luau/tools/codesizeprediction.py
Vighnesh-V f31232d301
Sync to upstream/release/608 (#1145)
# Old Solver:

- Fix a bug in the old solver where a user could use the keyword
`typeof` as the name of a type alias.
- Fix stringification of scientific notation to omit a trailing decimal
place when not followed by a digit e.g. `1.e+20` -> `1e+20`
# New Solver
- Continuing work on the New non-strict mode
- Introduce `keyof` and `rawkeyof` type function for acquiring the type
of all keys in a table or class
(https://github.com/luau-lang/rfcs/pull/16)

---
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
Co-authored-by: Vighnesh Vijay <vvijay@roblox.com>

---------

Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: David Cope <dcope@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
2024-01-12 14:25:27 -08:00

186 lines
6.3 KiB
Python

#!/usr/bin/python3
# This file is part of the Luau programming language and is licensed under MIT License; see LICENSE.txt for details
# NOTE: This script is experimental. This script uses a linear regression to construct a model for predicting native
# code size from bytecode. Some initial work has been done to analyze a large corpus of Luau scripts, and while for
# most functions the model predicts the native code size quite well (+/-25%), there are many cases where the predicted
# size is off by as much as 13x. Notably, the predicted size is generally better for smaller functions and worse for
# larger functions. Therefore, in its current form this analysis is probably not suitable for use as a basis for
# compilation heuristics. A nonlinear model may produce better results. The script here exists as a foundation for
# further exploration.
import json
import glob
from pathlib import Path
import pandas as pd
import numpy as np
from sklearn.linear_model import LinearRegression
import matplotlib.pyplot as plt
import argparse
def readStats(statsFileGlob):
'''Reads files matching the supplied glob.
Files should be generated by the Compile.cpp CLI'''
statsFiles = glob.glob(statsFileGlob, recursive=True)
print("Reading %s files." % len(statsFiles))
df_dict = {
"statsFile": [],
"script": [],
"name": [],
"line": [],
"bcodeCount": [],
"irCount": [],
"asmCount": [],
"bytecodeSummary": []
}
for statsFile in statsFiles:
stats = json.loads(Path(statsFile).read_text())
for script, filestats in stats.items():
for funstats in filestats["lowerStats"]["functions"]:
df_dict["statsFile"].append(statsFile)
df_dict["script"].append(script)
df_dict["name"].append(funstats["name"])
df_dict["line"].append(funstats["line"])
df_dict["bcodeCount"].append(funstats["bcodeCount"])
df_dict["irCount"].append(funstats["irCount"])
df_dict["asmCount"].append(funstats["asmCount"])
df_dict["bytecodeSummary"].append(
tuple(funstats["bytecodeSummary"][0]))
return pd.DataFrame.from_dict(df_dict)
def addFunctionCount(df):
df2 = df.drop_duplicates(subset=['asmCount', 'bytecodeSummary'], ignore_index=True).groupby(
['bytecodeSummary']).size().reset_index(name='functionCount')
return df.merge(df2, on='bytecodeSummary', how='left')
# def deduplicateDf(df):
# return df.drop_duplicates(subset=['bcodeCount', 'asmCount', 'bytecodeSummary'], ignore_index=True)
def randomizeDf(df):
return df.sample(frac=1)
def splitSeq(seq):
n = len(seq) // 2
return (seq[:n], seq[n:])
def trainAsmSizePredictor(df):
XTrain, XValidate = splitSeq(
np.array([list(seq) for seq in df.bytecodeSummary]))
YTrain, YValidate = splitSeq(np.array(df.asmCount))
reg = LinearRegression(
positive=True, fit_intercept=False).fit(XTrain, YTrain)
YPredict1 = reg.predict(XTrain)
YPredict2 = reg.predict(XValidate)
trainRmse = np.sqrt(np.mean((np.array(YPredict1) - np.array(YTrain))**2))
predictRmse = np.sqrt(
np.mean((np.array(YPredict2) - np.array(YValidate))**2))
print(f"Score: {reg.score(XTrain, YTrain)}")
print(f"Training RMSE: {trainRmse}")
print(f"Prediction RMSE: {predictRmse}")
print(f"Model Intercept: {reg.intercept_}")
print(f"Model Coefficients:\n{reg.coef_}")
df.loc[:, 'asmCountPredicted'] = np.concatenate(
(YPredict1, YPredict2)).round().astype(int)
df['usedForTraining'] = np.concatenate(
(np.repeat(True, YPredict1.size), np.repeat(False, YPredict2.size)))
df['diff'] = df['asmCountPredicted'] - df['asmCount']
df['diffPerc'] = (100 * df['diff']) / df['asmCount']
df.loc[(df["diffPerc"] == np.inf), 'diffPerc'] = 0.0
df['diffPerc'] = df['diffPerc'].round()
return (reg, df)
def saveModel(reg, file):
f = open(file, "w")
f.write(f"Intercept: {reg.intercept_}\n")
f.write(f"Coefficients: \n{reg.coef_}\n")
f.close()
def bcodeVsAsmPlot(df, plotFile=None, minBcodeCount=None, maxBcodeCount=None):
if minBcodeCount is None:
minBcodeCount = df.bcodeCount.min()
if maxBcodeCount is None:
maxBcodeCount = df.bcodeCount.max()
subDf = df[(df.bcodeCount <= maxBcodeCount) &
(df.bcodeCount >= minBcodeCount)]
plt.scatter(subDf.bcodeCount, subDf.asmCount)
plt.title("ASM variation by Bytecode")
plt.xlabel("Bytecode Instruction Count")
plt.ylabel("ASM Instruction Count")
if plotFile is not None:
plt.savefig(plotFile)
return plt
def predictionErrorPlot(df, plotFile=None, minPerc=None, maxPerc=None, bins=200):
if minPerc is None:
minPerc = df['diffPerc'].min()
if maxPerc is None:
maxPerc = df['diffPerc'].max()
plotDf = df[(df["usedForTraining"] == False) & (
df["diffPerc"] >= minPerc) & (df["diffPerc"] <= maxPerc)]
plt.hist(plotDf["diffPerc"], bins=bins)
plt.title("Prediction Error Distribution")
plt.xlabel("Prediction Error %")
plt.ylabel("Function Count")
if plotFile is not None:
plt.savefig(plotFile)
return plt
def parseArgs():
parser = argparse.ArgumentParser(
prog='codesizeprediction.py',
description='Constructs a linear regression model to predict native instruction count from bytecode opcode distribution')
parser.add_argument("fileglob",
help="glob pattern for stats files to be used for training")
parser.add_argument("modelfile",
help="text file to save model details")
parser.add_argument("--nativesizefig",
help="path for saving the plot showing the variation of native code size with bytecode")
parser.add_argument("--predictionerrorfig",
help="path for saving the plot showing the distribution of prediction error")
return parser.parse_args()
if __name__ == "__main__":
args = parseArgs()
df0 = readStats(args.fileglob)
df1 = addFunctionCount(df0)
df2 = randomizeDf(df1)
plt = bcodeVsAsmPlot(df2, args.nativesizefig, 0, 100)
plt.show()
(reg, df4) = trainAsmSizePredictor(df2)
saveModel(reg, args.modelfile)
plt = predictionErrorPlot(df4, args.predictionerrorfig, -200, 200)
plt.show()