Guohao’s Engineering Notes
Systems software, GPU computing, and compilers
I’m Guohao, a software engineer. These notes connect architecture and source code to practical engineering decisions, keeping the applicable conditions and supporting evidence close to each conclusion.
Topics
AMD GPU Architecture Notes
Architecture comparisons and LLVM dependency analysis for RDNA and CDNA. Explore matrix instructions, lane layouts, memory movement, synchronization, and compiler behavior.
Choose a reading path. Start with an architecture comparison for hardware behavior, or go directly to a compiler article for LLVM handling. Each article states its targets, source revisions, and validation limits.
RDNA
RDNA 3, RDNA 3.5, and RDNA 4
Compare WMMA layouts, cross-lane operations, data paths, and synchronization.
LLVM on RDNA: WMMA dependencies and waits
Trace register-overlap handling, memory waits, and the limits of automatic repair.
CDNA
CDNA 3 and CDNA 4
Compare MFMA, numeric formats, lane layouts, LDS resources, and data movement.
LLVM on CDNA: MFMA dependencies and waits
Follow physical-register dependencies, wait-state accounting, and developer checks.
Source inspection, compilation, and GPU measurements are recorded separately.
Corrections and discussion are welcome through the source repository.