TM4C123G LaunchPad DAQ

ADC0 SS3, sampled by uDMA in a ping-pong pair, hits 400,687 S/s on command — getting there needs the ADC_INT_DMA_SS3 completion interrupt enabled and cleared, not the sequence-completion interrupt ADCSequenceDMAEnable implies, but once that's right the ADC keeps up with anything asked of it.

Finished samples land in a 16 KB ring buffer. From there the current design differs from the first one: main.c now keeps its own 4 KB transmit ring instead of calling into usblib's USBBufferWrite, filling and draining it with whole-packet memcpy calls. That rewrite is Section 04's subject — the fix for a bottleneck that first looked like the USB bus, and wasn't.

USBBufferWrite calls down into TivaWare's USBRingBufWrite, which copies a batch one byte at a time through UpdateIndexAtomic — a helper that globally disables and re-enables interrupts on every single byte. A 768-byte batch is 768 function calls and 768 CPSID/CPSIE pairs. Measured directly: **1,564 µs per call and 69.9% of the CPU**, against 17.4% for the entire USB interrupt handler combined. Double-packet buffering was never going to help a device that was CPU-bound before a sample ever reached the endpoint.

The fix was to stop calling USBBufferWrite at all. main.c keeps a private 4 KB transmit ring and hands the CDC class whole 64-byte packets directly, filling and draining both ends with memcpy. Only after that rewrite does the original diagnosis become true: the transmit ring now measures never empty (0%), and 94% of send attempts find the class still busy with the previous packet — the device is genuinely serialized on one packet in flight, for the first time. Double-packet buffering is the correct next step now, not the dead end it looked like the first time around.

Burst capture fills SRAM at full ADC rate, with an edge trigger and a pre-trigger buffer, then drains the finished frame over the same USB link afterward, off the acquisition critical path — sidestepping the link's limit entirely, whatever that limit measures on a given day. Streaming and burst are mutually exclusive and share the same buffer; while burst is armed, the ADC ISR stops filling the ring and the streaming drain is skipped entirely (measured: zero stray samples).

Verified at all five ADC rate presets — 100 k, 200 k, 250 k, 333 k and 400 k S/s — each returning a complete 8192-sample frame with the trigger at exactly the requested offset. Checked against the acquired rate measured from the data itself, against the known PWM period: 100 k, 200 k, 250 k and 400 k land at +0.00%, and 333 k at 333,320 S/s (−0.004%). The top two rate presets exceed what the link can stream continuously — exactly the case burst capture exists for.

Burst capture had a trigger from the start; the live view did not. It drew the last N samples of the ring on every refresh wherever the edges happened to fall, so a 10 kHz square wave at a 2000-sample window was a scrolling smear. The fix is the same thing every oscilloscope does: on each refresh, search the freshest few windows of the ring for the most recent level crossing in the chosen slope — with a small hysteresis band so noise cannot re-arm it — and draw the window aligned to that edge, pre-trigger fraction to its left.

It reuses the level, slope and pre-trigger controls the burst engine already had, so one setting drives both, and it switches the x-axis to milliseconds with t=0 on the edge to match the burst frame. The search is vectorised numpy: a 500 k-sample window costs under 25 ms, a normal one tens of microseconds. Nothing in the firmware or the protocol changed. The space bar is Run/Stop, with the ring still filling underneath so that releasing shows the present rather than the past.

The transmit ring has its own producer/consumer discipline to get right: ControlHandler runs in USB interrupt context, and the main loop is the ring's only producer. A flush landing between the producer's memcpy and its head update puts the stale head back, and the device then transmits kilobytes of stale bytes — which, with no sync marker in the protocol, the host never recovers from. Free space must also be re-read after servicing the status queue, since those writes share the same ring; sizing a batch from a figure taken before them overruns it by up to a full status queue.