Re: [Intel-gfx] [PATCH 1/2] drm/i915/sseu: Don't overallocate subslice storage

[Matt Roper <matthew.d.roper@intel.com>]

On Fri, Mar 11, 2022 at 11:00:09AM -0800, Lucas De Marchi wrote:
> On Thu, Mar 10, 2022 at 10:15:42PM -0800, Matt Roper wrote:
> > Xe_HP removed "slice" as a first-class unit in the hardware design.
> > Instead we now have a single pool of subslices (which are now referred
> > to as "DSS") that different hardware units have different ways of
> > grouping ("compute slices," "geometry slices," etc.).  For the purposes
> > of topology representation, we treat Xe_HP-based platforms as having a
> > single slice that contains all of the platform's DSS.  There's no need
> > to allocate storage space for (max legacy slices * max dss); let's
> > update some of our macros to minimize the storage requirement for sseu
> > topology.  We'll also document some of the constants to make it a little
> > bit more clear what they represent.
> > 
> > Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
> > ---
> > drivers/gpu/drm/i915/gt/intel_engine_types.h |  2 +-
> > drivers/gpu/drm/i915/gt/intel_sseu.h         | 47 +++++++++++++++-----
> > 2 files changed, 36 insertions(+), 13 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/gt/intel_engine_types.h b/drivers/gpu/drm/i915/gt/intel_engine_types.h
> > index 4fbf45a74ec0..f9e246004bc0 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_engine_types.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_engine_types.h
> > @@ -645,7 +645,7 @@ intel_engine_has_relative_mmio(const struct intel_engine_cs * const engine)
> > 
> > #define for_each_instdone_gslice_dss_xehp(dev_priv_, sseu_, iter_, gslice_, dss_) \
> > 	for ((iter_) = 0, (gslice_) = 0, (dss_) = 0; \
> > -	     (iter_) < GEN_MAX_SUBSLICES; \
> > +	     (iter_) < GEN_SS_MASK_SIZE; \
> > 	     (iter_)++, (gslice_) = (iter_) / GEN_DSS_PER_GSLICE, \
> > 	     (dss_) = (iter_) % GEN_DSS_PER_GSLICE) \
> > 		for_each_if(intel_sseu_has_subslice((sseu_), 0, (iter_)))
> > diff --git a/drivers/gpu/drm/i915/gt/intel_sseu.h b/drivers/gpu/drm/i915/gt/intel_sseu.h
> > index 8a79cd8eaab4..4f59eadbb61a 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_sseu.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_sseu.h
> > @@ -15,26 +15,49 @@ struct drm_i915_private;
> > struct intel_gt;
> > struct drm_printer;
> > 
> > -#define GEN_MAX_SLICES		(3) /* SKL upper bound */
> > -#define GEN_MAX_SUBSLICES	(32) /* XEHPSDV upper bound */
> > -#define GEN_SSEU_STRIDE(max_entries) DIV_ROUND_UP(max_entries, BITS_PER_BYTE)
> > -#define GEN_MAX_SUBSLICE_STRIDE GEN_SSEU_STRIDE(GEN_MAX_SUBSLICES)
> > -#define GEN_MAX_EUS		(16) /* TGL upper bound */
> > -#define GEN_MAX_EU_STRIDE GEN_SSEU_STRIDE(GEN_MAX_EUS)
> > +/*
> > + * Maximum number of legacy slices.  Legacy slices no longer exist starting on
> > + * Xe_HP ("gslices," "cslices," etc. on Xe_HP and beyond are a different
> > + * concept and are not expressed through fusing).
> > + */
> > +#define GEN_MAX_LEGACY_SLICES		3
> > +
> > +/*
> > + * Maximum number of subslices that can exist within a legacy slice.  This is
> > + * only relevant to pre-Xe_HP platforms (Xe_HP and beyond use the GEN_MAX_DSS
> > + * value below).
> > + */
> > +#define GEN_MAX_LEGACY_SUBSLICES	6
> 
> instead of calling the old legacy, maybe just add the XEHP_ prefix to
> the new ones?

Maybe a "HSW_" prefix on the old ones would be better?  People still use
the termm 'subslice' in casual discussion when talking about DSS, so I
want to somehow distinguish that what we're talking about here is a
different, older design than we have on modern platforms.


Matt

> 
> > +
> > +/* Maximum number of DSS on newer platforms (Xe_HP and beyond). */
> > +#define GEN_MAX_DSS			32
> > +
> > +/* Maximum number of EUs that can exist within a subslice or DSS. */
> > +#define GEN_MAX_EUS_PER_SS		16
> > +
> > +#define MAX(a, b)			((a) > (b) ? (a) : (b))
> 
> what's worse, include kernel.h in another header file or redefine MAX
> everywhere? Re-defining it in headers we risk situations in which the
> include order may create warnings about defining it in multiple places.
> 
> 
> > +
> > +/* The maximum number of bits needed to express each subslice/DSS independently */
> > +#define GEN_SS_MASK_SIZE		MAX(GEN_MAX_DSS, \
> > +					    GEN_MAX_LEGACY_SLICES * GEN_MAX_LEGACY_SUBSLICES)
> > +
> > +#define GEN_SSEU_STRIDE(max_entries)	DIV_ROUND_UP(max_entries, BITS_PER_BYTE)
> > +#define GEN_MAX_SUBSLICE_STRIDE		GEN_SSEU_STRIDE(GEN_SS_MASK_SIZE)
> > +#define GEN_MAX_EU_STRIDE		GEN_SSEU_STRIDE(GEN_MAX_EUS_PER_SS)
> > 
> > #define GEN_DSS_PER_GSLICE	4
> > #define GEN_DSS_PER_CSLICE	8
> > #define GEN_DSS_PER_MSLICE	8
> > 
> > -#define GEN_MAX_GSLICES		(GEN_MAX_SUBSLICES / GEN_DSS_PER_GSLICE)
> > -#define GEN_MAX_CSLICES		(GEN_MAX_SUBSLICES / GEN_DSS_PER_CSLICE)
> > +#define GEN_MAX_GSLICES		(GEN_MAX_DSS / GEN_DSS_PER_GSLICE)
> > +#define GEN_MAX_CSLICES		(GEN_MAX_DSS / GEN_DSS_PER_CSLICE)
> > 
> > struct sseu_dev_info {
> > 	u8 slice_mask;
> > -	u8 subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
> > -	u8 geometry_subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
> > -	u8 compute_subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
> > -	u8 eu_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICES * GEN_MAX_EU_STRIDE];
> > +	u8 subslice_mask[GEN_SS_MASK_SIZE];
> > +	u8 geometry_subslice_mask[GEN_SS_MASK_SIZE];
> > +	u8 compute_subslice_mask[GEN_SS_MASK_SIZE];
> > +	u8 eu_mask[GEN_SS_MASK_SIZE * GEN_MAX_EU_STRIDE];
> 
> 
> Aside the minor things above, everything look correct.
> 
> Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
> 
> thanks
> Lucas De Marchi
> 
> > 	u16 eu_total;
> > 	u8 eu_per_subslice;
> > 	u8 min_eu_in_pool;
> > -- 
> > 2.34.1
> > 

-- 
Matt Roper
Graphics Software Engineer
VTT-OSGC Platform Enablement
Intel Corporation
(916) 356-2795