TheRock/rocm-libraries/dnn-providers/miopen-provider/config/MIOPEN_ENGINE.toml

# Integration test configuration for MIOPEN_ENGINE
# Filters use GTest-style globs: * matches any sequence of characters
# Later entries take precedence over earlier ones when multiple filters match

[meta]
version = 1

# -- Tolerance Overrides --
# Override default tolerances for specific tests or groups of tests.
# Each entry requires:
#   filters: array of glob patterns matched against the full GTest name
#   atol: absolute tolerance
#   rtol: relative tolerance

# BatchnormBackwardCalcStats 3D /0 (shape [1,16,8,112,112], NCDHW) — DSCALE
# runs ~0.94-1.08% low vs CPU reference (uniform across 16 channels and all 3
# dtypes; impl/ref ratio ~0.9906 to 4 sig figs). Bumped tolerance to keep the
# fixture running outside the gfx11+windows skip scope while epic 1913 is
# investigated. Same fixture is skipped on gfx11+windows above; the skip
# fires in SetUp before this tolerance applies, so the two coexist cleanly.
[[tolerance_overrides]]
filters = ["Full/*BatchnormBackwardCalcStats3d*.Correctness/0"]
atol = 0.002
rtol = 0.012

# ROCm/rocm-libraries#8029 — Due to split-k, conv WRW can cause reduced accuracy
# ROCm/rocm-libraries#8030 — conv-wgrad switched from dynamic per-shape
# tolerances to a fixed 0.2 → flaky CI failures
# Both failing cases share the same shape (1x1x1 filter, batched, ref=0.691406
# at DW[0,0,0,0,0]) and produce a nondeterministic ~2x-wrong value at that
# element only. Across 100-pass sweep: case 14 fails ~45% with peak abs diff
# 1.105; case 5 fails ~3% with peak abs diff 0.438. atol bumped to 1.19
# (rtol unchanged at 0.2) gives a threshold of ~1.328 at the failing element
# — ~20% margin over worst observed — while only marginally loosening the
# per-element check on the (correct) larger-magnitude channels.
[[tolerance_overrides]]
filters = ["Smoke/IntegrationGpuConvWrw3dBfp16.Correctness/14"]
atol = 1.19
rtol = 0.2

# Case 5 recurred in CI exceeding the 1.328 threshold above. A ~1300-run sweep
# on gfx942 showed the excursion is multi-element (worst channels DW[0,0,0,0,0]
# ref=1.39 maxdiff=1.52, DW[0,14,...] ref=-2.56 maxdiff=1.58, DW[0,8,...]
# ref=-4.94 maxdiff=2.00) and that atol=1.6 still failed ~0.2%. Cover the tail
# via rtol instead: atol reverted to 1.19, rtol raised to 0.3 (threshold
# 1.19 + 0.3*|ref|) so the loosening scales with each channel's magnitude.
[[tolerance_overrides]]
filters = ["Smoke/IntegrationGpuConvWrw3dBfp16.Correctness/5"]
atol = 1.19
rtol = 0.3

# 3D fp32 conv forward accumulates a ~432-term dot product in fp32 on both GPU
# and CPU reference, giving an ~1-2e-5 absolute error floor independent of
# output magnitude. The default 1e-5 atol dips below it for small-magnitude
# outputs, flaking IntegrationGpuConvFwd3dFp32 (~19% on Windows gfx1151). Raise
# atol to 5e-5 (rtol unchanged at the 1e-5 default) for the 3D fp32 forward
# cases only; 5e-5 clears the floor while staying ~50x below the worst-case
# bound (~2.5e-3).
[[tolerance_overrides]]
filters = ["Smoke/IntegrationGpuConvFwd3dFp32.Correctness/*"]
atol = 5e-5
rtol = 1e-5

# -- Test Skips --
# Skip tests on specific architectures (or globally if 'archs' is omitted).
# Each entry requires:
#   archs:   optional array of substrings matched against the device's raw
#            gcnArchName (e.g. "gfx942" matches "gfx942:sramecc+:xnack-").
#            Omit or leave empty to skip on every arch.
#   filters: array of GTest globs matched against the full GTest name
#   reason:  human-readable string surfaced in the GTEST_SKIP message
# First matching entry wins.

# ROCm/rocm-libraries#6979 — MIOpen Conv-Bias-Activation (CBA) fusion has
# limited solver coverage: gfx90a returns wrong results, and RDNA archs
# (gfx10*, gfx11*, gfx12*) have no CK fusion kernels. Skip the
# integration-tests CBA fixtures on those archs until the upstream issue
# is resolved.
#
# Replaces the same-arch carve-out previously enforced at runtime by
# IS_WORKAROUND_ISSUE_6979 / shouldSkipOnEngineConfigResult in
# dnn-providers/miopen-provider/tests/common/TestWorkarounds.hpp.
[[test_skips]]
archs   = ["gfx90a", "gfx10", "gfx11", "gfx12"]
filters = ["*ConvFwdBiasActiv*"]
reason  = "ROCm/rocm-libraries#6979 — no engine has an applicable solution for ConvBiasActiv fusion (gfx90a returns wrong results; RDNA archs have no CK fusion kernels)"