Quoting Chris Wilson (2019-06-11 20:41:59)
> Quoting Mika Kuoppala (2019-06-11 18:27:24)
> > struct i915_page_table {
> > struct i915_page_dma base;
> > - atomic_t used_ptes;
> > + atomic_t used;
> > };
> >
> > struct i915_page_directory {
> > struct i915_page_dma base;
> > -
> > - struct i915_page_table *page_table[I915_PDES]; /* PDEs */
> > - atomic_t used_pdes;
> > - spinlock_t lock;
> > -};
> > -
> > -struct i915_page_directory_pointer {
> > - struct i915_page_dma base;
> > - struct i915_page_directory **page_directory;
> > - atomic_t used_pdpes;
> > - spinlock_t lock;
> > -};
> > -
> > -struct i915_pml4 {
> > - struct i915_page_dma base;
> > - struct i915_page_directory_pointer *pdps[GEN8_PML4ES_PER_PML4];
> > + atomic_t used;
> > spinlock_t lock;
> > + void *entry[0];
> > };
>
> And always (albeit with a single bsw discrepancy) 512. At the very least
> you can alias a fixed sized variant over the top to remove the extra
> pointer chasing you added.
I would float your make bsw behave identically patch. In the grand
scheme of things, no one will ever notice that bsw alone saved a couple
of pages per GTT.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
