[Intel-xe] [PATCH 00/10] Add OA functionality to Xe
Dixit, Ashutosh
ashutosh.dixit at intel.com
Tue Aug 8 02:41:04 UTC 2023
On Mon, 07 Aug 2023 18:31:49 -0700, Ashutosh Dixit wrote:
>
Hi Umesh/Lionel,
> This patchset is the initial port of i915 perf/OA functionality to the Xe
> driver. It has been tested against the following IGT patch:
>
> https://patchwork.freedesktop.org/series/121082/
>
> The following features in i915 have not been ported and will be added (as
> new patches) if/as they are needed:
>
> * Inline batch submission on stream exec_queue/hw_engine
> * NOA wait
About these two in the "not-yet-ported" feature list for Xe: In
i915 when a stream engine is specified, requests are submitted "inline" to
that engine whenever the OA stream is opened or reconfigured. This feature
is at present not ported to Xe (see xe_oa_submit_bb()).
I have a couple of questions about this "inline submission":
* Is inline submission a hard requirement? That is, in Xe can UMD's say
(completely outside OA uapi) fence till a segment of the workload
completes, then open/reconfigure the OA stream and then continue? If
UMD's are able to do this we don't need inline submission in Xe. If UMD's
cannot do this please explain why not.
* If inline submission is indeed required, can we do away with NOA wait (as
implemented in daed3e44396d ("drm/i915/perf: implement active wait for
noa configurations")) or implement NOA wait differently in Xe? E.g. can
we expose a fence in the OA uapi and UMD's would be required to wait on
this fence/sync-object before continuing workload submissions after any
OA stream open/reconfiguration.
The reason I am brining this up is because in i915 there are a couple of
scary bugs around this which are still unresolved. E.g.:
https://gitlab.freedesktop.org/drm/intel/-/issues/8746
https://gitlab.freedesktop.org/drm/intel/-/issues/8190
There is therefore reason to think if this functionality can be done
differently in Xe (not to say these bugs will be carried over to Xe).
Thanks.
--
Ashutosh
> * GuC ctx id (guc_sw_ctx_id)
> * CTX_R_PWR_CLK_STATE/GEN8_R_PWR_CLK_STATE
> * hold_preemption (DRM_XE_OA_PROP_HOLD_PREEMPTION)
> * sseu_config (DRM_XE_OA_PROP_GLOBAL_SSEU)
> * Override gucrc (override_gucrc_mode)
> * MTL bios_c6_setup
> * ratelimits
> * compat ioctl
>
> I am providing the following additional HAX patch (not part of this series)
> to help review these patches:
>
> https://patchwork.freedesktop.org/patch/551683/?series=120100&rev=4
>
> The commit message in the above patch explains how it can be useful for
> reviewing this series.
>
> Test-with: 20230808010017.37819-1-ashutosh.dixit at intel.com
>
> v2: Fix build
> v3: Rebase, due to s/xe_engine/xe_exec_queue/
>
> Ashutosh Dixit (10):
> drm/xe/oa: Introduce OA uapi
> drm/xe/oa: Add OA types
> drm/xe/oa: Add registers and GPU commands used by OA
> drm/xe/oa: Module init/exit and probe/remove
> drm/xe/oa: Add/remove config ioctl's
> drm/xe/oa: Start implementing OA stream open ioctl
> drm/xe/oa: OA stream initialization
> drm/xe/oa: Expose OA stream fd
> drm/xe/oa: Read file_operation
> drm/xe/oa: Implement queries
>
> drivers/gpu/drm/xe/Makefile | 1 +
> drivers/gpu/drm/xe/regs/xe_engine_regs.h | 5 +
> drivers/gpu/drm/xe/regs/xe_gpu_commands.h | 27 +
> drivers/gpu/drm/xe/regs/xe_oa_regs.h | 173 ++
> drivers/gpu/drm/xe/xe_device.c | 16 +
> drivers/gpu/drm/xe/xe_device_types.h | 4 +
> drivers/gpu/drm/xe/xe_gt_types.h | 4 +
> drivers/gpu/drm/xe/xe_hw_engine_types.h | 2 +
> drivers/gpu/drm/xe/xe_module.c | 5 +
> drivers/gpu/drm/xe/xe_oa.c | 2356 +++++++++++++++++++++
> drivers/gpu/drm/xe/xe_oa.h | 28 +
> drivers/gpu/drm/xe/xe_oa_types.h | 295 +++
> drivers/gpu/drm/xe/xe_query.c | 5 +-
> include/uapi/drm/xe_drm.h | 257 ++-
> 14 files changed, 3176 insertions(+), 2 deletions(-)
> create mode 100644 drivers/gpu/drm/xe/regs/xe_oa_regs.h
> create mode 100644 drivers/gpu/drm/xe/xe_oa.c
> create mode 100644 drivers/gpu/drm/xe/xe_oa.h
> create mode 100644 drivers/gpu/drm/xe/xe_oa_types.h
>
> --
> 2.41.0
>
More information about the Intel-xe
mailing list