Re: Version after Version

"J French" <erewhon@xxxxxxxxxx> wrote in message
news:43636f02.109826154@xxxxxxxxxxxxxxxxxxxxxxx


> I doubt that Maarten would do that
> - but if they are 4 byte integers there would be a /loss/ using a 64
> bit chip


Baloney.

Even MMX (64-bits, as you are well aware) does two 32-bit additions at a
time.
If you have a hundred 32-bit additions, you can do them twice as fast using
MMX, by doing 50 times two additions at a time.
I don't see what so hard to understand about this.

> - and 1000 8 byte numbers is a bit silly if they don't need to be
> Int64s

The whole point is that you can put TWO of those in a 64-bit register!

> That is only the case if you have everything 64 bit aligned
> - which it ain't

That's your own fault then.
It's trivial to align arrays on 64-bit boundaries.
If you have 100 32-bit values (50 64-bit values), you loose a few bytes,
worst-case, aligning them on a 64-bit boundary.
Many compilers & memory managers have at least the option to align data on
64-bit boundaries.
If they don't move to another compiler! Don't say 64-bit's is not fast when
your real problem is your compiler or the way YOU align YOUR data!


> - and you'll probably find that it will seldom be, as that is just
> gobbling (doubling) memory usage.

Utter nonsense.
Of course you PACK two 32-bit values in ONE 64-bit variable!

> Don't count on it
> - compression and decompression works at the Bit level

So? What's your point?
If I can do manipulations on 64 bits at a time without the need to load 32
more bits, I save 50% of the load/store/manipulation time.


> LZW is the until recently proprietory algorithm of Uinisys

So?
Does that mean it is not faster when 64 bits is taken advantage of, in the
implementation of the algorith?
ALL compression/decompression becomes faster when using 64-bits algo's, it's
irrelevant who had some patent on some algo.

> The ZIP format uses Huffman compression

So? WHo cares? What does that have to do with a speedup using 64-bit
algorithms?
Stick to the topic.
I can post the biography of Mister Lempel or Mr. Ziv or Mr. Welsh or their
dog but it's beside the point.


> The 'Z' in LZW is, I believe, nothing to do with the Z in ZIP

So? What else is new?

> No it does not
> - some things benefit, huge memory addressing and copying large chunks
> of memory- but most things don't benefit

Huge memory adressing is nice but it's a nice benefit.
I am wholly unexited by it. The exiting aspect of 64-bits processing is the
enormous speedup it gives in almost all algoritms, when rewritten for 64-bit
processing.



.

Relevant Pages

  • Re: Turing Machines and Physical Computation
    ... rather than adumbrated in that 1936 Turing paper On Computable ... ... any actual computer has finite memory, ... computer will run out of memory for some huge calculation. ... It is generally assumed that an algorithm must satisfy the following ...
    (comp.theory)
  • Re: Turing Machines and Physical Computation
    ... rather than adumbrated in that 1936 Turing paper On Computable ... ... any actual computer has finite memory, ... computer will run out of memory for some huge calculation. ... It is generally assumed that an algorithm must satisfy the following ...
    (sci.math)
  • Solving the memory issue of lock-free algorithms
    ... I read a lot about lock-free algorithm and tried to find an "elegant" ... solution to the memory problem that the algorithms poses. ... LFStackNode *Next; ... xor eax, eax ...
    (comp.programming.threads)
  • Re: Sans memory barrier native thread communication upon AMD64?
    ... producer/consumer or a distributed algorithm w/ mutual ... that may be garbage collected independent of any other native thread. ... microthread allocates its environment (including dynamic stacks) from the ... system provides a thread safe memory allocator. ...
    (comp.programming.threads)
  • Re: How much does it take to execute MMX instruction?
    ... Your code is using MMX ... That means you need temporary registers for that. ... A bit better is if you know the memory reading pattern, ... At end of the day you will much more likely to be memory bandwith ...
    (comp.lang.asm.x86)