Dive into deep learning compilers

Related: https://NN-512.com is a compiler that generates C99 code for neural net inference. It takes as input a simple text description of a convolutional neural net inference graph. It produces as output a human-readable, stand-alone (no dependencies) C99 implementation of that graph. The generated C99 code uses AVX-512 vector instructions to perform inference.

For those surprised at the dot product benchmark on 3.1 [0] where np.dot() is substantially faster when using floats over ints, the things is simply that numpy can use BLAS libraries for floats, but not for ints (because BLAS libs don't implement integer data types).

[0] https://tvm.d2l.ai/chapter_cpu_schedules/arch.html

[1] https://stackoverflow.com/questions/11856293/numpy-dot-produ...

Hasn't been updated since 2020: https://github.com/d2l-ai/d2l-tvm, may be severely outdated, since deep learning compilers is a rapidly moving field.

Legit question, what on earth is this (for)?

I tried reading the book a bit and it seems like a library for turning ml models into C code?

But its really not clear what the point of this book is about...Is this written by space aliens or is this a technical manual for chemistry?

I have coming up on a decade of Python, Java, JS experience, I know what programming is, I don't know what this...thing...is for.

This goes over network construction very well, but I'd like to know about more about how modern compilers go about doing autograd.

Dive into DL (d2l) is becoming an incredibly good one stop reference for theory, code and now machine / architecture internals.

Seems like automatic differentiation should be covered.

cool! is this your work?