org.nd4j.autodiff.listeners.profiler.comparison.ProfileAnalyzer Maven / Gradle / Ivy
The newest version!
/*
* ******************************************************************************
* *
* *
* * This program and the accompanying materials are made available under the
* * terms of the Apache License, Version 2.0 which is available at
* * https://www.apache.org/licenses/LICENSE-2.0.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * Unless required by applicable law or agreed to in writing, software
* * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
* * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
* * License for the specific language governing permissions and limitations
* * under the License.
* *
* * SPDX-License-Identifier: Apache-2.0
* *****************************************************************************
*/
package org.nd4j.autodiff.listeners.profiler.comparison;
import lombok.NonNull;
import lombok.extern.slf4j.Slf4j;
import org.apache.commons.io.FileUtils;
import org.apache.commons.io.IOUtils;
import org.nd4j.autodiff.functions.DifferentialFunction;
import org.nd4j.autodiff.listeners.profiler.ProfilingListener;
import org.nd4j.autodiff.listeners.profiler.data.Phase;
import org.nd4j.autodiff.listeners.profiler.data.TraceEvent;
import org.nd4j.autodiff.listeners.profiler.data.TraceEvents;
import org.nd4j.common.base.Preconditions;
import org.nd4j.imports.converters.DifferentialFunctionClassHolder;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.common.primitives.Pair;
import org.nd4j.list.NDArrayList;
import org.nd4j.shade.jackson.databind.ObjectMapper;
import java.io.*;
import java.nio.charset.Charset;
import java.nio.charset.StandardCharsets;
import java.util.*;
@Slf4j
public class ProfileAnalyzer {
/**
* Chrome profiler supports 2 formats:
* SameDiff == JSON Array Format
* TensorFlow == JSON Object Format
*/
public enum ProfileFormat {SAMEDIFF, TENSORFLOW}
/**
* Only applicable for profile comparisons.
* PROFILE1_PC - sort by profile 1 percentage of total time
* PROFILE2_PC - sort by profile 2 percentage of total time
* RATIO - sort by highest ratio (mean op time profile 1 / mean op time profile 2)
*/
public enum SortBy {PROFILE1_PC, PROFILE2_PC, RATIO}
/**
* TEXT: Human readable, columns padded for alignment
* CSV: CSV format, comma separated
*/
public enum OutputFormat {TEXT,CSV}
/**
* Summarize and print to stdout the specified profile file
*
* @param file Profile file
* @param profileFormat Format of the profiler file
*/
public static void summarizeProfile(File file, ProfileFormat profileFormat) {
System.out.println(summarizeProfileStr(file, profileFormat));
}
/**
* Summarize and return as a string the specified profile file
*
* @param file Profile file
* @param profileFormat Format of the profiler file
*/
public static String summarizeProfileStr(File file, ProfileFormat profileFormat) {
TraceEvent[] events = getTraceEvents(file, profileFormat);
return summarizeTraceEvents(events);
}
/**
* Aggregate, summarize and print to stdout all .json profile files in the specified directory (not recursive)
*
* @param dir Directory containing the profiles
* @param profileFormat Profile format
*/
public static void summarizeProfileDirectory(File dir, ProfileFormat profileFormat) {
System.out.println(summarizeProfileDirectoryStr(dir, profileFormat));
}
/**
* Aggregate, summarize and return as a String all .json profile files in the specified directory (not recursive)
*
* @param dir Directory containing the profiles
* @param profileFormat Profile format
*/
public static String summarizeProfileDirectoryStr(File dir, ProfileFormat profileFormat) {
return summarizeTraceEvents(getTraceEventsDir(dir, profileFormat));
}
/**
* Load, aggregate and return the TraceEvent object from all profiles in the specified directory
*
* @param dir Directory containing the profiles
* @param profileFormat Profile format
*/
public static TraceEvent[] getTraceEventsDir(File dir, ProfileFormat profileFormat) {
File[] files = dir.listFiles();
Preconditions.checkState(files != null && files.length > 0, "No profiles found in directory: %s", dir);
List l = new ArrayList<>();
for (File f : files) {
if (!f.getName().endsWith(".json")) {
log.info("Skipping non-JSON file in directory - {}", f.getAbsolutePath());
continue;
}
TraceEvent[] e = getTraceEvents(f, profileFormat);
Collections.addAll(l, e);
}
return l.toArray(new TraceEvent[0]);
}
/**
* Load and return the TraceEvent object from the specified profile file
*
* @param file Profile file
* @param profileFormat Profile format
*/
public static TraceEvent[] getTraceEvents(File file, ProfileFormat profileFormat) {
return getTraceEvents(file, profileFormat, true);
}
public static TraceEvent[] getTraceEvents(File file, ProfileFormat profileFormat, boolean aggregateTFSubOps) {
ObjectMapper json = ProfilingListener.jsonMapper();
String content = null;
try(BufferedInputStream bufferedInputStream = new BufferedInputStream(new FileInputStream(file))) {
try {
content = IOUtils.toString(bufferedInputStream, Charset.defaultCharset());
} catch (IOException e) {
throw new RuntimeException(e);
}
} catch (FileNotFoundException e) {
e.printStackTrace();
} catch (IOException e) {
e.printStackTrace();
}
if (!content.matches(".*]\\s*")) {
if (content.endsWith(",")) {
//Has comma, missing ]
content = content.substring(0, content.length() - 1) + "]";
} else if (content.endsWith(",\n")) {
//Has comma and newline, missing ]
content = content.substring(0, content.length() - 2) + "]";
} else {
content = content + "]";
}
}
TraceEvent[] events;
if (profileFormat == ProfileFormat.SAMEDIFF) {
try {
events = json.readValue(content, TraceEvent[].class);
} catch (IOException e) {
throw new RuntimeException(e);
}
} else {
//TF format
TraceEvents traceEvents;
try {
traceEvents = json.readValue(content, TraceEvents.class);
} catch (IOException e) {
throw new RuntimeException(e);
}
events = traceEvents.getTraceEvents().toArray(new TraceEvent[0]);
//Clean up TF format - sometimes things like "Softmax" are actually profiled as "_MklSoftmax"
//And we'll align TF names to SameDiff names
for (TraceEvent te : events) {
if (TF_PROFILE_ALIASES.containsKey(te.getName())) {
te.setName(TF_PROFILE_ALIASES.get(te.getName()));
}
DifferentialFunction df = DifferentialFunctionClassHolder.getInstance().getOpWithTensorflowName(te.getName());
if (df != null) {
te.setName(df.opName());
}
}
if(aggregateTFSubOps) {
//For CUDA ops, TF will log sub-ops like:
//fire2/e1x1/Conv2D:Conv2D#id=74,device=/job:localhost/replica:0/task:0/device:GPU:0,async=false#@@cudnn::maxwell::gemm::computeOffsetsKernel(cudnn::maxwell::gemm::ComputeOffsetsParams)
//fire2/e1x1/Conv2D:Conv2D#id=74,device=/job:localhost/replica:0/task:0/device:GPU:0,async=false#@@maxwell_scudnn_128x64_relu_interior_nn
//fire2/e1x1/Conv2D:Conv2D#id=74,device=/job:localhost/replica:0/task:0/device:GPU:0,async=false#@@void tensorflow::functor::ShuffleInTensor3Simple(int, float const*, tensorflow::functor::Dimension<3>, float*)
//We'll join these into one op, then strip everything after the ":" to recover the op name
//Also, TF has multiple sub-ops like this, sequentially, that need to be joined:
//19 = {TraceEvent@3157} "TraceEvent(name=Conv2D#id=80,device=/job, categories=null, ph=X, ts=1576896601259742, dur=466, tts=null, pid=5, tid=0, args={name=conv1/Conv2D, op=Conv2D#id=80,device=/job}, cname=null)"
//20 = {TraceEvent@3181} "TraceEvent(name=Conv2D#id=80,device=/job, categories=null, ph=X, ts=1576896601260229, dur=29, tts=null, pid=5, tid=0, args={name=conv1/Conv2D, op=Conv2D#id=80,device=/job}, cname=null)"
//21 = {TraceEvent@3206} "TraceEvent(name=Conv2D#id=80,device=/job, categories=null, ph=X, ts=1576896601260329, dur=31, tts=null, pid=5, tid=0, args={name=conv1/Conv2D, op=Conv2D#id=80,device=/job}, cname=null)"
//22 = {TraceEvent@3247} "TraceEvent(name=Conv2D#id=80,device=/job, categories=null, ph=X, ts=1576896601260390, dur=4998, tts=null, pid=5, tid=0, args={name=conv1/Conv2D, op=Conv2D#id=80,device=/job}, cname=null)"
Map map = new HashMap<>(); //Key: Op name with ID
List out = new ArrayList<>();
TraceEvent last = null;
for(TraceEvent te : events){
if(last != null && last.getPh() == Phase.X && te.getPh() == Phase.X &&
last.getName().equals(te.getName()) &&
last.getArgs() != null && te.getArgs() != null &&
last.getArgs().get("name").equals(te.getArgs().get("name")) &&
last.getArgs().get("op").equals(te.getArgs().get("op"))){
//Aggregate - same names, ops, etc
last.setDur(last.getDur() + te.getDur());
continue;
}
last = te;
if(te.getArgs() == null || te.getArgs().isEmpty()) {
out.add(te);
continue;
}
String n = (String) te.getArgs().get("name");
//Aggregate by op name...
//"fire2/e1x1/Conv2D:Conv2D#id=74,device=/job:localhost/replica:0/..." -> "fire2/e1x1/Conv2D"
//We're relying on TF's "one iteration per json file" here
if(n.matches("[\\w/_-]+:[\\w/_-]+#id=\\d+.*")) {
int idx = n.indexOf("#");
String sub1 = n.substring(0, idx);
String sub;
if (sub1.contains(":")) {
sub = sub1.substring(0, sub1.lastIndexOf(":"));
} else {
sub = sub1;
}
if (map.containsKey(sub)) {
TraceEvent t = map.get(sub);
Long dur = t.getDur();
if (dur == null && te.getDur() == null)
continue;
t.setDur(dur == null ? te.getDur() : dur + (te.getDur() == null ? 0 : te.getDur()));
} else {
map.put(sub, te);
out.add(te);
}
} else {
if(map.containsKey(n)){
TraceEvent t = map.get(n);
t.setDur(t.getDur() + te.getDur());
} else {
map.put(n, te);
out.add(te);
}
}
}
//Strip everything after ":" in "fire2/e1x1/Conv2D:Conv2D#id=74,device=/job:localhost/..."
for( int i = 0; i < out.size(); i++) {
TraceEvent te = out.get(i);
if(te.getArgs() == null || te.getArgs().isEmpty()){
continue;
}
String n = (String) te.getArgs().get("name");
if(n.matches("[\\w/_-]+:[\\w/_-]+#id=\\d+.*")){
int idx = n.indexOf(':');
String sub = n.substring(0,idx);
te.getArgs().put("name", sub);
}
}
events = out.toArray(new TraceEvent[0]);
}
}
return events;
}
/**
* Summarize the specified TraceEvents as a String
*
* @param events Events to summarize
*/
public static String summarizeTraceEvents(TraceEvent[] events) {
Pair> p = aggregateTraceEvents(events);
final Map stats = p.getSecond();
long allOpsUs = p.getFirst();
//Summarize by op type:
List l = new ArrayList<>(stats.keySet());
Collections.sort(l, new Comparator() {
@Override
public int compare(String o1, String o2) {
return -Long.compare(stats.get(o1).getSumUs(), stats.get(o2).getSumUs());
}
});
//Work out longest name and op name:
int longestName = 30;
int longestOpName = 30;
for (String s : l) {
longestName = Math.max(longestName, s.length() + 1);
longestOpName = Math.max(longestOpName, stats.get(s).getOpName().length() + 1);
}
StringBuilder sb = new StringBuilder();
String headerFormat = "%-" + longestName + "s%-" + longestOpName + "s%-10s%-10s%-10s%-10s%-10s%-10s\n";
sb.append(String.format(headerFormat, "Op Name", "Op", "Count", "Total uS", "%", "Min", "Max", "Std"));
String format = "%-" + longestName + "s%-" + longestOpName + "s%-10d%-10d%-10.2f%-10d%-10d%-10.2f\n";
for (String s : l) {
OpStats st = stats.get(s);
double pc = (100.0 * st.getSumUs()) / allOpsUs;
INDArray arr = st.getTimesUs().array();
long min = arr.minNumber().longValue();
long max = arr.maxNumber().longValue();
double std = arr.stdNumber().doubleValue();
sb.append(String.format(format, s, st.getOpName(), st.getCount(), st.getSumUs(), pc, min, max, std));
}
return sb.toString();
}
private static Pair> aggregateTraceEvents(TraceEvent[] events) {
//Summarize by op (instance) name:
final Map stats = new HashMap<>();
for (TraceEvent e : events) {
if (e.getPh() != Phase.X || e.getDur() == null) {
continue;
}
OpStats s;
String instanceName = (String) e.getArgs().get("name");
if (stats.containsKey(instanceName)) {
s = stats.get(instanceName);
} else {
s = new OpStats(instanceName, e.getName(), 0, new NDArrayList(DataType.LONG, 0), null);
stats.put(instanceName, s);
}
s.setCount(s.getCount() + 1);
s.getTimesUs().add((double) e.getDur());
}
long allOpsUs = 0;
for (OpStats s : stats.values()) {
s.setSumUs( s.getTimesUs().array().sumNumber().longValue());
allOpsUs += s.getSumUs();
}
return new Pair<>(allOpsUs, stats);
}
/**
* Compare the specified profile files, sorted by profile 1 % of total time
*
* @param file1 First profile file
* @param file2 Second profile file
* @param format1 Format of first profile
* @param format2 Format of second profile
* @return Comparison summary as a String
*/
public static String compareProfiles(@NonNull File file1, @NonNull File file2, @NonNull ProfileFormat format1, @NonNull ProfileFormat format2) {
return compareProfiles(file1, file2, format1, format2, false, false, null, null, SortBy.PROFILE1_PC);
}
/**
* Compare the specified profile files or directory
*
* @param file1 First profile file or directory of profiles
* @param file2 Second profile file or directory of profiles
* @param format1 Format for first profile file/s
* @param format2 Format for second profile file/s
* @param firstIsDir True if the first File object is a directory
* @param secondIsDir True if the second File object is a directory
* @param name1 Name of the first profile (just for display purposes). Optional
* @param name2 Name of the second profile (just for display purposes). Optional
* @param sortBy What to sort the summary results by
* @return Comparison summary as a String
*/
public static String compareProfiles(@NonNull File file1, @NonNull File file2, @NonNull ProfileFormat format1, @NonNull ProfileFormat format2,
boolean firstIsDir, boolean secondIsDir, String name1, String name2, final SortBy sortBy) {
return compareProfiles(Config.builder()
.profile1(file1)
.profile2(file2)
.profile1Format(format1)
.profile2Format(format2)
.profile1IsDir(firstIsDir)
.profile2IsDir(secondIsDir)
.p1Name(name1)
.p2Name(name2)
.sortBy(sortBy)
.build());
}
public static String compareProfiles(final Config c){
TraceEvent[] t1 = c.profile1IsDir() ? getTraceEventsDir(c.profile1(), c.profile1Format()) : getTraceEvents(c.profile1(), c.profile1Format());
TraceEvent[] t2 = c.profile2IsDir() ? getTraceEventsDir(c.profile2(), c.profile2Format()) : getTraceEvents(c.profile2(), c.profile2Format());
final Pair> p1 = aggregateTraceEvents(t1);
final Pair> p2 = aggregateTraceEvents(t2);
List l = new ArrayList<>(c.sortBy() != SortBy.PROFILE2_PC ? p1.getSecond().keySet() : p2.getSecond().keySet());
Collections.sort(l, new Comparator() {
@Override
public int compare(String o1, String o2) {
switch (c.sortBy()) {
case PROFILE1_PC:
return -Long.compare(p1.getSecond().get(o1).getSumUs(), p1.getSecond().get(o2).getSumUs());
case PROFILE2_PC:
return -Long.compare(p2.getSecond().get(o1).getSumUs(), p2.getSecond().get(o2).getSumUs());
case RATIO:
double m1a = meanTime(p1, o1);
double m1b = meanTime(p1, o2);
double m2a = meanTime(p2, o1);
double m2b = meanTime(p2, o2);
double ratio1 = m1a / m2a;
double ratio2 = m1b / m2b;
return -Double.compare(ratio1, ratio2);
default:
throw new RuntimeException();
}
}
});
Set set = new HashSet<>(l);
StringBuilder sb = new StringBuilder();
sb.append("1 = ").append(c.p1Name() == null ? "Profile 1" : c.p1Name()).append("\n")
.append("2 = ").append(c.p2Name() == null ? "Profile 2" : c.p2Name()).append("\n");
//Work out longest name and op name:
int longestName = 30;
int longestOpName = 30;
Map stats = c.sortBy() == SortBy.PROFILE2_PC ? p2.getSecond() : p1.getSecond();
for (String s : l) {
longestName = Math.max(longestName, s.length() + 1);
longestOpName = Math.max(longestOpName, stats.get(s).getOpName().length() + 1);
}
String headerFormat;
String format;
if(c.format() == null || c.format() == OutputFormat.TEXT){
headerFormat = "%-" + longestName + "s%-" + longestOpName + "s%-10s%-10s%-16s%-13s%-13s%-14s%-14s%-12s%-12s%-14s%-14s%-10s%-10s%-10s%-10s\n";
format = "%-" + longestName + "s%-" + longestOpName + "s%-10d%-10d%-16.2f%-13.2f%-13.2f%-14d%-14d%-12.2f%-12.2f%-14d%-14d%-10d%-10d%-10.2f%-10.2f\n";
} else {
headerFormat = "%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
format = "%s,%s,%d,%d,%.2f,%.2f,%.2f,%d,%d,%.2f,%.2f,%d,%d,%d,%d,%.2f,%.2f\n";
}
sb.append(String.format(headerFormat, "Op Name", "Op", "Count (1)", "Count (2)", "Mean Ratio 1/2", "Mean (1)", "Mean (2)", "Total uS (1)", "Total uS (2)", "% (1)", "% (2)", "Min (1)", "Min (2)", "Max (1)", "Max (2)", "Std (1)", "Std (2)"));
for (String s : l) {
OpStats s1 = p1.getSecond().get(s);
OpStats s2 = p2.getSecond().get(s);
if(c.filter() != null && !c.filter().apply(s1, s2))
continue;
double m1 = s1 == null ? 0 : s1.getTimesUs().array().meanNumber().doubleValue();
double m2 = s2 == null ? 0 : s2.getTimesUs().array().meanNumber().doubleValue();
double ratio = m1 / m2;
double pc1 = s1 == null ? 0 : 100.0 * s1.getSumUs() / p1.getFirst();
double pc2 = s2 == null ? 0 : 100.0 * s2.getSumUs() / p2.getFirst();
sb.append(String.format(format, s, s1 != null ? s1.getOpName() : s2.getOpName(),
s1 != null ? s1.getCount() : 0,
s2 != null ? s2.getCount() : 0,
//Ratio of means, means
ratio,
m1, m2,
//Total us, percent of op total
s1 != null ? s1.getSumUs() : 0,
s2 != null ? s2.getSumUs() : 0,
pc1, pc2,
//Min, max, std
s1 != null ? s1.getTimesUs().array().minNumber().longValue() : 0,
s2 != null ? s2.getTimesUs().array().minNumber().longValue() : 0,
s1 != null ? s1.getTimesUs().array().maxNumber().longValue() : 0,
s2 != null ? s2.getTimesUs().array().maxNumber().longValue() : 0,
s1 != null ? s1.getTimesUs().array().stdNumber().doubleValue() : 0.0,
s2 != null ? s2.getTimesUs().array().stdNumber().doubleValue() : 0.0));
}
boolean header = false;
String headerFormat2 = null;
String format3 = null;
List toAppend = null;
for (String s : (c.sortBy() == SortBy.PROFILE2_PC ? p1.getSecond().keySet() : p2.getSecond().keySet())) {
if (!set.contains(s)) {
Map m = c.sortBy() == SortBy.PROFILE2_PC ? p1.getSecond() : p2.getSecond();
OpStats st = m.get(s);
if(c.filter() != null){
OpStats other = c.sortBy() == SortBy.PROFILE2_PC ? p1.getSecond().get(s) : p2.getSecond().get(s);
boolean keep = c.filter().apply(other, st);
if(!keep)
continue;
}
if (!header) {
toAppend = new ArrayList<>();
longestName = 30;
longestOpName = 30;
for(String s2 : m.keySet()){
longestName = Math.max(longestName, s2.length()+1);
longestOpName = Math.max(longestOpName, m.get(s2).getOpName().length()+1);
}
if(c.format() == null || c.format() == OutputFormat.TEXT) {
headerFormat2 = "%-" + longestName + "s%-" + longestOpName + "s%-10s%-10s%-10s%-10s%-10s%-10s\n";
format3 = "%-" + longestName + "s%-" + longestOpName + "s%-10d%-10d%-10.2f%-10d%-10d%-10.2f\n";
} else {
headerFormat2 = "%s,%s,%s,%s,%s,%s,%s,%s\n";
format3 = "%s,%s,%d,%d,%.2f,%d,%d,%.2f\n";
}
sb.append(" *** Operations not in profile ").append(c.sortBy() == SortBy.PROFILE2_PC ? "1" : "2").append(" but in profile ")
.append(c.sortBy() == SortBy.PROFILE2_PC ? "2" : "1").append(" ***\n");
sb.append(String.format(headerFormat2, "Op Name", "Op", "Count", "Total uS", "%", "Min", "Max", "Std"));
header = true;
}
long allOpsUs = c.sortBy() == SortBy.PROFILE2_PC ? p1.getFirst() : p2.getFirst();
double pc = (100.0 * st.getTimesUs().array().sumNumber().longValue()) / allOpsUs;
INDArray arr = st.getTimesUs().array();
long min = arr.minNumber().longValue();
long max = arr.maxNumber().longValue();
double std = arr.stdNumber().doubleValue();
toAppend.add(String.format(format3, s, st.getOpName(), st.getCount(), st.getSumUs(), pc, min, max, std));
}
}
if(toAppend != null){
Collections.sort(toAppend);
for(String s : toAppend){
sb.append(s);
}
}
return sb.toString();
}
private static double meanTime(Pair> p, String name) {
if (!p.getSecond().containsKey(name)) {
return 0.0;
}
return p.getSecond().get(name).getTimesUs().array().meanNumber().doubleValue();
}
private static Map TF_PROFILE_ALIASES = new HashMap<>();
static {
TF_PROFILE_ALIASES.put("_MklSoftmax", "Softmax");
}
}