{ "metadata": { "platform": "macOS-26.6.2-arm64-arm-64bit-Mach-O", "date": "2026-09-08T02:58:06.989479+00:00", "rounds": 2, "samples": 7, "timeout_seconds": 60, "sample_ms": 30, "warmup_ms": 100, "packet_width": 8, "workers": 8, "timing": "synchronized_host_wall", "kernel_only": false, "cases": [ [ "rmsnorm", [ 17, 7 ] ], [ "rmsnorm", [ 17, 127 ] ], [ "rmsnorm", [ 64, 256 ] ], [ "rmsnorm", [ 1024, 256 ] ], [ "rmsnorm", [ 64, 513 ] ], [ "layernorm", [ 64, 256 ] ], [ "masked_softmax", [ 17, 65 ] ], [ "swiglu", [ 17, 65 ] ] ], "artifacts_sha256": { "baseline": { "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir": "79e517adfc84707ea3987341a84782388c001fef775e973e4de6f2ab2b4f4b7a", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libSPIRV.16.3.0.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libSPIRV.16.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libSPIRV.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang-default-resource-limits.16.3.0.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang-default-resource-limits.16.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang-default-resource-limits.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang.16.3.0.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang.16.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libglslang.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-ast.dylib": "3593c30c2c40059772688aaa0512f51e8e7b9dd84a659dedabeeef4f5d447190", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-backend-metal.so": "292c088e42be06f6324bf733a43bfe00a47698d6789ef45461b6c2c6b3dfeb63", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-backend-simd.so": "87111d1581dfcdc92685cf775964c70928cb3bdcd07d1a23eab7ab6b6929058f", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-benchmark-metal-timing.dylib": "989657bf3419937efd8de38e17681574ee73f28455cc2548184818f589cf7ae8", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-core.dylib": "55f31daf2e618e78bfa9e4686c47e6d7cb4a8f3727e54c97952d719e69f663b4", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-coro.dylib": "9b6b2ed9528c537b3236d3f2e1a9e0dcbe68e501a0ca5da7b9c47b19a88f65b3", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-dsl.dylib": "192fa9ea1ee2c49caf241ca79b93c776490c553ecc215046e8e245ec2158b7a5", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-osl.dylib": "d6f0121521dd0f25dc5fc1c37afdca6dff8f0280a1da85450da36449a0261fb1", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-runtime.dylib": "97165ca9efb0a1ee57eee076fd145bc0a95f373f578cd595dd1c0613cabcb4ac", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-tile-bridge-tirx.dylib": "de6a123adaf2a678e1e74bfd29f0c87789b71df94229c2728b00434b3704fd99", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-tile-bridge-xir.dylib": "0631c149e01ea6a6f74f8825447db7ade3e97f60fe1e5646ac2bca1304db3874", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-tile.dylib": "634878d1d9d5d27ca5a3b585916f2f602718cb2d1b327abde37ea6c9a29283e2", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-validation-layer.so": "1b791d7f116d07b095ff13925bf7e475c1eebb7ba02424f3411aa9b02f37f7eb", "/private/tmp/luisa-xir-indexable.21Tuda/baseline/libluisa-xir.dylib": "91fbf5814302f48263f0618798e5ec8cd3d24fdecdd19b62f49a551bb5812925" }, "candidate": { "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir": "006bd324cfcbd001184fe16a003c704724c62683f52954ec701f3f43534210b6", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libSPIRV.16.3.0.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libSPIRV.16.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libSPIRV.dylib": "3bb5958e5e53e17c57ec342b4aef3783d46eb172be8611ee79941b28044e2d63", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang-default-resource-limits.16.3.0.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang-default-resource-limits.16.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang-default-resource-limits.dylib": "1337e79b9de0888ffaec2b955f8a9248e560bb26b8897fe99b94e080d346e991", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang.16.3.0.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang.16.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libglslang.dylib": "2a11cbfe17b841c8d2dc55880d1c20141bd32549ffc1b328a73413b921894bde", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-ast.dylib": "3593c30c2c40059772688aaa0512f51e8e7b9dd84a659dedabeeef4f5d447190", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-backend-metal.so": "292c088e42be06f6324bf733a43bfe00a47698d6789ef45461b6c2c6b3dfeb63", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-backend-simd.so": "2ef4f4d7cbbb50e41adbe2ccae42e8ecebf65f2c32098bf717f1a8809b1af477", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-benchmark-metal-timing.dylib": "989657bf3419937efd8de38e17681574ee73f28455cc2548184818f589cf7ae8", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-core.dylib": "55f31daf2e618e78bfa9e4686c47e6d7cb4a8f3727e54c97952d719e69f663b4", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-coro.dylib": "9b6b2ed9528c537b3236d3f2e1a9e0dcbe68e501a0ca5da7b9c47b19a88f65b3", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-dsl.dylib": "192fa9ea1ee2c49caf241ca79b93c776490c553ecc215046e8e245ec2158b7a5", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-osl.dylib": "d6f0121521dd0f25dc5fc1c37afdca6dff8f0280a1da85450da36449a0261fb1", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-runtime.dylib": "97165ca9efb0a1ee57eee076fd145bc0a95f373f578cd595dd1c0613cabcb4ac", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-tile-bridge-tirx.dylib": "de6a123adaf2a678e1e74bfd29f0c87789b71df94229c2728b00434b3704fd99", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-tile-bridge-xir.dylib": "5e8928b27aec89cd0f6c8dd2c4c64103018762cd69ed66b87597b2a7c8bed04e", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-tile.dylib": "634878d1d9d5d27ca5a3b585916f2f602718cb2d1b327abde37ea6c9a29283e2", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-validation-layer.so": "1b791d7f116d07b095ff13925bf7e475c1eebb7ba02424f3411aa9b02f37f7eb", "/private/tmp/luisa-xir-indexable.21Tuda/candidate/libluisa-xir.dylib": "91fbf5814302f48263f0618798e5ec8cd3d24fdecdd19b62f49a551bb5812925" } }, "artifacts_unchanged": true }, "results": [ { "operation": "rmsnorm", "dimensions": [ 17, 7 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "17,7", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/00-rmsnorm-17x7-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 7 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 7 ], [ 1, 7 ], [ 1, 7 ] ], "output_shape": [ 17, 7 ], "fixture_ms": 0.313709, "runtime_init_ms": 884.211375, "compile_ms": 49.202958, "allocation_upload_ms": 0.009584, "cold_call_ms": 0.023125, "warmup_ms": 100.001542, "repetitions": 60000, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=7; exact search 6 candidates, uncalibrated cost 1754.000 (arithmetic 240.000, memory 1386.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 119, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77402201373e-07 }, "throughput_us": [ 0.39513265, 0.389261116667, 0.388590283333, 0.39679235, 0.399134716667, 0.390661816667, 0.3926632 ], "latency_us": [ 11.542, 0.583, 0.5, 0.458, 0.458, 0.458, 0.416 ], "throughput_us_p50": 0.3926632, "throughput_us_p90": 0.3977292966668, "latency_us_p50": 0.458, "latency_us_p90": 4.966600000000004 }, "oracle": { "elements": 119, "max_abs_error": 1.7740220137341112e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "07bc71a8cf9163f1ac5de64dda27ee4d39011d011235e870cdee6b18360672f6", "6e14f711133e45f6bf4c4767f18156da393b086c830f8dc93155f904e33cdcb5", "d32b7b498cd711103df6f469538e2f0a239b854da52d096911b81073f1b46405" ], "output_sha256": "1e98af37d7437fd8326a8aee44ec714b80c7f2afa6fd63becc77bd05cd641919", "process_and_validation_wall_seconds": 5.666455917060375 }, { "operation": "rmsnorm", "dimensions": [ 17, 7 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "17,7", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/00-rmsnorm-17x7-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 7 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 7 ], [ 1, 7 ], [ 1, 7 ] ], "output_shape": [ 17, 7 ], "fixture_ms": 0.419542, "runtime_init_ms": 911.182209, "compile_ms": 140.475042, "allocation_upload_ms": 0.00725, "cold_call_ms": 0.028375, "warmup_ms": 100.000167, "repetitions": 71942, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=7; exact search 6 candidates, uncalibrated cost 3161.000 (arithmetic 303.000, memory 2730.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 119, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77402201373e-07 }, "throughput_us": [ 0.346935726002, 0.331965333185, 0.33121935726, 0.333284103861, 0.332854354897, 0.329505017931, 0.331477092658 ], "latency_us": [ 15.792, 0.5, 0.459, 0.375, 0.375, 0.459, 0.375 ], "throughput_us_p50": 0.331965333185, "throughput_us_p90": 0.3387447527174, "latency_us_p50": 0.459, "latency_us_p90": 6.616800000000006 }, "oracle": { "elements": 119, "max_abs_error": 1.7740220137341112e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "07bc71a8cf9163f1ac5de64dda27ee4d39011d011235e870cdee6b18360672f6", "6e14f711133e45f6bf4c4767f18156da393b086c830f8dc93155f904e33cdcb5", "d32b7b498cd711103df6f469538e2f0a239b854da52d096911b81073f1b46405" ], "output_sha256": "1e98af37d7437fd8326a8aee44ec714b80c7f2afa6fd63becc77bd05cd641919", "process_and_validation_wall_seconds": 3.417393375071697 }, { "operation": "rmsnorm", "dimensions": [ 17, 7 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "17,7", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/00-rmsnorm-17x7-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 7 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 7 ], [ 1, 7 ], [ 1, 7 ] ], "output_shape": [ 17, 7 ], "fixture_ms": 0.072916, "runtime_init_ms": 11.491334, "compile_ms": 124.679125, "allocation_upload_ms": 0.006292, "cold_call_ms": 0.018416, "warmup_ms": 100.000083, "repetitions": 71942, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=7; exact search 6 candidates, uncalibrated cost 3161.000 (arithmetic 303.000, memory 2730.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 119, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77402201373e-07 }, "throughput_us": [ 0.346893441939, 0.342035361819, 0.331828639738, 0.337916891385, 0.332551458119, 0.352379917155, 0.331367629479 ], "latency_us": [ 17.5, 0.458, 0.459, 0.417, 0.417, 0.375, 0.375 ], "throughput_us_p50": 0.337916891385, "throughput_us_p90": 0.3490880320254, "latency_us_p50": 0.417, "latency_us_p90": 7.275400000000006 }, "oracle": { "elements": 119, "max_abs_error": 1.7740220137341112e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "07bc71a8cf9163f1ac5de64dda27ee4d39011d011235e870cdee6b18360672f6", "6e14f711133e45f6bf4c4767f18156da393b086c830f8dc93155f904e33cdcb5", "d32b7b498cd711103df6f469538e2f0a239b854da52d096911b81073f1b46405" ], "output_sha256": "1e98af37d7437fd8326a8aee44ec714b80c7f2afa6fd63becc77bd05cd641919", "process_and_validation_wall_seconds": 0.5014677500585094 }, { "operation": "rmsnorm", "dimensions": [ 17, 7 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "17,7", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/00-rmsnorm-17x7-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 7 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 7 ], [ 1, 7 ], [ 1, 7 ] ], "output_shape": [ 17, 7 ], "fixture_ms": 0.073042, "runtime_init_ms": 12.264959, "compile_ms": 35.154542, "allocation_upload_ms": 0.004666, "cold_call_ms": 0.014375, "warmup_ms": 100.001542, "repetitions": 60000, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=7; exact search 6 candidates, uncalibrated cost 1754.000 (arithmetic 240.000, memory 1386.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 119, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77402201373e-07 }, "throughput_us": [ 0.395752783333, 0.39231945, 0.390378466667, 0.392884033333, 0.391315283333, 0.391164583333, 0.390422916667 ], "latency_us": [ 10.708, 0.583, 0.5, 0.5, 0.458, 0.459, 0.417 ], "throughput_us_p50": 0.391315283333, "throughput_us_p90": 0.394031533333, "latency_us_p50": 0.5, "latency_us_p90": 4.633000000000004 }, "oracle": { "elements": 119, "max_abs_error": 1.7740220137341112e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "07bc71a8cf9163f1ac5de64dda27ee4d39011d011235e870cdee6b18360672f6", "6e14f711133e45f6bf4c4767f18156da393b086c830f8dc93155f904e33cdcb5", "d32b7b498cd711103df6f469538e2f0a239b854da52d096911b81073f1b46405" ], "output_sha256": "1e98af37d7437fd8326a8aee44ec714b80c7f2afa6fd63becc77bd05cd641919", "process_and_validation_wall_seconds": 0.37274162506219 }, { "operation": "rmsnorm", "dimensions": [ 17, 127 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "17,127", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/01-rmsnorm-17x127-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 127 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 127 ], [ 1, 127 ], [ 1, 127 ] ], "output_shape": [ 17, 127 ], "fixture_ms": 0.085333, "runtime_init_ms": 11.375583, "compile_ms": 887.210042, "allocation_upload_ms": 0.014625, "cold_call_ms": 0.077958, "warmup_ms": 100.028167, "repetitions": 1045, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=127; exact search 6 candidates, uncalibrated cost 75194.000 (arithmetic 49920.000, memory 25146.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 2159, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 3.65690357107e-07 }, "throughput_us": [ 29.0464909091, 28.7425043062, 29.4960526316, 29.13975311, 29.2050239234, 28.944337799, 29.1453349282 ], "latency_us": [ 30.958, 28.625, 28.75, 28.666, 28.667, 28.667, 28.541 ], "throughput_us_p50": 29.13975311, "throughput_us_p90": 29.32143540668, "latency_us_p50": 28.667, "latency_us_p90": 29.6332 }, "oracle": { "elements": 2159, "max_abs_error": 3.656903571069847e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a9838844e1513fa997383f98f9f08213ddc852f89c78b866dbae0151362112b2", "5806b289d94b7186437c69569f9265c500779b439e60473a5bb3c9a4a61468e9", "473fd2074594aab5418727e7f3d8689b2e47124bb72709c0c1f5117000106e73" ], "output_sha256": "f49a82a712a0edd3981f412f24a59d322f50879d759b94d8c55594ef66f9c40d", "process_and_validation_wall_seconds": 1.3078479580581188 }, { "operation": "rmsnorm", "dimensions": [ 17, 127 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "17,127", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/01-rmsnorm-17x127-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 127 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 127 ], [ 1, 127 ], [ 1, 127 ] ], "output_shape": [ 17, 127 ], "fixture_ms": 0.09525, "runtime_init_ms": 12.349333, "compile_ms": 1406.907708, "allocation_upload_ms": 0.011333, "cold_call_ms": 0.062667, "warmup_ms": 100.010708, "repetitions": 3692, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=127; exact search 6 candidates, uncalibrated cost 55001.000 (arithmetic 5343.000, memory 49530.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 2159, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 3.65690357107e-07 }, "throughput_us": [ 8.0923507584, 8.23713434453, 8.0552659805, 8.19626895991, 8.12469528711, 8.1331031961, 8.08623374865 ], "latency_us": [ 9.916, 8.25, 8.167, 8, 8.125, 8.083, 8.208 ], "throughput_us_p50": 8.12469528711, "throughput_us_p90": 8.212615113758, "latency_us_p50": 8.167, "latency_us_p90": 8.916400000000001 }, "oracle": { "elements": 2159, "max_abs_error": 3.656903571069847e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a9838844e1513fa997383f98f9f08213ddc852f89c78b866dbae0151362112b2", "5806b289d94b7186437c69569f9265c500779b439e60473a5bb3c9a4a61468e9", "473fd2074594aab5418727e7f3d8689b2e47124bb72709c0c1f5117000106e73" ], "output_sha256": "f49a82a712a0edd3981f412f24a59d322f50879d759b94d8c55594ef66f9c40d", "process_and_validation_wall_seconds": 1.8070465419441462 }, { "operation": "rmsnorm", "dimensions": [ 17, 127 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "17,127", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/01-rmsnorm-17x127-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 127 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 127 ], [ 1, 127 ], [ 1, 127 ] ], "output_shape": [ 17, 127 ], "fixture_ms": 0.102, "runtime_init_ms": 11.415792, "compile_ms": 1351.295083, "allocation_upload_ms": 0.009916, "cold_call_ms": 0.070625, "warmup_ms": 100.050834, "repetitions": 3636, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=127; exact search 6 candidates, uncalibrated cost 55001.000 (arithmetic 5343.000, memory 49530.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 2159, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 3.65690357107e-07 }, "throughput_us": [ 8.15589493949, 8.07902447745, 8.11619911991, 8.14930555556, 8.20238586359, 8.22520159516, 8.30824862486 ], "latency_us": [ 11.958, 8.166, 8.041, 7.958, 7.917, 8, 8 ], "throughput_us_p50": 8.15589493949, "throughput_us_p90": 8.25842040704, "latency_us_p50": 8.0, "latency_us_p90": 9.682800000000002 }, "oracle": { "elements": 2159, "max_abs_error": 3.656903571069847e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a9838844e1513fa997383f98f9f08213ddc852f89c78b866dbae0151362112b2", "5806b289d94b7186437c69569f9265c500779b439e60473a5bb3c9a4a61468e9", "473fd2074594aab5418727e7f3d8689b2e47124bb72709c0c1f5117000106e73" ], "output_sha256": "f49a82a712a0edd3981f412f24a59d322f50879d759b94d8c55594ef66f9c40d", "process_and_validation_wall_seconds": 1.7175036249682307 }, { "operation": "rmsnorm", "dimensions": [ 17, 127 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "17,127", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/01-rmsnorm-17x127-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 17, 127 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 127 ], [ 1, 127 ], [ 1, 127 ] ], "output_shape": [ 17, 127 ], "fixture_ms": 0.08025, "runtime_init_ms": 10.6175, "compile_ms": 887.405958, "allocation_upload_ms": 0.010708, "cold_call_ms": 0.073875, "warmup_ms": 100.071625, "repetitions": 965, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=127; exact search 6 candidates, uncalibrated cost 75194.000 (arithmetic 49920.000, memory 25146.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 2159, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 3.65690357107e-07 }, "throughput_us": [ 30.7327284974, 29.6645937824, 29.8440414508, 29.5725388601, 29.4493523316, 29.9124352332, 30.1136010363 ], "latency_us": [ 34.125, 30.583, 118.875, 34.542, 30.875, 30.833, 31.375 ], "throughput_us_p50": 29.8440414508, "throughput_us_p90": 30.36125202074, "latency_us_p50": 31.375, "latency_us_p90": 68.27520000000004 }, "oracle": { "elements": 2159, "max_abs_error": 3.656903571069847e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a9838844e1513fa997383f98f9f08213ddc852f89c78b866dbae0151362112b2", "5806b289d94b7186437c69569f9265c500779b439e60473a5bb3c9a4a61468e9", "473fd2074594aab5418727e7f3d8689b2e47124bb72709c0c1f5117000106e73" ], "output_sha256": "f49a82a712a0edd3981f412f24a59d322f50879d759b94d8c55594ef66f9c40d", "process_and_validation_wall_seconds": 1.2577240420505404 }, { "operation": "rmsnorm", "dimensions": [ 64, 256 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/02-rmsnorm-64x256-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.184792, "runtime_init_ms": 10.998667, "compile_ms": 3610.755375, "allocation_upload_ms": 0.027958, "cold_call_ms": 0.324666, "warmup_ms": 101.179541, "repetitions": 130, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 300172.000 (arithmetic 266252.000, memory 33792.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 5.44328237373e-07 }, "throughput_us": [ 224.152561538, 222.827246154, 218.267630769, 221.490061538, 218.425638462, 217.193269231, 223.486538462 ], "latency_us": [ 236.625, 210.333, 235.084, 230.459, 226.792, 230.5, 233.083 ], "throughput_us_p50": 221.490061538, "throughput_us_p90": 223.7529476924, "latency_us_p50": 230.5, "latency_us_p90": 235.7004 }, "oracle": { "elements": 16384, "max_abs_error": 5.443282355965806e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "f5df81115d4639e9fe1580dd8626442b010ed52ae10dad22b6fb890a4a9bb272", "process_and_validation_wall_seconds": 4.0252622080734 }, { "operation": "rmsnorm", "dimensions": [ 64, 256 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/02-rmsnorm-64x256-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.216459, "runtime_init_ms": 11.959792, "compile_ms": 5676.649417, "allocation_upload_ms": 0.010708, "cold_call_ms": 0.198209, "warmup_ms": 100.298084, "repetitions": 373, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 81036.000 (arithmetic 14348.000, memory 66560.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 5.44328237373e-07 }, "throughput_us": [ 87.3642761394, 91.3679624665, 90.1346085791, 85.8660643432, 87.8818150134, 89.7484343164, 86.2368176944 ], "latency_us": [ 83.042, 80.209, 81.167, 87.667, 89.833, 78.25, 83.5 ], "throughput_us_p50": 87.8818150134, "throughput_us_p90": 90.62795013406, "latency_us_p50": 83.042, "latency_us_p90": 88.5334 }, "oracle": { "elements": 16384, "max_abs_error": 5.443282355965806e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "f5df81115d4639e9fe1580dd8626442b010ed52ae10dad22b6fb890a4a9bb272", "process_and_validation_wall_seconds": 6.073460250045173 }, { "operation": "rmsnorm", "dimensions": [ 64, 256 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/02-rmsnorm-64x256-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.18825, "runtime_init_ms": 11.37575, "compile_ms": 5690.426958, "allocation_upload_ms": 0.011042, "cold_call_ms": 0.203333, "warmup_ms": 100.657416, "repetitions": 298, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 81036.000 (arithmetic 14348.000, memory 66560.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 5.44328237373e-07 }, "throughput_us": [ 86.4032416107, 87.2943221477, 86.3042483221, 87.6226241611, 87.1259765101, 89.5064295302, 87.1458322148 ], "latency_us": [ 79.459, 91.708, 89.792, 78.458, 101.417, 93.708, 84.792 ], "throughput_us_p50": 87.1458322148, "throughput_us_p90": 88.37614630874, "latency_us_p50": 89.792, "latency_us_p90": 96.7916 }, "oracle": { "elements": 16384, "max_abs_error": 5.443282355965806e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "f5df81115d4639e9fe1580dd8626442b010ed52ae10dad22b6fb890a4a9bb272", "process_and_validation_wall_seconds": 6.052980749984272 }, { "operation": "rmsnorm", "dimensions": [ 64, 256 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/02-rmsnorm-64x256-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.189083, "runtime_init_ms": 11.661875, "compile_ms": 3678.637084, "allocation_upload_ms": 0.009916, "cold_call_ms": 0.313917, "warmup_ms": 100.582708, "repetitions": 127, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 300172.000 (arithmetic 266252.000, memory 33792.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 5.44328237373e-07 }, "throughput_us": [ 239.070866142, 240.328086614, 228.695204724, 230.153543307, 231.117456693, 230.045929134, 230.470472441 ], "latency_us": [ 281.417, 225.541, 228.125, 232.042, 228.792, 227.083, 228.041 ], "throughput_us_p50": 230.470472441, "throughput_us_p90": 239.5737543308, "latency_us_p50": 228.125, "latency_us_p90": 251.792 }, "oracle": { "elements": 16384, "max_abs_error": 5.443282355965806e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "f5df81115d4639e9fe1580dd8626442b010ed52ae10dad22b6fb890a4a9bb272", "process_and_validation_wall_seconds": 4.046370875090361 }, { "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "1024,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/03-rmsnorm-1024x256-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 1024, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 1024, 256 ], "fixture_ms": 2.255292, "runtime_init_ms": 10.945333, "compile_ms": 3615.872541, "allocation_upload_ms": 0.282708, "cold_call_ms": 1.112625, "warmup_ms": 102.227708, "repetitions": 37, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 128 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 1200304.000 (arithmetic 1065008.000, memory 135168.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 262144, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 8.18452875473e-07 }, "throughput_us": [ 842.809675676, 914.588945946, 1081.12048649, 1223.51013514, 951.563054054, 849.862594595, 850.132864865 ], "latency_us": [ 1081.542, 811.042, 804.459, 808.334, 804.542, 812, 810.541 ], "throughput_us_p50": 914.588945946, "throughput_us_p90": 1138.0763459500001, "latency_us_p50": 810.541, "latency_us_p90": 919.8168000000001 }, "oracle": { "elements": 262144, "max_abs_error": 8.184528770271982e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "6e25aafd454395b8f93026f28cdec71afa650492bbc4cb387824dabfedf024c3", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "141fa71af3aa30d7c7f13d106a800cb3421dcbf0dc38b3861b27c7b59a790ac0", "process_and_validation_wall_seconds": 4.084601207985543 }, { "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "1024,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/03-rmsnorm-1024x256-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 1024, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 1024, 256 ], "fixture_ms": 2.29675, "runtime_init_ms": 11.213916, "compile_ms": 5593.685375, "allocation_upload_ms": 0.118042, "cold_call_ms": 0.370875, "warmup_ms": 102.33525, "repetitions": 115, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 128 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 323760.000 (arithmetic 57392.000, memory 266240.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 262144, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 8.18452875473e-07 }, "throughput_us": [ 284.790582609, 268.378982609, 268.78406087, 267.807243478, 268.235869565, 268.367026087, 267.436591304 ], "latency_us": [ 302.709, 262.458, 280.875, 251.083, 257.291, 252.458, 254.833 ], "throughput_us_p50": 268.367026087, "throughput_us_p90": 275.1866695656, "latency_us_p50": 257.291, "latency_us_p90": 289.6086 }, "oracle": { "elements": 262144, "max_abs_error": 8.184528770271982e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "6e25aafd454395b8f93026f28cdec71afa650492bbc4cb387824dabfedf024c3", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "141fa71af3aa30d7c7f13d106a800cb3421dcbf0dc38b3861b27c7b59a790ac0", "process_and_validation_wall_seconds": 5.985231625032611 }, { "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "1024,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/03-rmsnorm-1024x256-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 1024, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 1024, 256 ], "fixture_ms": 2.23275, "runtime_init_ms": 11.013917, "compile_ms": 5770.148959, "allocation_upload_ms": 0.15575, "cold_call_ms": 0.419209, "warmup_ms": 101.797125, "repetitions": 120, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 128 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 323760.000 (arithmetic 57392.000, memory 266240.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 262144, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 8.18452875473e-07 }, "throughput_us": [ 269.217358333, 270.365275, 269.672566667, 268.829166667, 270.652083333, 268.802083333, 269.519441667 ], "latency_us": [ 263.375, 254.416, 262.334, 254.792, 258.875, 254.542, 256.875 ], "throughput_us_p50": 269.519441667, "throughput_us_p90": 270.4799983332, "latency_us_p50": 256.875, "latency_us_p90": 262.7504 }, "oracle": { "elements": 262144, "max_abs_error": 8.184528770271982e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "6e25aafd454395b8f93026f28cdec71afa650492bbc4cb387824dabfedf024c3", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "141fa71af3aa30d7c7f13d106a800cb3421dcbf0dc38b3861b27c7b59a790ac0", "process_and_validation_wall_seconds": 6.2144060409627855 }, { "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "1024,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/03-rmsnorm-1024x256-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 1024, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 1024, 256 ], "fixture_ms": 2.3645, "runtime_init_ms": 11.693083, "compile_ms": 3620.827333, "allocation_upload_ms": 0.252584, "cold_call_ms": 0.894666, "warmup_ms": 101.019, "repetitions": 37, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 128 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=256; exact search 6 candidates, uncalibrated cost 1200304.000 (arithmetic 1065008.000, memory 135168.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 262144, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 8.18452875473e-07 }, "throughput_us": [ 834.292783784, 833.283783784, 830.864864865, 853.397513514, 854.676810811, 834.440324324, 832.930162162 ], "latency_us": [ 796.708, 807.875, 797.208, 1369.709, 1363.25, 1292.417, 813.167 ], "throughput_us_p50": 834.292783784, "throughput_us_p90": 853.9092324328, "latency_us_p50": 813.167, "latency_us_p90": 1365.8336 }, "oracle": { "elements": 262144, "max_abs_error": 8.184528770271982e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "6e25aafd454395b8f93026f28cdec71afa650492bbc4cb387824dabfedf024c3", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "141fa71af3aa30d7c7f13d106a800cb3421dcbf0dc38b3861b27c7b59a790ac0", "process_and_validation_wall_seconds": 4.065495749935508 }, { "operation": "rmsnorm", "dimensions": [ 64, 513 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "64,513", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/04-rmsnorm-64x513-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 513 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 513 ], [ 1, 513 ], [ 1, 513 ] ], "output_shape": [ 64, 513 ], "fixture_ms": 0.344083, "runtime_init_ms": 11.945625, "compile_ms": 14786.518458, "allocation_upload_ms": 0.028, "cold_call_ms": 0.935083, "warmup_ms": 104.705625, "repetitions": 38, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=513; exact search 6 candidates, uncalibrated cost 1128740.000 (arithmetic 1060896.000, memory 67716.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 32832, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.05428874586e-06 }, "throughput_us": [ 787.182026316, 771.060315789, 772.747789474, 765.334421053, 770.652421053, 765.549342105, 772.901315789 ], "latency_us": [ 762.875, 747.166, 738.458, 751.417, 751.041, 770.583, 771.625 ], "throughput_us_p50": 771.060315789, "throughput_us_p90": 778.6135999998, "latency_us_p50": 751.417, "latency_us_p90": 770.9997999999999 }, "oracle": { "elements": 32832, "max_abs_error": 1.0542887467490658e-06, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "08fa141135e2b280a82a4e07c578e6b87c374a1867ad4a3fb1487bd5ceffe46b", "89c042ec7683b7994a5c571a487b552381cc1017ebdd832156ad08a797180a17", "48f816376b7ce301ab9ab297af5cd74fb3e66ee84066e4fdb31aab9828da1cc4" ], "output_sha256": "7a2c80a911d8147b540b336763e9d1fca1ffe2c787d4b8604cabd9ff5a6c2434", "process_and_validation_wall_seconds": 15.213927458971739 }, { "operation": "rmsnorm", "dimensions": [ 64, 513 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "64,513", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/04-rmsnorm-64x513-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 513 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 513 ], [ 1, 513 ], [ 1, 513 ] ], "output_shape": [ 64, 513 ], "fixture_ms": 0.334416, "runtime_init_ms": 12.032166, "compile_ms": 24098.003959, "allocation_upload_ms": 0.023792, "cold_call_ms": 0.348208, "warmup_ms": 100.383209, "repetitions": 206, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=513; exact search 6 candidates, uncalibrated cost 162248.000 (arithmetic 28740.000, memory 133380.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 32832, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.05428874586e-06 }, "throughput_us": [ 147.23907767, 149.858820388, 146.986446602, 144.893004854, 145.801174757, 144.571601942, 143.782160194 ], "latency_us": [ 138.458, 138.459, 145.959, 144.167, 156.708, 139.792, 153.167 ], "throughput_us_p50": 145.801174757, "throughput_us_p90": 148.2869747572, "latency_us_p50": 144.167, "latency_us_p90": 154.5834 }, "oracle": { "elements": 32832, "max_abs_error": 1.0542887467490658e-06, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "08fa141135e2b280a82a4e07c578e6b87c374a1867ad4a3fb1487bd5ceffe46b", "89c042ec7683b7994a5c571a487b552381cc1017ebdd832156ad08a797180a17", "48f816376b7ce301ab9ab297af5cd74fb3e66ee84066e4fdb31aab9828da1cc4" ], "output_sha256": "7a2c80a911d8147b540b336763e9d1fca1ffe2c787d4b8604cabd9ff5a6c2434", "process_and_validation_wall_seconds": 24.48146270797588 }, { "operation": "rmsnorm", "dimensions": [ 64, 513 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "rmsnorm", "64,513", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/04-rmsnorm-64x513-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 513 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 513 ], [ 1, 513 ], [ 1, 513 ] ], "output_shape": [ 64, 513 ], "fixture_ms": 0.300792, "runtime_init_ms": 11.053167, "compile_ms": 24127.410417, "allocation_upload_ms": 0.030875, "cold_call_ms": 0.343292, "warmup_ms": 100.751917, "repetitions": 220, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=6, direct CFG=true, contiguous reads=0, broadcasts=513; exact search 6 candidates, uncalibrated cost 162248.000 (arithmetic 28740.000, memory 133380.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 32832, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.05428874586e-06 }, "throughput_us": [ 142.636931818, 141.666854545, 143.843559091, 143.365340909, 143.298672727, 144.667804545, 140.867231818 ], "latency_us": [ 142.917, 142.125, 140.75, 149.292, 150.292, 142.5, 165.791 ], "throughput_us_p50": 143.298672727, "throughput_us_p90": 144.1732572726, "latency_us_p50": 142.917, "latency_us_p90": 156.4916 }, "oracle": { "elements": 32832, "max_abs_error": 1.0542887467490658e-06, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "08fa141135e2b280a82a4e07c578e6b87c374a1867ad4a3fb1487bd5ceffe46b", "89c042ec7683b7994a5c571a487b552381cc1017ebdd832156ad08a797180a17", "48f816376b7ce301ab9ab297af5cd74fb3e66ee84066e4fdb31aab9828da1cc4" ], "output_sha256": "7a2c80a911d8147b540b336763e9d1fca1ffe2c787d4b8604cabd9ff5a6c2434", "process_and_validation_wall_seconds": 24.512369959033094 }, { "operation": "rmsnorm", "dimensions": [ 64, 513 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "rmsnorm", "64,513", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/04-rmsnorm-64x513-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "rmsnorm", "dimensions": [ 64, 513 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 513 ], [ 1, 513 ], [ 1, 513 ] ], "output_shape": [ 64, 513 ], "fixture_ms": 0.3135, "runtime_init_ms": 11.569792, "compile_ms": 14916.039334, "allocation_upload_ms": 0.021458, "cold_call_ms": 0.919417, "warmup_ms": 104.044667, "repetitions": 39, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=4, direct CFG=true, contiguous reads=0, broadcasts=513; exact search 6 candidates, uncalibrated cost 1128740.000 (arithmetic 1060896.000, memory 67716.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 32832, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.05428874586e-06 }, "throughput_us": [ 763.558769231, 765.516025641, 760.575846154, 776.511769231, 769.509615385, 759.637820513, 805.838666667 ], "latency_us": [ 772.209, 792.5, 790.542, 891.25, 764.458, 777.291, 779.75 ], "throughput_us_p50": 765.516025641, "throughput_us_p90": 788.2425282054, "latency_us_p50": 779.75, "latency_us_p90": 832.0 }, "oracle": { "elements": 32832, "max_abs_error": 1.0542887467490658e-06, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "08fa141135e2b280a82a4e07c578e6b87c374a1867ad4a3fb1487bd5ceffe46b", "89c042ec7683b7994a5c571a487b552381cc1017ebdd832156ad08a797180a17", "48f816376b7ce301ab9ab297af5cd74fb3e66ee84066e4fdb31aab9828da1cc4" ], "output_sha256": "7a2c80a911d8147b540b336763e9d1fca1ffe2c787d4b8604cabd9ff5a6c2434", "process_and_validation_wall_seconds": 15.314606958068907 }, { "operation": "layernorm", "dimensions": [ 64, 256 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "layernorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/05-layernorm-64x256-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "layernorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.201542, "runtime_init_ms": 10.983958, "compile_ms": 5158.898625, "allocation_upload_ms": 0.01425, "cold_call_ms": 0.461667, "warmup_ms": 102.249916, "repetitions": 75, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=7, direct CFG=true, contiguous reads=0, broadcasts=512; exact search 6 candidates, uncalibrated cost 566416.000 (arithmetic 531472.000, memory 34816.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 7.72739510158e-07 }, "throughput_us": [ 387.498333333, 389.021106667, 387.602786667, 385.551666667, 387.261666667, 387.11944, 386.686666667 ], "latency_us": [ 405.25, 394.625, 391.125, 382.417, 365.584, 395.542, 384.709 ], "throughput_us_p50": 387.261666667, "throughput_us_p90": 388.170114667, "latency_us_p50": 391.125, "latency_us_p90": 399.42519999999996 }, "oracle": { "elements": 16384, "max_abs_error": 7.727395097134604e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "e840b31e253898fc7ae7b37cca579f1b4e8f1ad8bb5bc5fdebd64bb25d183269", "process_and_validation_wall_seconds": 5.538193208980374 }, { "operation": "layernorm", "dimensions": [ 64, 256 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "layernorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/05-layernorm-64x256-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "layernorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.201125, "runtime_init_ms": 11.492, "compile_ms": 8568.195709, "allocation_upload_ms": 0.021708, "cold_call_ms": 0.24475, "warmup_ms": 100.221167, "repetitions": 278, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=11, direct CFG=true, contiguous reads=0, broadcasts=512; exact search 6 candidates, uncalibrated cost 128144.000 (arithmetic 27664.000, memory 100352.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 7.72739510158e-07 }, "throughput_us": [ 109.363910072, 106.619003597, 106.51828777, 111.85926259, 109.234859712, 109.276978417, 113.214179856 ], "latency_us": [ 134.333, 115.417, 111.042, 125.791, 112.75, 108.042, 109.166 ], "throughput_us_p50": 109.276978417, "throughput_us_p90": 112.4012294964, "latency_us_p50": 112.75, "latency_us_p90": 129.2078 }, "oracle": { "elements": 16384, "max_abs_error": 7.727395097134604e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "e840b31e253898fc7ae7b37cca579f1b4e8f1ad8bb5bc5fdebd64bb25d183269", "process_and_validation_wall_seconds": 8.943834083038382 }, { "operation": "layernorm", "dimensions": [ 64, 256 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "layernorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/05-layernorm-64x256-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "layernorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.204166, "runtime_init_ms": 11.195708, "compile_ms": 8498.531375, "allocation_upload_ms": 0.015709, "cold_call_ms": 0.212417, "warmup_ms": 100.87225, "repetitions": 266, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=11, direct CFG=true, contiguous reads=0, broadcasts=512; exact search 6 candidates, uncalibrated cost 128144.000 (arithmetic 27664.000, memory 100352.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 7.72739510158e-07 }, "throughput_us": [ 125.400691729, 127.04887218, 126.816571429, 122.640507519, 118.549342105, 120.904921053, 127.056390977 ], "latency_us": [ 136.125, 133.958, 105.417, 135.375, 144.417, 119.916, 112.375 ], "throughput_us_p50": 125.400691729, "throughput_us_p90": 127.05187969880001, "latency_us_p50": 133.958, "latency_us_p90": 139.4418 }, "oracle": { "elements": 16384, "max_abs_error": 7.727395097134604e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "e840b31e253898fc7ae7b37cca579f1b4e8f1ad8bb5bc5fdebd64bb25d183269", "process_and_validation_wall_seconds": 8.901846667053178 }, { "operation": "layernorm", "dimensions": [ 64, 256 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "layernorm", "64,256", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/05-layernorm-64x256-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "layernorm", "dimensions": [ 64, 256 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 64, 256 ], [ 1, 256 ], [ 1, 256 ] ], "output_shape": [ 64, 256 ], "fixture_ms": 0.202125, "runtime_init_ms": 10.65025, "compile_ms": 5282.311791, "allocation_upload_ms": 0.009375, "cold_call_ms": 0.504084, "warmup_ms": 100.21325, "repetitions": 80, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=7, direct CFG=true, contiguous reads=0, broadcasts=512; exact search 6 candidates, uncalibrated cost 566416.000 (arithmetic 531472.000, memory 34816.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 16384, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 7.72739510158e-07 }, "throughput_us": [ 392.8041625, 393.8510375, 392.372925, 391.928125, 391.1213625, 394.5427, 393.910425 ], "latency_us": [ 422.625, 389.75, 408.625, 382.583, 397.833, 398.459, 397.083 ], "throughput_us_p50": 392.8041625, "throughput_us_p90": 394.163335, "latency_us_p50": 397.833, "latency_us_p90": 414.225 }, "oracle": { "elements": 16384, "max_abs_error": 7.727395097134604e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "d7dc9eae83723e1637127522ec2528b82293849489494c146627ed555066d5e4", "fc3c674c6fcf5b27b7d69080bd8674cc1d0f8a6a1f3e42973934c57a39f5a898", "ebaf91fb4f1d9da279189c8ffa230f28082282dbb343cb15049c04656c731717" ], "output_sha256": "e840b31e253898fc7ae7b37cca579f1b4e8f1ad8bb5bc5fdebd64bb25d183269", "process_and_validation_wall_seconds": 5.672157124965452 }, { "operation": "masked_softmax", "dimensions": [ 17, 65 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "masked_softmax", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/06-masked_softmax-17x65-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "masked_softmax", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.101667, "runtime_init_ms": 10.85725, "compile_ms": 217.918584, "allocation_upload_ms": 0.009709, "cold_call_ms": 0.032959, "warmup_ms": 100.025917, "repetitions": 2562, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=7, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 39716.000 (arithmetic 27108.000, memory 12480.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 2.98437049517e-08 }, "throughput_us": [ 11.5610362998, 11.6793844653, 11.7486502732, 11.5484160812, 11.9549180328, 11.4297751756, 11.5940671351 ], "latency_us": [ 13.916, 11.333, 11.666, 11.625, 11.666, 11.584, 11.667 ], "throughput_us_p50": 11.5940671351, "throughput_us_p90": 11.83115737704, "latency_us_p50": 11.666, "latency_us_p90": 12.566600000000001 }, "oracle": { "elements": 1105, "max_abs_error": 2.984370495173749e-08, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "4f9470a40b2a46632cd472551c200badd4190265f2ad700135a8a0ce60c51939", "process_and_validation_wall_seconds": 0.60306645790115 }, { "operation": "masked_softmax", "dimensions": [ 17, 65 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "masked_softmax", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/06-masked_softmax-17x65-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "masked_softmax", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.09525, "runtime_init_ms": 11.267375, "compile_ms": 433.197084, "allocation_upload_ms": 0.008041, "cold_call_ms": 0.043417, "warmup_ms": 100.016375, "repetitions": 7500, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=11, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 43226.000 (arithmetic 5658.000, memory 37440.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 2.98437049517e-08 }, "throughput_us": [ 3.97006666667, 3.93545546667, 3.9355112, 3.8941056, 3.94108333333, 3.77356106667, 3.88282773333 ], "latency_us": [ 7.792, 4.166, 4.041, 3.959, 3.958, 4, 4.042 ], "throughput_us_p50": 3.93545546667, "throughput_us_p90": 3.952676666666, "latency_us_p50": 4.041, "latency_us_p90": 5.616400000000001 }, "oracle": { "elements": 1105, "max_abs_error": 2.984370495173749e-08, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "4f9470a40b2a46632cd472551c200badd4190265f2ad700135a8a0ce60c51939", "process_and_validation_wall_seconds": 0.8458958750125021 }, { "operation": "masked_softmax", "dimensions": [ 17, 65 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "masked_softmax", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/06-masked_softmax-17x65-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "masked_softmax", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.106583, "runtime_init_ms": 10.807833, "compile_ms": 439.765459, "allocation_upload_ms": 0.011125, "cold_call_ms": 0.041625, "warmup_ms": 100.014375, "repetitions": 7500, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=11, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 43226.000 (arithmetic 5658.000, memory 37440.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 2.98437049517e-08 }, "throughput_us": [ 3.7670944, 3.84260546667, 3.81451106667, 3.82693333333, 3.86256666667, 3.7854112, 3.7952944 ], "latency_us": [ 7.833, 4.166, 3.875, 3.916, 4.042, 3.959, 3.916 ], "throughput_us_p50": 3.81451106667, "throughput_us_p90": 3.85058994667, "latency_us_p50": 3.959, "latency_us_p90": 5.632800000000001 }, "oracle": { "elements": 1105, "max_abs_error": 2.984370495173749e-08, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "4f9470a40b2a46632cd472551c200badd4190265f2ad700135a8a0ce60c51939", "process_and_validation_wall_seconds": 0.8507387090940028 }, { "operation": "masked_softmax", "dimensions": [ 17, 65 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "masked_softmax", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/06-masked_softmax-17x65-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "masked_softmax", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.090625, "runtime_init_ms": 10.866459, "compile_ms": 220.065125, "allocation_upload_ms": 0.005542, "cold_call_ms": 0.033583, "warmup_ms": 100.035625, "repetitions": 2589, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=7, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 39716.000 (arithmetic 27108.000, memory 12480.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 2.98437049517e-08 }, "throughput_us": [ 11.8649412901, 11.9252927771, 11.8674036307, 11.7826056392, 11.4797056779, 11.668629973, 11.6018895326 ], "latency_us": [ 14.875, 12.125, 12.042, 12.042, 11.958, 11.959, 12 ], "throughput_us_p50": 11.7826056392, "throughput_us_p90": 11.89055928926, "latency_us_p50": 12.042, "latency_us_p90": 13.225000000000001 }, "oracle": { "elements": 1105, "max_abs_error": 2.984370495173749e-08, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "4f9470a40b2a46632cd472551c200badd4190265f2ad700135a8a0ce60c51939", "process_and_validation_wall_seconds": 0.6134027080843225 }, { "operation": "swiglu", "dimensions": [ 17, 65 ], "round": 0, "variant": "baseline", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "swiglu", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/07-swiglu-17x65-0-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "swiglu", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.069833, "runtime_init_ms": 11.000541, "compile_ms": 1435.356084, "allocation_upload_ms": 0.010375, "cold_call_ms": 0.048541, "warmup_ms": 100.010167, "repetitions": 6923, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=1, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 19823.000 (arithmetic 975.000, memory 18720.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77788147537e-07 }, "throughput_us": [ 4.35838870432, 4.34360104001, 4.31503076701, 4.39649833887, 4.39149689441, 4.35681785353, 4.31195536617 ], "latency_us": [ 8.417, 4.541, 4.5, 4.375, 4.458, 4.375, 4.5 ], "throughput_us_p50": 4.35681785353, "throughput_us_p90": 4.393497472194, "latency_us_p50": 4.5, "latency_us_p90": 6.091400000000002 }, "oracle": { "elements": 1105, "max_abs_error": 1.7778814753732775e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "628c3cc4c7a5005337f7f76cbd7bbc6d08034cad68229005c6e222a956f443d8", "process_and_validation_wall_seconds": 1.839129374944605 }, { "operation": "swiglu", "dimensions": [ 17, 65 ], "round": 0, "variant": "candidate", "order": [ "baseline", "candidate" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "swiglu", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/07-swiglu-17x65-0-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "swiglu", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.080542, "runtime_init_ms": 10.908541, "compile_ms": 1432.61525, "allocation_upload_ms": 0.009458, "cold_call_ms": 0.052208, "warmup_ms": 100.011834, "repetitions": 7058, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=1, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 19823.000 (arithmetic 975.000, memory 18720.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77788147537e-07 }, "throughput_us": [ 4.29169028053, 4.27505426466, 4.19474950411, 4.3404528195, 4.42480759422, 4.29889246245, 4.28024936243 ], "latency_us": [ 8.917, 4.583, 4.459, 4.334, 4.375, 4.375, 4.333 ], "throughput_us_p50": 4.29169028053, "throughput_us_p90": 4.374194729388, "latency_us_p50": 4.375, "latency_us_p90": 6.316600000000001 }, "oracle": { "elements": 1105, "max_abs_error": 1.7778814753732775e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "628c3cc4c7a5005337f7f76cbd7bbc6d08034cad68229005c6e222a956f443d8", "process_and_validation_wall_seconds": 1.8530774579849094 }, { "operation": "swiglu", "dimensions": [ 17, 65 ], "round": 1, "variant": "candidate", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/candidate/benchmark_tile_xir", "llm", "swiglu", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/07-swiglu-17x65-1-candidate.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "swiglu", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.0675, "runtime_init_ms": 11.188416, "compile_ms": 1432.031542, "allocation_upload_ms": 0.010209, "cold_call_ms": 0.050542, "warmup_ms": 100.010667, "repetitions": 6857, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=1, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 19823.000 (arithmetic 975.000, memory 18720.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77788147537e-07 }, "throughput_us": [ 4.29342277964, 4.28872568179, 4.22969841038, 4.32612046084, 4.33395916582, 4.26728161003, 4.33894808225 ], "latency_us": [ 9.291, 4.583, 4.334, 4.334, 4.375, 4.375, 4.417 ], "throughput_us_p50": 4.29342277964, "throughput_us_p90": 4.335954732392, "latency_us_p50": 4.375, "latency_us_p90": 6.466200000000002 }, "oracle": { "elements": 1105, "max_abs_error": 1.7778814753732775e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "628c3cc4c7a5005337f7f76cbd7bbc6d08034cad68229005c6e222a956f443d8", "process_and_validation_wall_seconds": 1.8201822499977425 }, { "operation": "swiglu", "dimensions": [ 17, 65 ], "round": 1, "variant": "baseline", "order": [ "candidate", "baseline" ], "valid": true, "command": [ "/private/tmp/luisa-xir-indexable.21Tuda/baseline/benchmark_tile_xir", "llm", "swiglu", "17,65", "1", "1", "7", "30", "100", "/tmp/luisa-xir-indexable.21Tuda/tensors-ab/07-swiglu-17x65-1-baseline.f32" ], "returncode": 0, "measurement": { "implementation": "tile_xir_simd", "backend": "cpu", "precision": "fp32", "fast_math": false, "relaxed_precision": false, "runtime": "luisa", "timing": "synchronized_host_wall", "batch_policy": "one_runtime_command_list_per_batch", "operation": "swiglu", "dimensions": [ 17, 65 ], "reduction_tree": false, "requested_input_views": false, "attention_qk": "not_applicable", "source_reduction_policy": "unordered_tree", "reduction_candidate_setting": "not_applicable", "requested_group_threads": 0, "attention_block": [ 1, 1 ], "input_shapes": [ [ 17, 65 ], [ 17, 65 ], [ 17, 65 ] ], "output_shape": [ 17, 65 ], "fixture_ms": 0.071625, "runtime_init_ms": 10.834334, "compile_ms": 1431.371166, "allocation_upload_ms": 0.009708, "cold_call_ms": 0.051208, "warmup_ms": 100.019792, "repetitions": 7058, "realization": "TileIR -> XIR SSA -> SIMD Schedule -> LLVM; W8, 32 workers/block, 8 CPU workers; ordered CPU pipeline; Schedule blocks=1, direct CFG=true, contiguous reads=0, broadcasts=0; exact search 6 candidates, uncalibrated cost 19823.000 (arithmetic 975.000, memory 18720.000, dispatch 128.000, imbalance 0.000), root order [0]; fast_math=false; ordered_reduction=false", "correctness": { "checks": 2, "elements_per_check": 1105, "guard_elements_per_check": 34, "atol": 5e-05, "rtol": 5e-05, "max_abs_error": 1.77788147537e-07 }, "throughput_us": [ 4.35731907056, 4.34647449702, 4.3948651176, 4.27041413998, 4.32467894588, 4.35466846132, 4.4024274582 ], "latency_us": [ 34.625, 5.125, 4.584, 4.542, 4.542, 38.625, 7.209 ], "throughput_us_p50": 4.35466846132, "throughput_us_p90": 4.39789005384, "latency_us_p50": 5.125, "latency_us_p90": 36.225 }, "oracle": { "elements": 1105, "max_abs_error": 1.7778814753732775e-07, "atol": 5e-05, "rtol": 5e-05 }, "input_sha256": [ "a6e048a8e0295a7c4252b3d419aaf58873fae4b20624e96bf190070d5db6a4c6", "06d75d893dd487843afb83c85f2c5f331f042823d27645db2c57bc4e300c04bc", "02ee90ac3ffd8c46e914e926c89fb82551ffaa1e236b0eb354695ccab376dfd9" ], "output_sha256": "628c3cc4c7a5005337f7f76cbd7bbc6d08034cad68229005c6e222a956f443d8", "process_and_validation_wall_seconds": 1.853909375029616 } ], "summaries": [ { "operation": "rmsnorm", "dimensions": [ 17, 7 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 1.1828439922706444, 1.1580222631935893 ], "median_ratio": 1.170433127732117 }, { "operation": "rmsnorm", "dimensions": [ 17, 127 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 3.58656565941997, 3.6591988582758996 ], "median_ratio": 3.622882258847935 }, { "operation": "rmsnorm", "dimensions": [ 64, 256 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 2.5203173319102223, 2.6446528374751024 ], "median_ratio": 2.5824850846926624 }, { "operation": "rmsnorm", "dimensions": [ 1024, 256 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 3.407978093588539, 3.0954827548759756 ], "median_ratio": 3.2517304242322576 }, { "operation": "rmsnorm", "dimensions": [ 64, 513 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 5.288436921541202, 5.342101298449524 ], "median_ratio": 5.315269109995363 }, { "operation": "layernorm", "dimensions": [ 64, 256 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 3.5438540878135636, 3.13239231047368 ], "median_ratio": 3.338123199143622 }, { "operation": "masked_softmax", "dimensions": [ 17, 65 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 2.9460547154686423, 3.0888901443104224 ], "median_ratio": 3.0174724298895326 }, { "operation": "swiglu", "dimensions": [ 17, 65 ], "complete_pairs": 2, "baseline_over_candidate_throughput_ratios": [ 1.0151752733172434, 1.0142650013342351 ], "median_ratio": 1.0147201373257393 } ] }