The attached patch implements a "linesize" cache attribute. You now get
things like
L3Cache #0 (8192KB line=64)
L2Cache #0 (256KB line=64)
L1Cache #0 (32KB line=64)
>> Talking about caches, one thing we need to think about is Instruction
>> caches (we only gather Data and Unified caches on Linux so far).
>>
> Why is runtime icache information important? :)
>
Some people manually optimizing their kernels want this kind of info as
well as TLB size for instance...
Brice
>From 51c2b86df81b17ac5cdc4cfe6bbdaf42eb554b4f Mon Sep 17 00:00:00 2001
From: Brice Goglin <bgog...@debian.org>
List-Post: hwloc-devel@lists.open-mpi.org
Date: Tue, 25 May 2010 22:55:26 +0200
Subject: [PATCH] Add linesize cache attribute
---
include/hwloc.h | 1 +
src/topology-aix.c | 1 +
src/topology-darwin.c | 1 +
src/topology-linux.c | 11 +++++++++++
src/topology-synthetic.c | 1 +
src/topology-windows.c | 2 ++
src/topology-x86.c | 1 +
src/topology-xml.c | 10 ++++++++++
src/topology.c | 2 ++
src/traversal.c | 5 +++--
10 files changed, 33 insertions(+), 2 deletions(-)
diff --git a/include/hwloc.h b/include/hwloc.h
index 1f275d5..c3a1b38 100644
--- a/include/hwloc.h
+++ b/include/hwloc.h
@@ -302,6 +302,7 @@ union hwloc_obj_attr_u {
struct hwloc_cache_attr_s {
uint64_t size; /**< \brief Size of cache in bytes */
unsigned depth; /**< \brief Depth of cache */
+ unsigned linesize; /**< \brief Cache-line size in bytes */
} cache;
/** \brief Machine-specific Object Attributes */
struct hwloc_machine_attr_s {
diff --git a/src/topology-aix.c b/src/topology-aix.c
index b6bfe8d..886aed6 100644
--- a/src/topology-aix.c
+++ b/src/topology-aix.c
@@ -214,6 +214,7 @@ look_rset(int sdl, hwloc_obj_type_t type, struct hwloc_topology *topology, int l
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = 0; /* TODO: ? */
+ obj->attr->cache.linesize = 0; /* TODO: ? */
obj->attr->cache.depth = 2;
break;
case HWLOC_OBJ_GROUP:
diff --git a/src/topology-darwin.c b/src/topology-darwin.c
index 5d2d157..6a9f7b5 100644
--- a/src/topology-darwin.c
+++ b/src/topology-darwin.c
@@ -122,6 +122,7 @@ hwloc_look_darwin(struct hwloc_topology *topology)
i, j, obj->cpuset);
obj->attr->cache.depth = i;
obj->attr->cache.size = cachesize[i];
+ obj->attr->cache.linesize = 0; /* TODO */
} else {
hwloc_debug_1arg_cpuset("node %u has cpuset %s\n",
j, obj->cpuset);
diff --git a/src/topology-linux.c b/src/topology-linux.c
index fcc8274..c35327b 100644
--- a/src/topology-linux.c
+++ b/src/topology-linux.c
@@ -1365,6 +1365,7 @@ look_sysfscpu(struct hwloc_topology *topology, const char *path)
struct hwloc_obj *cache;
hwloc_cpuset_t cacheset;
unsigned long kB = 0;
+ unsigned linesize = 0;
int depth; /* 0 for L1, .... */
/* get the cache level depth */
@@ -1403,6 +1404,15 @@ look_sysfscpu(struct hwloc_topology *topology, const char *path)
fclose(fd);
}
+ /* get the line size */
+ sprintf(mappath, "%s/cpu%d/cache/index%d/coherency_line_size", path, i, j);
+ fd = hwloc_fopen(mappath, "r", topology->backend_params.sysfs.root_fd);
+ if (fd) {
+ if (fgets(str2,sizeof(str2), fd))
+ linesize = atol(str2); /* in bytes */
+ fclose(fd);
+ }
+
sprintf(mappath, "%s/cpu%d/cache/index%d/shared_cpu_map", path, i, j);
cacheset = hwloc_parse_cpumap(mappath, topology->backend_params.sysfs.root_fd);
if (cacheset) {
@@ -1415,6 +1425,7 @@ look_sysfscpu(struct hwloc_topology *topology, const char *path)
cache = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, -1);
cache->attr->cache.size = kB << 10;
cache->attr->cache.depth = depth+1;
+ cache->attr->cache.linesize = linesize;
cache->cpuset = cacheset;
hwloc_debug_1arg_cpuset("cache depth %d has cpuset %s\n",
depth, cacheset);
diff --git a/src/topology-synthetic.c b/src/topology-synthetic.c
index fe37d20..99e64c4 100644
--- a/src/topology-synthetic.c
+++ b/src/topology-synthetic.c
@@ -274,6 +274,7 @@ hwloc__look_synthetic(struct hwloc_topology *topology,
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.depth = topology->backend_params.synthetic.depth[level];
+ obj->attr->cache.linesize = 64;
if (obj->attr->cache.depth == 1)
/* 32Kb in L1 */
obj->attr->cache.size = 32*1024;
diff --git a/src/topology-windows.c b/src/topology-windows.c
index 3193652..31313b0 100644
--- a/src/topology-windows.c
+++ b/src/topology-windows.c
@@ -284,6 +284,7 @@ hwloc_look_windows(struct hwloc_topology *topology)
}
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = procInfo[i].Cache.Size;
+ obj->attr->cache.linesize = procInfo[i].Cache.LineSize;
obj->attr->cache.depth = procInfo[i].Cache.Level;
break;
case HWLOC_OBJ_GROUP:
@@ -395,6 +396,7 @@ hwloc_look_windows(struct hwloc_topology *topology)
}
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = procInfo->Cache.CacheSize;
+ obj->attr->cache.linesize = procInfo->Cache.LineSize;
obj->attr->cache.depth = procInfo->Cache.Level;
break;
default:
diff --git a/src/topology-x86.c b/src/topology-x86.c
index 653353c..8946685 100644
--- a/src/topology-x86.c
+++ b/src/topology-x86.c
@@ -333,6 +333,7 @@ static void summarize(hwloc_topology_t topology, struct procinfo *infos, unsigne
cache = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, cacheid);
cache->attr->cache.depth = level;
cache->attr->cache.size = infos[i].cache[l].size;
+ cache->attr->cache.linesize = infos[i].cache[l].linesize;
cache->cpuset = cache_cpuset;
hwloc_debug_2args_cpuset("os L%u cache %u has cpuset %s\n",
level, cacheid, cache_cpuset);
diff --git a/src/topology-xml.c b/src/topology-xml.c
index ae11fee..0fec2ea 100644
--- a/src/topology-xml.c
+++ b/src/topology-xml.c
@@ -137,6 +137,14 @@ hwloc__xml_import_object_attr(struct hwloc_topology *topology __hwloc_attribute_
fprintf(stderr, "ignoring cache_size attribute for non-cache object type\n");
}
+ else if (!strcmp(name, "cache_linesize")) {
+ unsigned long lvalue = strtoul(value, NULL, 10);
+ if (obj->type == HWLOC_OBJ_CACHE)
+ obj->attr->cache.linesize = lvalue;
+ else
+ fprintf(stderr, "ignoring cache_linesize attribute for non-cache object type\n");
+ }
+
else if (!strcmp(name, "local_memory"))
obj->memory.local_memory = strtoull(value, NULL, 10);
@@ -498,6 +506,8 @@ hwloc__xml_export_object (hwloc_topology_t topology, hwloc_obj_t obj, xmlNodePtr
xmlNewProp(node, BAD_CAST "cache_size", BAD_CAST tmp);
sprintf(tmp, "%u", obj->attr->cache.depth);
xmlNewProp(node, BAD_CAST "depth", BAD_CAST tmp);
+ sprintf(tmp, "%u", (unsigned) obj->attr->cache.linesize);
+ xmlNewProp(node, BAD_CAST "cache_linesize", BAD_CAST tmp);
break;
case HWLOC_OBJ_GROUP:
sprintf(tmp, "%u", obj->attr->group.depth);
diff --git a/src/topology.c b/src/topology.c
index 267c7e3..9b67cbb 100644
--- a/src/topology.c
+++ b/src/topology.c
@@ -730,6 +730,8 @@ hwloc__insert_object_by_cpuset(struct hwloc_topology *topology, hwloc_obj_t cur,
case HWLOC_OBJ_CACHE:
merge_sizes(obj, child, attr->cache.size);
check_sizes(obj, child, attr->cache.size);
+ merge_sizes(obj, child, attr->cache.linesize);
+ check_sizes(obj, child, attr->cache.linesize);
break;
default:
break;
diff --git a/src/traversal.c b/src/traversal.c
index f9d5346..9c36d9a 100644
--- a/src/traversal.c
+++ b/src/traversal.c
@@ -225,9 +225,10 @@ hwloc_obj_attr_snprintf(char * __hwloc_restrict string, size_t size, hwloc_obj_t
switch (obj->type) {
case HWLOC_OBJ_CACHE:
- hwloc_snprintf(specific, sizeof(specific), "%lu%s",
+ hwloc_snprintf(specific, sizeof(specific), "%lu%s%sline=%u",
(unsigned long) hwloc_memory_size_printf_value(obj->attr->cache.size, verbose),
- hwloc_memory_size_printf_unit(obj->attr->cache.size, verbose));
+ hwloc_memory_size_printf_unit(obj->attr->cache.size, verbose),
+ separator, obj->attr->cache.linesize);
break;
default:
break;
--
1.7.0
