Re: Lua 5 and lua2c

["Leandro Candido" <enclle@...>]

Hello Jeff,

    As I don't work with this platform/processor, I will only trust you ;-).

                                                    God bless you,


Leandro.

----- Original Message -----
From: "Jeff Koftinoff" <jeffk@jdkoftinoff.com>
To: "Lua list" <lua@bazar2.conectiva.com.br>
Sent: Thursday, December 04, 2003 4:57 PM
Subject: Re: Lua 5 and lua2c


> On Wed, 3 Dec 2003, Leandro Candido wrote:
>
> > Hello Florian,
> >
> >     The answer, in my opinion, is not so if the functions have simple
and
> > little code, and yes, always gain speed, if functions with big/normal
code.
> > I'm researching on this subject, I'm trying to create a "lua2c"
translator,
> > to translate lua code either to lua C API calls (lua_pushnumber, etc...)
and
> > to mimics the behavior of lvm, but the code isn't useful yet, as I said,
I'm
> > researching yet, learning the opcodes, etc...
> >
> >     Of course, native code always beat interpreted code/bytecode.
> >
> >                                                                 God
bless
> > you,
> >
> >
> > Leandro.
> >
>
>
> I have been trying out lua specifically because i am using a platform
> where I believe that the lua byte-codes would execute much, much faster
> than the equivalent native C code.
>
> The processor is a DSP - specifically the Texas Instruments TMS320C6701
>
> It has on-chip, quad ported 256 bit wide memory for VLIW execution of
> program opcodes, very fast.  It also has on-chip memory for data storage
> which is mostly being used for digital audio channels.  We also have 32
> megabytes of SDRAM hooked up to it.
>
> Problem is that the SDRAM runs really slowly as the DSP does not do any
> memory bursting unless it is using DMA.  how slow? 20 to 40
> processor clock cycles to read one 32 bit word from SDRAM.
>
> In one processor clock cycle, the DSP could load and execute up to 8
> instructions in parallel.  So this means that executing code from SDRAM is
> up to 320 times slower.
>
> As an experiment I took lua-5.0 and hacked it into the dsp's PMEM (Program
> Memory).  The idea is that lua's code will run fast from PMEM - a whole
> bunch of code is triggered by a much smaller bytecode in sdram.  I want to
> see if I can write a caching mechanism so that the bytecodes could be
> DMA's into the chip's DMEM temporarily when it is being executed.  I've
> not completed this experment (other things are taking priority), but I
> expect that using lua I can get on average 10 times faster system
> performance compared to native C/C++ code.
>
> jeff
>
>