-
Inside Mirage (3) - Megakernel Persistent Runtime
Deep dive into Mirage's MPK runtime - worker-scheduler model, event-driven synchronization, and multi-GPU support
-
Paper Reading - Mirage Persistent Kernel (MPK)
Notes on the Mirage MPK paper - a compiler and runtime for mega-kernelizing tensor programs
-
Inside Mirage (2) - Transpiler from MuGraph to CUDA
Deep dive into Mirage's transpiler - the journey from MuGraph to CUDA through fusion resolution, layout planning, scheduling, and code generation
-
Inside Mirage (1) - Superoptimizing Tensor Programs via MuGraph and Search
Diving into Mirage's MuGraph IR and search algorithms - how a superoptimizer finds GPU kernels that hand-written fusion rules could never dream of
-
A First Look at PyPTO. "Tile & Human"
Exploring Huawei's PyPTO - a tile-based programming paradigm that puts humans back in the optimization loop, because sometimes NP-Hard problems need human wisdom
-
(Paper Reading) ACT, Automatically Generating Compiler Backends from Tensor Accelerator ISA Descriptions
Paper reading on ACT - a compiler-compiler that auto-generates compiler backends from tensor accelerator ISA descriptions, solving the eternal pain of custom accelerator adoption
-
LLM in very daily life? Enjoy Claude Code and Gemini-cli's super power!
Share an expierence of using LLM agents for life (rather than for work)
-
What if You Never Had to Choose? A Journey into E-graphs and Equality Saturation
How e-graphs turn the optimizer's dilemma into a solved problem by keeping all the good ideas at once, featuring real examples, bad jokes, and the magic of egg
-
Glancing at the Simulation Landscape
Survey on simulation tools for computer architecture research