On 09/20/2011 10:11 AM, Xin Tong wrote:
One of the research projects I am doing involves disassembling machine
instructions (such as PTX ) to TCG IR and use the QEMU backends to
generate
PTX, do you mean Parallel Thread Execution, a pseudo-assembly language
for NVIDIA GPUs?
code I need. Basically, I am planning to use QEMU as a dynamic
translators, instead of the full system emulation it is developed for.
I need a way to hack
In TCG, the whole translation has two parts: blocks of target code be rewritten
in tcg ops, a kind of machine independent intermediate notation, then this
notation
be compiled for the host's architecture by TCG.
TCG requires that there be dedicated code written to support every architecture
it is being run on. It also requires that the target instruction translation be
rewritten to take advantage of TCG ops.
As I know, PTX is not machine code but a typed intermediate language, to some
extend,
target dependent(it contains a lots of GPU-specific instructions), and till now
TCG did
not support such target.
If you wanna translate PTX to the code you need, the only idea pops into my
mind is:
maybe you have to write tcg ops code for PTX yourself... :-)
out the piece of QEMU used specifically for dyn translations so that i
do not need a machine image to start the dyn trans parts of QEMU. I am
fairly familiar with the QEMU code base. I would like to hear
suggestions.
Thanks
Xin
--
Lei
