- Home
- Skills
- Testing & Debugging
- FPGA AXI Performance Monitor Bring-Up
FPGA AXI Performance Monitor Bring-Up
Your DMA is not keeping up, frames drop at peak rate, and nobody knows which master on the bus is the problem.
$30
/moSubscription · cancel anytime
Secure checkout via Stripe
- Always the latest version
- Delivered live to your agent via MCP
- Cancel anytime, access ends at period end
Works with the AI Agents you already use
FPGA AXI Performance Monitor Bring-Up
Example session with this skill installed
Calculate metrics for a 128-bit bus at 200MHz core clock using the sample file at assets/example/apm_samples.csv. The theoretical peak is 19,200 MB/s.
- Read your context and instructions
- Compiled the fpga axi performance
Raw counters processed. Applied 200 MHz core clock frequency for time-domain conversion on a 128-bit (16-byte) bus width.
AXI Performance Metrics Report
| Slot | Write BW | Read BW | Utilization | Avg Latency | Burst Eff | Verdict |
|---|---|---|---|---|---|---|
| 0 | 3200.0 MB/s | 0.0 MB/s | 16.7% | 265.0 ns | 1.00 | [OK] No dominant stall detected; bursts are full. |
| 1 | 0.0 MB/s | 2000.0 MB/s | 10.4% | 420.0 ns | 0.50 | [WARN] 50% beat waste; check master burst config. |
Observations
- Slot 1 shows low bus efficiency (0.31). The master read idle count is high, indicating the consumer is not draining the bus.
- Total system utilization is well below the 19,200 MB/s peak, ruling out interconnect saturation.
Next steps
- Check the master configuration on Slot 1 to increase burst size from 32B to 64B.
- Verify the devicetree using
apm_devicetree_check.pyto ensure the 200MHz core clock is correctly mapped. - Apply the interrupt handler patch if
/proc/interruptsshows high spurious counts.
Subscription · always the latest version
- Secure checkout via Stripe
- Cancel anytime
- Delivered live via MCP
About this skill
Your DMA is not keeping up, frames drop at peak rate, and nobody knows which master on the bus is the problem. The usual answer is an ILA, hand-written RTL counters, and days of narrowing down. The measurement is already in the silicon: the AXI Performance Monitor sits parallel to the bus — no added latency, no timing impact — and on Zynq UltraScale+ MPSoC four instances are hardened into the PS interconnect and need no RTL at all.
Getting a CORRECT number out of it is the hard part, and this skill is built around the three failures that each happen silently:
-
The counters are in cycles of a clock the vendor UIO driver does not export. It hands user space the AXI4-Lite clock, not the core clock the internals count with — so every bandwidth and latency figure comes out wrong by a constant factor, with no error anywhere. The skill carries the driver change and the devicetree fix for the related case where the platform clock controller reports a frequency of zero.
-
The documented programming sequence does not work as written. The Load bit for the sample interval acts on the FALLING edge: it must be set AND cleared. Follow the document and you get a sampler that runs, produces samples, and ignores your interval — every downstream number averaged over the wrong window, dashboard looking healthy.
-
The mode you pick may not be able to see your bus. Of the three operation modes only Advanced handles AXI4-Stream. And the counter budget runs opposite to the names: the most capable mode has the FEWEST counters.
Also covered: the second driver patch that stops the handler claiming every interrupt on a shared line (which corrupts interrupt statistics and disables the kernel's spurious-interrupt protection); the hardened PS instances with their exact slot and counter counts and per-instance clocks; the hardware histogram you get for free from the range registers; why per-master ID filtering stops working across an interconnect; and the full set of derived metrics with formulas — bandwidth, bus utilisation, average latency in nanoseconds, bus efficiency, burst size, burst efficiency, throughput efficiency.
And the counter nobody expects to be the answer: what engineers call "latency" is usually an idle counter. Slave write idle names the destination, master read idle names the requester. One number, and you know which side is stalling.
Three tools ship with it, all pure-stdlib Python, offline, no telemetry, read-only: a metric calculator that turns a CSV of sampled counters into every derived metric plus a per-slot verdict; a mode advisor that refuses impossible flavour/mode combinations before you instantiate anything; and a devicetree checker that catches the missing core clock. All three carry self-tests against hand-computed vectors (26 checks), and each was run with a deliberately corrupted copy to prove the tests can fail.
Who it is for: FPGA and embedded-Linux teams on Zynq UltraScale+ MPSoC running data paths with DMA, video or measurement streams who are hunting throughput problems at system level — and firmware/driver people who want hardware metrics in Prometheus or Grafana.
Local-first and quiet: no network access, no telemetry, nothing phoned home. Python 3.8+ standard library only — nothing to install.
Cannot attribute traffic to masters post-interconnect. PS instances have a restricted metric subset. PL-only event logging. Does not perform live register R/W.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 22 days ago
Requires