Re: [Intel-xe] [PATCH v4 5/9] drm/xe/xe2: Update chunk size for each iteration of ccs copy

["Ghimiray, Himal Prasad" <himal.prasad.ghimiray@intel.com>]

On 07-12-2023 05:31, Matt Roper wrote:
> On Wed, Dec 06, 2023 at 10:01:22AM +0530, Himal Prasad Ghimiray wrote:
>> In xe2 platform XY_CTRL_SURF_COPY_BLT can handle ccs copy for
>> max of 1024 main surface pages.
>>
>> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>> Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_migrate.c | 34 ++++++++++++++++++++++++++++-----
>>   1 file changed, 29 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
>> index b4dd1b6d78f0..98dca906a023 100644
>> --- a/drivers/gpu/drm/xe/xe_migrate.c
>> +++ b/drivers/gpu/drm/xe/xe_migrate.c
>> @@ -672,11 +672,24 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>>   		u32 update_idx;
>>   		u64 ccs_ofs, ccs_size;
>>   		u32 ccs_pt;
>> +
>>   		bool usm = xe->info.supports_usm;
>> +		u32 avail_pts = NUM_PT_PER_BLIT;
>>   
>>   		src_L0 = xe_migrate_res_sizes(&src_it);
>>   		dst_L0 = xe_migrate_res_sizes(&dst_it);
>>   
>> +		/* In IGFX the XY_CTRL_SURF_COPY_BLT can handle max of 1024
>> +		 * pages. Hence limit the processing size to SZ_4M per
>> +		 * iteration.
>> +		 */
>> +		if (!IS_DGFX(xe) && GRAPHICS_VER(xe) >= 20) {
> Where is the igpu limitation coming from?  The change to expressing copy
> size in terms of pages seems to be a general Xe2 IP change that we'd
> expect all future platforms, both igpu and dgpu to follow.

I added the limitation considering dgfx can have 64 K pages. And then 
limiting size should be decided by

xe_migrate_res_sizes. Will try to make this change more generic in next version.

>
>
>> +			src_L0 = min_t(u64, src_L0, SZ_4M);
>> +			dst_L0 = min_t(u64, dst_L0, SZ_4M);
>> +
>> +			avail_pts = SZ_4M / SZ_2M;
> What does the SZ_2M here represent?  It's not super obvious.
>
> Also, even though you have a comment above, it still might be nicer to
> "show the work" for SZ_4M instead of using a magic number.  E.g.,
>
>          XE_PAGE_SIZE * (FIELD_MAX(XE2_CCS_SIZE_MASK) + 1) / ...
>
>
> Matt
>
>> +		}
>> +
>>   		drm_dbg(&xe->drm, "Pass %u, sizes: %llu & %llu\n",
>>   			pass++, src_L0, dst_L0);
>>   
>> @@ -684,18 +697,18 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m,
>>   
>>   		batch_size += pte_update_size(m, src_is_vram, src, &src_it, &src_L0,
>>   					      &src_L0_ofs, &src_L0_pt, 0, 0,
>> -					      NUM_PT_PER_BLIT);
>> +					      avail_pts);
>>   
>>   		batch_size += pte_update_size(m, dst_is_vram, dst, &dst_it, &src_L0,
>>   					      &dst_L0_ofs, &dst_L0_pt, 0,
>> -					      NUM_PT_PER_BLIT, NUM_PT_PER_BLIT);
>> +					      avail_pts, avail_pts);
>>   
>>   		if (copy_system_ccs) {
>>   			ccs_size = xe_device_ccs_bytes(xe, src_L0);
>>   			batch_size += pte_update_size(m, false, NULL, &ccs_it, &ccs_size,
>>   						      &ccs_ofs, &ccs_pt, 0,
>> -						      2 * NUM_PT_PER_BLIT,
>> -						      NUM_PT_PER_BLIT);
>> +						      2 * avail_pts,
>> +						      avail_pts);
>>   		}
>>   
>>   		/* Add copy commands size here */
>> @@ -923,8 +936,19 @@ struct dma_fence *xe_migrate_clear(struct xe_migrate *m,
>>   		struct xe_bb *bb;
>>   		u32 batch_size, update_idx;
>>   		bool usm = xe->info.supports_usm;
>> +		u32 avail_pts = NUM_PT_PER_BLIT;
>>   
>>   		clear_L0 = xe_migrate_res_sizes(&src_it);
>> +
>> +		/* In IGFX the XY_CTRL_SURF_COPY_BLT can handle max of 1024
>> +		 * pages. Hence limit the processing size to SZ_4M per
>> +		 * iteration.
>> +		 */
>> +		if (!IS_DGFX(xe) && GRAPHICS_VER(xe) >= 20) {
>> +			clear_L0 = min_t(u64, clear_L0, SZ_4M);
>> +			avail_pts = SZ_4M / SZ_2M;
>> +		}
>> +
>>   		drm_dbg(&xe->drm, "Pass %u, size: %llu\n", pass++, clear_L0);
>>   
>>   		/* Calculate final sizes and batch size.. */
>> @@ -932,7 +956,7 @@ struct dma_fence *xe_migrate_clear(struct xe_migrate *m,
>>   			pte_update_size(m, clear_vram, src, &src_it,
>>   					&clear_L0, &clear_L0_ofs, &clear_L0_pt,
>>   					emit_clear_cmd_len(gt), 0,
>> -					NUM_PT_PER_BLIT);
>> +					avail_pts);
>>   		if (xe_device_has_flat_ccs(xe) && clear_vram)
>>   			batch_size += EMIT_COPY_CCS_DW;
>>   
>> -- 
>> 2.25.1
>>