| Message ID | 20170301185602.6873-1-eric@anholt.net (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, 1 Mar 2017 10:56:01 -0800 Eric Anholt <eric@anholt.net> wrote: > The from_cache flag was actually "the BO is invisible to userspace", > so we can repurpose it to just zero out a cached BO and return it to > userspace. > > Improves wall time for a loop of 5 glsl-algebraic-add-add-1 by > -1.44989% +/- 0.862891% (n=28, 1 outlier removed from each that > appeared to be other system noise) > > Note that there's an intel-gpu-tools test to check for the proper > zeroing behavior here, which we continue to pass. > > Signed-off-by: Eric Anholt <eric@anholt.net> Reviewed-by: Boris Brezillon <boris.brezillon@free-electrons.com> > --- > drivers/gpu/drm/vc4/vc4_bo.c | 13 +++++++------ > 1 file changed, 7 insertions(+), 6 deletions(-) > > diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c > index 7abcd9c5dbe2..e5c7aa935b4b 100644 > --- a/drivers/gpu/drm/vc4/vc4_bo.c > +++ b/drivers/gpu/drm/vc4/vc4_bo.c > @@ -211,21 +211,22 @@ struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size) > } > > struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size, > - bool from_cache) > + bool allow_unzeroed) > { > size_t size = roundup(unaligned_size, PAGE_SIZE); > struct vc4_dev *vc4 = to_vc4_dev(dev); > struct drm_gem_cma_object *cma_obj; > + struct vc4_bo *bo; > > if (size == 0) > return ERR_PTR(-EINVAL); > > /* First, try to get a vc4_bo from the kernel BO cache. */ > - if (from_cache) { > - struct vc4_bo *bo = vc4_bo_get_from_cache(dev, size); > - > - if (bo) > - return bo; > + bo = vc4_bo_get_from_cache(dev, size); > + if (bo) { > + if (!allow_unzeroed) > + memset(bo->base.vaddr, 0, bo->base.base.size); > + return bo; > } > > cma_obj = drm_gem_cma_create(dev, size);
diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c index 7abcd9c5dbe2..e5c7aa935b4b 100644 --- a/drivers/gpu/drm/vc4/vc4_bo.c +++ b/drivers/gpu/drm/vc4/vc4_bo.c @@ -211,21 +211,22 @@ struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size) } struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size, - bool from_cache) + bool allow_unzeroed) { size_t size = roundup(unaligned_size, PAGE_SIZE); struct vc4_dev *vc4 = to_vc4_dev(dev); struct drm_gem_cma_object *cma_obj; + struct vc4_bo *bo; if (size == 0) return ERR_PTR(-EINVAL); /* First, try to get a vc4_bo from the kernel BO cache. */ - if (from_cache) { - struct vc4_bo *bo = vc4_bo_get_from_cache(dev, size); - - if (bo) - return bo; + bo = vc4_bo_get_from_cache(dev, size); + if (bo) { + if (!allow_unzeroed) + memset(bo->base.vaddr, 0, bo->base.base.size); + return bo; } cma_obj = drm_gem_cma_create(dev, size);
The from_cache flag was actually "the BO is invisible to userspace", so we can repurpose it to just zero out a cached BO and return it to userspace. Improves wall time for a loop of 5 glsl-algebraic-add-add-1 by -1.44989% +/- 0.862891% (n=28, 1 outlier removed from each that appeared to be other system noise) Note that there's an intel-gpu-tools test to check for the proper zeroing behavior here, which we continue to pass. Signed-off-by: Eric Anholt <eric@anholt.net> --- drivers/gpu/drm/vc4/vc4_bo.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)