{
  "recorded_at": "2026-10-02T15:33:51.276436+00:00",
  "platform": "macOS-26.2-arm64-arm-64bit",
  "compiler_implementation": "ce77f7952f45d48ccb6e958524e4999b3f7aaf994b6b0c14d738b9119131411f",
  "benchmark_source_sha256": "e002a169fd595a106b48f97df716bb81ae7205aa10eb8a960b4263853a04b65b",
  "device": "Apple M5",
  "metal_compiler": "Apple metal version 32023.864 (metalfe-32023.864)\nTarget: air64-apple-darwin25.2.0\nThread model: posix\nInstalledDir: /private/var/run/com.apple.security.cryptexd/mnt/com.apple.MobileAsset.MetalToolchain-v17.3.7003.10.GzPOjD/Metal.xctoolchain/usr/metal/current/bin",
  "mlx_version": "0.32.0",
  "environment": {
    "MLX_ENABLE_TF32": "0"
  },
  "metric": "alternating, per-invocation synchronized wall latency; setup excluded",
  "precision_comparison": {
    "class": "same_storage_precision",
    "same_weight_representation": true,
    "baseline_weights": "float32",
    "optimized_weights": "float32",
    "bitwise_exact": false,
    "storage": "FP32 inputs, intermediate and outputs; no FP16 rounding boundary",
    "metile": "FP32 accumulation, RELAXED_PRECISION=False",
    "mlx": "mx.compile with FP32 inputs and MLX_ENABLE_TF32=0",
    "reference": "NumPy FP32 matmul and epilogue; tolerance, not bitwise equivalence",
    "rtol": 0.0003,
    "atol": 3e-05
  },
  "configuration": {
    "sizes": [
      64,
      256,
      1024
    ],
    "seed": 0,
    "backend": "auto",
    "warmup_ms": 100,
    "rep_ms": 500,
    "output": "benchmarks/results/m5-schedule-fusion.json"
  },
  "cases": [
    {
      "shape": [
        64,
        64,
        64
      ],
      "seed": 64,
      "max_absolute_error": {
        "metile_fused": 5.960464477539063e-08,
        "metile_two_launches": 5.960464477539063e-08,
        "mlx_compile": 0.0
      },
      "fusion_pair_us": {
        "fused": 323.271,
        "two_launches": 332.646,
        "speedup": 1.02900043616656
      },
      "mlx_pair_us": {
        "fused": 322.29200000000003,
        "mlx_compile": 331.45900000000006,
        "speedup": 1.0284431509314536
      },
      "execution_reports": {
        "fused": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 1,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "product": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "separate_epilogue": {
          "plan": {
            "backend": "elementwise",
            "threadgroup_size": [
              256,
              1,
              1
            ],
            "simdgroup_grid": null,
            "tile_shape": [
              256
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Lane geometry follows the traced arange/BLOCK contract."
            ],
            "cooperative_layout": "scalar lanes",
            "outer_bounds_proven": false
          },
          "passes": [
            "split_elementwise_loops",
            "vectorize_elementwise",
            "fold_constants"
          ],
          "loops": [],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        }
      }
    },
    {
      "shape": [
        256,
        256,
        256
      ],
      "seed": 256,
      "max_absolute_error": {
        "metile_fused": 5.960464477539063e-08,
        "metile_two_launches": 5.960464477539063e-08,
        "mlx_compile": 8.344650268554688e-07
      },
      "fusion_pair_us": {
        "fused": 225.7705,
        "two_launches": 231.39600000000002,
        "speedup": 1.0249168956971793
      },
      "mlx_pair_us": {
        "fused": 372.959,
        "mlx_compile": 392.083,
        "speedup": 1.0512764137612982
      },
      "execution_reports": {
        "fused": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 1,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "product": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "separate_epilogue": {
          "plan": {
            "backend": "elementwise",
            "threadgroup_size": [
              256,
              1,
              1
            ],
            "simdgroup_grid": null,
            "tile_shape": [
              256
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Lane geometry follows the traced arange/BLOCK contract."
            ],
            "cooperative_layout": "scalar lanes",
            "outer_bounds_proven": false
          },
          "passes": [
            "split_elementwise_loops",
            "vectorize_elementwise",
            "fold_constants"
          ],
          "loops": [],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        }
      }
    },
    {
      "shape": [
        1024,
        1024,
        1024
      ],
      "seed": 1024,
      "max_absolute_error": {
        "metile_fused": 5.960464477539063e-08,
        "metile_two_launches": 5.960464477539063e-08,
        "mlx_compile": 0.0
      },
      "fusion_pair_us": {
        "fused": 770.7710000000001,
        "two_launches": 790.354,
        "speedup": 1.0254070275088192
      },
      "mlx_pair_us": {
        "fused": 786.1460000000001,
        "mlx_compile": 850.729,
        "speedup": 1.082151406990559
      },
      "execution_reports": {
        "fused": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 1,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "product": {
          "plan": {
            "backend": "tensor_ops",
            "threadgroup_size": [
              128,
              1,
              1
            ],
            "simdgroup_grid": [
              2,
              2
            ],
            "tile_shape": [
              64,
              64,
              16
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Metal tensor operations support the dtype, shape and selected SIMDgroup grid.",
              "Inputs load directly from device into register-resident cooperative tensors.",
              "Native cooperative-tensor lane ownership is opaque to meTile."
            ],
            "cooperative_layout": "native SDK cooperative_tensor (opaque lane ownership)",
            "outer_bounds_proven": true
          },
          "passes": [
            "optimize_tile_schedules",
            "decompose_nax_fragments",
            "fold_constants"
          ],
          "loops": [
            {
              "induction_variable": "k",
              "kind": "loop",
              "lanes": 128,
              "elements_per_lane": 1,
              "iteration_step": 16
            }
          ],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        },
        "separate_epilogue": {
          "plan": {
            "backend": "elementwise",
            "threadgroup_size": [
              256,
              1,
              1
            ],
            "simdgroup_grid": null,
            "tile_shape": [
              256
            ],
            "staging": "device",
            "vector_width": null,
            "double_buffer": null,
            "reasons": [
              "Lane geometry follows the traced arange/BLOCK contract."
            ],
            "cooperative_layout": "scalar lanes",
            "outer_bounds_proven": false
          },
          "passes": [
            "split_elementwise_loops",
            "vectorize_elementwise",
            "fold_constants"
          ],
          "loops": [],
          "allocations": [],
          "double_buffered": false,
          "vectorized_loads": 0,
          "epilogue_regions": 0,
          "notes": [
            "Layout widths describe emitted Metal operations, not final machine instructions.",
            "Masked tails may use scalar accesses even when aligned interiors are vectorized.",
            "No software double buffering materialized."
          ]
        }
      }
    }
  ]
}