Agreed! As funny as it sounds, I usually watch Tsoding's videos from my TV in bed at the end of the day to wind down, I can still focus on what's going on (where as if I watch in the day time normally my attention is split between whatever billion things is going on during the day) and the text size that Tsoding uses in Emacs makes it perfectly easy to read everything even from the TV as well. It sounds strange, I know. But it's become a ritual of mine, and one I look forward to!

​@@russjr08 Same!

it got used lol

Ah, that's a good point! I keep forgetting C++ exists :)

@@TsodingDaily What do you think about C++?

It's a thin wrapper around OpenGL. So you still have all the OpenGL bs. For instance, different version of the shader language. In fact that problem popped up immediately after I published the source code. github.com/tsoding/musializer/issues/1

Wow, is that going well? I was experimenting with performence of basic neutral networks and implementing them with different api/methods... standard single thread implementation, one with SSE, one AVX, oneDNN, Cuda(runtime and driver), openCL and last but not the least Vulkan. Setting up vulkan and doing just processing requires 100x boiler code unfortunately(don't know how that would scale). Vulkan's memory usage was sweet compared to runtime Cuda but comes very short compared to cublas acceleration. The memory advantage isn't there when compared to driver api, but that also requires a bit of boilerplate.

@@sids911 Interesting to hear about your experience. I wrote a basic single threaded version in C++ to make sure I understood the algorithms, and instead of going the intermediate step of multi threaded SIMD on the cpu, to just jump straight to the GPU and see what I could do. I don't have a lot of experience with Vulkan yet, so I'm still learning how to set up everything. I don't even have the first test code complete yet. I think I've set up the pipeline correctly, so today I'm working on setting up memory allocation, buffers, and descriptor sets. But my strategy is, at first, to just use the compute shader as basically a dot product calculator for forward propagation, and on back prop it will just handle a single layer at a time. That will require host synchronization, which I'm sure is not the fastest way to do it. From there I might try to dynamically generate the pipeline to include barriers between layers so it can all be done with one dispatch, but I'll have to scratch my head a little more first, lol. It's not surprising that Nvidia would have special optimizations in CUDA for doing AI work, but I portability is one of the goals for me. Interesting to hear about the memory usage. Currently in the CPU code I'm using a bump allocator to place the weights, biases, and activations, so I can easily get the exact size of bulk memory I need before hand. I'm hoping that for now I can put everything in host visible memory, but if I want to run large models I might have to handle the case where the physical device needs to move things into device-only memory. That's gonna be "fun." And yes, it's a LOT of boilerplate, but the idea with Vulkan is that it's more work to set up, but faster and easier to call a dispatch. But anyway, it's still very much a WIP, so no stats to share yet. But I'll try to remember to post again when I get something working.

It's issue 2410 if you want an in depth thingy (can't post a link here unfortunately)

I like watching shader showdowns

The future is now old man

raylib is the graphics library used here

@@jacobwilson8275 if i learned with him i can understand it ?

highschool math and programming stuff

@@----__--- i am engineering