The issue was introduced by a Threadripper2 commit back in
ce927e15e0. This boils down to threads inheriting affinity
from the parent thread. It is a question how this slipped
through the review (we definitely run benchmark round).
Quick fix could have been to always set CPU group affinity
in Cycles, and it would work for Windows. On other platforms
we did not have CPU groups API finished.
Ended up making Cycles aware of NUMA topology, so now we
bound threads to a specific NUMA node. This required adding
an external dependency to Cycles, but made some code there
shorter.
277 lines
8.5 KiB
C
277 lines
8.5 KiB
C
// Copyright (c) 2016, libnumaapi authors
|
|
//
|
|
// Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
// of this software and associated documentation files (the "Software"), to
|
|
// deal in the Software without restriction, including without limitation the
|
|
// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
|
|
// sell copies of the Software, and to permit persons to whom the Software is
|
|
// furnished to do so, subject to the following conditions:
|
|
//
|
|
// The above copyright notice and this permission notice shall be included in
|
|
// all copies or substantial portions of the Software.
|
|
//
|
|
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
|
// IN THE SOFTWARE.
|
|
//
|
|
// Author: Sergey Sharybin (sergey.vfx@gmail.com)
|
|
|
|
#include "build_config.h"
|
|
|
|
#if OS_LINUX
|
|
|
|
#include "numaapi.h"
|
|
|
|
#include <stdlib.h>
|
|
|
|
#ifndef WITH_DYNLOAD
|
|
# include <numa.h>
|
|
#else
|
|
# include <dlfcn.h>
|
|
#endif
|
|
|
|
#include <stdio.h>
|
|
|
|
#ifdef WITH_DYNLOAD
|
|
|
|
// Descriptor numa library.
|
|
static void* numa_lib;
|
|
|
|
// Types of all symbols which are read from the library.
|
|
struct bitmask;
|
|
typedef int tnuma_available(void);
|
|
typedef int tnuma_max_node(void);
|
|
typedef int tnuma_node_to_cpus(int node, struct bitmask* mask);
|
|
typedef long tnuma_node_size(int node, long* freep);
|
|
typedef int tnuma_run_on_node(int node);
|
|
typedef void* tnuma_alloc_onnode(size_t size, int node);
|
|
typedef void* tnuma_alloc_local(size_t size);
|
|
typedef void tnuma_free(void* start, size_t size);
|
|
typedef struct bitmask* tnuma_bitmask_clearall(struct bitmask *bitmask);
|
|
typedef int tnuma_bitmask_isbitset(const struct bitmask *bitmask,
|
|
unsigned int n);
|
|
typedef struct bitmask* tnuma_bitmask_setbit(struct bitmask *bitmask,
|
|
unsigned int n);
|
|
typedef unsigned int tnuma_bitmask_nbytes(struct bitmask *bitmask);
|
|
typedef void tnuma_bitmask_free(struct bitmask *bitmask);
|
|
typedef struct bitmask* tnuma_allocate_cpumask(void);
|
|
typedef struct bitmask* tnuma_allocate_nodemask(void);
|
|
typedef void tnuma_free_cpumask(struct bitmask* bitmask);
|
|
typedef void tnuma_free_nodemask(struct bitmask* bitmask);
|
|
typedef int tnuma_run_on_node_mask(struct bitmask *nodemask);
|
|
typedef int tnuma_run_on_node_mask_all(struct bitmask *nodemask);
|
|
typedef void tnuma_set_interleave_mask(struct bitmask *nodemask);
|
|
typedef void tnuma_set_localalloc(void);
|
|
|
|
// Actual symbols.
|
|
static tnuma_available* numa_available;
|
|
static tnuma_max_node* numa_max_node;
|
|
static tnuma_node_to_cpus* numa_node_to_cpus;
|
|
static tnuma_node_size* numa_node_size;
|
|
static tnuma_run_on_node* numa_run_on_node;
|
|
static tnuma_alloc_onnode* numa_alloc_onnode;
|
|
static tnuma_alloc_local* numa_alloc_local;
|
|
static tnuma_free* numa_free;
|
|
static tnuma_bitmask_clearall* numa_bitmask_clearall;
|
|
static tnuma_bitmask_isbitset* numa_bitmask_isbitset;
|
|
static tnuma_bitmask_setbit* numa_bitmask_setbit;
|
|
static tnuma_bitmask_nbytes* numa_bitmask_nbytes;
|
|
static tnuma_bitmask_free* numa_bitmask_free;
|
|
static tnuma_allocate_cpumask* numa_allocate_cpumask;
|
|
static tnuma_allocate_nodemask* numa_allocate_nodemask;
|
|
static tnuma_free_nodemask* numa_free_nodemask;
|
|
static tnuma_free_cpumask* numa_free_cpumask;
|
|
static tnuma_run_on_node_mask* numa_run_on_node_mask;
|
|
static tnuma_run_on_node_mask_all* numa_run_on_node_mask_all;
|
|
static tnuma_set_interleave_mask* numa_set_interleave_mask;
|
|
static tnuma_set_localalloc* numa_set_localalloc;
|
|
|
|
static void* findLibrary(const char** paths) {
|
|
int i = 0;
|
|
while (paths[i] != NULL) {
|
|
void* lib = dlopen(paths[i], RTLD_LAZY);
|
|
if (lib != NULL) {
|
|
return lib;
|
|
}
|
|
++i;
|
|
}
|
|
return NULL;
|
|
}
|
|
|
|
static void numaExit(void) {
|
|
if (numa_lib == NULL) {
|
|
return;
|
|
}
|
|
dlclose(numa_lib);
|
|
numa_lib = NULL;
|
|
}
|
|
|
|
static NUMAAPI_Result loadNumaSymbols(void) {
|
|
// Prevent multiple initializations.
|
|
static bool initialized = false;
|
|
static NUMAAPI_Result result = NUMAAPI_NOT_AVAILABLE;
|
|
if (initialized) {
|
|
return result;
|
|
}
|
|
initialized = true;
|
|
// Find appropriate .so library.
|
|
const char* numa_paths[] = {
|
|
"libnuma.so.1",
|
|
"libnuma.so",
|
|
NULL};
|
|
// Register de-initialization.
|
|
const int error = atexit(numaExit);
|
|
if (error) {
|
|
result = NUMAAPI_ERROR_ATEXIT;
|
|
return result;
|
|
}
|
|
// Load library.
|
|
numa_lib = findLibrary(numa_paths);
|
|
if (numa_lib == NULL) {
|
|
result = NUMAAPI_NOT_AVAILABLE;
|
|
return result;
|
|
}
|
|
// Load symbols.
|
|
|
|
#define _LIBRARY_FIND(lib, name) \
|
|
do { \
|
|
name = (t##name *)dlsym(lib, #name); \
|
|
} while (0)
|
|
#define NUMA_LIBRARY_FIND(name) _LIBRARY_FIND(numa_lib, name)
|
|
|
|
NUMA_LIBRARY_FIND(numa_available);
|
|
NUMA_LIBRARY_FIND(numa_max_node);
|
|
NUMA_LIBRARY_FIND(numa_node_to_cpus);
|
|
NUMA_LIBRARY_FIND(numa_node_size);
|
|
NUMA_LIBRARY_FIND(numa_run_on_node);
|
|
NUMA_LIBRARY_FIND(numa_alloc_onnode);
|
|
NUMA_LIBRARY_FIND(numa_alloc_local);
|
|
NUMA_LIBRARY_FIND(numa_free);
|
|
NUMA_LIBRARY_FIND(numa_bitmask_clearall);
|
|
NUMA_LIBRARY_FIND(numa_bitmask_isbitset);
|
|
NUMA_LIBRARY_FIND(numa_bitmask_setbit);
|
|
NUMA_LIBRARY_FIND(numa_bitmask_nbytes);
|
|
NUMA_LIBRARY_FIND(numa_bitmask_free);
|
|
NUMA_LIBRARY_FIND(numa_allocate_cpumask);
|
|
NUMA_LIBRARY_FIND(numa_allocate_nodemask);
|
|
NUMA_LIBRARY_FIND(numa_free_cpumask);
|
|
NUMA_LIBRARY_FIND(numa_free_nodemask);
|
|
NUMA_LIBRARY_FIND(numa_run_on_node_mask);
|
|
NUMA_LIBRARY_FIND(numa_run_on_node_mask_all);
|
|
NUMA_LIBRARY_FIND(numa_set_interleave_mask);
|
|
NUMA_LIBRARY_FIND(numa_set_localalloc);
|
|
|
|
#undef NUMA_LIBRARY_FIND
|
|
#undef _LIBRARY_FIND
|
|
|
|
result = NUMAAPI_SUCCESS;
|
|
return result;
|
|
}
|
|
#endif
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
// Initialization.
|
|
|
|
NUMAAPI_Result numaAPI_Initialize(void) {
|
|
#ifdef WITH_DYNLOAD
|
|
NUMAAPI_Result result = loadNumaSymbols();
|
|
if (result != NUMAAPI_SUCCESS) {
|
|
return result;
|
|
}
|
|
#endif
|
|
if (numa_available() < 0) {
|
|
return NUMAAPI_NOT_AVAILABLE;
|
|
}
|
|
return NUMAAPI_SUCCESS;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
// Topology query.
|
|
|
|
int numaAPI_GetNumNodes(void) {
|
|
return numa_max_node() + 1;
|
|
}
|
|
|
|
bool numaAPI_IsNodeAvailable(int node) {
|
|
return numaAPI_GetNumNodeProcessors(node) > 0;
|
|
}
|
|
|
|
int numaAPI_GetNumNodeProcessors(int node) {
|
|
struct bitmask* cpu_mask = numa_allocate_cpumask();
|
|
numa_node_to_cpus(node, cpu_mask);
|
|
const unsigned int num_bytes = numa_bitmask_nbytes(cpu_mask);
|
|
const unsigned int num_bits = num_bytes *8;
|
|
// TODO(sergey): There might be faster way calculating number of set bits.
|
|
int num_processors = 0;
|
|
for (unsigned int bit = 0; bit < num_bits; ++bit) {
|
|
if (numa_bitmask_isbitset(cpu_mask, bit)) {
|
|
++num_processors;
|
|
}
|
|
}
|
|
#ifdef WITH_DYNLOAD
|
|
if (numa_free_cpumask != NULL) {
|
|
numa_free_cpumask(cpu_mask);
|
|
} else {
|
|
numa_bitmask_free(cpu_mask);
|
|
}
|
|
#else
|
|
numa_free_cpumask(cpu_mask);
|
|
#endif
|
|
return num_processors;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
// Affinities.
|
|
|
|
bool numaAPI_RunProcessOnNode(int node) {
|
|
numaAPI_RunThreadOnNode(node);
|
|
return true;
|
|
}
|
|
|
|
bool numaAPI_RunThreadOnNode(int node) {
|
|
// Construct bit mask from node index.
|
|
struct bitmask* node_mask = numa_allocate_nodemask();
|
|
numa_bitmask_clearall(node_mask);
|
|
numa_bitmask_setbit(node_mask, node);
|
|
numa_run_on_node_mask_all(node_mask);
|
|
// TODO(sergey): The following commands are based on x265 code, we might want
|
|
// to make those optional, or require to call those explicitly.
|
|
//
|
|
// Current assumption is that this is similar to SetThreadGroupAffinity().
|
|
if (numa_node_size(node, NULL) > 0) {
|
|
numa_set_interleave_mask(node_mask);
|
|
numa_set_localalloc();
|
|
}
|
|
#ifdef WITH_DYNLOAD
|
|
if (numa_free_nodemask != NULL) {
|
|
numa_free_nodemask(node_mask);
|
|
} else {
|
|
numa_bitmask_free(node_mask);
|
|
}
|
|
#else
|
|
numa_free_nodemask(node_mask);
|
|
#endif
|
|
return true;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
// Memory management.
|
|
|
|
void* numaAPI_AllocateOnNode(size_t size, int node) {
|
|
return numa_alloc_onnode(size, node);
|
|
}
|
|
|
|
void* numaAPI_AllocateLocal(size_t size) {
|
|
return numa_alloc_local(size);
|
|
}
|
|
|
|
void numaAPI_Free(void* start, size_t size) {
|
|
numa_free(start, size);
|
|
}
|
|
|
|
#endif // OS_LINUX
|