Coder?
Jhubei City, Hsinchu County,
Joined January 2009
- Tweets1.8K
- Following482
- Followers557
- Likes2.4K
ANE INT8 on A20 Pro is about 105 TOPS for INT8. Measured with my little MPSGraph based app, github.com/freedomtan/measur…. Note that these numbers are likely Winograd 1D F(2, 3) "accelerated".
Dug into ANECompiler.framework of iOS/macOS 27 RC and found “bonded networks” on H18g/M6 and H19/A20 Pro. Before, multi-ANE chips like the M-Ultra just split up the work at the driver level across separate devices. (1/2)
Finally figured out how to read Apple ANE's PMU counters. In short, you have to change boot-args. I put down some example cdoe at github.com/freedomtan/ane_pm…. Tested on M1 and M4 only.
If you ever wondered how to run Apple's ANE .hwx file on non-jailbroken phones, check my little program github.com/freedomtan/coreml…
It turns out it's possible to retrieve per-op profiling information w/o using `MLComputePlan`. Some undocumented classes and methods are required, though.
the output: github.com/freedomtan/coreml…
simple objective-c code: github.com/freedomtan/coreml…
@fleetwood___: look what I found
1
4
230
According to the name of its firmware, M3's ANE is h15/themis, the one from A16 (A17 Pro's ANE is h16/iaso). So its compute capacity stays roughly the same as A16.
248
Wow, I didn't know that the 2nd of "The AWK Programming Language" is coming. AWK is my favorite language in Android. Yes, in Android.
awk.dev/
3
241
Wow
SnapFusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds
paper page: huggingface.co/papers/2306.0…
present a generic approach that, for the first time, unlocks running text-to-image diffusion models on mobile devices in less than 2 seconds. We achieve so by introducing efficient network architecture and improving step distillation. Specifically, we propose an efficient UNet by identifying the redundancy of the original model and reducing the computation of the image decoder via data distillation. Further, we enhance the step distillation by exploring training strategies and introducing regularization from classifier-free guidance. Our extensive experiments on MS-COCO show that our model with 8 denoising steps achieves better FID and CLIP scores than Stable Diffusion v1.5 with 50 steps. Our work democratizes content creation by bringing powerful text-to-image diffusion models to the hands of users.
based on Keras CV python code, I wrote some C++ code so that I can run Stable Diffusion on Android with tflite models I converted.
github.com/freedomtan/keras_…
wrote a script to show Stable Diffusion tflite models I converted work as expected
mostly of the code is borrowed from Keras CV implementation
@RisingSayak
github.com/freedomtan/keras_…
3
609
wrote a python script to parse all iOS 16 Espresso models (*.espresso.{net,shape}), print i/o tensor names and shape, and dump model graph.
github.com/freedomtan/coreml…
2
5
668
Converting Keras CV Stable Diffusion models to .tflite is a bit complicated than I thought.
github.com/freedomtan/keras_…
444
if image quality matters, remember to set the dct_method when using tf.io.decode_jpeg().
tensorflow.org/api_docs/pyth…
Ha, I used Cuecat before
Sometime in 2001, I walked into a Radio Shack on San Francisco's Market Street and asked for a Cuecat: a handheld barcode scanner that looked a bit like a cat and a bit like a sex toy. The clerk handed one over to me and I left, feeling a little giddy. I didn't pay a cent. 1/
ALT A Cuecat scanner with a bundled cable and PS/2 adapter; it resembles a plastic cat and also, slightly, a sex toy. It is posed on a Matrix movie 'code waterfall' background and limned by a green 'supernova' light effect. Image: Jerry Whiting (modified) https://en.wikipedia.org/wiki/File:CueCat_barcode_scanner.jpg CC BY-SA 3.0: https://creativecommons.org/licenses/by-sa/3.0/deed.en
According a recent LLVM commit, Apple A16 is ARMv8.6a based.
github.com/apple/llvm-projec…
According to its device tree, the Everest (the big core of Apple A16 bionic) cluster in iPhone 14 Pro has 0x1000000 = 16 MiB L2 cache.
Note that M1 L2 size was wrong in its device tree. So this is a reference only.
gist.github.com/freedomtan/5…
1
1
Reports of TensorFlow's Death Have Been Greatly Exaggerated
Today, the results of the 2022 global developer survey (run by StackOverflow) just went out.
TensorFlow is used more than any other ML tool -- 12.95% of all developers use TensorFlow (slightly more than sklearn at 12.59%, and 1.5x more than PyTorch at 8.61%).
3
