Re: [Intel-gfx] [PATCH libdrm 1/2] intel: align reuse buffer's size on page size instead

[Xiong, James]

On Wed, Feb 21, 2018 at 09:43:55PM +, Chris Wilson wrote:
> Quoting James Xiong (2018-02-20 17:48:03)
> > From: "Xiong, James" 
> > 
> > With gem_reuse enabled, when a buffer size is different than
> > the sizes of buckets, it is aligned to the next bucket's size,
> > which means about 25% more memory than the requested is allocated
> > in the worst senario. For example:
> > 
> > Orignal sizeActual
> > 32KB+1Byte  40KB
> > .
> > .
> > .
> > 8MB+1Byte   10MB
> > .
> > .
> > .
> > 96MB+1Byte  112MB
> > 
> > This is very memory expensive and make the reuse feature less
> > favorable than it deserves to be.
> 
> The reuse feature also misses one important source: reusing temporaries
> within a batch.
>  
> > This commit aligns the reuse buffer size on page size instead,
> > the buffer whose size falls between bucket[n] and bucket[n+1] is
> > put in bucket[n] when it's done; And when searching for a cached
> > buffer for reuse, it goes through the cached buffers list in the
> > bucket until a cached buffer, whose size is larger than or equal
> > to the requested size, is found.
> 
> So how many times do you have to allocate a new buffer because you
> refused to hand back a slightly larger buffer? Have you checked the
> impact on !llc? With mmaps? On how wide a workload?
bucket[n] contains a list of buffers with size between bucket[n].size
and bucket[n+1].size - 1, a larger cached buffer could still be reused
if available.
I managed to run some performance tests on GEN9 without noticeable
performance impact but didn't have chance to test for more coverages.
> 
> > Signed-off-by: Xiong, James 
> > ---
> >  intel/intel_bufmgr_gem.c | 180 
> > +--
> >  libdrm_lists.h   |   6 ++
> >  2 files changed, 101 insertions(+), 85 deletions(-)
> > 
> > diff --git a/intel/intel_bufmgr_gem.c b/intel/intel_bufmgr_gem.c
> > index 386da30..5b2d0d0 100644
> > --- a/intel/intel_bufmgr_gem.c
> > +++ b/intel/intel_bufmgr_gem.c
> > @@ -402,11 +402,10 @@ drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_gem 
> > *bufmgr_gem,
> >  {
> > int i;
> >  
> > -   for (i = 0; i < bufmgr_gem->num_buckets; i++) {
> > -   struct drm_intel_gem_bo_bucket *bucket =
> > -   _gem->cache_bucket[i];
> > -   if (bucket->size >= size) {
> > -   return bucket;
> > +   for (i = 0; i < bufmgr_gem->num_buckets - 1; i++) {
> > +   if (size >= bufmgr_gem->cache_bucket[i].size &&
> > +   size < bufmgr_gem->cache_bucket[i+1].size) {
> > +   return _gem->cache_bucket[i];
> 
> So you never return the last bucket?
The last bucket's size is 112MB, I can add a 128M bucket at the end to
cache and reuse allocations between [112MB, 128M-1] if you think it's
important.
> 
> Given the ordered set of buckets, the test remains correct even when
> every bo within the bucket is not the full size (each bo is still at
> least bigger than the previous bucket).
The function returns an bucket according to the requested buffer size
at allocating.
if (buffer_size in [bucket[n].size, bucket[n+1].size))
return [n];
it also gets called at freeing. In both cases, the correct bucket is returned.
> 
> > }
> > }
> >  
> > @@ -685,25 +684,95 @@ drm_intel_gem_bo_madvise(drm_intel_bo *bo, int madv)
> >  madv);
> >  }
> >  
> > -/* drop the oldest entries that have been purged by the kernel */
> > +/* drop the entries that are older than the given time */
> >  static void
> >  drm_intel_gem_bo_cache_purge_bucket(drm_intel_bufmgr_gem *bufmgr_gem,
> > -   struct drm_intel_gem_bo_bucket *bucket)
> > +   struct drm_intel_gem_bo_bucket *bucket,
> > +   time_t time)
> >  {
> > -   while (!DRMLISTEMPTY(>head)) {
> > -   drm_intel_bo_gem *bo_gem;
> > -
> > -   bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
> > - bucket->head.next, head);
> > -   if (drm_intel_gem_bo_madvise_internal
> > -   (bufmgr_gem, bo_gem, I915_MADV_DONTNEED))
> > -   break;
> > -
> > -   DRMLISTDEL(_gem->head);
> > -   drm_intel_gem_bo_free(_gem->bo);
> > +   drm_intel_bo_gem *bo_gem, *temp;
> > +   DRMLISTFOREACHENTRYSAFE(bo_gem, temp, >head, head) {
> > +   if (bo_gem->free_time >= time) {
> > +   drm_intel_gem_bo_madvise_internal
> > +   (bufmgr_gem, bo_gem, I915_MADV_DONTNEED);
> > +   DRMLISTDEL(_gem->head);
> > +   drm_intel_gem_bo_free(_gem->bo);
> > +   }
> 
> This function is called after the kernel reports that it purged a
> buffer, and the intent here is that we find all the buffers that the
> 

Re: [Intel-gfx] [PATCH libdrm 1/2] intel: align reuse buffer's size on page size instead

[Chris Wilson]

Quoting James Xiong (2018-02-20 17:48:03)
> From: "Xiong, James" 
> 
> With gem_reuse enabled, when a buffer size is different than
> the sizes of buckets, it is aligned to the next bucket's size,
> which means about 25% more memory than the requested is allocated
> in the worst senario. For example:
> 
> Orignal sizeActual
> 32KB+1Byte  40KB
> .
> .
> .
> 8MB+1Byte   10MB
> .
> .
> .
> 96MB+1Byte  112MB
> 
> This is very memory expensive and make the reuse feature less
> favorable than it deserves to be.

The reuse feature also misses one important source: reusing temporaries
within a batch.
 
> This commit aligns the reuse buffer size on page size instead,
> the buffer whose size falls between bucket[n] and bucket[n+1] is
> put in bucket[n] when it's done; And when searching for a cached
> buffer for reuse, it goes through the cached buffers list in the
> bucket until a cached buffer, whose size is larger than or equal
> to the requested size, is found.

So how many times do you have to allocate a new buffer because you
refused to hand back a slightly larger buffer? Have you checked the
impact on !llc? With mmaps? On how wide a workload?

> Signed-off-by: Xiong, James 
> ---
>  intel/intel_bufmgr_gem.c | 180 
> +--
>  libdrm_lists.h   |   6 ++
>  2 files changed, 101 insertions(+), 85 deletions(-)
> 
> diff --git a/intel/intel_bufmgr_gem.c b/intel/intel_bufmgr_gem.c
> index 386da30..5b2d0d0 100644
> --- a/intel/intel_bufmgr_gem.c
> +++ b/intel/intel_bufmgr_gem.c
> @@ -402,11 +402,10 @@ drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_gem 
> *bufmgr_gem,
>  {
> int i;
>  
> -   for (i = 0; i < bufmgr_gem->num_buckets; i++) {
> -   struct drm_intel_gem_bo_bucket *bucket =
> -   _gem->cache_bucket[i];
> -   if (bucket->size >= size) {
> -   return bucket;
> +   for (i = 0; i < bufmgr_gem->num_buckets - 1; i++) {
> +   if (size >= bufmgr_gem->cache_bucket[i].size &&
> +   size < bufmgr_gem->cache_bucket[i+1].size) {
> +   return _gem->cache_bucket[i];

So you never return the last bucket?

Given the ordered set of buckets, the test remains correct even when
every bo within the bucket is not the full size (each bo is still at
least bigger than the previous bucket).

> }
> }
>  
> @@ -685,25 +684,95 @@ drm_intel_gem_bo_madvise(drm_intel_bo *bo, int madv)
>  madv);
>  }
>  
> -/* drop the oldest entries that have been purged by the kernel */
> +/* drop the entries that are older than the given time */
>  static void
>  drm_intel_gem_bo_cache_purge_bucket(drm_intel_bufmgr_gem *bufmgr_gem,
> -   struct drm_intel_gem_bo_bucket *bucket)
> +   struct drm_intel_gem_bo_bucket *bucket,
> +   time_t time)
>  {
> -   while (!DRMLISTEMPTY(>head)) {
> -   drm_intel_bo_gem *bo_gem;
> -
> -   bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
> - bucket->head.next, head);
> -   if (drm_intel_gem_bo_madvise_internal
> -   (bufmgr_gem, bo_gem, I915_MADV_DONTNEED))
> -   break;
> -
> -   DRMLISTDEL(_gem->head);
> -   drm_intel_gem_bo_free(_gem->bo);
> +   drm_intel_bo_gem *bo_gem, *temp;
> +   DRMLISTFOREACHENTRYSAFE(bo_gem, temp, >head, head) {
> +   if (bo_gem->free_time >= time) {
> +   drm_intel_gem_bo_madvise_internal
> +   (bufmgr_gem, bo_gem, I915_MADV_DONTNEED);
> +   DRMLISTDEL(_gem->head);
> +   drm_intel_gem_bo_free(_gem->bo);
> +   }

This function is called after the kernel reports that it purged a
buffer, and the intent here is that we find all the buffers that the
kernel purged and evict them. It is not about discarding old buffers,
just throwing out the empty.

Honestly, it's pointless. The cost of having the empty purged bo around
is insignificant (we reclaim a little bit of mmap virtual space
quicker).

> }
>  }
>  
> +static drm_intel_bo_gem *
> +drm_intel_gem_bo_cached_for_size(drm_intel_bufmgr_gem *bufmgr_gem,
> + unsigned long size,
> + uint32_t tiling_mode,
> + unsigned long stride,
> + unsigned long alignment,
> + bool for_render)
> +{
> +   struct drm_intel_gem_bo_bucket *bucket =
> +   drm_intel_gem_bo_bucket_for_size(bufmgr_gem, size);
> +
> +   if(bucket != NULL) {
> +   drm_intel_bo_gem *bo_gem, *temp_bo_gem;
> +retry:
> +   bo_gem = NULL;
> +   if (for_render) {
> +