[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Rep:Re: [f-cpu] registers



hello,

Nicolas Boulay wrote:

>i maid a misatake when i speak about waste.
>
no : it's part of the whole equation. If some part of the die is underused,
then it means that it can be removed or the same function is already
implemented somewhere else. Then the die can be smaller, the power
consumption and the price can be reduced, and the cost/performance
ratio enhanced.

>The true point is performance. We can't access more than 64 registers
>because of the instruction word size.
>
On the other extreme, we could easily make a "move machine" (TTA) with 
64K registers ;-P
Some computers with 32-bit instruction words like MMIX or SX-6 have a 4-byte
organisation (8 bits for opcode, 8 for each operand) and 256 registers. 
I have already
explored the world of 128-register machines (leaving 11 bits for the 
opcode) but this was
rejected. Fortunately there are more than 32 registers and we can still 
make very
flexible opcodes.

Increasing the instruction word size creates another world of problem.
But we currently use it because it's the best compromise yet. Maybe other
"extensions" will be used for a superscalar core ?...

> How using more register (i
>remember you that's the main point to decrease memory bandwith pressure) ?
>
may i remember you that you don't have to convince me on this ?
The real problem is "how"... and at what cost.

>So the answer of the sparc designer is to ADD an other register bank.
>They choose to put it for fp numbers.
>  
>
This problem has been taken into account since the beginning of FC0,
because in 1999, "classical" RISC machines had already reached their limits.
However, the solutions they used were not always very good and only
moved the real problems around (that is : adding a new register set does not
change the fact that in the beginning, there were not enough registers).

>The true killer point is all the change in the register set to use it in
>fcpu. I admit it could be a hard stuff. 
>  
>
do you think we don't have enough registers ?

>An other proposal for making Whygee cry : use only 16 256 bits vector
>register. So you leave 2 bits to direct access each of the 64 bits chunk
>and you will gain a lot of flexibility to access them (compare to now)
>ok it's too few compare to the pipeline size...
>
good point...

>[ok je sors ;p] 
>
welcome to the english f-cpu bouchot...
n'oublie pas d'ailleurs de participer à Da Manual Bouchot, ce soir...
et encore bravo pour le troll : très réussi !

YG


*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/