[PATCH 4/6] drm/xe/xe_migrate: Handle migration logic for xe2+ dgfx
Nirmoy Das
nirmoy.das at linux.intel.com
Tue Jul 16 08:12:01 UTC 2024
Hi Akshata,
On 7/15/2024 10:56 PM, Jahagirdar, Akshata wrote:
>
>
> On 7/12/2024 5:55 AM, Nirmoy Das wrote:
>>
>> Hi Akshata,
>>
>> On 7/12/2024 8:39 AM, Akshata Jahagirdar wrote:
>>> From: "Jahagirdar, Akshata"<akshata.jahagirdar at intel.com>
>>>
>>> During eviction (vram->sysmem), we use the mapping from compressed -> uncompressed.
>>> During restore (sysmem->vram), we need the mapping from uncompressed -> uncompressed.
>>> Handle logic for selecting the compressed identity map for eviction,
>>> and selecting uncompressed map for restore operations.
>>>
>>> Signed-off-by: Akshata Jahagirdar<akshata.jahagirdar at intel.com>
>>> Reviewed-by: Matthew Auld<matthew.auld at intel.com>
>>> Reviewed-by: Himal Prasad Ghimiray<himal.prasad.ghimiray at intel.com>
>>> ---
>>> drivers/gpu/drm/xe/xe_migrate.c | 7 ++++---
>>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
>>> index 3b8a334fe08f..1eee6be24423 100644
>>> --- a/drivers/gpu/drm/xe/xe_migrate.c
>>> +++ b/drivers/gpu/drm/xe/xe_migrate.c
>>> @@ -715,7 +715,7 @@ static u32 xe_migrate_ccs_copy(struct xe_migrate *m,
>>> struct xe_gt *gt = m->tile->primary_gt;
>>> u32 flush_flags = 0;
>>>
>>> - if (xe_device_has_flat_ccs(gt_to_xe(gt)) && !copy_ccs && dst_is_indirect) {
>>> + if (xe_device_needs_ccs_emit(gt_to_xe(gt)) && !copy_ccs && dst_is_indirect) {
>>
>>
>> Do we need to call xe_migrate_ccs_copy() when
>> xe_device_needs_ccs_emit() is true ? If not then move the check for
>> xe_device_needs_ccs_emit() to xe_migrate_copy()
>>
I see that I've confused you here. I meant "Do we need to call
xe_migrate_ccs_copy() when xe_device_needs_ccs_emit() is *not* true ?"
>> to ensure that xe_migrate_ccs_copy() is not even called on such
>> platforms.
>>
>> Regards,
>> Nirmoy
>
> xe_device_needs_ccs_emit() is true for any non-BMG platform that has
> flat-ccs enabled.
> In such cases, we do need to call xe_migrate_ccs_copy() while clearing
> any ccs metadata in vram.
>
So let's use xe_device_needs_ccs_emit() in the upper layer to decide
whether to call xe_migrate_ccs_copy(), instead of checking
xe_device_needs_ccs_emit() within an if-else block of xe_migrate_ccs_copy()
Either:
if (xe_device_needs_ccs_emit())
xe_migrate_ccs_copy(...)
Or even*"
*
xe_migrate_ccs_copy() {
if ( !xe_device_needs_ccs_emit() )
return;
..........
}
Regards,
Nirmoy
>
> BR,
> Akshata
>
>>> /*
>>> * If the src is already in vram, then it should already
>>> * have been cleared by us, or has been populated by the
>>> @@ -791,6 +791,7 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>>> bool copy_ccs = xe_device_has_flat_ccs(xe) &&
>>> xe_bo_needs_ccs_pages(src_bo) && xe_bo_needs_ccs_pages(dst_bo);
>>> bool copy_system_ccs = copy_ccs && (!src_is_vram || !dst_is_vram);
>>> + bool use_comp_pat = GRAPHICS_VER(xe) >= 20 && IS_DGFX(xe) && src_is_vram && !dst_is_vram;
>>>
>>> /* Copying CCS between two different BOs is not supported yet. */
>>> if (XE_WARN_ON(copy_ccs && src_bo != dst_bo))
>>> @@ -833,7 +834,7 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>>>
>>> src_L0 = min(src_L0, dst_L0);
>>>
>>> - batch_size += pte_update_size(m, src_is_vram, false, src, &src_it, &src_L0,
>>> + batch_size += pte_update_size(m, src_is_vram, use_comp_pat, src, &src_it, &src_L0,
>>> &src_L0_ofs, &src_L0_pt, 0, 0,
>>> avail_pts);
>>>
>>> @@ -852,7 +853,7 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>>>
>>> /* Add copy commands size here */
>>> batch_size += ((copy_only_ccs) ? 0 : EMIT_COPY_DW) +
>>> - ((xe_device_has_flat_ccs(xe) ? EMIT_COPY_CCS_DW : 0));
>>> + ((xe_device_needs_ccs_emit(xe) ? EMIT_COPY_CCS_DW : 0));
>>>
>>> bb = xe_bb_new(gt, batch_size, usm);
>>> if (IS_ERR(bb)) {
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-xe/attachments/20240716/29c63297/attachment.htm>
More information about the Intel-xe
mailing list