Hi Kaijie,

On Sat, Mar 19, 2016 at 9:01 PM, 冯凯杰 <[email protected]> wrote:

> Dear Joel,
>   Thank you for your response.
>   Except the coherence messages, will CPU and GPU caches transfer data
> that is for actual computing(e.g. CPU cache hold a matrix and can GPU cache
> directly get that
> matrix through caches transfer rather than using shared memory)?
>

Yes, sharing can/will occur if your benchmark is set up to share memory
virtual addresses between CPU and GPU. You can see examples of such sharing
in the benchmarks directory: benchmarks/rodinia-nocopy.

  Joel


> -----原始邮件-----
> *发件人:*"Joel Hestness" <[email protected]>
> *发送时间:*2016-03-19 00:18:59 (星期六)
> *收件人:* "冯凯杰" <[email protected]>, "gem5-gpu developers" <
> [email protected]>
> *抄送:*
> *主题:* Re: Cache Coherence Quiz About Gem5-GPU
>
>
> Hi Kaijie,
>   I'm not exactly sure what you're asking here. A primary aim of gem5-gpu
> is to support the study of heterogeneous processor cache systems, so all of
> the gem5-gpu coherence protocols allow sharing data between CPU and GPU
> caches. To keep the caches coherent, some messages are sent between them.
> Also, depending on your simulated system configuration (e.g. gem5-gpu
> command line parameter: --access-host-pagetable), there are cases where CPU
> and GPU communicate directly through shared memory rather than through DMA
> (the GPUCopyEngine).
>
>   Are you asking something more specific here?
>
>   Thanks,
>   Joel
>
>
>
>
> On Fri, Mar 18, 2016 at 2:57 AM, 冯凯杰 <[email protected]> wrote:
>
>> Dear Joel,
>>   Recently my team and I are concentrating on researching of Gem5-GPU and
>> we found some strange situations while tracing the GPU memory access
>> requests from Shader-lsq to DRAM-ctrl.
>>   Some GPU ST requests which causes the miss of GPU L1$ and L2$ will be
>> matched in CPU L1$. That is sounded a little strange and can not match the
>> structure of Gem5-GPU in which only L2$ are connected by NoC.
>>   We used a simple program to do experiment and DPRINTF all the
>> information of requests in RubySlicc and RubyCache. Then we traced the
>> address of the requests and deduced the data path. Here are two pictures
>> attached to this mail which can haply describe the strange situation we
>> found.
>>   Is Gem5-GPU actually supports that kind of access model or it is just a
>> display bug(we also found that some requests from CPU that is matched in
>> L1$ will still be
>> transferred to DRAM and causes a DRAM response)?
>>   And we really want to know whether a cache can get data from other
>> cache directly? If a GPU request is missed in GPU L1$ and L2$, will Ruby
>> get the data from DRAM or CPU
>> cache hierarchy(if CPU just owns a copy of that data)?
>>   Thank you for your reading and sincerely wish your response.
>>
>> NWPU School of Computing
>> Kaijie Feng
>>
>>
>>
>
>
>
> --
>   Joel Hestness
>   PhD Candidate, Computer Architecture
>   Dept. of Computer Science, University of Wisconsin - Madison
>   http://pages.cs.wisc.edu/~hestness/
>
>


-- 
  Joel Hestness
  PhD Candidate, Computer Architecture
  Dept. of Computer Science, University of Wisconsin - Madison
  http://pages.cs.wisc.edu/~hestness/
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev

Reply via email to