B03 Exercises

Three drills that turn "it's slow" into a named, measured bottleneck. Do them before opening ../solutions/. Write down the evidence, not just the verdict — an unsupported diagnosis is a guess, and guesses are what profiling exists to eliminate.

This directory has its own Makefile that runs the profiler invocations each drill needs, so you can work without leaving it:

make                          # list targets
make timeline                 # drill 1, level 1 (rocprofv3 --summary --sys-trace)
make counters KERNEL=kernel_beta   # drill 1, level 2 (rocprofv3 --pmc)
make parse                    # phase/kernel breakdown of the CSVs
make transpose-timeline       # drill 3, on A02's transpose ladder
make clean

Set GPU_ARCH to your device (gfx950 MI350, gfx942 MI300, gfx90a MI200); find it with /opt/rocm/bin/offload-arch. The equivalent raw commands, spelled out, are in each drill file.

#	File	Skill
1	01_diagnose_mystery.md	Profile four kernels; name each bottleneck and cite the counter that proves it.
2	02_amdahl.md	From a phase breakdown, compute the end-to-end ceiling of a 10× kernel win and decide whether to do it.
3	03_profile_your_transpose.md	Show from counters why A02's padded transpose beats the unpadded one.

Each drill maps to an interview competency: evidence-based diagnosis / prioritization under Amdahl's law / connecting a microarchitectural counter to a source-level cause.

---

Ground rules for every drill

These are the rules that separate a measurement from a story:

1. Hypothesis first. Before you run the profiler, write one sentence: "I think kernel X is limited by ___ because ___." Then measure to confirm or kill it. A profiler run with no hypothesis produces a wall of counters and no decision.
2. Timeline before counters. Establish that the thing you are about to micro-analyze actually matters. Drill 2 is the formal version of this rule.
3. Never profile a debug build or the first iteration. -O3 (the module Makefile already does this) and skip the warmup dispatches (--kernel-iteration-range on AMD, --launch-skip on NVIDIA).
4. One change at a time. If you change two things and it gets faster, you have learned nothing about which one did it.
5. Sanitize before sharing. Profiler logs carry hostnames, pids, and absolute paths. See the sed recipe in the module README, section 4.

What "evidence" looks like

A complete answer for one kernel has four parts:

Part	Example shape
Verdict	"kernel_X is limited by uncoalesced global access."
Primary counter	Name the counter and its value, e.g. MemUnitStalled high while VALUBusy is low.
Corroborating number	The derived quantity: effective GB/s vs. the copy ceiling you measured in A02 Lab 1.
Predicted fix + expected delta	"Make the innermost index stride-1; expect to approach the ceiling."

If you cannot fill all four rows, you have a hypothesis, not a diagnosis.