
to get the word
matmulr ( a b c n1 n2 n3 -- )
MATMULR computes C=AxB, where A, B, and C are matrixes: where A has n1 rows and n2 columns, B has n2 rows and n3 columns, and C has n1 rows and n3 columns. Matrixes are stored in row-major order, i.e., the elements of one row are adjacent, whereas the items of a column are separated by all the items in the first row.
To test whether this works as intended with your Forth system, type
create matmul-verbose include matmul.4th include test.4th
Details (irrelevant for most usage)
MATMUL calls a word FAXPY-NOSTRIDE, and the library contains several implementations of this word, and automatically selects the one expected to give the best performance. If you want to know which implementation is used, you can CREATE MATMUL-VERBOSE before including matmul.4th. You can also determine which implementation is used by defining a constant FAXPY-NOSTRIDE-VARIANT. The values for this constant are as follows:
unrolling no yes 1 use FAXPY (a primitive in Gforth) 4 2 separate-stack (Forth-2012) implementation 5 3 Forth-94 implementation by Joel Rees and Anton Ertl 15 13 Forth-94 implementation by pahihu
The Forth-94 implementations work on systems with separate FP stacks as well as on systems with a combined FP and data stack with any number of cells per FP item.
Use as Benchmark
For use as benchmark, there is benchmark.4th, which performs one
500x500 matrix multiplication. There is also the script runbenchs; by
default, runbenchs will run the benchmark on gforth-fast, vfxlin (with
various FP packages), and sf if available (and for VFX, if %vfxpath%
works), and it will use
perf (with a fallback to
time) for timing.
You can influence these choices and more by setting environment
variables before calling runbenchs. E.g., if you want to benchmark
iforth only, you can do (on a bourne-compatible shell):
SF=iforth GFORTH=xx VFXLIN=xx ./runbenchs
The environment variables used are:
TIMING command used for measuring, e.g., `time` RUN file used for running benchmark (RUN=test.4th for testing) GFORTH how to invoke Gforth (default: gforth-fast) SF how to invoke SwiftForth (default: sf) VFXLIN How to invoke VFX (default: vfxlin) VFXHOME where the VfxLin directory is (default: %vfxpath%) VFXLIB where the FP packages can be found (default: $VFXHOME/Lib) VFXX86 where the x86-specific FP packages can be found
The defaults are reasonable, but you may need to define VFXHOME yourself.
You can also use this script to test all the different variants on several systems with
TIMING=command RUN=test.4th ./runbenchs
This may not be the best-performing matrix multiplication written in Forth and certainly not outside of Forth, but it has the advantage of just working (at least that's the intention). It's also not too badly performing: Gforth performs 0.8 FLOP/cycle on a Core i5-6600K using the faxpy primitive, and ideally Forth compilers will generate as good code for FAXPY-NOSTRIDE as that for the FAXPY primitive.
Here are some results (using unrolled versions):
cycles Core i5-6600K Forth-2012 Forth-94 Forth-94 primitive Rees/Ertl pahihu faxpy *1046M 1202M 1517M vfxlin Ndp387.fth 2220M 2945M 2764M *316M gforth-fast *4844M 5067M 5017M SwiftForth HW stack, no WAIT
The default variants are marked with "*".