Is there any way that I can get sm active and sm occupancy programatically?

[DamonsJ]

I want to get sm active and sm occupancy in my code , like run a dependent process and collect these metrics and export metrics value.

so is there any way I can do that programatically?

I do not want to collect metrics for a single kernel or some kernels? just collect metrics like nsys

[Advait251206]

Short answer: not directly via nvbench itself, but yes—you can collect SM Active and SM Occupancy programmatically, just not through a simple “nvbench API switch” in the way you’re describing.

nvbench is a benchmark runner, not a full profiling/telemetry framework like Nsight Systems or CUPTI.

---

1. Why nvbench can’t do this directly

Metrics like:

• SM Active
• SM Occupancy
• warp execution efficiency
• achieved occupancy over time

come from hardware performance counters, and those are exposed via:

• CUPTI (CUDA Profiling Tools Interface)
• Nsight Compute (per-kernel)
• Nsight Systems (timeline + aggregated GPU metrics)

nvbench does not:

• attach CUPTI collectors globally
• run continuous GPU sampling
• export system-wide GPU utilization metrics

It only measures:

• kernel execution time
• throughput-style metrics you define
• benchmark-level statistics

---

2. What you actually want (system-wide SM metrics)

What you're describing (“like nsys, not per-kernel”) maps to:

✔ Nsight Systems (nsys)

• SM Active (%)
• SM Occupancy (%)
• GPU utilization over time
• context-wide aggregation

You can run it programmatically like:

nsys profile \
  --stats=true \
  --output=my_profile \
  ./your_app

or from code via subprocess:

system("nsys profile --stats=true ./your_app");

---

3. Programmatic access (CUPTI route)

If you truly want to integrate inside your code (no external tool), then:

You need CUPTI:

CUPTI exposes:

• activity API → SM occupancy sampling
• metric API → SM active, occupancy, etc.
• device-wide counters (depending on GPU + driver support)

You can:

1. Initialize CUPTI context
2. Enable metric collection:

cuptiSetEventCollectionMode(...)
cuptiMetricGetValue(...)

3. Query:

• sm__cycles_active.avg.pct_of_peak_sustained_active
• sm__inst_executed.sum
• occupancy derived metrics

But:

👉 This is low-level, complex, and intrusive compared to nvbench.

---

4. Hybrid approach (most practical)

Since you said:

“run a dependent process and collect these metrics and export metrics value”

The standard production approach is:

Option A (recommended)

Run nvbench normally + wrap with Nsight Systems:

nsys profile --metrics=smsp__cycles_active.avg.pct_of_peak_sustained_active ./bench

Then parse:

• .nsys-rep
• export to CSV/SQLite

---

Option B

Use nvbench for workload generation + CUPTI for telemetry

• nvbench runs kernels
• CUPTI hooks into runtime globally
• you aggregate SM occupancy over time

---

5. Important limitation

SM Occupancy is:

not a single instantaneous value

It is derived from:

• active warps per SM
• theoretical max warps
• scheduling state over time

So:

• Nsight Systems → averaged over time
• Nsight Compute → per kernel
• CUPTI → raw data stream (you compute it)

---

Final answer

• ❌ nvbench alone cannot give SM Active / SM Occupancy programmatically
• ✔ Nsight Systems can give it at process level (closest to what you want)
• ✔ CUPTI can give it programmatically (but complex)
• ✔ best practical solution = nsys + automation script

---

If you want, I can show:

• how to extract SM Active from .nsys-rep via Python
• or how to build a lightweight CUPTI-based occupancy sampler

---

If this answer helped or pointed you in the right direction, I'd appreciate it if you could mark it as the accepted answer so it's easier for others with the same issue to find.

Also, if you found my contribution useful, I'd appreciate it if you could check out my GitHub profile, follow me, and star any repositories you find interesting.

GitHub: https://github.com/Advait251206