---
include: hipblaslt_common.yaml
include: matmul_common.yaml


Tests:
- name: rocroller_predicate
  category: pre_checkin
  function:
    rocroller_predicate: *rocroller_predicate
  M: [128]
  N: [128]
  K: [96]
  transA: T
  transB: N
  alpha: 1.5
  beta: 2.0
  initialization: hpl
  unit_check: 0
  norm_check: 1
  gpu_arch: '950'

# A single rocRoller buffer descriptor's num_records field is 32 bits, so it cannot
# address an operand >= 4 GiB. Here input A = M*K (f8) = 65536*65536 = 4 GiB = 2^32
# bytes, which would otherwise wrap num_records and silently produce all-zero output.
# The kernel must instead be reported unsupported ("NO solution found!"). Output is kept
# tiny (N=256) so the case allocates ~8 GiB (A + host copy), not 4 GiB of fp32 output.
# nightly + gfx950: large allocation, matches the matmul_*_grid_limit large-size group.
- name: rocroller_predicate_4gib_buffer
  category: nightly
  function:
    rocroller_predicate:
    - { a_type: f8_r, b_type: f8_r, c_type: f32_r, d_type: f32_r, compute_type: c_f32_r, scaleA: 3, scaleB: 3, scale_type: f32_r }
  M: [65536]
  N: [256]
  K: [65536]
  transA: T
  transB: N
  alpha: 1
  beta: 0
  initialization: hpl
  unit_check: 0
  norm_check: 1
  gpu_arch: '950'

...