HANS Ascend 950 validation

Run the commands below from the repository root on Linux with a CANN toolchain and CPU-debug libraries that support Ascend 950. Record the exact CANN build used; this guide does not specify a minimum or first release.

Host tiling UT

bash build.sh -u --ophost --soc=ascend950 --ops=hans_encode,hans_decode --gtest_filter='*HansEncode950*:*HansDecode950*'

The filter selects 40 cases (28 encode, 12 decode), including parameterized cases. They exercise registered tiling functions with explicit platform resources. Coverage includes FP16/BF16/FP32, statistic/reshuff attributes, buffer and workspace sizes, core limits, partial tiles, invalid parameters, and the existing Ascend910B capacity branch. Omit the filter to include the other HANS host tests.

Kernel UT (CPU debug)

# Encode only
bash build.sh -u --opkernel --soc=ascend950 --ops=hans_encode --gtest_filter='HansEncode950KernelTest.*'

# Decode only
bash build.sh -u --opkernel --soc=ascend950 --ops=hans_decode --gtest_filter='HansDecode950KernelTest.*'

# Both suites
bash build.sh -u --opkernel --soc=ascend950 --ops=hans_encode,hans_decode --gtest_filter='HansEncode950KernelTest.*:HansDecode950KernelTest.*'

The targets compile the 950 apt entries and run through ICPU_RUN_KF. Running these commands on an NPU host still uses CPU debug, not the NPU.

• Encode registers 39 cases; 23 currently return GTEST_SKIP() because of unresolved CPU-debug failures. The other 16 exercise FP32 encoding, var-only output and boundary checks. Active encode work uses one core; names containing multicore do not establish multicore coverage.
• Decode registers 44 cases. Checks include exact output bytes for var-only and constructed single-tile fixed streams, both byte-width specializations, reshuffling, multiple cores and invalid metadata.
• Kernel fixtures may bypass host validation to test empty/small inputs and insufficient capacity. They do not extend the supported operator inputs. Invalid-stream checks do not guarantee rollback after partial output.

For commit 113ef4201, CI #3433 reported 60 passed and 23 skipped: encode 16 passed/23 skipped, decode 44 passed. This is not an 83-case pass or an NPU round-trip result. Revalidate after source or environment changes; do not count skips as passes.

Ascend 950 output capacity

Let N be the input element count (a multiple of 64, at least 32768), and B the element width in bytes. The mantissa buffer must contain N * (B - 1) bytes. Define:

C = min(floor(N / 32768), available AIV cores, 56)
V = 64 * (floor((N / 64) / C) + (N / 64) % C)
fixed-only bound (bytes) = 512 + C * (V + ceil(V / 4096) * 128 + 8448)

C must be positive; V is the longest core's element count. The bound includes the common header, per-tile metadata and state tails. It is independent of PDF contents and may exceed the space needed for a particular input.

• With reshuff enabled, fixed must meet this bound, which is also used for fixed staging in workspace.
• Without reshuff, fixed must be at least 512 bytes. Tiling checks the var capacity against the remaining symbols after the guaranteed fixed prefix of each core; total fixed + var capacity alone is insufficient. A 512-byte fixed buffer requires N var bytes.
• The implementation also checks signed header-field limits. These rules are specific to Ascend 950; other platforms retain their existing rules.

NPU validation

Build the production kernel separately:

bash build.sh --opkernel --soc=ascend950 --ops=hans_encode,hans_decode

A successful build is not a device test. Use the built operator package for encode/decode round trips on Ascend 950 and compare the recovered bytes against the original input, including NaN payloads and signed zero.

Cover FP16/BF16/FP32, supplied/recomputed PDF, both reshuffle modes, single/multiple cores, fixed-only/var-only/mixed outputs, partial tiles and multi-tile overflow. Decode CPU fixtures currently lack mixed fixed/var and multi-tile overflow coverage. Check malformed streams separately and measure performance on the NPU, not in CPU debug.

Attach the source commit, exact CANN build, driver/firmware, device model, commands and complete logs. For round trips, also record dtype, shape, PDF mode, buffer capacities, reshuffle mode and bytewise comparison results.