March Test

Overview

march_test runs a March C- sequence over VRAM. The bandwidth workloads read and write in whatever order goes fastest; a march applies a fixed, ordered read-then-write sequence to every cell, so the order of operations is itself part of the test.

That ordering is what makes it different. Bandwidth tests find cells that cannot hold a value. A march finds cells whose value depends on what happened to a neighbouring cell just before — coupling faults that are invisible to any test touching memory in parallel or in arbitrary order.

Execution Mechanics

March C- is six elements, run in both address directions. w0 writes zeros, r0 reads and expects zeros, ⇑ ascends and ⇓ descends:

#	Element	Catches
0	`⇑ (w0)`	— (initialisation)
1	`⇑ (r0, w1)`	Stuck-at-1, transition faults
2	`⇑ (r1, w0)`	Stuck-at-0, transition faults
3	`⇓ (r0, w1)`	Coupling faults, descending
4	`⇓ (r1, w0)`	Coupling faults, descending
5	`⇓ (r0)`	Final verification

Running both directions is what separates March C- from a simple write/read pass: a coupling fault between two cells is only exposed from one direction.

A march is inherently sequential and a GPU is not, so each thread is given a private contiguous chunk and marches it in order. The ordered sequence is preserved within a chunk while chunks run in parallel. Startup reports Chunk/thread so the ordered run length is visible; the host caps the thread count so no chunk falls below 4096 elements, since coupling faults need a run of addresses to appear.

Target Subsystems

Primary Target: DRAM cell integrity — stuck-at, transition, and coupling faults.
Secondary: Address ordering behaviour within a contiguous region.

Failure Symptoms

Critical Failures

March error: Reported with the element index, expected and actual values, and an XOR mask naming the flipped bits.
Which element failed tells you the fault class. A failure at element 1 or 2 is a stuck-at or transition fault. A failure only in the descending elements (3, 4) points at a coupling fault.
One bit position failing across many unrelated addresses indicates a stuck data lane rather than a defective cell.

Coverage boundary

Coupling faults between chunks are not covered. Sweeping --grid_size moves the chunk boundaries.

Usage

pantheon --test march_test --duration 60 --gpu 0 --mem 50

Record every failing address rather than the capped console sample:

./build/march_test 0 60 50 --verify --fault_map march_faults.csv

Result

Throughput reports march-ops/s — read and write operations across all march elements. It is a progress metric, not a bandwidth figure: the ordered access pattern is deliberately not the fastest way to move data.