Re: Support for amdgpu VM update via CPU on large-bar systems

From: zhoucm1 <david1.zhou-5C7GfCeVMHo@public.gmane.org>
To: "Christian König"
	<deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org>,
	"Kasiviswanathan,
	Harish" <Harish.Kasiviswanathan-5C7GfCeVMHo@public.gmane.org>,
	"amd-gfx-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org"
	<amd-gfx-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org>
Subject: Re: Support for amdgpu VM update via CPU on large-bar systems
Date: Fri, 12 May 2017 16:44:01 +0800	[thread overview]
Message-ID: <591575D1.3020909@amd.com> (raw)
In-Reply-To: <b82fc6cd-c19e-85c2-8628-89c2cb10c5c0-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org>


[-- Attachment #1.1: Type: text/plain, Size: 2628 bytes --]


On 2017年05月12日 16:43, Christian König wrote:
> Am 12.05.2017 um 10:37 schrieb zhoucm1:
>>
>>
>> On 2017年05月12日 16:33, Christian König wrote:
>>> Am 12.05.2017 um 10:25 schrieb zhoucm1:
>>>>
>>>> On 2017年05月10日 05:47, Kasiviswanathan, Harish wrote:
>>>>> Hi,
>>>>>
>>>>> Please review the patch set that supports amdgpu VM update via CPU. This feature provides improved performance for compute (HSA) where mapping / unmapping is carried out (by Kernel) independent of command submissions (done directly by user space). This version doesn't support shadow copy of VM page tables for CPU based update.
>>>> I think your improved performance is from less waiting for cs, 
>>>> generally, SDMA engine updating page table is faster than CPU, 
>>>> otherwise we don't need sdma for updating PT.
>>>> So whether your this improvement proves we have some redundant sync 
>>>> when mapping / unmapping, if yes, we should fix that, then not sure 
>>>> if CPU method is need or not.
>>>
>>> The problem is that the KFD is designed synchronously for page table 
>>> updates. In other words they need to wait for the update to finish 
>>> and that takes time.
>>>
>>> Apart from that your comment is absolutely correct, we found that 
>>> the SDMA is sometimes much faster to do the update than the CPU.
>> If the sdma is faster, even they wait for finish, which time is 
>> shorter than CPU, isn't it? Of course, the precondition is sdma is 
>> exclusive. They can reserve a sdma for PT updating.
>
> No, if I understood Felix numbers correctly the setup and wait time 
> for SDMA is a bit (but not much) longer than doing it with the CPU.
>
> What would really help is to fix the KFD design and work with async 
> page tables updates there as well.
OK, no problem, just curious.

Regards,
David Zhou
>
> Regards,
> Christian.
>
>>
>> Regards,
>> David Zhou
>>>
>>> Regards,
>>> Christian.
>>>
>>>>
>>>> Regards,
>>>> David Zhou
>>>>> Best Regards,
>>>>> Harish
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> amd-gfx mailing list
>>>>> amd-gfx-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org
>>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>
>>>
>>
>>
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>


[-- Attachment #1.2: Type: text/html, Size: 6179 bytes --]

[-- Attachment #2: Type: text/plain, Size: 154 bytes --]

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx