Magic-Mirror/llvm-capstone
1
0
Fork 0
mirror of https://github.com/capstone-engine/llvm-capstone.git synced 2024-10-09 12:24:08 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Allow machine hierarchy expansion

Browse Source 
This fix allows the machine hierarchy to be expanded in case it needs to handle 
more threads. It adds a resize function to accomplish this.

Differential Revision: http://reviews.llvm.org/D9900

llvm-svn: 240292
This commit is contained in:
Jonathan Peyton 2015-06-22 15:59:18 +00:00
parent 06407c0320
commit 7f09a98ab1
1 changed files with 78 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

88
openmp/runtime/src/kmp_affinity.cpp
View File

@ -312,12 +312,18 @@ __kmp_affinity_cmp_Address_child_num(const void *a, const void *b)
return 0;
}
/** A structure for holding machine-specific hierarchy info to be computed once at init. */
/** A structure for holding machine-specific hierarchy info to be computed once at init.
This structure represents a mapping of threads to the actual machine hierarchy, or to
our best guess at what the hierarchy might be, for the purpose of performing an
efficient barrier. In the worst case, when there is no machine hierarchy information,
it produces a tree suitable for a barrier, similar to the tree used in the hyper barrier. */
class hierarchy_info {
public:
/** Typical levels are threads/core, cores/package or socket, packages/node, nodes/machine,
etc. We don't want to get specific with nomenclature */
static const kmp_uint32 maxLevels=7;
/** Number of levels in the hierarchy. Typical levels are threads/core, cores/package
or socket, packages/node, nodes/machine, etc. We don't want to get specific with
nomenclature. When the machine is oversubscribed we add levels to duplicate the
hierarchy, doubling the thread capacity of the hierarchy each time we add a level. */
kmp_uint32 maxLevels;
/** This is specifically the depth of the machine configuration hierarchy, in terms of the
number of levels along the longest path from root to any leaf. It corresponds to the
@ -325,12 +331,13 @@ public:
kmp_uint32 depth;
kmp_uint32 base_num_threads;
volatile kmp_int8 uninitialized; // 0=initialized, 1=uninitialized, 2=initialization in progress
volatile kmp_int8 resizing; // 0=not resizing, 1=resizing
/** Level 0 corresponds to leaves. numPerLevel[i] is the number of children the parent of a
node at level i has. For example, if we have a machine with 4 packages, 4 cores/package
and 2 HT per core, then numPerLevel = {2, 4, 4, 1, 1}. All empty levels are set to 1. */
kmp_uint32 numPerLevel[maxLevels];
kmp_uint32 skipPerLevel[maxLevels];
kmp_uint32 *numPerLevel;
kmp_uint32 *skipPerLevel;
void deriveLevels(AddrUnsPair *adr2os, int num_addrs) {
int hier_depth = adr2os[0].first.depth;
@ -346,7 +353,11 @@ public:
}
}
hierarchy_info() : depth(1), uninitialized(1) {}
hierarchy_info() : maxLevels(7), depth(1), uninitialized(1), resizing(0) {}
// TO FIX: This destructor causes a segfault in the library at shutdown.
//~hierarchy_info() { if (!uninitialized && numPerLevel) __kmp_free(numPerLevel); }
void init(AddrUnsPair *adr2os, int num_addrs)
{
kmp_int8 bool_result = KMP_COMPARE_AND_STORE_ACQ8(&uninitialized, 1, 2);
@ -356,10 +367,14 @@ public:
}
KMP_DEBUG_ASSERT(bool_result==1);
/* Added explicit initialization of the depth here to prevent usage of dirty value
/* Added explicit initialization of the data fields here to prevent usage of dirty value
observed when static library is re-initialized multiple times (e.g. when
non-OpenMP thread repeatedly launches/joins thread that uses OpenMP). */
depth = 1;
resizing = 0;
maxLevels = 7;
numPerLevel = (kmp_uint32 *)__kmp_allocate(maxLevels*2*sizeof(kmp_uint32));
skipPerLevel = &(numPerLevel[maxLevels]);
for (kmp_uint32 i=0; i<maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
numPerLevel[i] = 1;
skipPerLevel[i] = 1;
@ -406,6 +421,56 @@ public:
uninitialized = 0; // One writer
}
void resize(kmp_uint32 nproc)
{
kmp_int8 bool_result = KMP_COMPARE_AND_STORE_ACQ8(&resizing, 0, 1);
if (bool_result == 0) { // Someone else is resizing
while (TCR_1(resizing) != 0) KMP_CPU_PAUSE();
return;
}
KMP_DEBUG_ASSERT(bool_result!=0);
KMP_DEBUG_ASSERT(nproc > base_num_threads);
// Calculate new max_levels
kmp_uint32 old_sz = skipPerLevel[depth-1];
kmp_uint32 incs = 0, old_maxLevels= maxLevels;
while (nproc > old_sz) {
old_sz *=2;
incs++;
}
maxLevels += incs;
// Resize arrays
kmp_uint32 *old_numPerLevel = numPerLevel;
kmp_uint32 *old_skipPerLevel = skipPerLevel;
numPerLevel = skipPerLevel = NULL;
numPerLevel = (kmp_uint32 *)__kmp_allocate(maxLevels*2*sizeof(kmp_uint32));
skipPerLevel = &(numPerLevel[maxLevels]);
// Copy old elements from old arrays
for (kmp_uint32 i=0; i<old_maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
numPerLevel[i] = old_numPerLevel[i];
skipPerLevel[i] = old_skipPerLevel[i];
}
// Init new elements in arrays to 1
for (kmp_uint32 i=old_maxLevels; i<maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
numPerLevel[i] = 1;
skipPerLevel[i] = 1;
}
// Free old arrays
__kmp_free(old_numPerLevel);
// Fill in oversubscription levels of hierarchy
for (kmp_uint32 i=old_maxLevels; i<maxLevels; ++i)
skipPerLevel[i] = 2*skipPerLevel[i-1];
base_num_threads = nproc;
resizing = 0; // One writer
}
};
static hierarchy_info machine_hierarchy;
@ -415,11 +480,14 @@ void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
// The test below is true if affinity is available, but set to "none". Need to init on first use of hierarchical barrier.
if (TCR_1(machine_hierarchy.uninitialized))
machine_hierarchy.init(NULL, nproc);
// Adjust the hierarchy in case num threads exceeds original
if (nproc > machine_hierarchy.base_num_threads)
machine_hierarchy.resize(nproc);
depth = machine_hierarchy.depth;
KMP_DEBUG_ASSERT(depth > 0);
// The loop below adjusts the depth in the case of oversubscription
while (nproc > machine_hierarchy.skipPerLevel[depth-1] && depth<machine_hierarchy.maxLevels-1)
// The loop below adjusts the depth in the case of a resize
while (nproc > machine_hierarchy.skipPerLevel[depth-1])
depth++;
thr_bar->depth = depth;