-----Original Message-----
From: Giovanni Parodi [mailto:g...@yahoo.it]
Sent: Monday, September 20, 2004 12:18 PM
To: c...@yahoogroups.com
Subject: [c6x] Differences between _nassert() and DATA_MEM_BANK

Hello everybody, I have a little problem understanding the way by which it's possible to use the instructions
#pragma DATA_MEM_BANK
and
_nassert
In effect in the examples pg_dotp_flt and pg_vecsum proposed by TI it seems that it's necessary to use _nassert with short int and int numbers and DATA_MEM_BANK with floating point number.
However, I don't understand that. The pragma instruction is relative to the preprocessor so when the compiler find that (and it find that before allocating the variables, that's true?), it take the trouble to allocate the variable in the correct way (aligned to memory).
However _nassert doesn't seem to be a preprocessor instruction (I think that all the "instruction" that start with # are of the preprocessor), and so I'm not able to understand how this instruction can say to the system "take the variable x and allocate it aligning to word boundary". It seems as this instruction say simply "this short variable is memory alig! ned", but how can be this always true?
What's the difference between DATA_MEM_BANK and _nassert?
Please help me, thanks a lot (and please sorry my English sometimes too similar to Italian language:-))

---

Scopri Mister Yahoo! - il fantatorneo sul calcio di Yahoo! Sport'

_____________________________________

Giovanni-

>If that's ok, I'm missing that: how is it possible to use _nassert
>without DATA_MEM_BANK.

Well you can! As you just proved in your code!

_nassert in yr code is simply asserting(as the name implies) that the last
the two bits of the address is 00. Then it becomes re-assuring to the
compiler that the data is aligned on a 32 bit word boundary. Period. Nothing
else to really do with DATA_MEM_BANK pragma. It is the DATA_ALIGN pragma
that deals with *forcing* data alignment in C & .align directive in linear
asm/asm.
(For alignment related issues you really should be behind the DATA_ALIGN
pragma)

Also you need to look up the default aligment rules of the c compiler for
various data types-For EX: shorts are by default aligned to half-word
boundaries but arrays of shorts are always aligned to word boundaries, ints
are by default aligned to word boundaries...

So, there you go, in your code the data is(in all probability) already
naturally word aligned by the compiler and you are merely asserting that
fact to the compiler through the _nassert

Ps: if you _nassert the data is word aligned and then if the data is not
really aligned, then the behaviour of your program is undefined!

--Bhooshan

If I have a situation like this:

void main()
{
int x[4]={1,2,3,4};
func(x);
}

void func(int y)
{
int i=0;
_nassert(((int)(y)&0x3)==0)
for(i=0;i<4;i++)
y[i]=y[i]*y[i];
}

I think that I cannot be sure that the compiler allocate the variables
aligning them to word boundaries, so when I use func() function it's
possible that the variable y isn't word aligned and so the result of the
func() cannot be predetermined.
In effect may be I'm missing something: I have the following doubt: when
the linker runs, it always tries to word align the arrays? If the
response is yes I understand the use of _nassert ('cause if I use this
instruction with a array, I can be sure that the array will always be
aligned unless there's not enough space to word align it), but I don't
understand its usefulness, because I will always have a array word
aligned (if it's possible).
However, if the linker doesn't automatically word align variables, I
think that using _nassert without DATA_MEM_BANK (that said to align to a
specific memory bank) is a Russian roulette, but this approach is used
in the help files of Texas Instruments, so it's impossible that it's wrong.
May be that my questions are a bit stupid, sorry for that, and thanks to
everybody will help me.

Mody, Mihir wrote:

>Hi Bhoosan,
>
>You are correct. The number of memory banks (4, 8 etc) and its existence
depends on particular device you are using. You may to refer particular
device manual to find out the details regarding this.
>
>Regards,
>Mihir Mody,
>Multimedia codecs group,
>Texas Instruments India, Ltd,
>Email : m...@ti.com
>Phone : +91-80-25099307
>
>-----Original Message-----
>From: Bhooshan iyer [mailto:b...@hotmail.com]
>Sent: Monday, September 20, 2004 3:30 PM
>To: Mody, Mihir; g...@yahoo.it; c...@yahoogroups.com
>Subject: RE: [c6x] Differences between _nassert() and DATA_MEM_BANK>
>Mihir->
>>In case of c6x, the memory consists of different banks. E.g. all
addresses
>>multiple of 8 forms bank >0. addresss like 8* K +1 forms bank 1, so on.
>>
>>
>
>Correct me if am wrong: You say C6x's memory consists of Bank Structure,
but
>i think it should read "C60x" device's memory consists of Bank Structure.
>
>As far as i know C6211 and C6711 doesnt have Banked Memory Structure. I
>think the L2 memory of C6211 and C6711 are Dual Access Memory, So doesnt
>matter which banks the datas are in, you are still guarenteed two loads
per
>clock without any stalls.
>
>But the 0x devices dont have Dual Access RAM's so if you want to achieve
two
>loads per clock, it is not guaranteed unless make sure those data points
are
>stored in two different banks.(The assumption is that you will always want
>to do something like x[n]*y[n] in DSP and hence will need two data reads
per
>clock.Though you have two data units(D1&D2) it is still not guarenteed
that
>you can achieve two loads per clock but by making sure x is in a
*different
>bank* compared to y,you make two loads possible in a cycle in banked
memory
>structures like the one available in C6201)
>
>So, in conclusion i dont think #pragma DATA_MEM_BANK is of any use to
C6711
>& C6211 variations but it is highly useful to C6201 and C6701 variations.
>
>--Bhooshan
>
>_________________________________________________________________
>All the news that matters. All the gossip from home.
>http://www.msn.co.in/NRI/ Specially for NRIs!>
>
>_____________________________________