jlebar’s gists

jlebar / run_resnet50.sh

Created November 14, 2018 21:27

Run resnet50 benchmark on GCE

	gcloud compute ssh $INSTANCE_NAME

	# Clone TensorFlow benchmark repository.
	git clone https://github.com/tensorflow/benchmarks.git && cd benchmarks
	git reset --hard 1e7d788042dfc6d5e5cd87410c57d5eccee5c664
	cd scripts/tf_cnn_benchmarks

	## Synthetic data test

	# 8 GPUs

jlebar / create_gce_instance.sh

Created November 14, 2018 21:27

Create a GCE instance with 8 V100 GPUs

	export INSTANCE_NAME="xla-benchmark-8xV100"
	export IMAGE_FAMILY="tf-1-12-cu100"
	export PROJECT_NAME="<your project name>"

	gcloud beta compute instances create $INSTANCE_NAME \
	--project=$PROJECT_NAME \
	--machine-type=n1-standard-64 \
	--maintenance-policy=TERMINATE \
	--accelerator=type=nvidia-tesla-v100,count=8 \
	--tags=http-server,https-server \

jlebar / unsupported_op.py

Created November 7, 2018 01:21

Example of TF op that's not supported by XLA

	def model_fn_unsupported_op(x):
	return tf.where(tf.cast(x, tf.bool))

jlebar / tf_dynamic_shapes.py

Created November 7, 2018 01:21

Example of TF function with dynamic but nonetheless inferrable shapes

	def model_fn_changing_shapes(x):
	return 2 * x

	def run_changing_shapes_model():
	with tf.Session() as sess:
	x = tf.placeholder(tf.float32, name='x')
	result = xla.compile(model_fn_changing_shapes, (x,))[0]
	a = sess.run(result, feed_dict={x: [1., 2.]})
	b = sess.run(result, feed_dict={x: [1., 2., 3.]})

jlebar / uninferrable_shapes.py

Created November 7, 2018 01:20

Example of TF functions with uninferrable shapes

	def model_fn_random_shape():
	random_dim_size = tf.random_uniform(
	shape=[], minval=0, maxval=5, dtype=tf.int32)
	# Return a vector with a random number of elements, all of them 42.0
	return tf.fill([random_dim_size], 42.)

	def run_random_shapes_model():
	with tf.Session() as sess:
	x = tf.placeholder(tf.float32, name='x')
	result = xla.compile(model_fn_random_shape)[0]

jlebar / xla_compile_switch.py

Created November 7, 2018 01:19

Example of switching xla.compile on or off via a flag

	if should_use_xla():
	result = xla.compile(model_fn, (x, y, z))[0]
	else:
	result = model_fn(x, y, z)

jlebar / xla_compile.py

Created November 7, 2018 01:18

Simple example of using xla.compile

	from tensorflow.contrib.compiler import xla

	def model_fn(x, y, z):
	return tf.reduce_sum(x + y * z)

	def create_and_run_graph():
	with tf.Session() as sess:
	x = tf.placeholder(tf.float32, name='x')
	y = tf.placeholder(tf.float32, name='y')
	z = tf.placeholder(tf.float32, name='z')

jlebar / tf_computation.py

Created November 7, 2018 01:17

A simple TensorFlow computation

	def model_fn(x, y, z):
	return tf.reduce_sum(x + y * z)

jlebar / gist:5367521

Last active December 16, 2015 03:09

Spawns one thread, then does a series of computations on the main thread and on the spawned thread. Periodically outputs the speed of those two computations.

	// Compile me with -lrt -lpthread -lm.
	//
	// To force this program to run on one core, use \|taskset 1\|.
	//
	// See post on http://jlebar.com for why this is relevant.

	#include <pthread.h>
	#include <math.h>
	#include <stdio.h>
	#include <unistd.h>

jlebar / gist:4731401

Created February 7, 2013 14:53

Convert hg patches to git patches. curl -L hg_patch | this-script | git am

	#!/usr/bin/env python

	r'''Convert an hg-exported patch to a patch suitable for use by git am.

	>>> hg_patch_to_git_patch(StringIO('# HG changeset patch\n# User Foo <[email protected]>\n# Node ID deadbeef\n# Parent cafebabe\nCommit\n\nMsg\n\ndiff -\ndiffdiff'))
	From: Foo <[email protected]>
	Subject: Commit
	<BLANKLINE>
	Msg
	<BLANKLINE>

Justin Lebar jlebar