colesbury

diff --git a/Makefile.pre.in b/Makefile.pre.in
index 21b122ae0f..37a8b06987 100644
--- a/Makefile.pre.in
+++ b/Makefile.pre.in
@@ -439,6 +439,7 @@ PYTHON_OBJS=	\
 		Python/modsupport.o \
 		Python/mysnprintf.o \
 		Python/mystrtoul.o \
+		Python/object_stack.o \
 		Python/optimizer.o \

colesbury

import time
import json
import math
import subprocess

LOOPS = 1000

def main():
    times = []
    for i in range(1):

colesbury

185.2%    logging_silent
122.9%    deltablue   
107.0%    richards    
 89.2%    unpack_sequence
 87.2%    go          
 71.8%    fannkuch    
 68.1%    scimark_sor 
 61.5%    unpickle_pure_python
 60.8%    pyflate     
 60.4%    logging_simple

colesbury

==================
WARNING: ThreadSanitizer: data race (pid=48373)
  Read of size 8 at 0x7b5000000080 by main thread:
    #0 memcpy <null> (python+0x449377)
    #1 _Py_qsbr_register /private/home/sgross/scratch/nogil/Python/qsbr.c:192:5 (python+0x68b550)
    #2 new_threadstate /private/home/sgross/scratch/nogil/Python/pystate.c:940:20 (python+0x67fc02)
    #3 _PyThreadState_Prealloc /private/home/sgross/scratch/nogil/Python/pystate.c:986:12 (python+0x67fdc7)
    #4 thread_PyThread_start_new_thread /private/home/sgross/scratch/nogil/./Modules/_threadmodule.c:1338:29 (python+0x71ba6e)
    #5 cfunction_call /private/home/sgross/scratch/nogil/Objects/methodobject.c:471:18 (python+0x791ac9)
    #6 _PyObject_MakeTpCall /private/home/sgross/scratch/nogil/Objects/call.c:189:18 (python+0x4e0d46)

colesbury

import threading
from queue import Queue
import numpy as np


def thread1(queue):
    queue.get().fill(5)

def thread2(queue):
    queue.get().resize((1, 1))

colesbury

INFO:root:World size is : 8
INFO:root:Running without GIL
INFO:root:Set start method of multiprocessing to spawn
INFO:root:Using the config 
{"seed": 0, "data": {"dataset": "iamdb", "data_path": "/datasets01/iamdb/060820/", "img_height": 64}, "criterion_type": "transducer", "criterion": {"blank": true, "allow_repeats": false, "ngram": 0}, "model_type": "tds2d", "model": {"depth": 4, "tds_groups": [{"channels": 4, "num_blocks": 3, "stride": [2, 2]}, {"channels": 16, "num_blocks": 3, "stride": [2, 2]}, {"channels": 32, "num_blocks": 3, "stride": [2, 1]}, {"channels": 64, "num_blocks": 3, "stride": [2, 1]}], "kernel_size": [5, 7], "dropout": 0.1}, "optim": {"batch_size": 32, "epochs": 400, "learning_rate": 0.1, "crit_learning_rate": 0.1, "step_size": 100, "max_grad_norm": 5}}
INFO:root:Loading dataset ...
INFO:root:Loading model ...
INFO:root:Training tds2d model with 2,644,859 parameters.
INFO:root:Starting training ...
INFO:root:Epoch 1 started. 

colesbury

import random
import os
import tempfile
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim
import torch.multiprocessing as mp

from torch.nn.parallel import DistributedDataParallel as DDP

colesbury

import gc
from time import sleep
import numpy as np

print('allocating memory')
pile=[]
for i in range(1500):
    for x in range(4):
        pile.append(np.ones((256, 64), dtype=np.float32))
        pile.append(np.ones((256, 64), dtype=np.float32))

colesbury

// Compile with g++ -O3 alloc.cpp -lnuma
#include <vector>
#include <memory>
#include <iostream>
#include <unistd.h>
#include <assert.h>

#include <numa.h>
#include <numaif.h>

colesbury

#include <vector>
#include <memory>
#include <iostream>
#include <unistd.h>
#include <assert.h>

#include <numa.h>
#include <numaif.h>

int USE_MBIND = 0;