{ "recorded_at": "2026-10-02T16:44:40.288463+00:00", "benchmark_source_sha256": "1f567742f6b2b71e94b90bcdf9c10e70c58952d53a061c3a8a110d5541fe2344", "metric": "warmed per-dispatch Metal GPU timestamps and synchronized wall latency; compilation and allocation excluded", "pair_order": [ "baseline", "current", "current", "baseline" ], "precision_comparison": { "class": "same_storage_precision", "same_weight_representation": true, "storage": "identical FP16 or FP32 inputs and outputs per case", "arithmetic": "SIMD-group FP32 accumulation; RELAXED_PRECISION=False for both variants", "reference": "seeded NumPy FP32 matmul checked before timing every case", "bitwise_exact": false }, "configuration": { "warmup_ms": 200.0, "rep_ms": 500.0, "seed": 431, "cooldown_seconds": 2.0 }, "samples": { "baseline": [ { "root": "/private/tmp/metile-layout-baseline.Jdz5Ug", "compiler_implementation_sha256": "a052f8823627e3c7880bb620fe2574677ec08383e7cbd3e1c766ec2eb8d13c38", "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", "platform": "macOS-26.2-arm64-arm-64bit", "cases": [ { "name": "float16_64x64x64", "dtype": "float16", "shape": [ 64, 64, 64 ], "seed": 431, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "f7743c34dfc78564a28e726c1924c410fbdb76f4158beba809e487c7fdc4d8b3", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1383, "gpu_us": 64.29170025512576, "wall_us": 365.084, "gpu_min_max_us": [ 64.1250517219305, 77.75000995025039 ], "wall_min_max_us": [ 141.791, 463.209 ] }, "status": "ok" }, { "name": "float16_256x256x256", "dtype": "float16", "shape": [ 256, 256, 256 ], "seed": 432, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "93ab6d6120e1756849446ff813e60fbd0616fa9eec32fbbd8c515ecaa3488b3d", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1005, "gpu_us": 191.8749767355621, "wall_us": 500.875, "gpu_min_max_us": [ 188.74998204410076, 271.9166805036366 ], "wall_min_max_us": [ 283.625, 587.792 ] }, "status": "ok" }, { "name": "float16_1024x1024x1024", "dtype": "float16", "shape": [ 1024, 1024, 1024 ], "seed": 433, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "a3cdef8bbc69c848e84c524701f359db6268c696607cea04283a35e3be9e32c1", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 483, "gpu_us": 817.4583199433982, "wall_us": 1031.667, "gpu_min_max_us": [ 801.7499931156635, 1262.7500109374523 ], "wall_min_max_us": [ 896.2090000000001, 1476.625 ] }, "status": "ok" }, { "name": "float16_257x255x259", "dtype": "float16", "shape": [ 257, 255, 259 ], "seed": 434, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "9b256a8e5e5aa03086c616dbb685bfb3b2a3838b8a2de16524f12307645110ef", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1494, "gpu_us": 57.95829929411411, "wall_us": 281.021, "gpu_min_max_us": [ 41.99997056275606, 222.5416828878224 ], "wall_min_max_us": [ 121.583, 571.208 ] }, "status": "ok" }, { "name": "float32_64x64x64", "dtype": "float32", "shape": [ 64, 64, 64 ], "seed": 435, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "5daffc014e3bf6c7425a716bc5a36f0a9f5ac3d06a994a5e9e8084b7cfb9cd1b", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1361, "gpu_us": 68.49999772384763, "wall_us": 371.375, "gpu_min_max_us": [ 68.29167250543833, 189.74998965859413 ], "wall_min_max_us": [ 150.708, 458.875 ] }, "status": "ok" }, { "name": "float32_256x256x256", "dtype": "float32", "shape": [ 256, 256, 256 ], "seed": 436, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "ceba1209c2607c35efa378403bf1bd3a583a3330944ee09239152ee064143b95", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 967, "gpu_us": 212.7499901689589, "wall_us": 520.25, "gpu_min_max_us": [ 210.3750011883676, 240.24996673688293 ], "wall_min_max_us": [ 303.666, 603.917 ] }, "status": "ok" }, { "name": "float32_1024x1024x1024", "dtype": "float32", "shape": [ 1024, 1024, 1024 ], "seed": 437, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "16ed3544359adfe6ed4ecc94e89ed08c70fd5aeef3243f80b2009c2096d05f8c", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 405, "gpu_us": 981.2916978262365, "wall_us": 1199.167, "gpu_min_max_us": [ 970.3333489596844, 1196.2083517573774 ], "wall_min_max_us": [ 1105.042, 7811.625 ] }, "status": "ok" }, { "name": "float32_257x255x259", "dtype": "float32", "shape": [ 257, 255, 259 ], "seed": 438, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "acff8f3e2b353c103ecee705731d3d2c8c26b98f0a80634d1fd4a06c38113a26", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1520, "gpu_us": 58.2499778829515, "wall_us": 280.29150000000004, "gpu_min_max_us": [ 45.41664384305477, 244.666647631675 ], "wall_min_max_us": [ 133.791, 588.417 ] }, "status": "ok" } ] }, { "root": "/private/tmp/metile-layout-baseline.Jdz5Ug", "compiler_implementation_sha256": "a052f8823627e3c7880bb620fe2574677ec08383e7cbd3e1c766ec2eb8d13c38", "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", "platform": "macOS-26.2-arm64-arm-64bit", "cases": [ { "name": "float16_64x64x64", "dtype": "float16", "shape": [ 64, 64, 64 ], "seed": 431, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "f7743c34dfc78564a28e726c1924c410fbdb76f4158beba809e487c7fdc4d8b3", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1371, "gpu_us": 64.87499922513962, "wall_us": 368.958, "gpu_min_max_us": [ 64.62499732151628, 78.62498750910163 ], "wall_min_max_us": [ 142.459, 449.208 ] }, "status": "ok" }, { "name": "float16_256x256x256", "dtype": "float16", "shape": [ 256, 256, 256 ], "seed": 432, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "93ab6d6120e1756849446ff813e60fbd0616fa9eec32fbbd8c515ecaa3488b3d", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 999, "gpu_us": 192.0000067912042, "wall_us": 504.166, "gpu_min_max_us": [ 188.9999839477241, 219.62495520710945 ], "wall_min_max_us": [ 273.333, 625.333 ] }, "status": "ok" }, { "name": "float16_1024x1024x1024", "dtype": "float16", "shape": [ 1024, 1024, 1024 ], "seed": 433, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "a3cdef8bbc69c848e84c524701f359db6268c696607cea04283a35e3be9e32c1", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 485, "gpu_us": 817.3332898877561, "wall_us": 1030.5, "gpu_min_max_us": [ 799.2916507646441, 889.7499646991491 ], "wall_min_max_us": [ 916.583, 1142.375 ] }, "status": "ok" }, { "name": "float16_257x255x259", "dtype": "float16", "shape": [ 257, 255, 259 ], "seed": 434, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "9b256a8e5e5aa03086c616dbb685bfb3b2a3838b8a2de16524f12307645110ef", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1495, "gpu_us": 57.75003228336573, "wall_us": 279.208, "gpu_min_max_us": [ 41.91667539998889, 224.87499518319964 ], "wall_min_max_us": [ 124.917, 574.125 ] }, "status": "ok" }, { "name": "float32_64x64x64", "dtype": "float32", "shape": [ 64, 64, 64 ], "seed": 435, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "5daffc014e3bf6c7425a716bc5a36f0a9f5ac3d06a994a5e9e8084b7cfb9cd1b", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1357, "gpu_us": 68.29161429777741, "wall_us": 374.25, "gpu_min_max_us": [ 68.12496576458216, 86.74996206536889 ], "wall_min_max_us": [ 156.75, 469.834 ] }, "status": "ok" }, { "name": "float32_256x256x256", "dtype": "float32", "shape": [ 256, 256, 256 ], "seed": 436, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "ceba1209c2607c35efa378403bf1bd3a583a3330944ee09239152ee064143b95", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 964, "gpu_us": 212.99999207258224, "wall_us": 521.625, "gpu_min_max_us": [ 208.37498595938087, 239.66666776686907 ], "wall_min_max_us": [ 302.708, 609.75 ] }, "status": "ok" }, { "name": "float32_1024x1024x1024", "dtype": "float32", "shape": [ 1024, 1024, 1024 ], "seed": 437, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "16ed3544359adfe6ed4ecc94e89ed08c70fd5aeef3243f80b2009c2096d05f8c", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 413, "gpu_us": 984.8750196397305, "wall_us": 1202.541, "gpu_min_max_us": [ 972.0000089146197, 1111.0000195913017 ], "wall_min_max_us": [ 1117.833, 1374.375 ] }, "status": "ok" }, { "name": "float32_257x255x259", "dtype": "float32", "shape": [ 257, 255, 259 ], "seed": 438, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "acff8f3e2b353c103ecee705731d3d2c8c26b98f0a80634d1fd4a06c38113a26", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ] }, "timing": { "samples": 1546, "gpu_us": 57.64584057033062, "wall_us": 275.64549999999997, "gpu_min_max_us": [ 45.58329237625003, 241.0417073406279 ], "wall_min_max_us": [ 140.625, 572.2090000000001 ] }, "status": "ok" } ] } ], "current": [ { "root": "/Users/andreslavescu/Documents/projects/meTile", "compiler_implementation_sha256": "d258683dcf0a92130bc92c14d306b75dfdb1194f300bbdc9aafe671cb6ffff65", "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", "platform": "macOS-26.2-arm64-arm-64bit", "cases": [ { "name": "float16_64x64x64", "dtype": "float16", "shape": [ 64, 64, 64 ], "seed": 431, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "f7743c34dfc78564a28e726c1924c410fbdb76f4158beba809e487c7fdc4d8b3", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1309, "gpu_us": 65.33332634717226, "wall_us": 367.41700000000003, "gpu_min_max_us": [ 64.62499732151628, 209.83332069590688 ], "wall_min_max_us": [ 142.958, 10539.292 ] }, "status": "ok" }, { "name": "float16_256x256x256", "dtype": "float16", "shape": [ 256, 256, 256 ], "seed": 432, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "93ab6d6120e1756849446ff813e60fbd0616fa9eec32fbbd8c515ecaa3488b3d", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1003, "gpu_us": 192.16665532439947, "wall_us": 501.875, "gpu_min_max_us": [ 189.33333922177553, 221.2083782069385 ], "wall_min_max_us": [ 283.708, 600.375 ] }, "status": "ok" }, { "name": "float16_1024x1024x1024", "dtype": "float16", "shape": [ 1024, 1024, 1024 ], "seed": 433, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "a3cdef8bbc69c848e84c524701f359db6268c696607cea04283a35e3be9e32c1", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 471, "gpu_us": 850.9583421982825, "wall_us": 1058.583, "gpu_min_max_us": [ 810.9583286568522, 931.1250178143382 ], "wall_min_max_us": [ 919.708, 1179.833 ] }, "status": "ok" }, { "name": "float16_257x255x259", "dtype": "float16", "shape": [ 257, 255, 259 ], "seed": 434, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "9b256a8e5e5aa03086c616dbb685bfb3b2a3838b8a2de16524f12307645110ef", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1500, "gpu_us": 57.999975979328156, "wall_us": 279.833, "gpu_min_max_us": [ 42.12500061839819, 219.2499814555049 ], "wall_min_max_us": [ 126.291, 555.25 ] }, "status": "ok" }, { "name": "float32_64x64x64", "dtype": "float32", "shape": [ 64, 64, 64 ], "seed": 435, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "5daffc014e3bf6c7425a716bc5a36f0a9f5ac3d06a994a5e9e8084b7cfb9cd1b", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1351, "gpu_us": 68.62496957182884, "wall_us": 376.709, "gpu_min_max_us": [ 68.41664435341954, 85.99995635449886 ], "wall_min_max_us": [ 146.0, 482.625 ] }, "status": "ok" }, { "name": "float32_256x256x256", "dtype": "float32", "shape": [ 256, 256, 256 ], "seed": 436, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "ceba1209c2607c35efa378403bf1bd3a583a3330944ee09239152ee064143b95", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 917, "gpu_us": 212.9583153873682, "wall_us": 519.708, "gpu_min_max_us": [ 210.5833264067769, 236.37496633455157 ], "wall_min_max_us": [ 300.208, 6823.292 ] }, "status": "ok" }, { "name": "float32_1024x1024x1024", "dtype": "float32", "shape": [ 1024, 1024, 1024 ], "seed": 437, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "16ed3544359adfe6ed4ecc94e89ed08c70fd5aeef3243f80b2009c2096d05f8c", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 396, "gpu_us": 987.1041984297335, "wall_us": 1200.979, "gpu_min_max_us": [ 970.5416741780937, 1175.4999868571758 ], "wall_min_max_us": [ 1071.666, 10793.042 ] }, "status": "ok" }, { "name": "float32_257x255x259", "dtype": "float32", "shape": [ 257, 255, 259 ], "seed": 438, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "acff8f3e2b353c103ecee705731d3d2c8c26b98f0a80634d1fd4a06c38113a26", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1502, "gpu_us": 58.85417340323329, "wall_us": 280.97950000000003, "gpu_min_max_us": [ 45.66670395433903, 242.24998196586967 ], "wall_min_max_us": [ 134.958, 579.792 ] }, "status": "ok" } ] }, { "root": "/Users/andreslavescu/Documents/projects/meTile", "compiler_implementation_sha256": "d258683dcf0a92130bc92c14d306b75dfdb1194f300bbdc9aafe671cb6ffff65", "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", "platform": "macOS-26.2-arm64-arm-64bit", "cases": [ { "name": "float16_64x64x64", "dtype": "float16", "shape": [ 64, 64, 64 ], "seed": 431, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "f7743c34dfc78564a28e726c1924c410fbdb76f4158beba809e487c7fdc4d8b3", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1370, "gpu_us": 64.9583525955677, "wall_us": 368.875, "gpu_min_max_us": [ 64.7500273771584, 79.24996316432953 ], "wall_min_max_us": [ 149.708, 455.958 ] }, "status": "ok" }, { "name": "float16_256x256x256", "dtype": "float16", "shape": [ 256, 256, 256 ], "seed": 432, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "93ab6d6120e1756849446ff813e60fbd0616fa9eec32fbbd8c515ecaa3488b3d", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1004, "gpu_us": 191.95833010599017, "wall_us": 500.9585, "gpu_min_max_us": [ 189.12501400336623, 216.95834584534168 ], "wall_min_max_us": [ 284.625, 592.9590000000001 ] }, "status": "ok" }, { "name": "float16_1024x1024x1024", "dtype": "float16", "shape": [ 1024, 1024, 1024 ], "seed": 433, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "a3cdef8bbc69c848e84c524701f359db6268c696607cea04283a35e3be9e32c1", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 471, "gpu_us": 849.9999530613422, "wall_us": 1057.709, "gpu_min_max_us": [ 807.3333301581442, 953.9999882690609 ], "wall_min_max_us": [ 922.0, 1158.958 ] }, "status": "ok" }, { "name": "float16_257x255x259", "dtype": "float16", "shape": [ 257, 255, 259 ], "seed": 434, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "9b256a8e5e5aa03086c616dbb685bfb3b2a3838b8a2de16524f12307645110ef", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.002, "atol": 0.002 }, "source_sha256": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "half", "elements": 1088, "bytes": 2176 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "half", "elements": 1040, "bytes": 2080 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "half", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "half", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1394, "gpu_us": 62.62504030019045, "wall_us": 291.04150000000004, "gpu_min_max_us": [ 42.041647247970104, 220.24998906999826 ], "wall_min_max_us": [ 133.875, 3756.583 ] }, "status": "ok" }, { "name": "float32_64x64x64", "dtype": "float32", "shape": [ 64, 64, 64 ], "seed": 435, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "5daffc014e3bf6c7425a716bc5a36f0a9f5ac3d06a994a5e9e8084b7cfb9cd1b", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1411, "gpu_us": 68.91664816066623, "wall_us": 371.66700000000003, "gpu_min_max_us": [ 68.6667044647038, 85.99995635449886 ], "wall_min_max_us": [ 137.667, 529.667 ] }, "status": "ok" }, { "name": "float32_256x256x256", "dtype": "float32", "shape": [ 256, 256, 256 ], "seed": 436, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "ceba1209c2607c35efa378403bf1bd3a583a3330944ee09239152ee064143b95", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 930, "gpu_us": 212.7499901689589, "wall_us": 521.979, "gpu_min_max_us": [ 210.20829444751143, 234.9583082832396 ], "wall_min_max_us": [ 307.834, 2700.0 ] }, "status": "ok" }, { "name": "float32_1024x1024x1024", "dtype": "float32", "shape": [ 1024, 1024, 1024 ], "seed": 437, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "16ed3544359adfe6ed4ecc94e89ed08c70fd5aeef3243f80b2009c2096d05f8c", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": true }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 415, "gpu_us": 976.9166354089975, "wall_us": 1198.625, "gpu_min_max_us": [ 967.8333881311119, 1122.9583178646863 ], "wall_min_max_us": [ 1100.625, 1354.584 ] }, "status": "ok" }, { "name": "float32_257x255x259", "dtype": "float32", "shape": [ 257, 255, 259 ], "seed": 438, "schedule": { "backend": "simdgroup", "double_buffer": true, "BLOCK_M": 64, "BLOCK_N": 64, "BLOCK_K": 16, "RELAXED_PRECISION": false }, "input_sha256": "acff8f3e2b353c103ecee705731d3d2c8c26b98f0a80634d1fd4a06c38113a26", "correctness": { "reference": "NumPy FP32 matmul rounded to the output storage dtype", "max_absolute_error": 0.0, "rtol": 0.0003, "atol": 3e-05 }, "source_sha256": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724", "execution_report": { "plan": { "backend": "simdgroup", "threadgroup_size": [ 128, 1, 1 ], "simdgroup_grid": [ 2, 2 ], "tile_shape": [ 64, 64, 16 ], "staging": "threadgroup", "vector_width": null, "double_buffer": true, "reasons": [ "The selected SIMDgroup grid partitions 8x8 accumulator fragments without overlap.", "Cooperative loads stage operand tiles in threadgroup memory." ], "cooperative_layout": "8x8 simdgroup_matrix fragments", "outer_bounds_proven": false }, "passes": [ "pad_shared_memory", "double_buffer_k_loop", "vectorize_loads", "serpentine_mma", "preload_mma_tiles", "block_swizzle", "fold_constants" ], "loops": [ { "induction_variable": "kb", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 16 }, { "induction_variable": "kk", "kind": "loop", "lanes": 128, "elements_per_lane": 1, "iteration_step": 8 } ], "allocations": [ { "name": "shared_a_0", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_a_1", "address_space": "threadgroup", "element_type": "float", "elements": 1088, "bytes": 4352 }, { "name": "shared_b_0", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 }, { "name": "shared_b_1", "address_space": "threadgroup", "element_type": "float", "elements": 1040, "bytes": 4160 } ], "double_buffered": true, "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." ], "value_layouts": [], "layout_conversions": [], "pipelines": [ { "buffers": [ { "logical_name": "shared_a", "slots": [ "shared_a_0", "shared_a_1" ], "element_type": "float", "elements": 1088, "current_pointer": "sa_curr", "next_pointer": "sa_next" }, { "logical_name": "shared_b", "slots": [ "shared_b_0", "shared_b_1" ], "element_type": "float", "elements": 1040, "current_pointer": "sb_curr", "next_pointer": "sb_next" } ], "phases": [ { "name": "load_first", "region": "prologue", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_first", "region": "prologue", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "recycles": [], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "load_next", "region": "steady", "kind": "copy", "reads": [], "writes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_current", "region": "steady", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "publish_and_recycle", "region": "steady", "kind": "barrier", "reads": [], "writes": [], "publishes": [ { "buffer": "shared_a", "slot": "next" }, { "buffer": "shared_b", "slot": "next" } ], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" }, { "name": "rotate_slots", "region": "steady", "kind": "rotate", "reads": [], "writes": [], "publishes": [], "recycles": [], "rotates": [ "shared_a", "shared_b" ], "barrier_scope": null, "barrier_flags": null }, { "name": "consume_last", "region": "drain", "kind": "compute", "reads": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "writes": [], "publishes": [], "recycles": [], "rotates": [], "barrier_scope": null, "barrier_flags": null }, { "name": "release_last", "region": "drain", "kind": "barrier", "reads": [], "writes": [], "publishes": [], "recycles": [ { "buffer": "shared_a", "slot": "current" }, { "buffer": "shared_b", "slot": "current" } ], "rotates": [], "barrier_scope": "threadgroup", "barrier_flags": "mem_threadgroup" } ], "stages": 2, "mechanism": "software" } ] }, "timing": { "samples": 1554, "gpu_us": 55.145821534097195, "wall_us": 277.60400000000004, "gpu_min_max_us": [ 45.874970965087414, 245.95833383500576 ], "wall_min_max_us": [ 142.167, 603.708 ] }, "status": "ok" } ] } ] }, "comparisons": [ { "name": "float16_64x64x64", "shape": [ 64, 64, 64 ], "dtype": "float16", "status": "ok", "gpu_us": { "baseline": 64.58269121238443, "current": 65.14556968129843, "speedup": 0.9913596815306447, "samples": { "baseline": [ 64.29170025512576, 64.87499922513962 ], "current": [ 65.33332634717226, 64.9583525955677 ] } }, "wall_us": { "baseline": 367.0158885824973, "current": 368.14527821907484, "speedup": 0.9969322175146698, "samples": { "baseline": [ 365.084, 368.958 ], "current": [ 367.41700000000003, 368.875 ] } }, "source_sha256": { "baseline": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "current": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1" } }, { "name": "float16_256x256x256", "shape": [ 256, 256, 256 ], "dtype": "float16", "status": "ok", "gpu_us": { "baseline": 191.9374815826498, "current": 192.0624644695707, "speedup": 0.9993492591732275, "samples": { "baseline": [ 191.8749767355621, 192.0000067912042 ], "current": [ 192.16665532439947, 191.95833010599017 ] } }, "wall_us": { "baseline": 502.5178059034324, "current": 501.41654060022813, "speedup": 1.0021963082866912, "samples": { "baseline": [ 500.875, 504.166 ], "current": [ 501.875, 500.9585 ] } }, "source_sha256": { "baseline": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "current": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1" } }, { "name": "float16_1024x1024x1024", "shape": [ 1024, 1024, 1024 ], "dtype": "float16", "status": "ok", "gpu_us": { "baseline": 817.3958025249793, "current": 850.4790126309393, "speedup": 0.9611004979374884, "samples": { "baseline": [ 817.4583199433982, 817.3332898877561 ], "current": [ 850.9583421982825, 849.9999530613422 ] } }, "wall_us": { "baseline": 1031.083334895875, "current": 1058.145909762448, "speedup": 0.9744245338786515, "samples": { "baseline": [ 1031.667, 1030.5 ], "current": [ 1058.583, 1057.709 ] } }, "source_sha256": { "baseline": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "current": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1" } }, { "name": "float16_257x255x259", "shape": [ 257, 255, 259 ], "dtype": "float16", "status": "ok", "gpu_us": { "baseline": 57.854072072102824, "current": 60.26815770467437, "speedup": 0.9599442603770795, "samples": { "baseline": [ 57.95829929411411, 57.75003228336573 ], "current": [ 57.999975979328156, 62.62504030019045 ] } }, "wall_us": { "baseline": 280.1130331990998, "current": 285.3822280197211, "speedup": 0.9815363596493571, "samples": { "baseline": [ 281.021, 279.208 ], "current": [ 279.833, 291.04150000000004 ] } }, "source_sha256": { "baseline": "925bd354b757114e81e8fe33b26e1821d01d90e9e0e26ac3dc4207f5dfc30f3b", "current": "775bf51864f7fadf5c0a51035e57a2f85c7857b382291cefd60dd69bf86f92d1" } }, { "name": "float32_64x64x64", "shape": [ 64, 64, 64 ], "dtype": "float32", "status": "ok", "gpu_us": { "baseline": 68.39572664981074, "current": 68.77065422851635, "speedup": 0.9945481458201667, "samples": { "baseline": [ 68.49999772384763, 68.29161429777741 ], "current": [ 68.62496957182884, 68.91664816066623 ] } }, "wall_us": { "baseline": 372.80972861501346, "current": 374.1795075936147, "speedup": 0.9963392464023206, "samples": { "baseline": [ 371.375, 374.25 ], "current": [ 376.709, 371.66700000000003 ] } }, "source_sha256": { "baseline": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "current": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724" } }, { "name": "float32_256x256x256", "shape": [ 256, 256, 256 ], "dtype": "float32", "status": "ok", "gpu_us": { "baseline": 212.87495442026582, "current": 212.8541272915813, "speedup": 1.0000978469572075, "samples": { "baseline": [ 212.7499901689589, 212.99999207258224 ], "current": [ 212.9583153873682, 212.7499901689589 ] } }, "wall_us": { "baseline": 520.9370463405346, "current": 520.8422622368504, "speedup": 1.0001819823592601, "samples": { "baseline": [ 520.25, 521.625 ], "current": [ 519.708, 521.979 ] } }, "source_sha256": { "baseline": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "current": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724" } }, { "name": "float32_1024x1024x1024", "shape": [ 1024, 1024, 1024 ], "dtype": "float32", "status": "ok", "gpu_us": { "baseline": 983.0817260883854, "current": 981.9972058657142, "speedup": 1.0011044025545013, "samples": { "baseline": [ 981.2916978262365, 984.8750196397305 ], "current": [ 987.1041984297335, 976.9166354089975 ] } }, "wall_us": { "baseline": 1200.8528150223067, "current": 1199.8014226841874, "speedup": 1.000876305293727, "samples": { "baseline": [ 1199.167, 1202.541 ], "current": [ 1200.979, 1198.625 ] } }, "source_sha256": { "baseline": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "current": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724" } }, { "name": "float32_257x255x259", "shape": [ 257, 255, 259 ], "dtype": "float32", "status": "ok", "gpu_us": { "baseline": 57.94712191529367, "current": 56.96983186767813, "speedup": 1.017154518726428, "samples": { "baseline": [ 58.2499778829515, 57.64584057033062 ], "current": [ 58.85417340323329, 55.145821534097195 ] } }, "wall_us": { "baseline": 277.95879310295254, "current": 279.2866504471703, "speedup": 0.9952455395125699, "samples": { "baseline": [ 280.29150000000004, 275.64549999999997 ], "current": [ 280.97950000000003, 277.60400000000004 ] } }, "source_sha256": { "baseline": "f3f104bd99ff3f71a2487b1d00cfd450cf4ed9085888a27b979798e443a98808", "current": "4be0fb2e516dd502b4eab90ce5d42bc755710fb3331f6f7e4ea2f7c96e11e724" } } ] }