From ceaf3d1bf448ae284b7e547d42edf346711ff848 Mon Sep 17 00:00:00 2001
From: Philippe Virouleau <philippe.virouleau@imag.fr>
Date: Tue, 17 Jan 2017 16:53:42 +0100
Subject: [PATCH] Improve affinity integration.
- Polish the non-affinity push/select
- Setup basic affinity select (only cores)
- Setup affinity push (thread/node/data/default)
- Remove Kaapi's affinity: no need for it anymore!
Note: right now it's still possible to end up with a deadlock if affinity(node:X) is used, we don't visit the nodes' queue yet!
Next commit will introduce this.
---
runtime/CMakeLists.txt | 4 -
runtime/src/CMakeLists.txt | 3 -
runtime/src/kaapi_numaffinity.c | 480 --------------------------------
runtime/src/kaapi_numaffinity.h | 91 ------
runtime/src/kmp.h | 12 +-
runtime/src/kmp_affinity.cpp | 4 +-
runtime/src/kmp_tasking.c | 173 +++++-------
7 files changed, 83 insertions(+), 684 deletions(-)
delete mode 100644 runtime/src/kaapi_numaffinity.c
delete mode 100644 runtime/src/kaapi_numaffinity.h
diff --git a/runtime/CMakeLists.txt b/runtime/CMakeLists.txt
index 046d5bc..2109ad8 100644
--- a/runtime/CMakeLists.txt
+++ b/runtime/CMakeLists.txt
@@ -325,10 +325,6 @@ if(LIBOMP_USE_THE_AGGREGATION AND (NOT LIBOMP_USE_THEQUEUE))
set(LIBOMP_USE_THEQUEUE TRUE)
endif()
-if(LIBOMP_USE_AFFINITY AND (NOT LIBOMP_USE_THEQUEUE))
- libomp_error_say("OpenMP affinity scheduler required T.H.E work stealing protocol")
-endif()
-
set(LIBOMP_USE_PAPI FALSE CACHE BOOL
"libOMP tracing based on PAPI")
diff --git a/runtime/src/CMakeLists.txt b/runtime/src/CMakeLists.txt
index d19fdd8..b6a0767 100644
--- a/runtime/src/CMakeLists.txt
+++ b/runtime/src/CMakeLists.txt
@@ -103,9 +103,6 @@ else()
if (${LIBOMP_USE_THEQUEUE})
libomp_append(LIBOMP_CFILES kaapi_sched_ccsync.c)
libomp_append(LIBOMP_CFILES kaapi_rt.c)
- if(${LIBOMP_USE_AFFINITY})
- libomp_append(LIBOMP_CFILES kaapi_numaffinity.c)
- endif()
endif()
# Get C++ files
set(LIBOMP_CXXFILES
diff --git a/runtime/src/kaapi_numaffinity.c b/runtime/src/kaapi_numaffinity.c
deleted file mode 100644
index 186a252..0000000
--- a/runtime/src/kaapi_numaffinity.c
+++ /dev/null
@@ -1,480 +0,0 @@
-/*
-** xkaapi
-**
-** Copyright 2009,2010,2011,2012 INRIA.
-**
-** Contributors :
-**
-** thierry.gautier@inrialpes.fr
-**
-** This software is a computer program whose purpose is to execute
-** multithreaded computation with data flow synchronization between
-** threads.
-**
-** This software is governed by the CeCILL-C license under French law
-** and abiding by the rules of distribution of free software. You can
-** use, modify and/ or redistribute the software under the terms of
-** the CeCILL-C license as circulated by CEA, CNRS and INRIA at the
-** following URL "http://www.cecill.info".
-**
-** As a counterpart to the access to the source code and rights to
-** copy, modify and redistribute granted by the license, users are
-** provided only with a limited warranty and the software's author,
-** the holder of the economic rights, and the successive licensors
-** have only limited liability.
-**
-** In this respect, the user's attention is drawn to the risks
-** associated with loading, using, modifying and/or developing or
-** reproducing the software by the user in light of its specific
-** status of free software, that may mean that it is complicated to
-** manipulate, and that also therefore means that it is reserved for
-** developers and experienced professionals having in-depth computer
-** knowledge. Users are therefore encouraged to load and test the
-** software's suitability as regards their requirements in conditions
-** enabling the security of their systems and/or data to be ensured
-** and, more generally, to use and operate it in the same conditions
-** as regards security.
-**
-** The fact that you are presently reading this means that you have
-** had knowledge of the CeCILL-C license and that you accept its
-** terms.
-**
-*/
-#include "kmp.h"
-#include "kaapi_numaffinity.h"
-#include "kaapi_atomic.h"
-#include <stdio.h>
-#include <ctype.h>
-
-#if LIBOMP_USE_AFFINITY
-#include <numa.h>
-#include <numaif.h>
-#ifndef _GNU_SOURCE
-# define _GNU_SOURCE
-#endif
-#include <sched.h> /* sched_getcpu */
-#endif
-
-/** All places as Kaapi exports: NUMA & core places only. Other entries are set to 0
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i] is the pointer to the i-th numa node.
-*/
-kaapi_place_t** kaapi_all_places[KAAPI_HWS_LEVELID_MAX] =
-{ 0, 0, 0, 0, 0 };
-
-/** sizes of each kaapi_all_places[level] */
-int kaapi_all_places_count[KAAPI_HWS_LEVELID_MAX] =
-{ 0, 0, 0, 0, 0 };
-
-/**
-*/
-char* kaapi_cpu2numa = 0;
-
-#if defined(__linux__)
-static bool kaapi_parse_range(
- char** str,
- unsigned int* index_low,
- unsigned int* index_high,
- int* stride
-);
-static bool
-kaapi_parse_unsigned_longlong(
- char** str,
- unsigned long long *pvalue
-);
-static bool
-kaapi_parse_unsigned_longlong(
- char** str,
- unsigned long long *pvalue
-);
-static bool
-kaapi_parse_unsigned_int(
- char** str,
- unsigned int *pvalue
-);
-static bool
-kaapi_parse_unsigned_long(
- char** str,
- unsigned long *pvalue
-);
-static bool
-kaapi_parse_int(
- char** str,
- int *pvalue
-);
-#endif
-
-
-/**
-*/
-static kaapi_atomic_t is_hwdetect_called = {0};
-int kaapi_mt_hwdetect(void)
-{
- if (KAAPI_ATOMIC_INCR(&is_hwdetect_called) >1)
- return EALREADY;
-
- /* */
- kaapi_all_places_count[KAAPI_HWS_LEVELID_MACHINE] = 1;
- kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE] = (kaapi_place_t**)__kmp_allocate(
- kaapi_all_places_count[KAAPI_HWS_LEVELID_MACHINE] * sizeof(kaapi_place_t*)
- );
- kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0] =
- (kaapi_place_t*)__kmp_allocate(sizeof(kaapi_place_t));
- kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0]->level = KAAPI_HWS_LEVELID_CORE;
- kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0]->queue =
- (kaapi_wsqueue_t*)__kmp_allocate(sizeof(kaapi_wsqueue_t));
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0]->queue,
- INITIAL_TASK_DEQUE_SIZE,
- -1
- );
- kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0]->private_queue =
- (kaapi_wsqueue_t*)__kmp_allocate(sizeof(kaapi_wsqueue_t));
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_MACHINE][0]->private_queue,
- INITIAL_TASK_DEQUE_SIZE,
- -1
- );
-
- /* compute the number of cpu of the system */
- kaapi_all_places_count[KAAPI_HWS_LEVELID_CORE] = __kmp_xproc;
- kaapi_all_places[KAAPI_HWS_LEVELID_CORE] = (kaapi_place_t**)__kmp_allocate(
- kaapi_all_places_count[KAAPI_HWS_LEVELID_CORE] * sizeof(kaapi_place_t*)
- );
- kaapi_cpu2numa = (char*)__kmp_allocate( kaapi_all_places_count[KAAPI_HWS_LEVELID_CORE] );
- for (int i =0; i<kaapi_all_places_count[KAAPI_HWS_LEVELID_CORE]; ++i)
- {
-#if LIBOMP_USE_AFFINITY
- int numa_node = numa_node_of_cpu( i );
-#else
- int numa_node = -1;
-#endif
- kaapi_cpu2numa[i] = numa_node;
-
- /* todo : better allocation place[i] on associated NUMA node ! */
- kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i] =
- (kaapi_place_t*)__kmp_allocate(sizeof(kaapi_place_t));
- kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i]->level = KAAPI_HWS_LEVELID_CORE;
- kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i]->queue = (kaapi_wsqueue_t*)
-#if LIBOMP_USE_AFFINITY
- numa_alloc_onnode( sizeof(kaapi_wsqueue_t), numa_node );
-#else
- __kmp_allocate(sizeof(kaapi_wsqueue_t));
-#endif
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i]->queue,
- INITIAL_TASK_DEQUE_SIZE, numa_node
- );
- kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i]->private_queue = (kaapi_wsqueue_t*)
-#if LIBOMP_USE_AFFINITY
- numa_alloc_onnode( sizeof(kaapi_wsqueue_t), numa_node );
-#else
- __kmp_allocate(sizeof(kaapi_wsqueue_t));
-#endif
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_CORE][i]->private_queue,
- INITIAL_TASK_DEQUE_SIZE, numa_node
- );
- }
-
-#if defined(__linux__)
- char filename[sizeof("/sys/devices/system/cpu/cpu1024/topology/thread_siblings_list") + 16];
- FILE *file;
-
-#if !defined(__MIC__)
- /* 1/ parse numa node informations - MIC does not have it (KNC). */
- memcpy(filename,
- "/sys/devices/system/node/online",
- sizeof("/sys/devices/system/node/online"));
- file = fopen(filename,"r");
- if (file !=0)
- {
- char *line = 0;
- char *str = 0;
- size_t len = 0;
- kaapi_assert( getline (&line, &len, file) > 0);
- str = line;
-
- /* scan num '-' num */
- unsigned int start, end;
- kaapi_assert(kaapi_parse_range(&str, &start, &end, 0));
- fclose( file );
- free(line);
-
- /* [OFFLOAD TG]extend count by the number of device memory view as the NUMA domain */
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] = 1+end-start;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA] = (kaapi_place_t**)__kmp_allocate(
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] * sizeof(kaapi_place_t*)
- );
- for (int i=0; i<kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA]; ++i)
- {
- /* todo : better allocation place[i] on i-th NUMA node ! */
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i] =
- (kaapi_place_t*)__kmp_allocate(sizeof(kaapi_place_t));
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->level = KAAPI_HWS_LEVELID_NUMA;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->queue = (kaapi_wsqueue_t*)
-#if LIBOMP_USE_AFFINITY
- numa_alloc_onnode( sizeof(kaapi_wsqueue_t), i );
-#else
- __kmp_allocate(sizeof(kaapi_wsqueue_t));
-#endif
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->queue,
- INITIAL_TASK_DEQUE_SIZE, i
- );
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->private_queue = (kaapi_wsqueue_t*)
-#if LIBOMP_USE_AFFINITY
- numa_alloc_onnode( sizeof(kaapi_wsqueue_t), i );
-#else
- __kmp_allocate(sizeof(kaapi_wsqueue_t));
-#endif
- kaapi_wsqueue_init( kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->private_queue,
- INITIAL_TASK_DEQUE_SIZE, i
- );
-#if 0
- sprintf(filename, "/sys/devices/system/node/node%i/cpulist", i);
- file = fopen(filename,"r");
- kaapi_assert( file != 0 );
- line = 0;
- kaapi_assert( getline (&line, &len, file) > 0);
- str = line;
- fclose(file);
-
- kaapi_cpuset_t* place_i = &affset[KAAPI_HWS_LEVELID_NUMA].cpuset[i];
- KAAPI_CPUSET_ZERO(place_i);
-
- unsigned int cpu;
- kaapi_assert( kaapi_parse_oneplace(&str, &cpu, place_i) );
- _kaapi_init_reversemap(KAAPI_HWS_LEVELID_NUMA, kaapi_default_param.map2affset, place_i);
- free(line);
-#endif
- }
- }
- else
- printf("[kaapi] Cannot access to topology information\n");
-#else
-#error "To redo"
- /* mic has one numa node with all cores inside */
- unsigned int start, end;
- start=0;
- end=1;
- /* [OFFLOAD TG]extend count by the number of device memory view as the NUMA domain */
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] = 1+end-start;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA] = (kaapi_place_t**)__kmp_allocate(
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] * sizeof(kaapi_place_t*)
- );
- for (int i=0; i<kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA]; ++i)
- {
- /* todo : better allocation place[i] on i-th NUMA node ! */
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i] =
- (kaapi_place_t*)__kmp_allocate(sizeof(kaapi_place_t));
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->level = KAAPI_HWS_LEVELID_NUMA;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->queue = 0;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->private_queue = 0;
- }
-#endif
-
-#else // not if defined(__linux__)
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] = 1;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA] = (kaapi_place_t**)__kmp_allocate(
- kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA] * sizeof(kaapi_place_t*)
- );
- for (int i=0; i<kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA]; ++i)
- {
- /* todo : better allocation place[i] on i-th NUMA node ! */
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i] =
- (kaapi_place_t*)__kmp_allocate(sizeof(kaapi_place_t));
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->level = KAAPI_HWS_LEVELID_NUMA;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->queue = 0;
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i]->private_queue = 0;
- }
-
-#endif
-
- return 0;
-}
-
-
-#if defined(__linux__)
-/*
- */
-static inline int is_range_delim(const int c)
-{
- return (c == ':') || (c == '-');
-}
-
-/*
- */
-static inline int is_delim_stride(const int c)
-{
- return (c == ':');
-}
-
-/* end of list
- */
-static inline int is_eol(const int c)
-{
- /* end of list */
- return (c == ',') || (c == 0);
-}
-
-/*
-*/
-static inline void eat_space(char** str)
-{
- while ((*str !=0) && isspace ((int)**str))
- ++*str;
-}
-
-/** Parse :
- [!] [ low ] [ -|: [ high ] [: stride] ]
-
- low -> num | <empty>
- high -> num | <empty>
- if empty detected for low, high then it set value to (unsigned int)-1.
- If stride pointer is not null and no stride is specified, then stride is set to 1.
- If stride pointer is null and a stride is specified, then it is an error
- If low '-' high then, low <= high.
- Else if format is low ':' high, then this is the OpenMP format with low:count.
- */
-static bool kaapi_parse_range(
- char** str,
- unsigned int* index_low,
- unsigned int* index_high,
- int* stride
-)
-{
- char range_delim = ' ';
- if (isdigit(**str))
- {
- if (!kaapi_parse_unsigned_int(str, index_low))
- return false;
- }
- else
- {
- if (!is_range_delim(**str)) /* - or : */
- return false;
- *index_low = (unsigned int)-1;
- }
-
- *index_high = *index_low;
- if (stride !=0)
- *stride = 1;
-
- if (is_range_delim(**str)) /* - or : */
- {
- range_delim = **str;
- ++*str;
-
- /* 1 token look ahead */
- if (is_eol(**str))
- *index_high = (unsigned int)-1;
- else
- {
- if (!kaapi_parse_unsigned_int(str, index_high))
- return false;
-
- if ((stride ==0) && is_delim_stride(**str))
- return false;
-
- if (is_delim_stride(**str))
- {
- ++*str;
- if (is_eol(**str))
- *stride = 1;
- if (!kaapi_parse_int(str, stride))
- return false;
- }
- }
- }
-
- /* test if empty set (index_low-index_high:stride)
- or *stride ==0.
- */
- if (range_delim == ':')
- {
- if (*index_high == -1) *index_high = 1;
- if (*index_high < 1) return false;
- *index_high = *index_low + (*index_high -1) * (stride == 0 ? 1 : *stride);
- }
-
- if ((*index_low > *index_high) || ((stride !=0) && (*stride ==0)))
- return false;
-
- return true;
-}
-
-/* Parse an unsigned long environment varible. Return true if one was
- present and it was successfully parsed. */
-static bool
-kaapi_parse_int (char** str, int *pvalue)
-{
- unsigned long value = 0;
- if (!kaapi_parse_unsigned_long(str, &value))
- return false;
- if ((value > INT_MAX) || ((long)value < INT_MIN))
- return false;
- *pvalue = (int)value;
- return true;
-}
-
-/* Parse an unsigned int environment varible. Return true if one was
- present and it was successfully parsed. */
-static bool
-kaapi_parse_unsigned_int(
- char** str,
- unsigned int *pvalue
-)
-{
- unsigned long long value = 0;
- if (!kaapi_parse_unsigned_longlong(str, &value))
- return false;
- if (value > UINT_MAX)
- return false;
- *pvalue = (unsigned int)value;
- return true;
-}
-
-/* Parse an unsigned long environment varible. Return true if one was
- present and it was successfully parsed. */
-static bool
-kaapi_parse_unsigned_long(
- char** str,
- unsigned long *pvalue
-)
-{
- unsigned long long value = 0;
- if (!kaapi_parse_unsigned_longlong(str, &value))
- return false;
- if (value > ULONG_MAX)
- return false;
- *pvalue = (unsigned long)value;
- return true;
-}
-
-
-/* Parse an unsigned long long environment varible. Return true if one was
- present and it was successfully parsed. */
-static bool
-kaapi_parse_unsigned_longlong(
- char** str,
- unsigned long long *pvalue
-)
-{
- char *end;
- unsigned long long value;
-
- if (*str == 0)
- return false;
-
- eat_space(str);
- if (**str == '\0')
- goto invalid;
-
- errno = 0;
- value = strtoul (*str, &end, 10);
- if (errno || value == ULLONG_MAX)
- goto invalid;
-
- *pvalue = (unsigned long long)value;
- *str = end;
- return true;
-
- invalid:
- return false;
-}
-#endif
diff --git a/runtime/src/kaapi_numaffinity.h b/runtime/src/kaapi_numaffinity.h
deleted file mode 100644
index 571d4ab..0000000
--- a/runtime/src/kaapi_numaffinity.h
+++ /dev/null
@@ -1,91 +0,0 @@
-/*
-** xkaapi
-**
-** Copyright 2009,2010,2011,2012 INRIA.
-**
-** Contributors :
-**
-** thierry.gautier@inrialpes.fr
-**
-** This software is a computer program whose purpose is to execute
-** multithreaded computation with data flow synchronization between
-** threads.
-**
-** This software is governed by the CeCILL-C license under French law
-** and abiding by the rules of distribution of free software. You can
-** use, modify and/ or redistribute the software under the terms of
-** the CeCILL-C license as circulated by CEA, CNRS and INRIA at the
-** following URL "http://www.cecill.info".
-**
-** As a counterpart to the access to the source code and rights to
-** copy, modify and redistribute granted by the license, users are
-** provided only with a limited warranty and the software's author,
-** the holder of the economic rights, and the successive licensors
-** have only limited liability.
-**
-** In this respect, the user's attention is drawn to the risks
-** associated with loading, using, modifying and/or developing or
-** reproducing the software by the user in light of its specific
-** status of free software, that may mean that it is complicated to
-** manipulate, and that also therefore means that it is reserved for
-** developers and experienced professionals having in-depth computer
-** knowledge. Users are therefore encouraged to load and test the
-** software's suitability as regards their requirements in conditions
-** enabling the security of their systems and/or data to be ensured
-** and, more generally, to use and operate it in the same conditions
-** as regards security.
-**
-** The fact that you are presently reading this means that you have
-** had knowledge of the CeCILL-C license and that you accept its
-** terms.
-**
-*/
-#include "kaapi_wsprotocol.h"
-
-#if defined(__cplusplus)
-extern "C" {
-#endif
-
-
-/** \ingroup HWS
- hierarchy level identifiers
- Duplicated in likomp.h
- */
-typedef enum kaapi_hws_levelid
-{
- KAAPI_HWS_LEVELID_THREAD = 0,
- KAAPI_HWS_LEVELID_CORE = 1,
- KAAPI_HWS_LEVELID_NUMA = 2,
- KAAPI_HWS_LEVELID_SOCKET = 3,
- KAAPI_HWS_LEVELID_MACHINE = 4,
- KAAPI_HWS_LEVELID_MAX
-} kaapi_hws_levelid_t;
-
-/** A place : queues with locks and steal context to receive steal request.
- Allocated on page boundary and initialized by one of the threads of the place.
-*/
-typedef struct kaapi_place_t {
- kaapi_hws_levelid_t level; /* kind */
- kaapi_wsqueue_t* queue; /* a queue of task */
- kaapi_wsqueue_t* private_queue; /* private queue to the locality domain */
-} kaapi_place_t;
-
-/** All places as Kaapi exports: NUMA & core places only. Other entries are set to 0
- kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][i] is the pointer to the i-th numa node.
-*/
-extern kaapi_place_t** kaapi_all_places[KAAPI_HWS_LEVELID_MAX];
-
-/** sizes of each kaapi_all_places[level] */
-extern int kaapi_all_places_count[KAAPI_HWS_LEVELID_MAX];
-
-/** Map cpu -> NUMA node
-*/
-extern char* kaapi_cpu2numa;
-
-/** Should be called after kmp_aux_affinity_initialize
-*/
-extern int kaapi_mt_hwdetect(void);
-
-#if defined(__cplusplus)
-}
-#endif
diff --git a/runtime/src/kmp.h b/runtime/src/kmp.h
index 1a2fc81..8744af0 100644
--- a/runtime/src/kmp.h
+++ b/runtime/src/kmp.h
@@ -2257,9 +2257,11 @@ struct kmp_taskdata { /* aligned during dynamic
kmp_int32 td_level; /* task nesting level */
kmp_int32 td_untied_count; /* untied task active parts counter */
ident_t* td_ident; /* task identifier */
+#if LIBOMP_USE_AFFINITY
kmp_uint8 td_aff_kind; // Affinity kind & tag & strict for the task
kmp_uint8 td_aff_strict;
kmp_uint64 td_aff_tag;
+#endif
// Taskwait data.
ident_t * td_taskwait_ident;
kmp_uint32 td_taskwait_counter;
@@ -2322,7 +2324,7 @@ typedef enum machine_levels
typedef struct kmp_base_thread_data {
kmp_info_p * td_thr; // Pointer back to thread info
// Used only in __kmp_execute_tasks_template, maybe not avail until task is queued?
- kmp_queue_data_t *td_tasks_queues[KMP_LEVEL_MAX];
+ kmp_queue_data_t *td_tasks_queues[KMP_LEVEL_MAX] = {0};
// GEH: shouldn't this be volatile since used in while-spin?
kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
#ifdef BUILD_TIED_TASK_STACK
@@ -2357,7 +2359,8 @@ typedef struct kmp_base_task_team {
#if OMP_45_ENABLED
kmp_int32 tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */
#endif
-#if KMP_AFFINITY_SUPPORTED
+ //TODO PV restore this once the initialization of stuff is correct!
+#if 1 || LIBOMP_USE_AFFINITY
// Mask of NUMA nodes active for this team
kmp_affin_mask_t *tt_nodes_mask;
#endif
@@ -2494,6 +2497,8 @@ typedef struct KMP_ALIGN_CACHE kmp_base_info {
kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
#if LIBOMP_USE_AFFINITY
+ // The following stores the affinity set up by the last "__kmpc_omp_set_task_affinity" call
+ // so that we can set the correct affinity in the taskdata on the next task creation call
kmp_uint8 th_aff_kind; // Affinity kind & tag for the next task
kmp_uint8 th_aff_strict;
kmp_uint64 th_aff_tag;
@@ -3232,9 +3237,6 @@ extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
extern void __kmp_balanced_affinity( int tid, int team_size );
-#if LIBOMP_USE_AFFINITY
-extern int kaapi_mt_hwdetect(void);
-#endif
#endif /* KMP_AFFINITY_SUPPORTED */
extern void __kmp_cleanup_hierarchy();
diff --git a/runtime/src/kmp_affinity.cpp b/runtime/src/kmp_affinity.cpp
index f39a748..06d0025 100644
--- a/runtime/src/kmp_affinity.cpp
+++ b/runtime/src/kmp_affinity.cpp
@@ -4095,10 +4095,10 @@ __kmp_affinity_initialize(void)
if (disabled) {
__kmp_affinity_type = affinity_disabled;
}
-#if LIBOMP_USE_AFFINITY
+ //TODO PV restore this when init is fixed
+#if 1 || LIBOMP_USE_AFFINITY
//Also init the places;
machine_places.init();
- kaapi_mt_hwdetect();
#endif
}
diff --git a/runtime/src/kmp_tasking.c b/runtime/src/kmp_tasking.c
index 670a324..69d9396 100644
--- a/runtime/src/kmp_tasking.c
+++ b/runtime/src/kmp_tasking.c
@@ -32,7 +32,6 @@
# define _GNU_SOURCE
#endif
#include <sched.h> /* sched_getcpu */
-#include "kaapi_numaffinity.h"
/*
*/
@@ -324,91 +323,65 @@ __kmp_push_task(kmp_int32 gtid, kmp_task_t * task )
// Find tasking deque specific to encountering thread
thread_data = & task_team -> tt.tt_threads_data[ tid ];
- /*TODO PV: hws push!*/
- kmp_base_queue_data_t *selected_queue = &thread_data->td.td_tasks_queues[KMP_LEVEL_CORE]->qd;
- KMP_DEBUG_ASSERT(selected_queue);
+ kmp_base_queue_data_t *selected_queue = NULL;
+ /*= &thread_data->td.td_tasks_queues[KMP_LEVEL_CORE]->qd;*/
+
- // No lock needed since only owner can allocate
-#if LIBOMP_USE_THEQUEUE
- kaapi_wsqueue_t* pqueue = 0;
-#if defined(__linux__)
- int cpu = sched_getcpu();
-#else
- int cpu = 0;
-#endif
- int local_node = 0;
#if LIBOMP_USE_AFFINITY
- local_node = kaapi_cpu2numa[cpu];
- kaapi_wsqueue_t** ppqueue = 0;
- int isremote = 1; /* always consider 'thief' concurrent push for non own queue */
- int select = taskdata->td_aff_kind;
-#if 0
- /* push at least a task for the current thread */
- if (kaapi_wsqueue_empty(&(selected_queue->td_wsdeque)) && (select !=0) && (taskdata->td_aff_strict==0))
- select = 0;
-#endif
- switch (select)
- {
+ int selected_nodeid = -1;
+ int selected_coreid = thread->th.th_team_nproc;
+ switch (taskdata->td_aff_kind) {
+ /*Default case, just push to our current thread queue!*/
case 0:
- pqueue = &(selected_queue->td_wsdeque);
- isremote = __kmp_get_gtid() == gtid ? 0 : 1;
- break;
-
- case 1: /* data */
- case 2: /* numa id */
- {
- int numaid;
- if (taskdata->td_aff_kind == 1)
- numaid = kaapi_numa_getpage_id( (void*)taskdata->td_aff_tag );
- else
- {
- numaid = (int)taskdata->td_aff_tag;
- if (numaid >= kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA])
- numaid %= kaapi_all_places_count[KAAPI_HWS_LEVELID_NUMA];
+ selected_queue = &thread_data->td.td_tasks_queues[KMP_LEVEL_CORE]->qd;
+ KA_TRACE(5, ( "__kmp_push_task: pushing to own node, no affinity" ) );
+ break;
+ /*affinity "data"*/
+ case 1:
+ /*TODO PV this is a system call, we should improve this!*/
+ selected_nodeid = kaapi_numa_getpage_id( (void*)taskdata->td_aff_tag );
+ KA_TRACE(5, ( "__kmp_push_task: selected node based on mem addr: %i\n", selected_nodeid ) );
+ /*Intentional no break, we need to get the queue!*/
+ /*affinity "node" => numa id. if the numa node is not in the team, push to the first node*/
+ case 2:
+ if (taskdata->td_aff_kind == 2)
+ selected_nodeid = (int)taskdata->td_aff_tag;
+ KA_TRACE(5, ( "__kmp_push_task: selected node based on number: %i\n", selected_nodeid ) );
+ if (KMP_CPU_ISSET(selected_nodeid, task_team->tt.tt_nodes_mask)) {
+ KA_TRACE(5, ( "__kmp_push_task: node is in team!\n" ) );
+ selected_queue = &(__kmp_queue_of_node(selected_nodeid)->qd);
+ } else {
+ KA_TRACE(5, ( "__kmp_push_task: node not in team, select our own node!\n" ) );
+ selected_nodeid = -1;
+ selected_queue = &thread_data->td.td_tasks_queues[KMP_LEVEL_NUMA]->qd;
}
- isremote = (numaid != local_node);
- if (taskdata->td_aff_strict)
- {
- ppqueue = &kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][numaid]->private_queue;
- }
- else
- {
- ppqueue = &kaapi_all_places[KAAPI_HWS_LEVELID_NUMA][numaid]->queue;
- }
-//printf("%i push on numa %i\n", cpu, numaid);
- } break;
-
- case 3: /* core id ~ tid of the thread in the team (%) */
- {
- int coreid = (int)taskdata->td_aff_tag;
- if (coreid >= thread->th.th_team_nproc)
- coreid %=(int)thread->th.th_team_nproc;
- if (taskdata->td_aff_strict)
- {
- ppqueue = &kaapi_all_places[KAAPI_HWS_LEVELID_CORE][coreid]->private_queue;
- }
- else
- {
- ppqueue = &kaapi_all_places[KAAPI_HWS_LEVELID_CORE][coreid]->queue;
- }
- isremote = (coreid != cpu);
- } break;
- };
- KMP_DEBUG_ASSERT ( (ppqueue != 0) || (pqueue !=0) );
-
- if (ppqueue)
- pqueue = *ppqueue;
-
- if (pqueue->deque == NULL ) { /* only for local queue */
- kaapi_wsqueue_init( pqueue, INITIAL_TASK_DEQUE_SIZE, local_node);
+ break;
+ /*affinity "core" => tid or cpu id on the machine, todo -> clarify this*/
+ case 3:
+ selected_coreid = (int)taskdata->td_aff_tag;
+ if (selected_coreid >= thread->th.th_team_nproc)
+ selected_coreid %=(int)thread->th.th_team_nproc;
+ KA_TRACE(5, ( "__kmp_push_task: selected core based on number: %i\n", selected_coreid ) );
+ selected_queue = &(task_team->tt.tt_threads_data[selected_coreid].td.td_tasks_queues[KMP_LEVEL_CORE]->qd);
+ break;
+ default:
+ /* *should* be impossible*/
+ KMP_DEBUG_ASSERT(0);
+ break;
}
#else
- pqueue = &(selected_queue->td_wsdeque);
+ /*If no affinity, we just use cores, that are always available!*/
+ selected_queue = &thread_data->td.td_tasks_queues[KMP_LEVEL_CORE]->qd;
+#endif/* USE_AFFINITY */
+ KMP_DEBUG_ASSERT(selected_queue);
+
+ // No lock needed since only owner can allocate
+#if LIBOMP_USE_THEQUEUE
+ kaapi_wsqueue_t* pqueue = &(selected_queue->td_wsdeque);
if (pqueue->deque == NULL ) {
- kaapi_wsqueue_init( pqueue, INITIAL_TASK_DEQUE_SIZE, -1);
+ kaapi_wsqueue_init( pqueue, INITIAL_TASK_DEQUE_SIZE, selected_nodeid);
}
int isremote = __kmp_get_gtid() == gtid ? 0 : 1;
-#endif
kmp_int32 err;
err = kaapi_wsqueue_locked_push_task(
pqueue,
@@ -1831,12 +1804,7 @@ __kmp_remove_my_task( kmp_info_t * thread, kmp_int32 gtid, kmp_task_team_t *task
if (taskdata) goto igotatask;
}
{ /* scope */
-#if defined(__linux__)
- int cpu = sched_getcpu();
-#else
- int cpu = 0;
-#endif
-#if LIBOMP_USE_AFFINITY /* use steal in place of pop because concurrency required if no aggregation */
+#if 0 && LIBOMP_USE_AFFINITY /* use steal in place of pop because concurrency required if no aggregation */
int local_node = kaapi_cpu2numa[cpu];
#if 0
@@ -1973,19 +1941,24 @@ __kmp_steal_task( kmp_info_t *victim, kmp_int32 gtid, kmp_task_team_t *task_team
taskdata = 0;
-#if LIBOMP_USE_AFFINITY
-#if defined(__linux__)
- int cpu = sched_getcpu();
-#else
- int cpu = 0;
-#endif
- int local_node = kaapi_cpu2numa[cpu];
- kmp_info_t* thread = __kmp_threads[ gtid ];
- for (int ii=0; ii<8; ++ii)
- {
- for (int i=0; i<8; ++i)
+/*#if LIBOMP_USE_AFFINITY*/
+/*#if defined(__linux__)*/
+ /*int cpu = sched_getcpu();*/
+/*#else*/
+ /*int cpu = 0;*/
+/*#endif*/
+ /*int local_node = kaapi_cpu2numa[cpu];*/
+ /*kmp_info_t* thread = __kmp_threads[ gtid ];*/
+ for (int ii=0; ii<8; ++ii)
{
- sched_yield();
+ for (int i=0; i<8; ++i)
+ {
+ sched_yield();
+ if (!kaapi_wsqueue_empty(&(victim_queue->td_wsdeque)))
+ taskdata = kaapi_wsqueue_steal_task( &(victim_queue->td_wsdeque) );
+ }
+ }
+#if 0
#if 0
if (!kaapi_wsqueue_empty(kaapi_all_places[KAAPI_HWS_LEVELID_CORE][cpu]->private_queue))
{
@@ -2046,8 +2019,7 @@ __kmp_steal_task( kmp_info_t *victim, kmp_int32 gtid, kmp_task_team_t *task_team
}
#endif
#endif
- if (!kaapi_wsqueue_empty(&(victim_queue->td_wsdeque)))
- taskdata = kaapi_wsqueue_steal_task( &(victim_queue->td_wsdeque) );
+/*#endif*//*old USE_AFFINITY*/
return_from:
if (taskdata ==0) return 0;
task = KMP_TASKDATA_TO_TASK( taskdata );
@@ -2069,7 +2041,7 @@ return_from:
KMP_COUNT_BLOCK(TASK_stolen);
return task;
-#else
+#else //LIBOMP_USE_THEQUEUE
KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: task_team=%p ntasks=%d "
"head=%u tail=%u\n",
gtid, __kmp_gtid_from_thread( victim ), task_team, victim_queue->td_deque_ntasks,
@@ -2232,7 +2204,10 @@ static inline int __kmp_execute_tasks_template(kmp_info_t *thread, kmp_int32 gti
#if 1//!LIBOMP_USE_THEQUEUE /* with THE queue, always use random ws */
// Try to steal from the last place I stole from successfully.
if (victim == -2) { // haven't stolen anything yet
- victim = own_queue->td_deque_last_stolen;
+ if (own_queue)
+ victim = own_queue->td_deque_last_stolen;
+ else
+ victim = -1;
if (victim != -1) // if we have a last stolen from victim, get the thread
other_thread = threads_data[victim].td.td_thr;
}
--
GitLab