--- # ============================================================ # REQUIRED FIELDS — must be provided by the user # ============================================================ # Matrix transpose options: 'T' or 'N' TRANSA: 'N' TRANSB: 'N' # Data types. Valid: H/B/I8/S/D/C/Z/F8/F8B8/B8F8/X/X1 DataType: "B" DestDataType: "B" ComputeDataType: "S" # Hardware target. Valid: gfx950, gfx942, gfx942_80cu, gfx942_38cu, gfx942_20cu, gfx942_228cu ARCH: "gfx950" # Sizes as [M, N, B, K] lists (required when SIZE_OPTION=0, the default) Sizes: - [1024, 1024, 1, 1024] - [4096, 4096, 1, 4096] # ============================================================ # OPTIONAL FIELDS — defaults shown below # ============================================================ # StreamK: True # True=StreamK (default), False=DataParallel # GA: True # True=Ductile/GA tuning, False=exhaustive # MACROTILE_OPT: False # Origami MT tuning (GA only) # MT_DU: None # Fixed [MT0, MT1, DU] for MACROTILE_OPT # USE_HEURISTICS: False # Refined heuristic param lists per size # SIZE_OPTION: 0 # 0=size list, 1=grid # ONE_SIZE_PER_CONFIG: True # One size per output config file # CMS: True # CMS optimization # CMS_PRIORITY: True # Prioritize CMS tiles # MI_FILTER: 2 # MI filtering aggressiveness (0=none, 1=moderate, 2=aggressive) # EPILOGUES: True # Include epilogues in tuning # CLUSTER: 0 # 0=all-in-one, 1=cluster by top MI # MAX_NUM_KERNELS_PER_CONFIG: 180000000 # Kernel cap per config (non-GA only; ignored in GA mode) # GRID_DENSITY: 4 # Grid density 1-5 (only for SIZE_OPTION=1)