[Mesa-dev] [PATCH 4/5] gallium/radeon: use pb_cache buckets for fewer pb_cache misses

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

This makes Bioshock Infinite with deferred flushing 2.2% faster.
---
 src/gallium/winsys/radeon/drm/radeon_drm_bo.c | 29 ++++++++++++++++++++-------
 1 file changed, 22 insertions(+), 7 deletions(-)

diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
index 82dec8e..897b536 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
@@ -503,7 +503,8 @@ static struct radeon_bo *radeon_create_bo(struct radeon_drm_winsys *rws,
                                           unsigned size, unsigned alignment,
                                           unsigned usage,
                                           unsigned initial_domains,
-                                          unsigned flags)
+                                          unsigned flags,
+                                          unsigned pb_cache_bucket)
 {
     struct radeon_bo *bo;
     struct drm_radeon_gem_create args;
@@ -551,7 +552,8 @@ static struct radeon_bo *radeon_create_bo(struct radeon_drm_winsys *rws,
     bo->va = 0;
     bo->initial_domain = initial_domains;
     pipe_mutex_init(bo->map_mutex);
-    pb_cache_init_entry(&rws->bo_cache, &bo->cache_entry, &bo->base, 0);
+    pb_cache_init_entry(&rws->bo_cache, &bo->cache_entry, &bo->base,
+                        pb_cache_bucket);
 
     if (rws->info.has_virtual_memory) {
         struct drm_radeon_gem_va va;
@@ -723,7 +725,7 @@ radeon_winsys_bo_create(struct radeon_winsys *rws,
 {
     struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
     struct radeon_bo *bo;
-    unsigned usage = 0;
+    unsigned usage = 0, pb_cache_bucket;
 
     /* Only 32-bit sizes are supported. */
     if (size > UINT_MAX)
@@ -742,19 +744,32 @@ radeon_winsys_bo_create(struct radeon_winsys *rws,
     if (domain == RADEON_DOMAIN_VRAM_GTT)
         usage = 1 << 2;
     else
-        usage = domain >> 1;
+        usage = (unsigned)domain >> 1;
     assert(flags < sizeof(usage) * 8 - 3);
     usage |= 1 << (flags + 3);
 
-    bo = radeon_bo(pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, usage, 0));
+    /* Determine the pb_cache bucket for minimizing pb_cache misses. */
+    pb_cache_bucket = 0;
+    if (size <= 4096) /* small buffers */
+       pb_cache_bucket += 1;
+    if (domain & RADEON_DOMAIN_VRAM) /* VRAM or VRAM+GTT */
+       pb_cache_bucket += 2;
+    if (flags == RADEON_FLAG_GTT_WC) /* WC */
+       pb_cache_bucket += 4;
+    assert(pb_cache_bucket < ARRAY_SIZE(ws->bo_cache.buckets));
+
+    bo = radeon_bo(pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment,
+                                           usage, pb_cache_bucket));
     if (bo)
         return &bo->base;
 
-    bo = radeon_create_bo(ws, size, alignment, usage, domain, flags);
+    bo = radeon_create_bo(ws, size, alignment, usage, domain, flags,
+                          pb_cache_bucket);
     if (!bo) {
         /* Clear the cache and try again. */
         pb_cache_release_all_buffers(&ws->bo_cache);
-        bo = radeon_create_bo(ws, size, alignment, usage, domain, flags);
+        bo = radeon_create_bo(ws, size, alignment, usage, domain, flags,
+                              pb_cache_bucket);
         if (!bo)
             return NULL;
     }
-- 
2.7.4