vptyp/gstreamer-pipelines
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[autotest] docker execution

Browse Source
This commit is contained in:
Artur 2025-10-10 21:08:12 +03:00
parent f6fd5e50c8
commit 7b92fb6073
5 changed files with 172 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
.gitignore vendored
View File

@ -1,2 +1,7 @@
*.log
container/Drivers/*
container/Drivers/*
__pycache__
*.yuv
*.mp4
*.csv
*log*txt

17
PyScripts/extra.py Normal file
View File

@ -0,0 +1,17 @@
from functools import wraps
def log_args_decorator(func):
"""
A decorator that logs the arguments passed to a function.
"""
@wraps(func)
def wrapper(*args, **kwargs):
arg_names = func.__code__.co_varnames[:func.__code__.co_argcount]
pos_args = dict(zip(arg_names, args))
all_args = {**pos_args, **kwargs}
print(f"Calling function '{func.__name__}' with arguments: {all_args}")
result = func(*args, **kwargs)
print(f"Function '{func.__name__}' returned: {result}")
return result
return wrapper

184
PyScripts/gstreamerAutotest.py
View File

@ -2,17 +2,19 @@
from itertools import product
import qa
from latencyParse import getLatencyTable
import os, stat, subprocess
import pandas as pd
from extra import log_args_decorator
options = {
"x264enc": {
"bitrate": ["10000", "20000", "5000"],
"speed-preset": ["ultrafast", "fast", "medium"],
"tune": ["zerolatency"],
"slices-threads": ["true", "false"],
"sliced-threads": ["true", "false"],
"b-adapt": ["true", "false"],
"rc-lookahead": ["40", "0"],
"ref": ["3", "0"]
},
"nvh264enc": {
"bitrate": ["10000", "20000", "5000"],
@ -20,43 +22,70 @@ options = {
"rc-lookahead": ["0"],
"rc-mode": ["2", "0", "5"],
"zerolatency": ["true", "false"],
},
"nvv4l2h264enc": {
"bitrate": ["10000000", "20000000", "5000000"],
"profile": ["0", "1", "2"],
"preset-id": ["1", "2", "3"],
"control-rate": ["1", "2"],
"tuning-info-id": ["4", "2", "3"]
}
# ,
# "nvv4l2h264enc": {
# "bitrate": ["10000000", "20000000", "5000000"],
# "profile": ["0", "1", "2"],
# "preset-id": ["1", "2", "3"],
# "control-id": ["1", "2"],
# "tuning-info-id": ["4", "2"]
# }
}
videos = [""]
videos = {
"base-daVinci": "./base-daVinci-stereo-left-10.yuv"
}
testsource = "videotestsrc pattern=smpte"
videosrc = ["filesrc location=", "! qtdemux ! h264parse ! avdec_h264"]
videosrc = {
"raw":["filesrc location=", " ! rawvideoparse "],
"h264": ["filesrc location=", " ! decodebin"]
}
psnr_check = {
"x264enc": "-pixel_format yuv420p -color_range pc",
"nvh264enc": "-pixel_format nv12 -color_range tv",
"nvv4l2h264enc": "-pixel_format yuv420p -color_range tv"
"nvv4l2h264enc": "-pixel_format nv12 -color_range tv"
}
with_docker = [ "nvv4l2h264enc" ]
formats = {
"x264enc": "I420",
"nvh264enc": "NV12",
"nvv4l2h264enc": "I420"
"nvv4l2h264enc": "NV12"
}
profiles = ["baseline", "main"]
encoder_prefix = {
"nvv4l2h264enc": " nvvideoconvert !",
"nvh264enc": "",
"x264enc": ""
}
video_info = {
"video1":"-video_size 1920x1080 -framerate 23.98"
"video1":"-video_size 1920x1080 -framerate 23.98",
"sample-surgery":"-video_size 1280x720 -framerate 29.97",
"base-daVinci": "-video_size 1280x720 -framerate 59.94"
}
gst_video_info = {
"video1":"format=I420,height=1080,width=1920,framerate=24000/1001",
"base-daVinci": "format=2 height=720 width=1280 colorimetry=bt601 framerate=60000/1001"
}
latency_filename = "latency-traces-autotest.log"
# Step-by-step:
# 1. Generate all combinations for each encoder
# 2. For each combination, create a GStreamer pipeline string
# 3. Start each pipeline with latency tracing enabled
# 3.1 Monitor CPU, GPU and memory usage during each pipeline run (nah, later, maybe)
# 4. Start latency parsing script after each pipeline and store results in a pandas dataframe:
# - two key columns: encoder name, parameters string
# 5. Run PSNR check after each pipeline and add results in the dataframe
# 6. Save dataframe to CSV file
class Pipeline:
def __init__(self):
self.pipeline = "gst-launch-1.0 -e "
@ -72,33 +101,34 @@ class Pipeline:
return self
def add_source(self, source):
self.pipeline += source + " ! videoconvert ! "
self.pipeline += source + " ! clocksync sync-to-first=true ! videoconvert ! "
return self
def __add_tee(self, encoder):
self.pipeline += "capsfilter caps=video/x-raw,format=" + formats[encoder] + " ! "
self.pipeline += "tee name=t t. ! queue ! filesink location=\"base-autotest.yuv\" "
self.pipeline += "tee name=t t. ! queue max-size-time=5000000000 max-size-bytes=100485760 max-size-buffers=1000 ! filesink location=\"base-autotest.yuv\" "
def add_encoder(self, encoder, params):
self.__add_tee(encoder)
self.options += " ".join(params) + " "
self.pipeline += "t. ! queue ! "
self.pipeline += "t. ! queue max-size-time=5000000000 max-size-bytes=100485760 max-size-buffers=1000 ! "
self.pipeline += encoder_prefix[encoder]
self.pipeline += encoder + " "
self.pipeline += " ".join(params) + " "
return self
def add_profile(self, profile):
self.pipeline += "capsfilter caps=\"video/x-h264,profile=" + profile + "\" ! "
self.pipeline += "! capsfilter caps=\"video/x-h264,profile=" + profile + "\" ! "
self.options += "profile=" + profile + " "
return self
def to_file(self, filename):
self.pipeline += "h264parse ! mp4mux ! filesink location=\"" + filename + "\""
self.pipeline += "h264parse ! mpegtsmux ! filesink location=\"" + filename + "\""
return self
def makeVideoSrc(idx):
return videosrc[0] + videos[idx] + videosrc[1]
def makeVideoSrc(videoName):
return videosrc["raw"][0] + videos[videoName] + videosrc["raw"][1] + gst_video_info[videoName]
def generateEncoderStrings():
@ -133,45 +163,85 @@ def generate_combinations(config_dict):
return combinations
# Step-by-step:
# 1. Generate all combinations for each encoder
# 2. For each combination, create a GStreamer pipeline string
# 3. Start each pipeline with latency tracing enabled
# 3.1 Monitor CPU, GPU and memory usage during each pipeline run (nah, later, maybe)
# 4. Start latency parsing script after each pipeline and store results in a pandas dataframe:
# - two key columns: encoder name, parameters string
# 5. Run PSNR check after each pipeline and add results in the dataframe
# 6. Save dataframe to CSV file
import pandas as pd
qualityDataframe = pd.DataFrame()
latencyDataframe = pd.DataFrame()
dockerRunString = "sudo -S docker container exec deepstream-gst bash"
def run_pipeline(pipeline):
import subprocess
print("Running pipeline:")
print(pipeline)
with open("pipeline-log.txt", "w") as f:
proc = subprocess.run(pipeline, shell=True, stdout=f, stderr=subprocess.STDOUT, text=True)
print(f"Pipeline finished with return code: {proc.returncode}")
with open("pipeline-log.txt", "r") as f:
def execPermissions(scriptFile = "to_exec.sh"):
current_permissions = os.stat(scriptFile).st_mode
new_permissions = current_permissions | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH
os.chmod(scriptFile, new_permissions)
def writeToExecFile(contents, file):
with open(file, "w") as f:
f.write(str(contents))
execPermissions(file)
def is_docker(func):
def wrapper(pipeline):
script_name = "to_exec.sh"
for encoder in with_docker:
if encoder in pipeline:
writeToExecFile(pipeline, script_name)
pipeline = dockerRunString + f" {script_name}"
func(pipeline)
return wrapper
def is_sudo(pipeline):
if pipeline.startswith("sudo"):
return True
return False
def passwordAuth(proc):
password = os.getenv("UAUTH")
if password is not None:
proc.communicate(password)
def printLog(file):
with open(file, "r") as f:
out = f.read()
print(out)
print(out)
@is_docker
@log_args_decorator
def run_pipeline(pipeline):
logfile = "pipeline-log.txt"
with open(logfile, "w") as f:
proc = subprocess.Popen(pipeline, shell=True,
stdin=subprocess.PIPE, stdout=f,
stderr=subprocess.STDOUT, text=True)
if is_sudo(pipeline):
passwordAuth(proc)
code = proc.wait()
printLog(logfile)
if proc.returncode != 0:
raise Exception("Pipeline failed, see log for details")
def time_trace(func):
def wrapper():
import time
start_time = time.time()
func()
end_time = time.time()
elapsed_time = end_time - start_time
print(f"Total execution time: {elapsed_time} seconds")
return wrapper
@time_trace
def run_autotest():
global qualityDataframe, latencyDataframe
encoders = generateEncoderStrings()
for encoder, combinations in encoders.items():
qualityDataframe = pd.DataFrame()
latencyDataframe = pd.DataFrame()
for params in combinations:
for profile in profiles:
for idx in range(len(videos)):
filename = "autotest-" + encoder + "-" + profile + "-test-" + str(idx) + ".mp4"
for videoName, videoPath in videos.items():
filename = "autotest-" + encoder + "-" + profile + "-test-" + videoName + ".mp4"
pipeline = Pipeline()
pipeline = (
pipeline.add_tracing()
.add_source(makeVideoSrc(idx))
.add_source(makeVideoSrc(videoName))
.add_encoder(encoder, params.split(" "))
.add_profile(profile)
.to_file(filename)
@ -183,18 +253,18 @@ def run_autotest():
print(f"Error occurred: {e}")
continue
psnr_metrics, ssim_metrics = qa.run_quality_check(
videos[idx],
videoPath,
filename,
video_info[videos[idx]] + " " + psnr_check[encoder]
video_info[videoName] + " " + psnr_check[encoder]
)
dfPsnr = qa.parse_quality_report(psnr_metrics, ssim_metrics)
print("-----")
dfLatency = getLatencyTable(latency_filename)
columnsQ = pd.MultiIndex.from_tuples(
[(encoder, profile, params, col) for col in dfPsnr.columns]
[(encoder, profile, videoName, params, col) for col in dfPsnr.columns]
)
columnsLatency = pd.MultiIndex.from_tuples(
[(encoder, profile, params, col) for col in dfLatency.columns]
[(encoder, profile, videoName, params, col) for col in dfLatency.columns]
)
dfPsnr.columns = columnsQ
dfLatency.columns = columnsLatency
@ -204,13 +274,7 @@ def run_autotest():
print("Current results:")
print(dfPsnr)
print(dfLatency)
qualityDataframe.to_csv(f"qualityResults{encoder}.csv")
latencyDataframe.to_csv(f"latencyDataframe{encoder}.csv")
def run_timetracer():
import time
start_time = time.time()
run_autotest()
end_time = time.time()
elapsed_time = end_time - start_time
print(f"Total execution time: {elapsed_time} seconds")
run_timetracer()
run_autotest()

30
PyScripts/latencyParse.py
View File

@ -4,16 +4,20 @@ import numpy as np
# Idea is next:
# on set of experiments we are calculating all latency information -> each element avg, std, max numbers, total is not calculated, because it requires
# additional parsing for parallel branches (from tee)
# Ideally we would write data to table
# Ideally we would write data to table
idxCache = dict()
def findWord(words, wordToSearch):
global idxCache
if wordToSearch in idxCache:
for idx in idxCache[wordToSearch]:
if words[idx].startswith(wordToSearch):
if idx < len(words) and words[idx].startswith(wordToSearch):
return words[idx]
else:
if idx >= len(words):
print(f"ERROR: trying to access index={idx} while: {words}")
for word in words:
if word.startswith(wordToSearch):
idx = words.index(word)
@ -24,9 +28,11 @@ def findWord(words, wordToSearch):
return ""
# taken with love from GStreamerLatencyPlotter implementation
def readAndParse(filename):
result = dict()
global idxCache
with open(filename, "r") as latencyFile:
lines = latencyFile.readlines()
for line in lines:
@ -35,12 +41,12 @@ def readAndParse(filename):
words = line.split()
if not words[len(words) - 1].startswith("ts="):
continue
def findAndRemove(wordToSearch):
res = findWord(words, wordToSearch)
res = res[res.find(")") + 1:len(res) - 1]
return res
name = findWord(words, "element=(string)")
if name == "":
name = findWord(words, "src-element=(string)")
@ -49,12 +55,15 @@ def readAndParse(filename):
src = findAndRemove("src=(string)")
name = name[name.find(")") + 1:len(name) - 1]
if name not in result:
result[name] = {"latency":[], "ts":[]}
result[name] = {"latency": [], "ts": []}
timeWord = findAndRemove("time=(guint64)")
tsWord = findAndRemove("ts=(guint64)")
result[name]["latency"].append(int(timeWord)/1e6) # time=(guint64)=14
result[name]["ts"].append(int(tsWord)/1e9) # ts=(guint64)=12
result[name]["latency"].append(
int(timeWord)/1e6) # time=(guint64)=14
result[name]["ts"].append(int(tsWord)/1e9) # ts=(guint64)=12
# drop cache for future runs
idxCache = dict()
return result
@ -78,9 +87,10 @@ def getLatencyTable(filename):
dt_max_latency[column] = dt
df_dt_max = pd.Series(dt_max_latency)
resultDf = pd.concat([df_dt_max, max_latency, avg_latency, median_latency, std_latency], axis=1)
resultDf = pd.concat(
[df_dt_max, max_latency, avg_latency, median_latency, std_latency], axis=1)
resultDf.columns = ['dTmax', 'max', 'avg', 'median', 'std']
print(resultDf)
return resultDf
getLatencyTable("latency_traces-x264enc-kpop-test-10.log")
# getLatencyTable("latency_traces-x264enc-kpop-test-10.log")

5
PyScripts/qa.py
View File

@ -4,6 +4,8 @@ import pandas as pd
def run_psnr_check(original, encoded, video_info):
out = ""
# bad practice, but idgaf
# -f rawvideo {video_info}
options = f"-f rawvideo {video_info} -i {original} -i {encoded} -filter_complex psnr -f null /dev/null"
with open("ffmpeg-log.txt", "w") as f:
proc = subprocess.run(["ffmpeg", *options.split()], stdout=f, stderr=subprocess.STDOUT, text=True)
@ -13,6 +15,9 @@ def run_psnr_check(original, encoded, video_info):
return out
def run_ssim_check(original, encoded, video_info):
# bad practice, but idgaf
# -f rawvideo {video_info}
# we don't need additional information with h264 encoded files
options = f"-f rawvideo {video_info} -i {original} -i {encoded} -filter_complex ssim -f null /dev/null"
with open("ffmpeg-log.txt", "w") as f:
proc = subprocess.run(["ffmpeg", *options.split()], stdout=f, stderr=subprocess.STDOUT, text=True)