Marcin Slusarz <[email protected]> writes: > add new relocation flag which allows to halving number of relocation > entries on >= nv50 cards > I suspect that this change is unnecessary. If you're actually seeing a performance improvement (I guess you are?), it's very likely to be masking the actual root of the problem -- last time I checked the nv50 gallium pipe driver was rather careless about reloc re-submission (it re-submits the whole reloc chain once for every draw call instead of once for every pushbuf flush), and nouveau_reloc_emit() and friends are very unoptimized.
I'm betting that using HILO or not will make pretty much no difference once those two places are fixed :). > Signed-off-by: Marcin Slusarz <[email protected]> > --- > drivers/gpu/drm/nouveau/nouveau_gem.c | 22 +++++++++++++++------- > include/drm/nouveau_drm.h | 1 + > 2 files changed, 16 insertions(+), 7 deletions(-) > > diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c > b/drivers/gpu/drm/nouveau/nouveau_gem.c > index 9a1fdcf..2e21412 100644 > --- a/drivers/gpu/drm/nouveau/nouveau_gem.c > +++ b/drivers/gpu/drm/nouveau/nouveau_gem.c > @@ -507,6 +507,7 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev, > struct drm_nouveau_gem_pushbuf_bo *b; > struct nouveau_bo *nvbo; > uint32_t data; > + uint32_t data2; > > if (unlikely(r->bo_index > req->nr_buffers)) { > NV_ERROR(dev, "reloc bo index invalid\n"); > @@ -544,17 +545,22 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev, > > if (r->flags & NOUVEAU_GEM_RELOC_LOW) > data = b->presumed.offset + r->data; > - else > - if (r->flags & NOUVEAU_GEM_RELOC_HIGH) > + else if (r->flags & NOUVEAU_GEM_RELOC_HIGH) > data = (b->presumed.offset + r->data) >> 32; > - else > + else if (r->flags & NOUVEAU_GEM_RELOC_HILO) { > + data = (b->presumed.offset + r->data) >> 32; > + data2 = b->presumed.offset + r->data; > + } else > data = r->data; > > if (r->flags & NOUVEAU_GEM_RELOC_OR) { > - if (b->presumed.domain == NOUVEAU_GEM_DOMAIN_GART) > - data |= r->tor; > - else > - data |= r->vor; > + if (b->presumed.domain == NOUVEAU_GEM_DOMAIN_GART) { > + data |= r->tor; > + data2 |= r->tor; > + } else { > + data |= r->vor; > + data2 |= r->vor; > + } > } > > spin_lock(&nvbo->bo.lock); > @@ -566,6 +572,8 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev, > } > > nouveau_bo_wr32(nvbo, r->reloc_bo_offset >> 2, data); > + if (r->flags & NOUVEAU_GEM_RELOC_HILO) > + nouveau_bo_wr32(nvbo, (r->reloc_bo_offset >> 2) + 1, > data2); > } > > kfree(reloc); > diff --git a/include/drm/nouveau_drm.h b/include/drm/nouveau_drm.h > index bc5590b..126ea8a 100644 > --- a/include/drm/nouveau_drm.h > +++ b/include/drm/nouveau_drm.h > @@ -137,6 +137,7 @@ struct drm_nouveau_gem_pushbuf_bo { > #define NOUVEAU_GEM_RELOC_LOW (1 << 0) > #define NOUVEAU_GEM_RELOC_HIGH (1 << 1) > #define NOUVEAU_GEM_RELOC_OR (1 << 2) > +#define NOUVEAU_GEM_RELOC_HILO (1 << 3) > #define NOUVEAU_GEM_MAX_RELOCS 1024 > struct drm_nouveau_gem_pushbuf_reloc { > uint32_t reloc_bo_index;
pgpn658v5svsE.pgp
Description: PGP signature
_______________________________________________ Nouveau mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/nouveau
