# A list of known bugs/failures. The bug can be specific or general.
# A test case is considered as matching if all indicated parameters match.
# Wildcards can be used for the function
#example
#- { function: a_type: bf16_r, b_type: bf16_r, c_type: bf16_r, d_type: bf16_r, compute_type: f32_r, transA: N, transB: N, M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 5.0, alphai: 0.0, beta: 0.0, betai: 0.0, known_bug_platforms: gfx908 }

#Known bugs:
#- { function: matmul, grouped_gemm: 1, known_bug_platforms: "gfx1100, gfx1101, gfx1102, gfx1200, gfx1201" }
#- { function: matmul, gradient: 1, known_bug_platforms: "gfx1100, gfx1101, gfx1102, gfx1200, gfx1201" }

Known bugs:
  # ROCM-1545: TF32 matmul with INF inputs returns NaN on gfx950 (MI355)
  - { name: matmul_tf32_inf_ROCM1545, function: matmul, initialization: inf, known_bug_platforms: "gfx950, gfx1250" }
  # ROCM-26455: bf16 matmul output-store address overflow -- an output element
  # offset > 2^31 (from an oversized output leading dimension ldc/ldd) wraps in the
  # 32-bit global-write address path and produces silent wrong results. The gate
  # matmul_size_t_output_overflow_bf16_ROCM26455 is the bit-exact reproducer
  # (FAIL-pre-fix / PASS-post-fix). It is registered here as a known failure on the
  # affected data-center arches so the reproducer can land ahead of the kernel fix.
  # REMOVE this entry once the output-store path is fixed so the case re-activates
  # as an enforced bit-exact regression gate.

  # ROCM-27082: gfx950 MXFP4 subtile scaleAlphaVec TN cases fail the norm/accuracy
  # check on asymmetric (M!=N) tiles. Square tiles (128x128, 192x192) still pass and
  # remain covered. Suspected TensileLite codegen regression; remove once root-caused.
  - { name: matmul_subtile_fp4_scaleAlphaVec, function: matmul, M: 128, N: 192, known_bug_platforms: gfx950 }
  - { name: matmul_subtile_fp4_scaleAlphaVec, function: matmul, M: 192, N: 128, known_bug_platforms: gfx950 }