Managed to get stable-diffusion.cpp integrated into sdkit v3 and Easy Diffusion. sdkit v3 wraps stable-diffusion.cpp with an API server. For now, the API server exposes an API compatible with Forge WebUI. This saves me time, and allows Easy Diffusion to work out-of-the-box with the new C++ based sdkit. It compiles and runs quite well. Ran it with Easy Diffusion’s UI. Tested with Vulkan and CUDA, on Windows.
Following up to the previous post on sdkit v3’s design: The initial experiments with generating ggml from onnx models were promising, and it looks like a fairly solid path forward. It produces numerically-identical results, and there’s a clear path to reach performance-parity with stable-diffusion.cpp with a few basic optimizations (since both will eventually generate the same underlying ggml graph). But I think it’s better to use the simpler option first, i.e. use stable-diffusion.cpp directly. It mostly meets the design goals for sdkit v3 (after a bit of performance tuning). Everything else is premature optimization and scope bloat.
Successfully compiled the VAE of Stable Diffusion 1.5 using graph-compiler. The compiled model is terribly slow because I haven’t written any performance optimizations, and it (conservatively) converts a lot of intermediate tensors to contiguous copies. But we don’t need any clever optimizations to get to decent performance, just basic ones. It’s pretty exciting because I was able to bypass the need to port the model to C++ manually. Instead, I was able to just compile the exported ONNX model and get the same output values as the original PyTorch implementation (given the same input and weights). I could compile to any platform supported by ggml by just changing one flag (e.g. CPU, CUDA, ROCm, Vulkan, Metal etc).
PolyBlocks is another interesting ML compiler, written using MLIR. It’s a startup incubated in IISc Bangalore, run by someone (Uday Bondhugula) who co-authored a paper on compiler optimizations for GPGPUs back in 2008 (17 years ago)! Some of the compiler passes to keep in mind: fusion tiling use hardware acceleration (like tensor cores) constant folding perform redundant computation to avoid global memory accesses where profitable pack into buffers loop transformation unroll-and-jam (register tiling?) vectorization reorder execution for better spatial, temporary and group reuse Scheduling approaches:
Following up to the deep-dive on ML compilers: sdkit v3 won’t use general-purpose ML compilers. They aren’t yet ready for sdkit’s target platforms, and need a lot of work (well beyond sdkit v3’s scope). But I’m quite certain that sdkit v4 will use them, and sdkit v3 will start making steps in that direction. For sdkit v3, I see two possible paths: Use an array of vendor-specific compilers (like TensorRT-RTX, MiGraphX, OpenVINO etc), one for each target platform. Auto-generate ggml code from onnx (or pytorch), and beat it on the head until it meets sdkit v3’s performance goals. Hand-tune kernels, contribute to ggml, and take advantage of ggml’s multi-backend kernels. Both approaches provide a big step-up from sdkit v2 in terms of install size and performance. So it makes sense to tap into these first, and leave ML compilers for v4 (as another leap forward).
This post concludes (for now) my ongoing deep-dive into ML compilers, while researching for sdkit v3. I’ve linked (at the end) to some of the papers that I read related to graph execution on GPUs. Some final takeaways: ML compilers might break CUDA’s moat (and fix AMD’s ROCm support). A single compiler is unlikely to fit every scenario. The scheduler needs to be grounded in truth. Simulators might be worth exploring more. ML compilers might break CUDA’s moat (and fix AMD’s ROCm support) It’s pretty clear that ML compilers are going to be a big deal. NVIDIA’s TensorRT is also an ML compiler, but it only targets their GPUs. Once the generated machine code (from cross-vendor ML compilers) is comparable in performance to hand-tuned kernels, these compilers are going to break the (in)famous moat of CUDA.
Great post on why a “work-in-progress” notes blog is useful - https://gregorygundersen.com/blog/2020/01/12/why-research-blog/ This is exactly why I (re)started this blog. This blog is mainly a way to share the notes that I take when working on problems. I’ve always written huge volumes of notes (privately) when working through problems, but making them public has forced me to: Work through them with more rigor and detail (since they’ll be public). Structure them better. Catch and fix biases. Tackle large topics through a series of posts over time. Write them in a way that I can revisit later on and remember what I was thinking (instead of a giant messy blob of notes). It is important though to avoid the trap of feeling productive by publishing notes, instead of finally “shipping” the actual thing that you were meant to finish.
It looks like ggml has recently added basic automatic operator fusion into their graph executor (example). It uses a hand-coded list of simple rule-based substitutions (e.g. fuse a matrix multiply followed by add into one op, or a matrix multiply followed by GLU activation into one op etc). Each fused op is a hand-written kernel. These fusion rules are specified per backend (e.g. separate rules for CUDA/ROCm, separate for Vulkan, separate for Metal etc), presumably people may not have written fused ops for certain backends (either due to the backend’s popularity, or lack of sufficient gain in performance).
The next major version of Freebird (i.e. v3) will use a new internal architecture that’s much easier to program with. In some ways, it’s an evolution of the architecture used in Freebird v2, but taken to its logical conclusion. The current version of Freebird (v2) uses a DOM-like model, and borrows a lot of programming patterns from browser-based programming. An underlying runtime abstracts away input events (like trigger_press, drag, enter, leave etc). It follows an event dispatch model (using add_event_listener and dispatch_event). Visual elements like menus, transform handles etc are DOM Nodes, which respond to events like drag and click. It also uses CSS-like styling to provide an easy way to style groups of related elements (like menu buttons).
A possible intuition for understanding GPU memory hierarchy (and the performance penalty for data transfer between various layers) is to think of it like a manufacturing logistics problem: CPU (host) to GPU (device) is like travelling overnight between two cities. The CPU city is like the “headquarters”, and contains a mega-sized warehouse of parts (think football field sizes), also known as ‘Host memory’. Each GPU is like a different city, containing its own warehouse outside the city, also known as ‘Global Memory’. This warehouse stockpiles whatever it needs from the headquarters city (CPU). Each SM/Core/Tile is a factory located in different areas of the city. Each factory contains a small warehouse for stockpiling whatever inventory it needs, also known as ‘Shared Memory’. Each warp is a bulk stamping machine inside the factory, producing 32 items in one shot. There’s a tray next to each machine, also known as ‘Registers’. This tray is used for keeping stuff temporarily for each stamping process. This analogy can help understand the scale and performance penalty for data transfers.