[Intel-gfx] [PATCH 1/2] drm/i915/selftests: Verify the LRC register layout between init and HW

Chris Wilson chris at chris-wilson.co.uk
Tue Sep 24 11:00:55 UTC 2019


Quoting Mika Kuoppala (2019-09-24 11:21:38)
> Chris Wilson <chris at chris-wilson.co.uk> writes:
> > +static u32 *set_offsets(u32 *regs,
> > +                     const u8 *data,
> > +                     const struct intel_engine_cs *engine)
> > +#define NOP(x) (BIT(7) | (x))
> > +#define LRI(count, flags) ((flags) << 6 | (count))
> > +#define POSTED BIT(0)
> > +#define REG(x) (((x) >> 2) | BUILD_BUG_ON_ZERO(x >= 0x200))
> > +#define REG16(x) \
> > +     (((x) >> 9) | BIT(7) | BUILD_BUG_ON_ZERO(x >= 0x10000)), \
> > +     (((x) >> 2) & 0x7f)
> 
> I am still not sure if the actual saving are worth the complexity.
> 
> > +#define END() 0
> > +{
> > +     const u32 base = engine->mmio_base;
> > +
> > +     while (*data) {
> > +             u8 count, flags;
> > +
> > +             if (*data & BIT(7)) { /* skip */
> > +                     regs += *data++ & ~BIT(7);
> > +                     continue;
> > +             }
> > +
> > +             count = *data & 0x3f;
> > +             flags = *data >> 6;
> > +             data++;
> > +
> > +             *regs = MI_LOAD_REGISTER_IMM(count);
> > +             if (flags & POSTED)
> > +                     *regs |= MI_LRI_FORCE_POSTED;
> > +             if (INTEL_GEN(engine->i915) >= 11)
> > +                     *regs |= MI_LRI_CS_MMIO;
> > +             regs++;
> > +
> > +             GEM_BUG_ON(!count);
> > +             do {
> > +                     u32 offset = 0;
> > +                     u8 v;
> > +
> > +                     do {
> > +                             v = *data++;
> > +                             offset <<= 7;
> > +                             offset |= v & ~BIT(7);
> > +                     } while (v & BIT(7));
> 
> ...but perhaps this amount of extra can be tolerated.
> 
> Did you check how this would play out with just REG being wide enough?

Function                                     old     new   delta
gen9_xcs_offsets                             122     145     +23
gen12_xcs_offsets                            136     157     +21
gen11_rcs_offsets                             44      60     +16
gen8_rcs_offsets                              41      55     +14
gen12_rcs_offsets                             47      60     +13
gen8_xcs_offsets                              40      51     +11
set_offsets.isra                             215     179     -36
-Chris


More information about the Intel-gfx mailing list