Improving language shootout results

From Free Pascal wiki
Revision as of 14:37, 24 March 2012 by Vincent (talk | contribs) (Text replace - "delphi>" to "syntaxhighlight>")
Jump to navigationJump to search

About

The computer language shootout (http://shootout.alioth.debian.org/) is a realistically flawed benchmarking system which compares many languages on illogical basis. See their homepage for more info.

Goals

Our goals are to be on the highest possible positions. The requirements to reach them can be categorized into two.

1: Optimizations on the assembler level. (core devel work)

2: Optimizations of the benchmarks. (junior devel work)

Way to go

It was decided that to keep memory requirements low it's best to use val() and str() and never include SysUtils unless really required. 32 bit integers should be used when possible to increase speed as the benchmarking machine is x86(athlon 1Gb, 256Mb ram). Use of inline and records instead of classes can also improve performance considerably, but additional testing and comparing with C and other languages should be the main factor. Special note: inlining of recursive functions is possible and sometimes increases speed.

Another thing to use is {$implicitexceptions off} to keep rtl hindering speed. Dropping down to pchar level in tight loops can be beneficial. Use also only native sized integers and floats if possible.

Benchmarks notes

SSE2 and comparions

Consider the following code compiled with -Cfsse2:

const 
  limit=4.0;
var 
  x : double;
begin
  x := 1.73;
  if x < limit then
    x := x + x;
end.

For the condition x<limit, no sse2 code will be generated, because the compiler assumes the const is an extended and therefore needs to perform the comparison with the x87 FPU unit.

const 
  limit = double(4.0);

... also doesn't work. However

const 
  limit: double = double(4.0);

does.

Binary trees

The binary trees benchmark can be improved using a TNonFreePooledMemoryManager available in fpc 2.1.1. That way the allocation of nodes occurs in blocks and all de-allocation happens all at once for all nodes of the tree. Look for an example here. Such an allocator probably won't be allowed though.

Mandelbrot

The mandelbrot benchmark can be improved with fpc 2.1.1 using the sse2 instructions. Note that sse2 instructions are volatile and the results of a calculation will be put back on the stack, if the same body contains a procedure call. Therefore, a sse2 calculation needs to be moved to a separate (nested) procedure. The mandelbrot benchmark would probably benefit from SSA in loops, because without it calculations are performed in a scratch register and the result moved to the final register. SSA will prevent this. At the moment fpc 2.1.1 supports SSA only in simple linear flow control without loops and branches.

The latest versions of the competition gain a lot (nearly a factor of 2) by using 128 bit xmm registers for cr, ci, tr, and ti and instructions to do 2 calculations at the same time. I do not see a way to do this with fpc. Please prove me wrong :-)

N-body

The n-body benchmark can be improved with fpc 2.1.1 using the sse2 instructions and factoring out common factors. Compiling with -Cfsse2 and fpc 2.0.4 gives an internal error.

Shootout criticism

Any page about benchmarking should discuss the use of the benchmarks. FPC does relatively well overall, but is relatively weak in its category.

Criticising the shootout is not difficult, just start with the "Flawed Benchmarks" page.

I mention a few points below.

  • The main problem is that the applications are relatively short and highly computational. This has a lot of consequences:
    • In general they favour lower level languages, well, that is not to Pascal's disadvantage.
    • these benchmarks also favour compilers that aggressively optimize tight heavy calculating loops and have code profiling. FPC is not in that league, but, like Delphi, more geared to overall application performance, and not numeric or visual computational work with lots of tight loops. This is also why frameworks like Java and .NET don't score _that_ bad. Simply because the JIT can do good work for these tight loops, and take really advantage of its profiling optimizations, and auto-selecting of the exact CPU architecture. Simply being able to use a system specific move is already a serious advantage in benchmarks.
  • Language usability and overall speed of the development process is not measured at all. (How would one do that?)
    • Systems with a heavy RTL are punished. Both in size and startup time. However for actually getting work done, as long as it is not extreme, a full RTL is nice. (Most of the programs run for seconds - startup time must be really long.)
    • ease of debugging is not measured. (How would one do that?)
    • Languages that have a lower and higher level mode (e.g. FPC with its Pascal and Delphi levels) typically choose the most beneficial mode (e.g. lowest possible version for lowlevel stuff, high level mode if linecount is important), even if that is not the one typically used for a certain job.
    • (not a problem of the shootout but its use) People are not used to reading boxplot charts with median and quartiles.
  • Currently the benchmark game uses a quad-core machine so programs that use all 4 cores obviously have the advantage.
  • Rules can change.