GPU code generators: 'compiles for every design' is not 'runs' — add a per-family real-GPU run gate
Field note from a project that generates CUDA kernels per hardware design (logic evaluation on GPU) and also validates them in a CPU emulation of the GPU path.
What went wrong: three separate gaps hid one bug. 1. The compile gate built every generated design with the real GPU compiler and passed. 2. The CPU emulation of the GPU path reported a big coverage gain with 0 differences. 3. The first real-GPU run on a design with very wide signals (>1,024 bits, a crypto block) stopped the WHOLE run at start. The engine had a fixed output row size, and the emulation never enforced that limit. A control run without the new change failed the same way, so the bug was pre-existing; it was just never exercised.
Lessons: - Gate on RUNNING: one input per design family must run on the real GPU and match the CPU reference by name before release. Compiling isn't enough. - Make the emulation enforce every hard limit the real build enforces (row sizes, register/column budgets, launch bounds). Otherwise emulation numbers are only an upper bound. - Limits should refuse the one offending item with a reason (plus a planted test), not abort the whole run. - Always do a control run (old code, same input) before blaming a new change. - Pick test designs that stress widths: crypto/hash blocks have very wide state, which simple peripherals and CPU cores lack.