[Intel-gfx] [PATCH] intel: Add more intermediate sizes of cache buckets between powers of 2.

[Xiang, Haihao]

On Sat, 2010-06-05 at 08:16 +0800, Eric Anholt 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_gem *bufmgr_gem,
>  {
>  	int i;
>  
> -	for (i = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
> +	for (i = 0; i < bufmgr_gem->num_buckets; i++) {
>  		struct drm_intel_gem_bo_bucket *bucket =
>  		    &bufmgr_gem->cache_bucket[i];
>  		if (bucket->size >= size) {
> @@ -822,7 +821,7 @@ drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem *bufmgr_gem, time_t time)
>  {
>  	int i;
>  
> -	for (i = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
> +	for (i = 0; i < bufmgr_gem->num_buckets; i++) {
>  		struct drm_intel_gem_bo_bucket *bucket =
>  		    &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 = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
> +	for (i = 0; i < bufmgr_gem->num_buckets; i++) {
>  		struct drm_intel_gem_bo_bucket *bucket =
>  		    &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 = 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 = size;
> +	bufmgr_gem->num_buckets++;
> +}
> +
> +static void
> +init_cache_buckets(drm_intel_bufmgr_gem *bufmgr_gem)
> +{
> +	unsigned long size, cache_max_size = 64 * 1024 * 1024;
> +
> +	/* Initialize the linked lists for BO reuse cache. */
> +	for (size = 4096; size <= cache_max_size; size *= 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 == 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);
> +		}
> +	}
> +}
> +
>  /**
>   * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
>   * and manage map buffer objections.
> @@ -1972,8 +2011,7 @@ drm_intel_bufmgr_gem_init(int fd, int batch_size)
>  	drm_intel_bufmgr_gem *bufmgr_gem;
>  	struct drm_i915_gem_get_aperture aperture;
>  	drm_i915_getparam_t gp;
> -	int ret, i;
> -	unsigned long size;
> +	int ret;
>  	int exec2 = 0;
>  
>  	bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
> @@ -2092,11 +2130,7 @@ drm_intel_bufmgr_gem_init(int fd, int batch_size)
>  	    drm_intel_gem_get_pipe_from_crtc_id;
>  	bufmgr_gem->bufmgr.bo_references = drm_intel_gem_bo_references;
>  
> -	/* Initialize the linked lists for BO reuse cache. */
> -	for (i = 0, size = 4096; i < DRM_INTEL_GEM_BO_BUCKETS; i++, size *= 2) {
> -		DRMINITLISTHEAD(&bufmgr_gem->cache_bucket[i].head);
> -		bufmgr_gem->cache_bucket[i].size = size;
> -	}
> +	init_cache_buckets(bufmgr_gem);
>  
>  	return &bufmgr_gem->bufmgr;
>  }

ACK.

This patch saves about 24M of GEM objects when playing a 1920x800, 4:2:0
HD video with VAAPI.