Further optimizations to the VM trace parser that brings the parse
time of ~3.8M lines of trace data down from >3.5seconds to ~1.08 seconds
on my Carrizo (A12-9800). The goal is to ensure that parsing large
amounts of mappings (for say large games/compute users) is always somewhat
reasonable.
In this case 3.8M lines represents running unigine-heaven on max settings
for about 4 hours with my Polaris10 card.
Signed-off-by: Tom St Denis <tom.stdenis at amd.com>
---
src/lib/read_vram.c | 26 ++++++++++++++++++++++----
1 file changed, 22 insertions(+), 4 deletions(-)
diff --git a/src/lib/read_vram.c b/src/lib/read_vram.c
index 667719a9705b..f2c3a15c27fe 100644
--- a/src/lib/read_vram.c
+++ b/src/lib/read_vram.c
@@ -35,7 +35,7 @@ static struct umr_map *find_map(struct umr_dma_maps *maps, uint64_t dma_addr, in
// so if we use an identity function on the search
// key we'll end up with a really unbalanced tree
// so mix up address a bit to randomize keys
- key = dma_addr ^ (dma_addr >> 9);
+ key = dma_addr ^ (dma_addr >> 11);
if (!n) {
maps->maps = calloc(1, sizeof(maps->maps[0]));
@@ -60,6 +60,7 @@ static struct umr_map *find_map(struct umr_dma_maps *maps, uint64_t dma_addr, in
return *nn;
}
}
+
return n;
}
@@ -71,9 +72,26 @@ static int insert_map(struct umr_dma_maps *maps,
// don't add a new node if we're marking it invalid
if (map) {
- map->dma_addr = dma_addr;
- map->phys_addr = phys_addr;
- map->valid = valid;
+ if (valid) {
+ map->dma_addr = dma_addr;
+ map->phys_addr = phys_addr;
+ map->valid = valid;
+ } else {
+ struct umr_map *tmap = NULL;
+
+ // if marking invalid see if we can prune
+ // the tree a little if the node we're marking
+ // as invalid only has one child
+ if (map->left == NULL && map->right) {
+ tmap = map->right;
+ *map = *(map->right);
+ } else if (map->right == NULL && map->left) {
+ tmap = map->left;
+ *map = *(map->left);
+ }
+ if (tmap)
+ free(tmap);
+ }
}
return 0;
--
2.12.0
