cuda-gdb에서 커널 단계별 실행하기

When a kernel misbehaves, cuda-gdb lets you pause it and look inside, just like gdb does for ordinary CPU programs. 🔍

cuda-gdb needs symbols, so compile your kernel with -G to add full device debug information before you debug.

The -G flag disables many optimizations so line numbers and variables stay accurate, which makes stepping through device code trustworthy.

Start a session by handing your binary to cuda-gdb, then type run to execute it under debugger control.

Place a breakpoint right inside the kernel by naming the function, so execution pauses the moment a thread reaches it.

Thousands of threads may hit that breakpoint, so cuda-gdb stops and focuses on one thread at a time for inspection.

Use next to run the current line and pause on the following one, walking through your kernel's logic step by step.

Check what a thread actually computed by asking cuda-gdb to print a local variable while you are paused inside the kernel.

Move attention to a specific thread with cuda thread, picking the exact block and lane you want to inspect.

You can print threadIdx and blockIdx just like normal variables, confirming exactly which data element this thread handles.

When you are done poking around, type continue to resume the program until the next breakpoint or until it finishes.

Which compiler flag adds device debug information for cuda-gdb?

You compiled with -G, set a kernel breakpoint, stepped lines, switched threads, and printed indices. That is the core cuda-gdb loop. 🎉