TTM limits system memory allocations to 1/2 system memory for regular
processes, or 3/4 for privileged processes.
For compute we opted to make the GART bigger than system memory to allow
for fragmentation. However, we argued that you don't really need GART
for most buffers anyways, since they are never accessed in VMID 0. They
need to be mapped in some VM page table, but never in the GART table. So
I think the long-term goal should be to remove the need to allocate GART
addresses for system memory BOs. Then you could get away with a quite
small GART size.
Regards,
Felix
On 16-07-21 08:04 AM, Christian König wrote:
>>
>> we can add people in if necessary here. so the initial though of this
>> change is with latest big memory GPU released, the remote memory goes
>> very large, like Polaris10.
>>
>> According to the windows WDDM, remote memory size is half of the
>> system memory, I don't know what was discussed before, Please let me
>> know if there're use case indeed need that big of remote memory.
>>
> Well that is at least something that makes sense.
>
> Previous discussions where always settled around making GART bigger
> than VRAM and not the other way around.
>
> Anyway I think a prerequisite to this is actually allowing half of
> system memory to be allocated for GART and I'm not sure if that is
> currently the case. TTM has a limitation for that as well, but I off
> hand don't remember how high much it was.
>
> I think we should use the same value which TTM uses for the upper
> limit here.
>
> Regards,
> Christian.
>
> Am 21.07.2016 um 13:56 schrieb Wang, Qingqing:
>>
>> alright, I didn't realize you're not in that thread.
>>
>> we can add people in if necessary here. so the initial though of this
>> change is with latest big memory GPU released, the remote memory goes
>> very large, like Polaris10.
>>
>> According to the windows WDDM, remote memory size is half of the
>> system memory, I don't know what was discussed before, Please let me
>> know if there're use case indeed need that big of remote memory.
>>
>> ------------------------------------------------------------------------
>> *å??件人:* Christian König <deathsimple at vodafone.de>
>> *å??é??æ?¶é?´:* 2016å¹´7æ??21æ?¥ 19:48:17
>> *�件人:* Wang, Qingqing; amd-gfx at lists.freedesktop.org
>> *主é¢?:* Re: ç?å¤?: ç?å¤?: [PATCH] drm/amdgpu: adjust gtt memory size
>>
>> Am 21.07.2016 um 13:15 schrieb Wang, Qingqing:
>>>
>>> Actually that discussion was held long ago internally but we never
>>> got a conclusion on this IIRC. So we should probably continue the
>>> discussion on this thread now.
>>>
>>> -there is a thread for this topic yesterday, some people are not in
>>> amd-gfx mail list, we'd better discuss in the internal thread.
>>>
>> No, exactly for this reason we have the public mailing list. I for
>> example wasn't part of the internal thread either.
>>
>> Christian.
>>
>>> ------------------------------------------------------------------------
>>> *å??件人:* Christian König <deathsimple at vodafone.de>
>>> *å??é??æ?¶é?´:* 2016å¹´7æ??21æ?¥ 17:48:41
>>> *�件人:* Wang, Qingqing; amd-gfx at lists.freedesktop.org
>>> *主é¢?:* Re: ç?å¤?: [PATCH] drm/amdgpu: adjust gtt memory size
>>>
>>> Am 21.07.2016 um 11:16 schrieb Wang, Qingqing:
>>>>
>>>> This doesn't work on 32bit systems since totalram is a long (IIRC) and
>>>> so this will certainly overflow.
>>>>
>>>> -- can you be more specific, how could the overflow happen?
>>>>
>>>
>>> Both values are 32bit on a 32bit system, so as soon as you have more
>>> than 4GB installed this will overflow.
>>>
>>> You need to cast the long to a 64bit value to avoid that.
>>>
>>>>
>>>> Additional if I remember correctly we didn't ended the discussion on
>>>> what to do here with a conclusion.
>>>>
>>>> -- ok, since it happens in anther mail list, we should stop talking
>>>> about it here, you can send out your ideas there.
>>>>
>>>
>>> Actually that discussion was held long ago internally but we never
>>> got a conclusion on this IIRC. So we should probably continue the
>>> discussion on this thread now.
>>>
>>> Regards,
>>> Christian.
>>>
>>>> ------------------------------------------------------------------------
>>>> *å??件人:* amd-gfx <amd-gfx-bounces at lists.freedesktop.org> 代表
>>>> Christian König <deathsimple at vodafone.de>
>>>> *å??é??æ?¶é?´:* 2016å¹´7æ??21æ?¥ 16:53:05
>>>> *�件人:* Wang, Qingqing; amd-gfx at lists.freedesktop.org
>>>> *主�:* Re: [PATCH] drm/amdgpu: adjust gtt memory size
>>>>
>>>> Am 21.07.2016 um 09:46 schrieb Ken Wang:
>>>> > Change-Id: If00d5b97ba9e30f9b7f68fdfc134a0f8b3ad2add
>>>> > Signed-off-by: Ken Wang <Qingqing.Wang at amd.com>
>>>> > ---
>>>> > drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 7 ++++++-
>>>> > drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 7 ++++++-
>>>> > 2 files changed, 12 insertions(+), 2 deletions(-)
>>>> >
>>>> > diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>>> b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>>> > index a3b6048..0b095d5 100644
>>>> > --- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>>> > +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c
>>>> > @@ -336,6 +336,11 @@ static int gmc_v7_0_mc_init(struct
>>>> amdgpu_device *adev)
>>>> > {
>>>> > u32 tmp;
>>>> > int chansize, numchan;
>>>> > + u64 sysmem_size;
>>>> > + struct sysinfo si;
>>>> > +
>>>> > + si_meminfo(&si);
>>>> > + sysmem_size = si.totalram * si.mem_unit;
>>>>
>>>> This doesn't work on 32bit systems since totalram is a long (IIRC) and
>>>> so this will certainly overflow.
>>>>
>>>> Additional if I remember correctly we didn't ended the discussion on
>>>> what to do here with a conclusion.
>>>>
>>>> Regards,
>>>> Christian.
>>>>
>>>> >
>>>> > /* Get VRAM informations */
>>>> > tmp = RREG32(mmMC_ARB_RAMCFG);
>>>> > @@ -392,7 +397,7 @@ static int gmc_v7_0_mc_init(struct
>>>> amdgpu_device *adev)
>>>> > * size equal to the 1024 or vram, whichever is larger.
>>>> > */
>>>> > if (amdgpu_gart_size == -1)
>>>> > - adev->mc.gtt_size = max((1024ULL << 20),
>>>> adev->mc.mc_vram_size);
>>>> > + adev->mc.gtt_size = max((1024ULL << 20),
>>>> min(adev->mc.mc_vram_size, sysmem_size/2));
>>>> > else
>>>> > adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
>>>> >
>>>> > diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>>> b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>>> > index 02b6872..790bf7a 100644
>>>> > --- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>>> > +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c
>>>> > @@ -413,6 +413,11 @@ static int gmc_v8_0_mc_init(struct
>>>> amdgpu_device *adev)
>>>> > {
>>>> > u32 tmp;
>>>> > int chansize, numchan;
>>>> > + u64 sysmem_size;
>>>> > + struct sysinfo si;
>>>> > +
>>>> > + si_meminfo(&si);
>>>> > + sysmem_size = si.totalram * si.mem_unit;
>>>> >
>>>> > /* Get VRAM informations */
>>>> > tmp = RREG32(mmMC_ARB_RAMCFG);
>>>> > @@ -469,7 +474,7 @@ static int gmc_v8_0_mc_init(struct
>>>> amdgpu_device *adev)
>>>> > * size equal to the 1024 or vram, whichever is larger.
>>>> > */
>>>> > if (amdgpu_gart_size == -1)
>>>> > - adev->mc.gtt_size = max((1024ULL << 20),
>>>> adev->mc.mc_vram_size);
>>>> > + adev->mc.gtt_size = max((1024ULL << 20),
>>>> min(adev->mc.mc_vram_size, sysmem_size/2));
>>>> > else
>>>> > adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
>>>> >
>>>>
>>>>
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>
>>>
>>
>
>
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20160721/0b989835/attachment-0001.html>
