Profile GPU or SIMD code

I need a reproducible profile of the shader and SSE code path on our Linux dev machine. Run GPU profiling while executing the demo scene for 60 seconds at 1920x1080 using the latest driver, capture the GPU timeline and shader invocation counts, and produce a report that pinpoints the top three shaders by GPU time. For the CPU side, run an instruction-level profile that shows hot loops using AVX2 and provide annotated source lines with cycles per iteration.

Before I hand the profiling report to the rendering lead, make it actionable: surface the single biggest shader by GPU time, show its input sizes and whether it branches, list the top CPU functions that feed it with precise cycle counts, and flag any data-layouts that prevent vectorization. Include one concrete change that will likely cut GPU time by at least 15% and the estimated implementation risk.

Running the scene on Linux looks fine, but on Windows machines with 4K monitors we see stutters and GPU spikes. I profiled briefly and saw one pixel shader with rising invocation time, but I don't understand if it's driver or our shadow atlas size. I'm afraid a change will only mask the issue. What diagnostics should I run next to distinguish driver scheduling, GPU memory thrash, and a shader algorithmic cost increase?

Across several releases we repeatedly hit last-minute performance problems traced to a few heavy shaders and un-vectorized CPU paths. We lose days rewriting code under time pressure. Which habits around profiling earlier, code ownership, and test scenes should we adopt so GPU and SIMD issues get found and fixed earlier without breaking features?