From: Eric Anholt Subject: Re: [PATCH] intel: Add more intermediate sizes of cache buckets between powers of 2. Date: Mon, 07 Jun 2010 10:35:38 -0700 Message-ID: <87wrua4ufp.fsf@pollan.anholt.net> References: <1275696983-15198-1-git-send-email-eric@anholt.net> <38414.115.186.195.130.1275882845.squirrel@webmail.exetel.com.au> Mime-Version: 1.0 Content-Type: multipart/mixed; boundary="===============0817282666==" Return-path: In-Reply-To: <38414.115.186.195.130.1275882845.squirrel@webmail.exetel.com.au> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Mime-version: 1.0 Sender: intel-gfx-bounces+gcfxdi-intel-gfx=m.gmane.org@lists.freedesktop.org Errors-To: intel-gfx-bounces+gcfxdi-intel-gfx=m.gmane.org@lists.freedesktop.org To: Robert Lowery Cc: intel-gfx@lists.freedesktop.org List-Id: intel-gfx@lists.freedesktop.org --===============0817282666== Content-Type: multipart/signed; boundary="=-=-="; micalg=pgp-sha1; protocol="application/pgp-signature" --=-=-= Content-Transfer-Encoding: quoted-printable On Mon, 7 Jun 2010 13:54:05 +1000 (EST), "Robert Lowery" wrote: > > We had two cases recently where the rounding to powers of two hurt > > badly: 4:2:0 YUV HD video frames would round up from 2.2MB to 4MB, > > Urban Terror was hitting aperture size limitations. Mipmap trees for > > power of two sizes will land right in the middle between two cache > > buckets. > > > > By giving a few more sizes between powers of two, Urban Terror on my > > 945 ends up consuming 207MB of GEM objects instead of 272MB. > > --- > > intel/intel_bufmgr_gem.c | 64 > > +++++++++++++++++++++++++++++++++++----------- > > 1 files changed, 49 insertions(+), 15 deletions(-) > > > > diff --git a/intel/intel_bufmgr_gem.c b/intel/intel_bufmgr_gem.c > > index b76fd7e..c3e189e 100644 > > --- a/intel/intel_bufmgr_gem.c > > +++ b/intel/intel_bufmgr_gem.c > > @@ -66,6 +66,8 @@ > > fprintf(stderr, __VA_ARGS__); \ > > } while (0) > > > > +#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0])) > > + > > typedef struct _drm_intel_bo_gem drm_intel_bo_gem; > > > > struct drm_intel_gem_bo_bucket { > > @@ -73,10 +75,6 @@ struct drm_intel_gem_bo_bucket { > > unsigned long size; > > }; > > > > -/* Only cache objects up to 64MB. Bigger than that, and the rounding = of > > the > > - * size makes many operations fail that wouldn't otherwise. > > - */ > > -#define DRM_INTEL_GEM_BO_BUCKETS 14 > > typedef struct _drm_intel_bufmgr_gem { > > drm_intel_bufmgr bufmgr; > > > > @@ -93,7 +91,8 @@ typedef struct _drm_intel_bufmgr_gem { > > int exec_count; > > > > /** Array of lists of cached gem objects of power-of-two sizes */ > > - struct drm_intel_gem_bo_bucket cache_bucket[DRM_INTEL_GEM_BO_BUCKETS]; > > + struct drm_intel_gem_bo_bucket cache_bucket[14 * 4]; > > + int num_buckets; > > > > uint64_t gtt_size; > > int available_fences; > > @@ -285,7 +284,7 @@ drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_g= em > > *bufmgr_gem, > > { > > int i; > > > > - for (i =3D 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) { > > + for (i =3D 0; i < bufmgr_gem->num_buckets; i++) { > > struct drm_intel_gem_bo_bucket *bucket =3D > > &bufmgr_gem->cache_bucket[i]; > > if (bucket->size >=3D size) { > > @@ -822,7 +821,7 @@ drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem > > *bufmgr_gem, time_t time) > > { > > int i; > > > > - for (i =3D 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) { > > + for (i =3D 0; i < bufmgr_gem->num_buckets; i++) { > > struct drm_intel_gem_bo_bucket *bucket =3D > > &bufmgr_gem->cache_bucket[i]; > > > > @@ -1250,7 +1249,7 @@ drm_intel_bufmgr_gem_destroy(drm_intel_bufmgr > > *bufmgr) > > pthread_mutex_destroy(&bufmgr_gem->lock); > > > > /* Free any cached buffer objects we were going to reuse */ > > - for (i =3D 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) { > > + for (i =3D 0; i < bufmgr_gem->num_buckets; i++) { > > struct drm_intel_gem_bo_bucket *bucket =3D > > &bufmgr_gem->cache_bucket[i]; > > drm_intel_bo_gem *bo_gem; > > @@ -1960,6 +1959,46 @@ drm_intel_gem_bo_references(drm_intel_bo *bo, > > drm_intel_bo *target_bo) > > return 0; > > } > > > > +static void > > +add_bucket(drm_intel_bufmgr_gem *bufmgr_gem, int size) > > +{ > > + unsigned int i =3D bufmgr_gem->num_buckets; > > + > > + assert(i < ARRAY_SIZE(bufmgr_gem->cache_bucket)); > > + > > + DRMINITLISTHEAD(&bufmgr_gem->cache_bucket[i].head); > > + bufmgr_gem->cache_bucket[i].size =3D size; > > + bufmgr_gem->num_buckets++; > > +} > > + > > +static void > > +init_cache_buckets(drm_intel_bufmgr_gem *bufmgr_gem) > > +{ > > + unsigned long size, cache_max_size =3D 64 * 1024 * 1024; > > + > > + /* Initialize the linked lists for BO reuse cache. */ > > + for (size =3D 4096; size <=3D cache_max_size; size *=3D 2) { > > + add_bucket(bufmgr_gem, size); > > + > > + /* OK, so power of two buckets was too wasteful of > > + * memory. Give 3 other sizes between each power of > > + * two, to hopefully cover things accurately enough. > > + * (The alternative is probably to just go for exact > > + * matching of sizes, and assume that for things like > > + * composited window resize the tiled width/height > > + * alignment and rounding of sizes to pages will get > > + * us useful cache hit rates anyway) > > + */ > > + if (size =3D=3D 8192) { > > + add_bucket(bufmgr_gem, size + size / 2); > > + } else if (size < cache_max_size) { > > + add_bucket(bufmgr_gem, size + size * 1 / 4); > > + add_bucket(bufmgr_gem, size + size * 2 / 4); > > + add_bucket(bufmgr_gem, size + size * 3 / 4); > > + } > > + } > > +} > > + >=20 > Are bucket sizes that are not a multiple of 4096 supported and/or worthwh= ile? >=20 > The above code will create buckets size 5120, 6144 and 7168 in the size = =3D > 4096 case. >=20 > Apologies for the noise if I'm off the mark here. Yeah, krh noted this bug on irc. I'm looking at the alternate patch to not use buckets at all currently. --=-=-= Content-Type: application/pgp-signature -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.10 (GNU/Linux) iEYEARECAAYFAkwNLeoACgkQHUdvYGzw6vcsVgCfcdU/e+eSSGlU1Flw+3PaYWih 28QAn2KPBgOqDKbqKCmrBV/7IaClY9yl =PDnw -----END PGP SIGNATURE----- --=-=-=-- --===============0817282666== Content-Type: text/plain; charset="us-ascii" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Disposition: inline _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx --===============0817282666==--