dl.khadas.com - Index of /firmware/Krescue/images/edge2/
Ubuntu 20.02 & Mali Drivers
glmark2 is a very synthetic benchmark, so while it can be useful for driver developers, it is not a good way of measuring the actual performance of a GPU. 2000 vs 4000 FPS means nothing for actual performance.
Why not fire up an actual game like Xonotic on that Khadas image and compare… wait no you can’t, you’ll only get llvmpipe. Xonotic does work with that PPA; surprisingly well if I say so myself. (At least if there were no mistakes in building the PPA; I haven’t got Ubuntu installed so I can’t test it.)
Yes, performance is sometimes bad even in real applications, such as SuperTuxKart
with the “Advanced pipeline” enabled, but that can be fixed in time. It will take a lot of effort to do so, though. More important to me is compatibility with applications that will never run with the blob.
One thing that often does help a lot is using the performance governor for the CPU and optionally DRAM:
$ glmark2-es2-wayland | tail -n3
=======================================================
glmark2 Score: 1129
=======================================================
$ echo performance | sudo tee /sys/bus/cpu/devices/cpu[046]/cpufreq/scaling_governor /sys/class/devfreq/dmc/governor
$ glmark2-es2-wayland | tail -n3
=======================================================
glmark2 Score: 2089
=======================================================
(That was using Sway.)
About stability: Yesterday I fixed a bug that caused crashes in applications that had multiple windows open, but that change is not in the PPA yet. Maybe it will fix the crashes you are seeing.
2 Likes
Here are the generic and emmc Ubuntu images for Edge2
I don’t think that Khadas are doing anything special, it looks to me like stock Ubuntu packages except for the mali blob.
One difference is the version of glmark2—I’m using Debian sid, which has a newer version that also runs quite a bit slower.
Using their glmark2-es2-wayland with blob mutter and adding taskset to only run on the big cores:
glmark2 Score: 4637
All you can do with synthetics is run like for like and when you run on a different distro and use taskset you are no longer like for like.
The only thing I can see in the cnx soft example is switching the performance governor to performance, but not using taskset.
Can we run a like for like? Also we don’t have to use repo versions as the compile of the master is very easy so we can be sure of like for like.
I have no idea with Xonotic as presume it requires X11 as we still have this transition or preference to deal with, but if it doesn’t run then pick something that does under Wayland than something that does not.
I have been using the ArmNN delegate for TFlite and the difference between CpuAcc & GpuAcc is just a slight reduction in load of all cores when running GpuAcc. (CpuAcc (Neon), CpuRef (Non Neon))
Dram, like all integrated graphics has much effect but pretty sure the Khadas is using the same 2112Mhz DDR4 (or is it?).
Yes you can change the environment of synthetics and greatly change results but that is not the point as all we need to do is create like for like environments as close as we can and use it purely as a rough metric as gaming benchmarks have for decades.
With what I was doing with ArmNN its got me scratching my head because for a GPU driver much work would seem to be CPU based?!
There are a load of apps we could choose doesn’t have to be ArmNN but could we choose one and get access to a stable version of Wayland on ubuntu 22.04 as you have pasted a figure and so far on any environment I have not been anywhere close to that?
I don’t really care about games as the GPU with OpenCL has more uses than that but a game that runs is as equally valid if its optimised to use a GPU.
I am trying to figure out my possible usages and that maybe things like Mesa and Vulkan may be needed and I am likely in for a long wait?
With OpenCl I can just point the icd loader to the driver without this confusion of wayland/x11 and trying to get display accelerated graphics and results don’t seem to be that great.
On the Radxa Debian image with the governor set to performance I get currently glmark2-es2
glmark2 Score: 986
Now what I need to do is set up a similar environment and also use the performance governor to get another approximate metric of a different GPU.
Even with taskset the only way I can get close to the above is by copying & pasting glmark2 Score: 4637
When using the userspace stack from the Khadas images, and the performance governor but not taskset, the result is just exactly the same as CNX Software got—just a tad above 4000. Using taskset -c 4-7 on gnome-shell and glmark2-es2-wayland improves the result by about 10%, and the rest of the gain is from switching the compositor to a newer mutter from Debian sid.
For the result you get with the Debian image, a lot of that is just because the mali blob is not well integrated with Xorg. There seem to be some hacks in the Wayland blob to improve performance beyond what should be possible given the kernel limitations, possibly at the cost of latency.
I can try to look to see why ArmNN doesn’t work as well as you expect it to…
Have you managed to add the fixes to ppa:liujianfeng1994/panfork-mesa as would love to get an image running with wayland and mutter.
Not so bothered about the 10% of the affinity as the cnx example doesn’t seem to do so but often its a case of robbing Paul to pay Peter as someway along the line something is going to lose out to the normal load balancing of the scheduler.
Yeah you can give ArmNN a go as prob its a really good bench as benches go with a finite model and the ability to flick between GpuAcc & CpuAcc on the cli.
Its https://developer.arm.com/documentation/102603/2108/Device-specific-installation/Install-on-Odroid-N2-Plus
Also like what seem to be all Arm software doc’s its slightly out of date and you need to debug as you go but you are obviously more than capable.
sudo apt-get install -y python3-pyarmnn armnn-latest-all than sudo apt-get install -y python3-pyarmnn libarmnn-latest-all
python3 run_audio_file.py --audio_file_path tests/testdata/quick_brown_fox_16000khz.wav --model_file_path tflite_int8/wav2letter_int8.tflite --preferred_backends CpuAcc CpuRef than python3 run_audio_file.py --audio_file_path tests/testdata/quick_brown_fox_16000khz.wav --model_file_path tflite_int8/wav2letter_int8.tflite --labels_file_path tests/testdata/wav2letter_labels.txt --preferred_backends CpuAcc CpuRef as the labels are in the python code.
I don’t think it will matter as linking to the x11 libmali-valhall-g610-g6p0-x11 just works bad as x11 does but presume libmali-valhall-g610-g6p0-wayland or if libmali-valhall-g610-g6p0-x11-wayland was linked to the icd loader instead then like wayland overall results would be better.
I am completely lost with the armbian image trying to work backwards what was actually driving Dri with everything that is installed in the base image is confusing.
It would be really interesting to be able to link against the wayland driver than x11
Thanks.
That ArmNN sample seems to spend a lot of time doing preprocessing of audio data. It’s written in Python, which isn’t exactly the fastest language in the world, but it also does a number of things which cause significant slowdown: doing element-wise calculation rather than doing operations on whole arrays at once, and recalculating constant data.
Fixing those significantly improves speed compared to realtime on a five-minute test file:
| CpuAcc | GpuAcc | |
|---|---|---|
| Before | 3.3x | 3.1x |
| After | 26.2x | 17.7x |
The GPU backend has some constant overhead, so testing again with a one-hour file to be more fair:
| CpuAcc | GpuAcc | |
|---|---|---|
| After | 28.1x | 25.6x |
(I’m not waiting forty minutes to test the hour-long file with the old code.)
There is also a C++ sample which is similar, but it also does a number of things that hurt performance: implementing a “fast” Fourier transform naïvely, and doing too high quality resampling.
Again, big speed improvements from fixing those:
One-minute file:
| CpuAcc | GpuAcc | |
|---|---|---|
| Before | 1.5x | 1.3x |
| After | 32.3x | 8.7x |
One-hour file:
| CpuAcc | GpuAcc | |
|---|---|---|
| After | 40.9x | 35.5x |
However, both CPU and GPU performance could still be improved—it appears that only four cores are used for CPU processing, and the GPU load is only 60% even after my tweaks. This could be fixed by processing multiple blocks of audio at the same time.
I haven’t yet investigated why the GPU backend is not faster than the CPU—maybe the model is just too small for the GPU to show much benefit?
Here are the patches I used:
armnn-patches.zip (2.5 KB)
3 Likes
The GPU backend has about a 5 sec load that the CPU doesn’t.
Plus yeah Python isn’t the fastest and it really sucks when iterating DSP like data.
I hacked the code so the model stays in memory and it runs twice and that takes off quite a bit, well the 5 sec delay that seems constant.
The code and model is actually pretty terrible but it was just purely testing ArmNN I should of done the same with the mfcc as load once run twice.
# Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
# SPDX-License-Identifier: MIT
"""Automatic speech recognition with PyArmNN demo for processing audio clips to text."""
import sys
import os
import numpy as np
import psutil
script_dir = os.path.dirname(__file__)
sys.path.insert(1, os.path.join(script_dir, '..', 'common'))
from argparse import ArgumentParser
from network_executor import ArmnnNetworkExecutor
from utils import prepare_input_data
from audio_capture import AudioCaptureParams, capture_audio
from audio_utils import decode_text, display_text
from wav2letter_mfcc import Wav2LetterMFCC, W2LAudioPreprocessor
from mfcc import MFCCParams
from datetime import datetime
# Model Specific Labels
labels = {0: 'a', 1: 'b', 2: 'c', 3: 'd', 4: 'e', 5: 'f', 6: 'g', 7: 'h', 8: 'i', 9: 'j', 10: 'k', 11: 'l', 12: 'm',
13: 'n',
14: 'o', 15: 'p', 16: 'q', 17: 'r', 18: 's', 19: 't', 20: 'u', 21: 'v', 22: 'w', 23: 'x', 24: 'y',
25: 'z',
26: "'", 27: ' ', 28: '$'}
def parse_args():
parser = ArgumentParser(description="ASR with PyArmNN")
parser.add_argument(
"--audio_file_path",
required=True,
type=str,
help="Path to the audio file to perform ASR",
)
parser.add_argument(
"--model_file_path",
required=True,
type=str,
help="Path to ASR model to use",
)
parser.add_argument(
"--preferred_backends",
type=str,
nargs="+",
default=["GpuAcc", "CpuAcc", "CpuRef"],
help="""List of backends in order of preference for optimizing
subgraphs, falling back to the next backend in the list on unsupported
layers. Defaults to [GpuAcc, CpuAcc, CpuRef]""",
)
return parser.parse_args()
def main(args, network):
# Read command line args
audio_file = args.audio_file_path
print(datetime.now() - starttime, psutil.cpu_percent())
print(datetime.now() - starttime, psutil.cpu_percent())
# Specify model specific audio data requirements
audio_capture_params = AudioCaptureParams(dtype=np.float32, overlap=31712, min_samples=47712, sampling_freq=16000,
mono=True)
buffer = capture_audio(audio_file, audio_capture_params)
print(datetime.now() - starttime, psutil.cpu_percent())
# Extract features and create the preprocessor
mfcc_params = MFCCParams(sampling_freq=16000, num_fbank_bins=128, mel_lo_freq=0, mel_hi_freq=8000,
num_mfcc_feats=13, frame_len=512, use_htk_method=False, n_fft=512)
print(datetime.now() - starttime, psutil.cpu_percent())
wmfcc = Wav2LetterMFCC(mfcc_params)
preprocessor = W2LAudioPreprocessor(wmfcc, model_input_size=296, stride=160)
current_r_context = ""
is_first_window = True
print("Processing Audio Frames...")
for audio_data in buffer:
# Prepare the input Tensors
input_data = prepare_input_data(audio_data, network.get_data_type(), network.get_input_quantization_scale(0),
network.get_input_quantization_offset(0), preprocessor)
# Run inference
output_result = network.run([input_data])
# Slice and Decode the text, and store the right context
current_r_context, text = decode_text(is_first_window, labels, output_result)
is_first_window = False
display_text(text)
print(datetime.now() - starttime, psutil.cpu_percent())
print(current_r_context, flush=True)
print(datetime.now() - starttime, psutil.cpu_percent())
print("Inference End", psutil.cpu_percent())
if __name__ == "__main__":
args = parse_args()
print("Inference Start", psutil.cpu_percent())
starttime = datetime.now()
# Create the ArmNN inference runner
network = ArmnnNetworkExecutor(args.model_file_path, args.preferred_backends)
print(datetime.now() - starttime, psutil.cpu_percent())
main(args, network)
starttime = datetime.now()
print(datetime.now() - starttime, psutil.cpu_percent())
main(args, network)
I was using these as longer but not so long from https://github.com/ggerganov/whisper.cpp which is a cpu version of OPenAI’s Whisper based on his own tensor lib which is interesting
#
# Audio samples
#
# download a few audio samples into folder "./samples":
.PHONY: samples
samples:
@echo "Downloading samples..."
@mkdir -p samples
@wget --quiet --show-progress -O samples/gb0.ogg https://upload.wikimedia.org/wikipedia/commons/2/22/George_W._Bush%27s_weekly_radio_address_%28November_1%2C_2008%29.oga
@wget --quiet --show-progress -O samples/gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
@wget --quiet --show-progress -O samples/hp0.ogg https://upload.wikimedia.org/wikipedia/en/d/d4/En.henryfphillips.ogg
@wget --quiet --show-progress -O samples/mm1.wav https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav
@echo "Converting to 16-bit WAV ..."
@ffmpeg -loglevel -0 -y -i samples/gb0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb0.wav
@ffmpeg -loglevel -0 -y -i samples/gb1.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb1.wav
@ffmpeg -loglevel -0 -y -i samples/hp0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/hp0.wav
@ffmpeg -loglevel -0 -y -i samples/mm1.wav -ar 16000 -ac 1 -c:a pcm_s16le samples/mm0.wav
@rm samples/mm1.wav
I will give them another go even though the model we have is pretty bad https://github.com/breizhn/DTLN might be better as its also 2 models and that would be really interesting to use GpuAcc on one and CpuAcc on another but from the load it looks almost like GpuAcc is more of a helper than removing most load from Cpu such as Cuda.
You should also be able to take a single big model and partition the layers and run with 2x delegates which is what I was wondering.
Its a shame RkNN didn’t go the delegate route.
I was lazy as it chunks the audio through the mfcc and very possible to just convert the whole audio and chunk the premade mfcc through the model, but didn’t bother.
May do as the interest is purely in the delegate and gpu vs cpu not some horrid mfcc.
Its strange as tensorflow has mfcc ops and you can subclass them into a model as does https://github.com/google-research/google-research/tree/master/kws_streaming or even librosa is more performant than the armnn example
Thanks though as now it makes so much sense why the cpu is being hit.
As is with samples/hp0.wav
GpuAcc CpuAcc CpuRef
0:01:29.798073 0.0
Inference End 0.0
CpuAcc CpuRef
0:01:33.897910 0.0
Inference End 0.0
sudo apt-get install irqbalance I did notice the irq affinity is near all on core0 and irqbalance does swap to core4 (big) [But not a lot of difference]
GpuAcc CpuAcc CpuRef
0:01:29.660175 0.0
Inference End 0.0
CpuAcc CpuRef
0:01:31.325590 50.0
Inference End 0.0
echo performance > /sys/devices/system/cpu/cpufreq/policy0/scaling_governor
echo performance > /sys/devices/system/cpu/cpufreq/policy4/scaling_governor
echo performance > /sys/devices/system/cpu/cpufreq/policy4/scaling_governor
GpuAcc CpuAcc CpuRef
0:01:24.576257 0.0
Inference End 0.0
CpuAcc CpuRef
0:01:29.003342 0.0
Inference End 0.0
I will stop being lazy and do a horrid hack out of the mffc and preprocess it.
Nope dunno what or where utils is and the internet is playing up so will be lazy and try something better.
From dumping the GPU shader assembly it does appear to be using the IDP instructions for 8-bit dot product, so there isn’t anything obviously wrong there.
This is I think the theoretical performance of 8-bit operations, in multiplications per nanosecond:
A55: 115 (four cores)
A76: 307 (four cores)
GPU: 1024
NPU: 3072
So the GPU should be faster than the CPU, but I guess that a larger model is needed to actually see that.
@icecream95 are the cpu calcs for cpu and not neon as we can test that also.
GpuAcc = Gpu, CpuAcc = Really is Neon, CpuRef = Cpu
I slept and had a look at the code and thought it would be easy.
It processes all audio into an array 1st so you may think has froze and then just runs the model.
python3 run_audio_file.py --audio_file_path samples/hp0.wav --model_file_path tflite_int8/wav2letter_int8.tflite --preferred_backends CpuAcc CpuRef
cpu seems to take approx 05.517623s and approx 45% load
gpu seems to take approx :06.291094 and approx 5% cpu load 75% gpu load
That is with a hacky python script feeding it but now very little in the inference loop.
There is some overhead of armnn and opencl but this is exactly what I wanted to check.
The rk3588 really is a power house for ml as its Mali mp4 is almost a perfect match for CPU
So we can run 2x models that are approx 2x the load of current and would be interesting to see how the mali also reacts with bigger models.
I had a go with tensorflow TTS as its a much heavier load but working out how to slot in armnn and the quantisation specifics is going to be bigger dev chore than what the above can demo.
# Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
# SPDX-License-Identifier: MIT
"""Automatic speech recognition with PyArmNN demo for processing audio clips to text."""
import sys
import os
import numpy as np
import psutil
import soundfile as sf
script_dir = os.path.dirname(__file__)
sys.path.insert(1, os.path.join(script_dir, '..', 'common'))
from argparse import ArgumentParser
from network_executor import ArmnnNetworkExecutor
from utils import prepare_input_data
from audio_capture import AudioCaptureParams, capture_audio
from audio_utils import decode_text, display_text
from wav2letter_mfcc import Wav2LetterMFCC, W2LAudioPreprocessor
from mfcc import MFCCParams
from datetime import datetime, timedelta
# Model Specific Labels
labels = {0: 'a', 1: 'b', 2: 'c', 3: 'd', 4: 'e', 5: 'f', 6: 'g', 7: 'h', 8: 'i', 9: 'j', 10: 'k', 11: 'l', 12: 'm',
13: 'n',
14: 'o', 15: 'p', 16: 'q', 17: 'r', 18: 's', 19: 't', 20: 'u', 21: 'v', 22: 'w', 23: 'x', 24: 'y',
25: 'z',
26: "'", 27: ' ', 28: '$'}
def time_float(result):
seconds = int(result)
microseconds = int((result * 1000000) % 1000000)
output = timedelta(0, seconds, microseconds)
return output
def parse_args():
parser = ArgumentParser(description="ASR with PyArmNN")
parser.add_argument(
"--audio_file_path",
required=True,
type=str,
help="Path to the audio file to perform ASR",
)
parser.add_argument(
"--model_file_path",
required=True,
type=str,
help="Path to ASR model to use",
)
parser.add_argument(
"--preferred_backends",
type=str,
nargs="+",
default=["GpuAcc", "CpuAcc", "CpuRef"],
help="""List of backends in order of preference for optimizing
subgraphs, falling back to the next backend in the list on unsupported
layers. Defaults to [GpuAcc, CpuAcc, CpuRef]""",
)
return parser.parse_args()
def main(args, network, input_data):
current_r_context = ""
is_first_window = True
avg_cpu = 0.0
for input_chunk in input_data:
# Run inference
output_result = network.run([input_chunk])
# Slice and Decode the text, and store the right context
current_r_context, text = decode_text(is_first_window, labels, output_result)
is_first_window = False
display_text(text)
runtime = datetime.now() - starttime
print(" " + str(runtime))
avg_cpu = avg_cpu + psutil.cpu_percent()
print(current_r_context, flush=True)
print("Inference End: Avg CPU%=" + str(avg_cpu / len(input_data)))
return runtime
if __name__ == "__main__":
args = parse_args()
# Create the ArmNN inference runner
network = ArmnnNetworkExecutor(args.model_file_path, args.preferred_backends)
# Read command line args
audio_file = args.audio_file_path
sf_data, samplerate = sf.read(audio_file)
sf_secs = time_float((len(sf_data) / samplerate))
# Specify model specific audio data requirements
audio_capture_params = AudioCaptureParams(dtype=np.float32, overlap=31712, min_samples=47712, sampling_freq=16000,
mono=True)
buffer = capture_audio(audio_file, audio_capture_params)
# Extract features and create the preprocessor
mfcc_params = MFCCParams(sampling_freq=16000, num_fbank_bins=128, mel_lo_freq=0, mel_hi_freq=8000,
num_mfcc_feats=13, frame_len=512, use_htk_method=False, n_fft=512)
wmfcc = Wav2LetterMFCC(mfcc_params)
preprocessor = W2LAudioPreprocessor(wmfcc, model_input_size=296, stride=160)
print("Processing Audio Frames...")
input_data = []
for audio_data in buffer:
# Prepare the input Tensors
input_data.append(prepare_input_data(audio_data, network.get_data_type(), network.get_input_quantization_scale(0),
network.get_input_quantization_offset(0), preprocessor))
starttime = datetime.now()
runtime = main(args, network, input_data)
print("Runtime=" + str(runtime))
print("Realtime=x" + str(sf_secs / runtime))
starttime = datetime.now()
runtime = main(args, network, input_data)
print("Runtime=" + str(runtime))
print("Realtime=x" + str(sf_secs / runtime))
rock@rock-5b:~/workspace/armnn/python/pyarmnn/examples/speech_recognition$ python3 run_audio_file.py --audio_file_path samples/hp0.wav --model_file_path tflite_int8/wav2letter_int8.tflite --preferred_backends CpuAcc CpuRef
Your ArmNN library instance does not support Onnx models parser functionality. Skipped IOnnxParser import.
Can't load libOpenCL.so: libOpenCL.so: cannot open shared object file: No such file or directory
Can't load libGLES_mali.so: libGLES_mali.so: cannot open shared object file: No such file or directory
arm_release_ver of this libmali is 'g6p0-01eac0', rk_so_ver is '5'.
Preferred backends: ['CpuAcc', 'CpuRef']
IDeviceSpec { supportedBackends: [CpuAcc, CpuRef, GpuAcc]}
Optimization warnings: ()
Processing Audio Frames...
henry f 0:00:00.039640
phillips 0:00:00.081647
from wic 0:00:00.111952
epedeia 0:00:00.134524
the free ind cy 0:00:00.153441
ycliopedioa 0:00:00.172588
at e 0:00:00.191404
and d 0:00:00.210939
ot we cud 0:00:00.230029
pedia tha 0:00:00.248594
t org 0:00:00.266827
0:00:00.285864
henry 0:00:00.305864
f philip 0:00:00.324896
s 0:00:00.344135
from 0:00:00.363412
wickepedia 0:00:00.381937
the freie andcyclo 0:00:00.401530
opedia 0:00:00.420985
0:00:00.439934
henry afh 0:00:00.458467
e philips 0:00:00.477728
eighteen ni 0:00:00.495774
ney to nin 0:00:00.513933
neteen fifty e 0:00:00.532492
ight 0:00:00.551160
a youas 0:00:00.570730
businessman 0:00:00.589955
from portland 0:00:00.608142
d or again 0:00:00.627386
a 0:00:00.645696
s the honor 0:00:00.664017
of having the phi 0:00:00.683096
lip's head scre 0:00:00.702685
w and sc 0:00:00.730444
crew driver 0:00:00.752651
named ater hi 0:00:00.774348
m 0:00:00.793570
0:00:00.812513
the importance 0:00:00.831017
of the cross had i 0:00:00.849545
screw design 0:00:00.868613
lies 0:00:00.887060
in its t self ce 0:00:00.908571
entering property 0:00:00.927869
y u 0:00:00.946793
useful on ad 0:00:00.964844
imated production 0:00:00.984788
n lines t 0:00:01.003689
he use powered 0:00:01.022886
d screw drivers 0:00:01.044580
0:00:01.065240
philip 0:00:01.085195
's major contrib 0:00:01.105548
bution was i 0:00:01.124491
n driving the cro 0:00:01.142652
oss head concept 0:00:01.163555
forward 0:00:01.182046
to the point 0:00:01.220041
where it was a 0:00:01.256361
dopted by scre 0:00:01.278730
ew makers an 0:00:01.298773
nd automabiele c 0:00:01.317458
companiese 0:00:01.336092
0:00:01.354528
although he recei 0:00:01.373344
ved patents for the 0:00:01.391744
design in 0:00:01.410014
nineten thirty 0:00:01.428450
six 0:00:01.449285
you as pate 0:00:01.467786
nt number 0:00:01.486576
to million 0:00:01.505696
forty six 0:00:01.535174
x thousand 0:00:01.563460
three hundred f 0:00:01.586442
forty three 0:00:01.606334
you es 0:00:01.625209
patents t 0:00:01.643606
oo million 0:00:01.661801
forty six thou 0:00:01.681349
usand aig 0:00:01.701606
he hundred thirty se 0:00:01.734885
ven to t 0:00:01.758276
two million 0:00:01.778433
forty six thou 0:00:01.797458
sand eight 0:00:01.815965
hundered forty 0:00:01.835066
0:00:01.853834
it was so wi 0:00:01.871975
dely copied 0:00:01.890931
that by nin 0:00:01.909609
neteen forty ni 0:00:01.928052
ne phili 0:00:01.946818
ips lost his p 0:00:01.965393
partent 0:00:01.983982
0:00:02.005255
the american 0:00:02.024593
screw company 0:00:02.045068
was respon 0:00:02.076243
nsible for devis 0:00:02.105305
sing a means 0:00:02.126591
of manufactu 0:00:02.145923
ring the screw 0:00:02.164145
and 0:00:02.182812
successfully pa 0:00:02.204005
tented and 0:00:02.222751
licence that the are 0:00:02.241392
method 0:00:02.260188
o 0:00:02.278339
others screw makers 0:00:02.298520
of the nineteen h th 0:00:02.316845
hirties dis 0:00:02.335092
missed the phillip's c 0:00:02.353498
concept 0:00:02.371793
since it calls for 0:00:02.389805
relatively co 0:00:02.411886
mplex r 0:00:02.430947
resisets ocke 0:00:02.450992
et sheepe i 0:00:02.470813
in the head of the scr 0:00:02.489585
rew 0:00:02.507690
as disti 0:00:02.526146
inct from the simple 0:00:02.544312
e milled slaught 0:00:02.562836
of a slaugh 0:00:02.583057
hted type scre 0:00:02.612682
w 0:00:02.642730
the philip' 0:00:02.666383
s screw compa 0:00:02.686970
any and 0:00:02.705801
the american sc 0:00:02.723954
rew company 0:00:02.743269
went on 0:00:02.761700
to devise 0:00:02.779779
the posi drive 0:00:02.799883
e screw 0:00:02.818531
which 0:00:02.837064
differs from the ph 0:00:02.855034
hilips 0:00:02.873806
in that it is de 0:00:02.897096
signed to accomo 0:00:02.917923
dae greaterd t 0:00:02.938485
ork than the phi 0:00:02.957239
lips 0:00:02.976965
0:00:02.995992
and image a 0:00:03.014475
companied this arti 0:00:03.033712
icol 0:00:03.071588
caption 0:00:03.109826
philips 0:00:03.139520
screw head 0:00:03.161752
0:00:03.180733
0:00:03.199879
the followin 0:00:03.218120
g is an infu bo 0:00:03.236750
x which ac 0:00:03.256835
ccompanies this ar 0:00:03.276173
rticle 0:00:03.294907
in 0:00:03.313449
fu box 0:00:03.332069
part o thes 0:00:03.350309
series un 0:00:03.369606
screw drive 0:00:03.388178
types 0:00:03.406281
0:00:03.425665
slaughted 0:00:03.444362
commonl 0:00:03.462749
y eroneou 0:00:03.481221
usly fla 0:00:03.500575
athead 0:00:03.519235
0:00:03.537542
phylips 0:00:03.560982
cross he 0:00:03.583342
ad 0:00:03.603654
0:00:03.623570
pasierive 0:00:03.643562
super 0:00:03.662293
drive 0:00:03.680999
0:00:03.700240
tokgs 0:00:03.718740
0:00:03.737211
ha 0:00:03.756135
cx a 0:00:03.776041
len 0:00:03.794814
0:00:03.814861
roberts son 0:00:03.836249
0:00:03.856562
0:00:03.875218
try wing 0:00:03.893283
0:00:03.915533
tark 0:00:03.936371
set 0:00:03.954823
0:00:03.973410
span er head 0:00:03.991776
0:00:04.012391
0:00:04.032344
triple square 0:00:04.051670
e ex 0:00:04.070651
sy nd 0:00:04.093896
0:00:04.116099
ot 0:00:04.137063
hers 0:00:04.155997
polly drives 0:00:04.174308
sp 0:00:04.193733
linmde drive 0:00:04.212975
double 0:00:04.231921
e hacks 0:00:04.250232
0:00:04.268701
many images ac 0:00:04.287337
ccompanyed this in 0:00:04.305945
pu box 0:00:04.326572
this 0:00:04.345428
page was last 0:00:04.363623
t modified 0:00:04.382075
on the ninth of va 0:00:04.400191
april two 0:00:04.418365
thousand aeight 0:00:04.437033
at s 0:00:04.455600
seventeen o 0:00:04.474167
for 0:00:04.492340
0:00:04.510894
all te 0:00:04.530033
xt as avaivlable 0:00:04.549138
under the term 0:00:04.595112
ms of thei ganew 0:00:04.630920
free document 0:00:04.650296
tation licens 0:00:04.668415
sea 0:00:04.686839
copyrites 0:00:04.705349
for details 0:00:04.725088
0:00:04.743654
wichpedia 0:00:04.766718
is aregister 0:00:04.786224
n trade mark 0:00:04.804714
of the wikie mmede 0:00:04.822999
ea foundation 0:00:04.841733
incorporated 0:00:04.859885
a 0:00:04.878413
eu as registrud 0:00:04.897591
d fival 0:00:04.921186
one sea 0:00:04.942560
three 0:00:04.961235
tax theductable 0:00:04.979391
non profhet c 0:00:04.997482
harity 0:00:05.016022
0:00:05.034412
this sound fi 0:00:05.054727
le and all 0:00:05.076378
text in the artic 0:00:05.097072
cle or li 0:00:05.115830
cense under 0:00:05.134143
the thenew fre 0:00:05.152953
e documentation 0:00:05.171724
n license 0:00:05.190068
availabl 0:00:05.212038
le at 0:00:05.231690
oubl you doubleyou 0:00:05.250553
dw do 0:00:05.270274
t g 0:00:05.290077
and you 0:00:05.308365
u dot 0:00:05.327549
horg 0:00:05.345873
slash 0:00:05.364095
cope left 0:00:05.382992
slash 0:00:05.401753
f dee 0:00:05.420065
d el 0:00:05.439199
dout each t 0:00:05.457707
ea m l 0:00:05.476005
0:00:05.494820
0:00:05.512901
Inference End: Avg CPU% 44.152573529411804
henry f 0:00:00.029676
phillips 0:00:00.054872
from wic 0:00:00.076020
epedeia 0:00:00.094366
the free ind cy 0:00:00.114368
ycliopedioa 0:00:00.133422
at e 0:00:00.152289
and d 0:00:00.183105
ot we cud 0:00:00.207808
pedia tha 0:00:00.229333
t org 0:00:00.250968
0:00:00.270178
henry 0:00:00.288826
f philip 0:00:00.308294
s 0:00:00.326518
from 0:00:00.347068
wickepedia 0:00:00.371150
the freie andcyclo 0:00:00.393486
opedia 0:00:00.414678
0:00:00.433286
henry afh 0:00:00.452419
e philips 0:00:00.479387
eighteen ni 0:00:00.505646
ney to nin 0:00:00.528157
neteen fifty e 0:00:00.549082
ight 0:00:00.568212
a youas 0:00:00.587044
businessman 0:00:00.610898
from portland 0:00:00.632699
d or again 0:00:00.652812
a 0:00:00.672068
s the honor 0:00:00.692313
of having the phi 0:00:00.711468
lip's head scre 0:00:00.730892
w and sc 0:00:00.749864
crew driver 0:00:00.768192
named ater hi 0:00:00.786391
m 0:00:00.804390
0:00:00.822623
the importance 0:00:00.840970
of the cross had i 0:00:00.860126
screw design 0:00:00.879214
lies 0:00:00.897920
in its t self ce 0:00:00.917412
entering property 0:00:00.935788
y u 0:00:00.954098
useful on ad 0:00:00.973072
imated production 0:00:00.992050
n lines t 0:00:01.010948
he use powered 0:00:01.030059
d screw drivers 0:00:01.048675
0:00:01.067221
philip 0:00:01.085260
's major contrib 0:00:01.104387
bution was i 0:00:01.122887
n driving the cro 0:00:01.141004
oss head concept 0:00:01.159672
forward 0:00:01.177971
to the point 0:00:01.196020
where it was a 0:00:01.214700
dopted by scre 0:00:01.233585
ew makers an 0:00:01.254226
nd automabiele c 0:00:01.273182
companiese 0:00:01.291665
0:00:01.309885
although he recei 0:00:01.328732
ved patents for the 0:00:01.347082
design in 0:00:01.365269
nineten thirty 0:00:01.384728
six 0:00:01.403558
you as pate 0:00:01.422254
nt number 0:00:01.440639
to million 0:00:01.458886
forty six 0:00:01.477751
x thousand 0:00:01.496483
three hundred f 0:00:01.515026
forty three 0:00:01.533806
you es 0:00:01.552109
patents t 0:00:01.570447
oo million 0:00:01.589292
forty six thou 0:00:01.608702
usand aig 0:00:01.628919
he hundred thirty se 0:00:01.647763
ven to t 0:00:01.666474
two million 0:00:01.685872
forty six thou 0:00:01.718653
sand eight 0:00:01.746268
hundered forty 0:00:01.765286
0:00:01.784792
it was so wi 0:00:01.803680
dely copied 0:00:01.827649
that by nin 0:00:01.848720
neteen forty ni 0:00:01.867413
ne phili 0:00:01.888695
ips lost his p 0:00:01.907449
partent 0:00:01.926010
0:00:01.945489
the american 0:00:01.963833
screw company 0:00:01.982087
was respon 0:00:02.003408
nsible for devis 0:00:02.023105
sing a means 0:00:02.042951
of manufactu 0:00:02.062428
ring the screw 0:00:02.083601
and 0:00:02.104128
successfully pa 0:00:02.123177
tented and 0:00:02.141869
licence that the are 0:00:02.160193
method 0:00:02.181144
o 0:00:02.200624
others screw makers 0:00:02.219218
of the nineteen h th 0:00:02.237288
hirties dis 0:00:02.255883
missed the phillip's c 0:00:02.274358
concept 0:00:02.294018
since it calls for 0:00:02.312455
relatively co 0:00:02.330471
mplex r 0:00:02.349534
resisets ocke 0:00:02.367540
et sheepe i 0:00:02.387564
in the head of the scr 0:00:02.405687
rew 0:00:02.423716
as disti 0:00:02.442485
inct from the simple 0:00:02.461059
e milled slaught 0:00:02.482157
of a slaugh 0:00:02.500962
hted type scre 0:00:02.521853
w 0:00:02.542050
the philip' 0:00:02.561338
s screw compa 0:00:02.580732
any and 0:00:02.607533
the american sc 0:00:02.630970
rew company 0:00:02.649956
went on 0:00:02.668940
to devise 0:00:02.687399
the posi drive 0:00:02.706255
e screw 0:00:02.737116
which 0:00:02.761253
differs from the ph 0:00:02.783470
hilips 0:00:02.801990
in that it is de 0:00:02.820875
signed to accomo 0:00:02.841162
dae greaterd t 0:00:02.860114
ork than the phi 0:00:02.878703
lips 0:00:02.897341
0:00:02.915621
and image a 0:00:02.934287
companied this arti 0:00:02.952481
icol 0:00:02.971039
caption 0:00:02.990690
philips 0:00:03.010994
screw head 0:00:03.029708
0:00:03.049006
0:00:03.067719
the followin 0:00:03.086644
g is an infu bo 0:00:03.105915
x which ac 0:00:03.124921
ccompanies this ar 0:00:03.143613
rticle 0:00:03.162389
in 0:00:03.180519
fu box 0:00:03.199188
part o thes 0:00:03.217988
series un 0:00:03.236804
screw drive 0:00:03.255463
types 0:00:03.273704
0:00:03.292290
slaughted 0:00:03.310921
commonl 0:00:03.329144
y eroneou 0:00:03.347206
usly fla 0:00:03.366471
athead 0:00:03.386305
0:00:03.405026
phylips 0:00:03.423253
cross he 0:00:03.445383
ad 0:00:03.466505
0:00:03.484779
pasierive 0:00:03.504122
super 0:00:03.522533
drive 0:00:03.541154
0:00:03.560328
tokgs 0:00:03.578647
0:00:03.599285
ha 0:00:03.618935
cx a 0:00:03.637462
len 0:00:03.656392
0:00:03.675186
roberts son 0:00:03.694608
0:00:03.713174
0:00:03.738853
try wing 0:00:03.761281
0:00:03.783589
tark 0:00:03.802532
set 0:00:03.821357
0:00:03.840611
span er head 0:00:03.859893
0:00:03.878397
0:00:03.896426
triple square 0:00:03.915424
e ex 0:00:03.934438
sy nd 0:00:03.952991
0:00:03.971123
ot 0:00:03.991752
hers 0:00:04.011104
polly drives 0:00:04.030115
sp 0:00:04.048891
linmde drive 0:00:04.067864
double 0:00:04.086988
e hacks 0:00:04.106561
0:00:04.125026
many images ac 0:00:04.143887
ccompanyed this in 0:00:04.162591
pu box 0:00:04.181225
this 0:00:04.200282
page was last 0:00:04.219203
t modified 0:00:04.237694
on the ninth of va 0:00:04.258145
april two 0:00:04.276569
thousand aeight 0:00:04.294845
at s 0:00:04.313113
seventeen o 0:00:04.332516
for 0:00:04.380070
0:00:04.416976
all te 0:00:04.438166
xt as avaivlable 0:00:04.457145
under the term 0:00:04.475252
ms of thei ganew 0:00:04.494399
free document 0:00:04.512919
tation licens 0:00:04.531204
sea 0:00:04.552007
copyrites 0:00:04.570451
for details 0:00:04.588813
0:00:04.608612
wichpedia 0:00:04.627385
is aregister 0:00:04.646055
n trade mark 0:00:04.682055
of the wikie mmede 0:00:04.709029
ea foundation 0:00:04.730148
incorporated 0:00:04.749153
a 0:00:04.767656
eu as registrud 0:00:04.786193
d fival 0:00:04.806515
one sea 0:00:04.839423
three 0:00:04.877059
tax theductable 0:00:04.903783
non profhet c 0:00:04.928306
harity 0:00:04.951226
0:00:04.971586
this sound fi 0:00:04.990772
le and all 0:00:05.009120
text in the artic 0:00:05.027202
cle or li 0:00:05.046101
cense under 0:00:05.065126
the thenew fre 0:00:05.084226
e documentation 0:00:05.102388
n license 0:00:05.122325
availabl 0:00:05.146357
le at 0:00:05.168419
oubl you doubleyou 0:00:05.188388
dw do 0:00:05.206907
t g 0:00:05.225644
and you 0:00:05.244077
u dot 0:00:05.262366
horg 0:00:05.280884
slash 0:00:05.326040
cope left 0:00:05.370482
slash 0:00:05.397159
f dee 0:00:05.419872
d el 0:00:05.441224
dout each t 0:00:05.460983
ea m l 0:00:05.479195
0:00:05.497677
0:00:05.516693
Inference End: Avg CPU% 44.149632352941204
The only overhead is there is some sort of load or processing with armnn & gpu as 2nd run always seems to be faster. Which actually for many models doesn’t matter as with code you could preload a trial run and just hold in memory.
rock@rock-5b:~/workspace/armnn/python/pyarmnn/examples/speech_recognition$ python3 run_audio_file.py --audio_file_path samples/hp0.wav --model_file_path tflite_int8/wav2letter_int8.tflite --preferred_backends GpuAcc CpuAcc CpuRef
Your ArmNN library instance does not support Onnx models parser functionality. Skipped IOnnxParser import.
Can't load libOpenCL.so: libOpenCL.so: cannot open shared object file: No such file or directory
Can't load libGLES_mali.so: libGLES_mali.so: cannot open shared object file: No such file or directory
arm_release_ver of this libmali is 'g6p0-01eac0', rk_so_ver is '5'.
Preferred backends: ['GpuAcc', 'CpuAcc', 'CpuRef']
IDeviceSpec { supportedBackends: [CpuAcc, CpuRef, GpuAcc]}
Optimization warnings: ()
Processing Audio Frames...
henry f 0:00:00.199833
phillips 0:00:00.263243
from wic 0:00:00.328341
epedeia 0:00:00.387643
the free ind cy 0:00:00.446891
ycliopedioa 0:00:00.506139
at e 0:00:00.568754
and d 0:00:00.633177
ot we cud 0:00:00.696274
pedia tha 0:00:00.760866
t org 0:00:00.825377
0:00:00.889125
henry 0:00:00.951952
f philip 0:00:01.014317
s 0:00:01.078718
from 0:00:01.142757
wickepedia 0:00:01.206237
the freie andcyclo 0:00:01.272636
opedia 0:00:01.335396
0:00:01.400698
henry afh 0:00:01.464624
e philips 0:00:01.530219
eighteen ni 0:00:01.594969
ney to nin 0:00:01.658931
neteen fifty e 0:00:01.725520
ight 0:00:01.790579
a youas 0:00:01.855143
businessman 0:00:01.921110
from portland 0:00:01.984640
d or again 0:00:02.046878
a 0:00:02.108508
s the honor 0:00:02.173074
of having the phi 0:00:02.235354
lip's head scre 0:00:02.295376
w and sc 0:00:02.356306
crew driver 0:00:02.419730
named ater hi 0:00:02.479308
m 0:00:02.541863
0:00:02.603097
the importance 0:00:02.665169
of the cross had i 0:00:02.729038
screw design 0:00:02.792687
lies 0:00:02.854305
in its t self ce 0:00:02.916365
entering property 0:00:02.981664
y u 0:00:03.045516
useful on ad 0:00:03.107087
imated production 0:00:03.169201
n lines t 0:00:03.231844
he use powered 0:00:03.294249
d screw drivers 0:00:03.357977
0:00:03.419971
philip 0:00:03.444255
's major contrib 0:00:03.470326
bution was i 0:00:03.493646
n driving the cro 0:00:03.516106
oss head concept 0:00:03.540040
forward 0:00:03.563859
to the point 0:00:03.587168
where it was a 0:00:03.614574
dopted by scre 0:00:03.639191
ew makers an 0:00:03.662893
nd automabiele c 0:00:03.686058
companiese 0:00:03.709900
0:00:03.733041
although he recei 0:00:03.756842
ved patents for the 0:00:03.779858
design in 0:00:03.803692
nineten thirty 0:00:03.826854
six 0:00:03.849560
you as pate 0:00:03.873317
nt number 0:00:03.896252
to million 0:00:03.919767
forty six 0:00:03.943308
x thousand 0:00:03.966859
three hundred f 0:00:03.989896
forty three 0:00:04.013471
you es 0:00:04.036481
patents t 0:00:04.058903
oo million 0:00:04.082610
forty six thou 0:00:04.105423
usand aig 0:00:04.129519
he hundred thirty se 0:00:04.151631
ven to t 0:00:04.175407
two million 0:00:04.198211
forty six thou 0:00:04.219682
sand eight 0:00:04.241892
hundered forty 0:00:04.265862
0:00:04.289995
it was so wi 0:00:04.313080
dely copied 0:00:04.334637
that by nin 0:00:04.357252
neteen forty ni 0:00:04.380913
ne phili 0:00:04.405236
ips lost his p 0:00:04.427298
partent 0:00:04.449701
0:00:04.472049
the american 0:00:04.495946
screw company 0:00:04.519409
was respon 0:00:04.541802
nsible for devis 0:00:04.565851
sing a means 0:00:04.588770
of manufactu 0:00:04.610094
ring the screw 0:00:04.632057
and 0:00:04.654730
successfully pa 0:00:04.678612
tented and 0:00:04.701875
licence that the are 0:00:04.724467
method 0:00:04.746392
o 0:00:04.772910
others screw makers 0:00:04.797385
of the nineteen h th 0:00:04.819947
hirties dis 0:00:04.842332
missed the phillip's c 0:00:04.865902
concept 0:00:04.890104
since it calls for 0:00:04.913179
relatively co 0:00:04.934973
mplex r 0:00:04.957153
resisets ocke 0:00:04.979518
et sheepe i 0:00:05.003380
in the head of the scr 0:00:05.027389
rew 0:00:05.051538
as disti 0:00:05.074455
inct from the simple 0:00:05.097022
e milled slaught 0:00:05.120876
of a slaugh 0:00:05.143777
hted type scre 0:00:05.166226
w 0:00:05.187945
the philip' 0:00:05.210199
s screw compa 0:00:05.234057
any and 0:00:05.257507
the american sc 0:00:05.279911
rew company 0:00:05.303471
went on 0:00:05.326708
to devise 0:00:05.350512
the posi drive 0:00:05.373730
e screw 0:00:05.397346
which 0:00:05.420818
differs from the ph 0:00:05.442951
hilips 0:00:05.466475
in that it is de 0:00:05.489528
signed to accomo 0:00:05.512900
dae greaterd t 0:00:05.535968
ork than the phi 0:00:05.558366
lips 0:00:05.580838
0:00:05.604397
and image a 0:00:05.628712
companied this arti 0:00:05.651715
icol 0:00:05.673942
caption 0:00:05.697683
philips 0:00:05.722421
screw head 0:00:05.744466
0:00:05.766550
0:00:05.788734
the followin 0:00:05.812560
g is an infu bo 0:00:05.835601
x which ac 0:00:05.856830
ccompanies this ar 0:00:05.879666
rticle 0:00:05.905828
in 0:00:05.928671
fu box 0:00:05.951345
part o thes 0:00:05.975262
series un 0:00:05.999708
screw drive 0:00:06.022731
types 0:00:06.045104
0:00:06.068814
slaughted 0:00:06.091912
commonl 0:00:06.113519
y eroneou 0:00:06.136937
usly fla 0:00:06.159848
athead 0:00:06.183362
0:00:06.206298
phylips 0:00:06.228803
cross he 0:00:06.251184
ad 0:00:06.275104
0:00:06.299545
pasierive 0:00:06.322722
super 0:00:06.345218
drive 0:00:06.368872
0:00:06.393003
tokgs 0:00:06.416137
0:00:06.438469
ha 0:00:06.460092
cx a 0:00:06.483427
len 0:00:06.506352
0:00:06.527695
roberts son 0:00:06.550262
0:00:06.573836
0:00:06.597164
try wing 0:00:06.619407
0:00:06.640919
tark 0:00:06.663290
set 0:00:06.689627
0:00:06.713785
span er head 0:00:06.736059
0:00:06.756870
0:00:06.779626
triple square 0:00:06.803074
e ex 0:00:06.826732
sy nd 0:00:06.848523
0:00:06.870672
ot 0:00:06.893005
hers 0:00:06.916677
polly drives 0:00:06.939797
sp 0:00:06.963346
linmde drive 0:00:06.986584
double 0:00:07.010035
e hacks 0:00:07.034139
0:00:07.057089
many images ac 0:00:07.079361
ccompanyed this in 0:00:07.100901
pu box 0:00:07.126865
this 0:00:07.150326
page was last 0:00:07.172780
t modified 0:00:07.196601
on the ninth of va 0:00:07.220068
april two 0:00:07.242628
thousand aeight 0:00:07.264406
at s 0:00:07.286862
seventeen o 0:00:07.313602
for 0:00:07.337783
0:00:07.359923
all te 0:00:07.386509
xt as avaivlable 0:00:07.409978
under the term 0:00:07.432496
ms of thei ganew 0:00:07.456288
free document 0:00:07.479422
tation licens 0:00:07.501876
sea 0:00:07.523476
copyrites 0:00:07.545940
for details 0:00:07.569659
0:00:07.592589
wichpedia 0:00:07.616525
is aregister 0:00:07.639508
n trade mark 0:00:07.663383
of the wikie mmede 0:00:07.686577
ea foundation 0:00:07.710205
incorporated 0:00:07.734723
a 0:00:07.757550
eu as registrud 0:00:07.779905
d fival 0:00:07.802303
one sea 0:00:07.823395
three 0:00:07.845993
tax theductable 0:00:07.869594
non profhet c 0:00:07.892780
harity 0:00:07.916367
0:00:07.939468
this sound fi 0:00:07.963100
le and all 0:00:07.986285
text in the artic 0:00:08.009850
cle or li 0:00:08.034208
cense under 0:00:08.057152
the thenew fre 0:00:08.079343
e documentation 0:00:08.103166
n license 0:00:08.126288
availabl 0:00:08.149965
le at 0:00:08.173247
oubl you doubleyou 0:00:08.196931
dw do 0:00:08.220100
t g 0:00:08.242015
and you 0:00:08.264490
u dot 0:00:08.288011
horg 0:00:08.310951
slash 0:00:08.333130
cope left 0:00:08.356892
slash 0:00:08.379863
f dee 0:00:08.403358
d el 0:00:08.426098
dout each t 0:00:08.449922
ea m l 0:00:08.472750
0:00:08.495137
0:00:08.516571
Inference End: Avg CPU% 6.319485294117635
henry f 0:00:00.025098
phillips 0:00:00.048059
from wic 0:00:00.069682
epedeia 0:00:00.092882
the free ind cy 0:00:00.115887
ycliopedioa 0:00:00.139495
at e 0:00:00.162664
and d 0:00:00.186114
ot we cud 0:00:00.209049
pedia tha 0:00:00.231531
t org 0:00:00.253739
0:00:00.275907
henry 0:00:00.299566
f philip 0:00:00.322653
s 0:00:00.346243
from 0:00:00.368908
wickepedia 0:00:00.390658
the freie andcyclo 0:00:00.413116
opedia 0:00:00.436799
0:00:00.459868
henry afh 0:00:00.482093
e philips 0:00:00.503667
eighteen ni 0:00:00.525942
ney to nin 0:00:00.549909
neteen fifty e 0:00:00.573129
ight 0:00:00.595368
a youas 0:00:00.619186
businessman 0:00:00.642221
from portland 0:00:00.666204
d or again 0:00:00.688944
a 0:00:00.712865
s the honor 0:00:00.735683
of having the phi 0:00:00.759400
lip's head scre 0:00:00.782900
w and sc 0:00:00.805186
crew driver 0:00:00.828966
named ater hi 0:00:00.851664
m 0:00:00.875356
0:00:00.898004
the importance 0:00:00.921846
of the cross had i 0:00:00.944766
screw design 0:00:00.966324
lies 0:00:00.989620
in its t self ce 0:00:01.012966
entering property 0:00:01.036775
y u 0:00:01.059720
useful on ad 0:00:01.082092
imated production 0:00:01.103954
n lines t 0:00:01.126182
he use powered 0:00:01.149641
d screw drivers 0:00:01.172723
0:00:01.196173
philip 0:00:01.219340
's major contrib 0:00:01.242948
bution was i 0:00:01.265901
n driving the cro 0:00:01.289600
oss head concept 0:00:01.312517
forward 0:00:01.336215
to the point 0:00:01.359559
where it was a 0:00:01.383004
dopted by scre 0:00:01.405795
ew makers an 0:00:01.429504
nd automabiele c 0:00:01.453834
companiese 0:00:01.478208
0:00:01.501158
although he recei 0:00:01.523546
ved patents for the 0:00:01.545980
design in 0:00:01.570002
nineten thirty 0:00:01.593172
six 0:00:01.616883
you as pate 0:00:01.639729
nt number 0:00:01.662191
to million 0:00:01.685943
forty six 0:00:01.708600
x thousand 0:00:01.732211
three hundred f 0:00:01.756375
forty three 0:00:01.779461
you es 0:00:01.803172
patents t 0:00:01.826351
oo million 0:00:01.848640
forty six thou 0:00:01.872611
usand aig 0:00:01.895845
he hundred thirty se 0:00:01.919687
ven to t 0:00:01.943920
two million 0:00:01.967219
forty six thou 0:00:01.991258
sand eight 0:00:02.013351
hundered forty 0:00:02.035500
0:00:02.059017
it was so wi 0:00:02.082262
dely copied 0:00:02.106268
that by nin 0:00:02.129780
neteen forty ni 0:00:02.153251
ne phili 0:00:02.176276
ips lost his p 0:00:02.198535
partent 0:00:02.222286
0:00:02.245231
the american 0:00:02.269256
screw company 0:00:02.292370
was respon 0:00:02.316087
nsible for devis 0:00:02.339033
sing a means 0:00:02.362709
of manufactu 0:00:02.386082
ring the screw 0:00:02.407517
and 0:00:02.429844
successfully pa 0:00:02.453563
tented and 0:00:02.476469
licence that the are 0:00:02.500109
method 0:00:02.523004
o 0:00:02.545306
others screw makers 0:00:02.569281
of the nineteen h th 0:00:02.592558
hirties dis 0:00:02.614155
missed the phillip's c 0:00:02.635826
concept 0:00:02.658427
since it calls for 0:00:02.682316
relatively co 0:00:02.705207
mplex r 0:00:02.727857
resisets ocke 0:00:02.751483
et sheepe i 0:00:02.774697
in the head of the scr 0:00:02.796381
rew 0:00:02.819614
as disti 0:00:02.842916
inct from the simple 0:00:02.864951
e milled slaught 0:00:02.887507
of a slaugh 0:00:02.911580
hted type scre 0:00:02.934264
w 0:00:02.956632
the philip' 0:00:02.980450
s screw compa 0:00:03.003790
any and 0:00:03.026059
the american sc 0:00:03.049843
rew company 0:00:03.072993
went on 0:00:03.095393
to devise 0:00:03.119325
the posi drive 0:00:03.142464
e screw 0:00:03.165982
which 0:00:03.189129
differs from the ph 0:00:03.212906
hilips 0:00:03.235966
in that it is de 0:00:03.259806
signed to accomo 0:00:03.282958
dae greaterd t 0:00:03.306649
ork than the phi 0:00:03.329702
lips 0:00:03.351889
0:00:03.375620
and image a 0:00:03.399796
companied this arti 0:00:03.422819
icol 0:00:03.446313
caption 0:00:03.469297
philips 0:00:03.491633
screw head 0:00:03.514018
0:00:03.537596
0:00:03.560754
the followin 0:00:03.582819
g is an infu bo 0:00:03.606589
x which ac 0:00:03.629573
ccompanies this ar 0:00:03.651937
rticle 0:00:03.675519
in 0:00:03.698401
fu box 0:00:03.722268
part o thes 0:00:03.745215
series un 0:00:03.768986
screw drive 0:00:03.793040
types 0:00:03.815889
0:00:03.837987
slaughted 0:00:03.859728
commonl 0:00:03.882820
y eroneou 0:00:03.906015
usly fla 0:00:03.929665
athead 0:00:03.953453
0:00:03.975363
phylips 0:00:03.998611
cross he 0:00:04.022581
ad 0:00:04.045314
0:00:04.069004
pasierive 0:00:04.092172
super 0:00:04.115807
drive 0:00:04.139027
0:00:04.162463
tokgs 0:00:04.185378
0:00:04.208838
ha 0:00:04.231554
cx a 0:00:04.255428
len 0:00:04.278669
0:00:04.302285
roberts son 0:00:04.325458
0:00:04.348964
0:00:04.371958
try wing 0:00:04.395594
0:00:04.419418
tark 0:00:04.446092
set 0:00:04.470235
0:00:04.492581
span er head 0:00:04.516097
0:00:04.540015
0:00:04.563000
triple square 0:00:04.585351
e ex 0:00:04.608908
sy nd 0:00:04.631941
0:00:04.655673
ot 0:00:04.678754
hers 0:00:04.702643
polly drives 0:00:04.725733
sp 0:00:04.749337
linmde drive 0:00:04.772469
double 0:00:04.796209
e hacks 0:00:04.818715
0:00:04.842505
many images ac 0:00:04.865688
ccompanyed this in 0:00:04.889308
pu box 0:00:04.912474
this 0:00:04.934762
page was last 0:00:04.958719
t modified 0:00:04.981706
on the ninth of va 0:00:05.005422
april two 0:00:05.027216
thousand aeight 0:00:05.049110
at s 0:00:05.071463
seventeen o 0:00:05.095495
for 0:00:05.119464
0:00:05.142777
all te 0:00:05.165063
xt as avaivlable 0:00:05.188987
under the term 0:00:05.212004
ms of thei ganew 0:00:05.234286
free document 0:00:05.256753
tation licens 0:00:05.278964
sea 0:00:05.302575
copyrites 0:00:05.325955
for details 0:00:05.348070
0:00:05.369810
wichpedia 0:00:05.393193
is aregister 0:00:05.416493
n trade mark 0:00:05.439150
of the wikie mmede 0:00:05.462763
ea foundation 0:00:05.485545
incorporated 0:00:05.509280
a 0:00:05.532497
eu as registrud 0:00:05.556309
d fival 0:00:05.579172
one sea 0:00:05.601571
three 0:00:05.623904
tax theductable 0:00:05.647556
non profhet c 0:00:05.671749
harity 0:00:05.694886
0:00:05.717290
this sound fi 0:00:05.741046
le and all 0:00:05.763774
text in the artic 0:00:05.786158
cle or li 0:00:05.809879
cense under 0:00:05.833817
the thenew fre 0:00:05.855954
e documentation 0:00:05.878282
n license 0:00:05.901719
availabl 0:00:05.924833
le at 0:00:05.947234
oubl you doubleyou 0:00:05.971071
dw do 0:00:05.995365
t g 0:00:06.018369
and you 0:00:06.042143
u dot 0:00:06.064688
horg 0:00:06.088104
slash 0:00:06.111014
cope left 0:00:06.133253
slash 0:00:06.156871
f dee 0:00:06.180072
d el 0:00:06.202211
dout each t 0:00:06.223746
ea m l 0:00:06.246271
0:00:06.269943
0:00:06.293052
Inference End: Avg CPU% 6.539338235294104
CpuRef runs on the calling core only and not going to wait for this to end 
rock@rock-5b:~/workspace/armnn/python/pyarmnn/examples/speech_recognition$ python3 run_audio_file.py --audio_file_path samples/hp0.wav --model_file_path tflite_int8/wav2letter_int8.tflite --preferred_backends CpuRef
Your ArmNN library instance does not support Onnx models parser functionality. Skipped IOnnxParser import.
Can't load libOpenCL.so: libOpenCL.so: cannot open shared object file: No such file or directory
Can't load libGLES_mali.so: libGLES_mali.so: cannot open shared object file: No such file or directory
arm_release_ver of this libmali is 'g6p0-01eac0', rk_so_ver is '5'.
Preferred backends: ['CpuRef']
IDeviceSpec { supportedBackends: [CpuAcc, CpuRef, GpuAcc]}
Optimization warnings: ()
Processing Audio Frames...
henry f 0:00:17.428529
phillips 0:00:34.844938
from wi 0:00:52.256472
epedeia 0:01:09.669870
the free ind cy 0:01:27.090357
yclopedio 0:01:44.510629
at e 0:02:01.929057
nd d 0:02:19.346602
ot we ckuld 0:02:36.765552
pedia tha 0:02:54.186937
t org 0:03:11.611450
0:03:29.036892
henry 0:03:46.463853
f philip 0:04:03.888195
s 0:04:21.314803
from 0:04:38.740595
wickepedia 0:04:56.169110
the freie andcyclo 0:05:13.594531
opedia 0:05:31.026083
0:05:48.457700
henry fh 0:06:05.888977
e philips 0:06:23.323438
eighteen ni 0:06:40.754814
nety to nin 0:06:58.187929
neteen fifty e 0:07:15.619196
ight 0:07:33.052838
a yous 0:07:50.487450
businessman 0:08:07.922139
from portland 0:08:25.360068
d or again 0:08:42.795694
a 0:09:00.233274
s the honor 0:09:17.666807
of having the phi 0:09:35.106635
lip's head scre 0:09:52.540795
w and sc 0:10:09.976951
crewe driver 0:10:27.415814
named after hi 0:10:44.852451
m 0:11:02.292572
0:11:19.729870
the importance 0:11:37.170950
of the cross head 0:11:54.611701
screw design 0:12:12.051614
lies 0:12:29.492745
in its t self ce 0:12:46.934857
entering property 0:13:04.377058
y 0:13:21.816384
useful on ad 0:13:39.257484
imnated production 0:13:56.696562
n lines t 0:14:14.136365
he use powered 0:14:31.577671
d screw drivers 0:14:49.021440
0:15:06.462271
philip 0:15:23.903284
's madjour contrib 0:15:41.344776
bution was i 0:15:58.786934
n driving the cro 0:16:16.227175
oss head concept 0:16:33.665869
forward 0:16:51.105858
to the point 0:17:08.546104
where it was a 0:17:25.988356
dopted by scre 0:17:43.433439
ew makers an 0:18:00.875144
nd automabiele c 0:18:18.316596
companiese 0:18:35.762228
0:18:53.207648
although he recei 0:19:10.653780
ved patents for the 0:19:28.100305
design in 0:19:45.545399
nineteen thirty 0:20:02.988668
six 0:20:20.431421
you as pate 0:20:37.879302
nt number 0:20:55.325021
Going back to Jammy & Wayland
With https://github.com/armbian/build/releases/download/22.11.0-trunk.0096/Armbian_22.11.0-trunk.0096_Rock-5b_jammy_legacy_5.10.72.img.xz
Plus current ppa still crashes and wondering if any updates on that? Still seems to be holding some packages back so maybe the ddr update is not in place.
@icecream95 If I build glmark2-es2-wayland and install it runs complete and ends with a Glmark2 Score 866
That is the normal like for like env I usually run glmark with no affinity or scheduler changes and less than I would hope for but hey it didn’t crash!
gg @icecream95 as things looking nearer
echo performance | sudo tee /sys/bus/cpu/devices/cpu[046]/cpufreq/scaling_governor /sys/class/devfreq/dmc/governor
1764
Bad bash script time
#!/bin/bash
glmark2-es2-wayland&
while :
do
LOAD=$(cat /sys/devices/platform/fb000000.gpu/utilisation)
echo "Current Load:"$LOAD
TOTAL=$((TOTAL+LOAD))
COUNT=$((COUNT+1))
echo "Avg Load:"$((TOTAL/COUNT))
sleep 1
done
Load on avg 41%
What crashes? How? When you do what? Does anything appear in dmesg, the console, or journalctl?
Yup, that’s about what’s expected. It will take some work to change the driver so that work can be done in parallel between the CPU and GPU. With CSF the kernel doesn’t help with doing that as much as it did previously.
I didn’t check that much as the repo glmark2-es2-wayland seems a catastrophic crash and locked up the wayland client.
I just compiled glmark2-es2-wayland with .waf and runs fine, so apols didn’t do any bug hunting as mostly used the Jammy Radxa image due to ArmNN as not sure if 22.04 is part of the repo (18.04/20.04 is)
The RK3588 is a low wattage powerhouse SBC for ML with the CPU & GPU and the 3rd option of the NPU that haven’t tried yet as not to optimistic about quite a few dmesg entries with the NPU.
With the amount of vendors and variations the rk3588 & rk3588s could become an extremely common platform.
Have you reached out to Alyssa and Mesa like Asahi Linya did with the M1 as a herd sharing and contributing to a mainline is likely going to be quicker and stronger than many all soloing?
In a perfect world Mfcc & Spectrogram 16Khz would be part of codecs and near load free but both Tensorflow & Pytorch have optimised C routines and strange why Arm opted for Python that can not be stressed how much it sucks for any DSP
Many thanks for your time and input.
Starting with Armbian_22.11.0-trunk.0096_Rock-5b_jammy_legacy_5.10.72.img
My sort of legacy thing I do as a check its all there sudo apt-get update && upgrade sudo apt-get install software-properties-common apt-utils git cmake build-essential autotools-dev autoconf libtool pkg-config reboot
Get the csf bin to lib/firmware
sudo add-apt-repository ppa:liujianfeng1994/panfork-mesa
update/upgrade
sudo apt install ubuntu-desktop
reboot
Log in to wayland
sudo apt-get install glmark2-es2-wayland
glmark2-es2-wayland
This time runs no prob but before on 2x different try I got a total catastrophic lockup and decided to compile by waf myself.
Guess not waf but just a spurious crash.
Still get artefacts on the background image but always ignored that.
So dunno as tried it again expecting a crash and worked fine GLmark2 Score 872
On radxas debian, I cannot see mouse cursor on wayland gnome using @icecream95 's panfork, but the cursor is visible on KDE wayland(and weston), and I’m not seeing the graphic glitch that @stuartiannaylor is seeing on either desktop. I compiled panfork myself.
KDE wayland is having performance issue using OpenGL based compositors, the desktop will hang from time to time, and back to normal after 1 second or 2, but XRender works just fine. Firefox is also having such hanging issue even on XRender compositor. However, Gnome does not have such performance issue at all, although I cannot see the cursor, firefox is very smooth.
Weston works perfectly, performance is good, cursor is visible, just Weston itself is not so user-friendly and I don’t want to use it every day.
X11 works fine on both KDE and Gnome, although less performant than wayland on my 4K60 displays and I can see frame drops, but at least it does not have serious bugs that affects normal usage.
There is one thing that I’m using the mali_csffw.bin come with radxas debian, not the one in JeffyCN’s repo.
mali_csffw.bin didn’t know there was a diff as using JeffyCN’s repo.
The Radxas version seems to be an older version than the Jeff’s Repo ones, and I’m have some problems with Jeff’s in early days of testing so I used Radxas one.