)]}' { "log": [ { "commit": "f0b3a2996d6fff82cd3f004c0b29f5ccb0ab947d", "tree": "ba6249799e9f160261c03489a76b1329f08cfa3b", "parents": [ "0a6f23d2984b3685de3b6e1da3fb4bfb96b2b478" ], "author": { "name": "Richard Townsend", "email": "ritownsend@google.com", "time": "Sat Jun 13 01:03:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sat Jun 13 01:04:58 2026" }, "message": "[gn] Experimental CI for x64 Windows\n\nAlso switches off AVX-512 assembly kernels for now on this platform.\nTakes care of another cluster of Chromium failures related to Windows.\nhttps://chromium-review.googlesource.com/c/chromium/src/+/7615883\n\nPiperOrigin-RevId: 931424107\n" }, { "commit": "0a6f23d2984b3685de3b6e1da3fb4bfb96b2b478", "tree": "a70a8591168449972822f95dc9a981734c328f60", "parents": [ "8084a7f94753a5ded5ecf1ed1e657ee01c43b77b" ], "author": { "name": "Gerardo Carranza", "email": "gcarranza@google.com", "time": "Fri Jun 12 20:58:33 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 20:59:29 2026" }, "message": "Add support for asymmetric quantization and weights format in FullyConnected.\n\nPiperOrigin-RevId: 931318618\n" }, { "commit": "8084a7f94753a5ded5ecf1ed1e657ee01c43b77b", "tree": "27aff3d771da7af33574facbcde9b95b7fc1dc25", "parents": [ "7f2b83e6687914524fe48fb04222fcfd3eb37d56" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Fri Jun 12 20:20:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 20:21:35 2026" }, "message": "Update ISA guards for F16-VLOG, F16-VSIN, and F16-VCOS in XNNPACK.\n\nPiperOrigin-RevId: 931300962\n" }, { "commit": "7f2b83e6687914524fe48fb04222fcfd3eb37d56", "tree": "99b76511afbe24b070b6a93bb381361fe1ffadc8", "parents": [ "15e4a6e6472dfaf65aaa595f02c31940543341e4" ], "author": { "name": "Gerardo Carranza", "email": "gcarranza@google.com", "time": "Fri Jun 12 19:51:02 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 19:51:55 2026" }, "message": "Add FullyConnected Generator to ATS.\n\nPiperOrigin-RevId: 931286477\n" }, { "commit": "15e4a6e6472dfaf65aaa595f02c31940543341e4", "tree": "09cf872f3cdbdce1a3aa812ec0db2ca7f1dcc4a8", "parents": [ "cc3f6ba9c156fba27edd86118feb7ea413226cf9" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri Jun 12 19:13:19 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 19:20:10 2026" }, "message": "Fix attention benchmark to be more realistic and test longer sequences\n\n- Fix batching for QD8 case\n- Fix QD8 parameters differing from F32\n- Most of the intermediate tensor sizes are wrong. (This doesn\u0027t matter because XNNPACK overwrites these with inferred shapes.)\n\nPiperOrigin-RevId: 931269422\n" }, { "commit": "cc3f6ba9c156fba27edd86118feb7ea413226cf9", "tree": "7716d3b306e15ce2bade36e225dbf12823829db0", "parents": [ "afa8f9c039d8e705a69f1f0d9b69d277a7e031a1" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri Jun 12 19:09:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 19:13:09 2026" }, "message": "Add `ynn_define_gather` and remove `ynn_define_lut`\n\nThis generalizes the LUT operation to more types, and allows multiple dimensions in the \"LUT\" (now the input).\n\nThis should perform well in both the gather dimension 0 case (provided we have an appropriate LUT kernel), and for other dimensions, because `slinky::copy` should boil down to a memcpy of large chunks in most cases.\n\nPiperOrigin-RevId: 931267865\n" }, { "commit": "afa8f9c039d8e705a69f1f0d9b69d277a7e031a1", "tree": "70a4a1d3ee32fb2d6e380c1cbccbca1f1a23803e", "parents": [ "ff27209ac527ec70bc93d76ea5ed29ba90fdc73d" ], "author": { "name": "Gerardo Carranza", "email": "gcarranza@google.com", "time": "Fri Jun 12 18:56:26 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 18:57:17 2026" }, "message": "Add BatchMatmul Generator to ATS.\n\nPiperOrigin-RevId: 931261281\n" }, { "commit": "ff27209ac527ec70bc93d76ea5ed29ba90fdc73d", "tree": "be6997efc9b3459463398de9f407f5c9d2d957ea", "parents": [ "f774d42bbdf2551308e36ccf423eae6633a8755a" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Fri Jun 12 01:51:39 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 12 01:52:25 2026" }, "message": "f16-f32acc-approxgelu for WebAssembly and native\n\n```gemma4 12b on AMD Zen4 is 7.38% faster end to end\nWas scalar\nInference (avg): 133760.34 ms (2 runs)\n 3.32% xnn_f16_vapproxgelu_ukernel__scalar_rational_6_4_div_u4\n\nNow avx512skx\nInference (avg): 124561.43 ms (2 runs)\n 0.03% xnn_f16_f32acc_vapproxgelu_ukernel__avx512f_rational_6_4_div_u32\n```\n\n| Architecture | Status | Details |\n| :---------------- | :------- | :------------------------------------ |\n| WASM relaxed SIMD | Added | New native FP16 and f32acc kernels. |\n| x86 avx512fp16 | Added | New native FP16 microkernel (u32). |\n| x86 avx512f | Added | New microkernel using f16_f32acc. |\n| x86 f16c | Added | New microkernel using f16_f32acc. |\n| Hexagon hvx | Added | New f16_f32acc microkernel (u128). |\n| ARM neonfp16 | Added | New fallback using f16_f32acc. |\n| RISC-V rvvfp16 | Improved | Increased loop unroll u2v to u8v. |\n| Fallback scalar | Improved | Switched to f16_f32acc for precision. |\n\nPiperOrigin-RevId: 930857873\n" }, { "commit": "f774d42bbdf2551308e36ccf423eae6633a8755a", "tree": "69edc635325c534e4d1099a6a4c2a1fbb761fadd", "parents": [ "199b0e5c9a90fc94dd67c4b324efe5fa8a157598" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu Jun 11 16:16:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 11 16:17:42 2026" }, "message": "Add `ynn_define_dynamic_quantization`\n\nThis adds a new node type that computes dynamic quantization parameters given a min_max result. It\u0027s a little unfortunate to need to define a node type for this (vs. just using arithmetic in a graph), but it\u0027s a lot of tricky arithmetic, and I think it isn\u0027t worth it. I think that this node (a unary operator that consumes a min_max and produces a scale and zero point) is a reasonable thing to have.\n\nPiperOrigin-RevId: 930567132\n" }, { "commit": "199b0e5c9a90fc94dd67c4b324efe5fa8a157598", "tree": "08c98d21d5d34da0cab0b8164889355b0fc68745", "parents": [ "e71d1ae431b2463735abc0fe59a327c949c04bbd" ], "author": { "name": "Byungchul Kim", "email": "byungchul@google.com", "time": "Thu Jun 11 06:48:41 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 11 06:49:35 2026" }, "message": "Internal changes only\n\nPiperOrigin-RevId: 930321019\n" }, { "commit": "e71d1ae431b2463735abc0fe59a327c949c04bbd", "tree": "4564647bb430d2b6d0af136ffd697ae441dfda15", "parents": [ "f0e859a4f219197fa9f8e007e54353a9386b45ca" ], "author": { "name": "Byungchul Kim", "email": "byungchul@google.com", "time": "Thu Jun 11 05:23:00 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 11 05:23:40 2026" }, "message": "Fix memory leak\n\nPiperOrigin-RevId: 930285500\n" }, { "commit": "f0e859a4f219197fa9f8e007e54353a9386b45ca", "tree": "6bc06060bb08684dcb3ff516d793a617d57ece45", "parents": [ "118524434ffeca4d5511cecebe9f33baa96b6e4f" ], "author": { "name": "Byungchul Kim", "email": "byungchul@google.com", "time": "Thu Jun 11 00:56:06 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 11 00:57:14 2026" }, "message": "XNNPACK runner allows external tensor to grow even when it was added without copy\n\nPiperOrigin-RevId: 930183355\n" }, { "commit": "118524434ffeca4d5511cecebe9f33baa96b6e4f", "tree": "03b5e0b57b9993bdfbaf0a0e68e4e9a80f91e390", "parents": [ "87730c125b45180acb1bb57010b8659ee8e9660b" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 22:03:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 22:04:03 2026" }, "message": "Add fp8 types and reference convert kernels to YNNPACK\n\nPiperOrigin-RevId: 930097979\n" }, { "commit": "87730c125b45180acb1bb57010b8659ee8e9660b", "tree": "270709612ebe707581e1fbb0f8af67470e3919dd", "parents": [ "1e14f367a23c628fdfaf4bf46890adec7873ec6a" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 21:44:43 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 21:45:17 2026" }, "message": "Fix bf16 gemm config when there are no kernels\n\nPiperOrigin-RevId: 930088216\n" }, { "commit": "1e14f367a23c628fdfaf4bf46890adec7873ec6a", "tree": "5405c8349c846ecbc8e83b50a3207d2cab96f6e6", "parents": [ "1f742eb0f94df33e6438ee03d41ca4a08474c85a" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 20:46:23 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 20:47:17 2026" }, "message": "Add tracing to tsl::profiler traces in tensorflow\n\nPiperOrigin-RevId: 930054643\n" }, { "commit": "1f742eb0f94df33e6438ee03d41ca4a08474c85a", "tree": "7a027ed2e30556d0a17a7200a81dca0bcc401bba", "parents": [ "508ca2d608a6040f4f431cce9fef8096f7862a10" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Wed Jun 10 18:46:45 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 18:47:43 2026" }, "message": "Remove packing output buffer scheduling.\n\nThis is mostly redundant now, because we pretty much always want to store it at the innermost level which is done automatically now by scheduler when the loop splits/extents are set. There are some cases though where specifying buffer output level was preventing it from being stored at the innermost level and caused TSAN issues (benign from what I can tell, but anyway).\n\nPiperOrigin-RevId: 929985655\n" }, { "commit": "508ca2d608a6040f4f431cce9fef8096f7862a10", "tree": "180e25ca0e60eda9ad82224a025e9a72f83b7a1a", "parents": [ "3912146d2278747e0e37d2995856ae390b5afde6" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 18:18:09 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 18:18:46 2026" }, "message": "Fix new tests in YNNPACK\n\nPiperOrigin-RevId: 929967518\n" }, { "commit": "3912146d2278747e0e37d2995856ae390b5afde6", "tree": "93de0adbc9ba129e4ff93791e89105dfdbeec596", "parents": [ "463600be62298d8e6ab11986244dfbc6fc07cd31" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 17:44:15 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 17:45:10 2026" }, "message": "Add tracing to tsl::profiler traces in tensorflow\n\nPiperOrigin-RevId: 929947868\n" }, { "commit": "463600be62298d8e6ab11986244dfbc6fc07cd31", "tree": "8319950f29ad6623fe7faa0d7ab14aa44ac160a1", "parents": [ "f42612c6fd2db39221bafb005a5cce0f4174eee5" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 17:10:00 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 17:15:45 2026" }, "message": "Add avx512 fp16 convert kernels\n\n```\n-----------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n-----------------------------------------------------------------------------------------------------------------\nbench/convert_fp16_to_fp32_avx512/m:1/n:4096/real_time 153 ns 153 ns 907802 Bytes\u003d160.259G/s Op\u003d26.7099G/s\nbench/convert_fp32_to_fp16_avx512/m:1/n:4096/real_time 183 ns 183 ns 785281 Bytes\u003d134.187G/s Op\u003d22.3645G/s\nbench/convert_fp16_to_fp32_f16c/m:1/n:4096/real_time 227 ns 227 ns 609680 Bytes\u003d108.185G/s Op\u003d18.0308G/s\nbench/convert_fp32_to_fp16_f16c/m:1/n:4096/real_time 239 ns 239 ns 593642 Bytes\u003d102.906G/s Op\u003d17.151G/s\n```\n\nPiperOrigin-RevId: 929926693\n" }, { "commit": "f42612c6fd2db39221bafb005a5cce0f4174eee5", "tree": "df27165ec1e987895e42e7c008da3d6b58546e3c", "parents": [ "71b2a887f7907a537867617b112eb8145e9d3b2d" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed Jun 10 17:09:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 17:12:10 2026" }, "message": "Fix SSE2 gemm being used when it should not be AVX/AVX512\n\nPiperOrigin-RevId: 929926446\n" }, { "commit": "71b2a887f7907a537867617b112eb8145e9d3b2d", "tree": "55eb62eca24e96c6891bba673a0ec5a53fbe93dc", "parents": [ "c6f7e32e4ef088c8226b3619bfcaf241989f4274" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 16:27:24 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 16:28:19 2026" }, "message": "Add tracing to tsl::profiler traces in tensorflow\n\nPiperOrigin-RevId: 929901110\n" }, { "commit": "c6f7e32e4ef088c8226b3619bfcaf241989f4274", "tree": "4d2b6ab201624232c79a1e58dd4e0f25e6feaf97", "parents": [ "85115d255214bcc82b2c1c91e49d797817c033e1" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 10 13:32:50 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 13:33:28 2026" }, "message": "Relax tolerance of batch_matrix_multiply test\n\nPiperOrigin-RevId: 929815744\n" }, { "commit": "85115d255214bcc82b2c1c91e49d797817c033e1", "tree": "a2da2b7c144e4a51c61a06e9bcc34db44b813e1d", "parents": [ "a80e3c43903fc46cdca04355d62048dc46df75d0", "337236cbc3cbb0e5ed0173f10b908539065f9f79" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 05:51:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 05:51:58 2026" }, "message": "Merge pull request #10412 from aizu-m:rope-tokens-weights-bound\n\nPiperOrigin-RevId: 929616593\n" }, { "commit": "a80e3c43903fc46cdca04355d62048dc46df75d0", "tree": "d77b8b3a4f37a386a7b7fc528c365c7f344c5a1a", "parents": [ "843528da3013dd382b05ef47dce0f8f76290f9d4", "a9c19098a1ec01747938276bfab046574f282320" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 01:20:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 10 01:20:55 2026" }, "message": "Merge pull request #10404 from aizu-m:expand-dims-axis-validation\n\nPiperOrigin-RevId: 929514261\n" }, { "commit": "843528da3013dd382b05ef47dce0f8f76290f9d4", "tree": "2c5e6d8b028672f46d918c41af57c03ac2fbe150", "parents": [ "9d225af71a8233bb804c5a90b9241537455db00d", "463c3e99ff7c9d6402f0fe7b90dfdd56ff5da8d8" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 23:38:39 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 23:38:39 2026" }, "message": "Merge pull request #10424 from aizu-m:softmax-convert-input-rank\n\nPiperOrigin-RevId: 929472951\n" }, { "commit": "9d225af71a8233bb804c5a90b9241537455db00d", "tree": "ea3739867b722afd7cb5c08460999c0a66f8cd9a", "parents": [ "13c32f5eb0270b5d9da80d1cbd7e889460fef53b", "a9a5a66201285e18b87d80ccd2c54710681ac273" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 22:29:01 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 22:29:01 2026" }, "message": "Merge pull request #10389 from aizu-m:igemm-indirection-overflow\n\nPiperOrigin-RevId: 929435380\n" }, { "commit": "13c32f5eb0270b5d9da80d1cbd7e889460fef53b", "tree": "ee5968691259b13b6edb80b08e2ef3441d19c4a9", "parents": [ "494ddd0620a88b2b8787822d1dde3f091be476c8" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Tue Jun 09 22:24:27 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 22:25:19 2026" }, "message": "Add AVX512SKX F16-F32ACC GEMM/IGEMM microkernels to XNNPACK\n - 7.4% on Gemma4 12B.\n - 22% speedup on MobileNet\n\nPiperOrigin-RevId: 929434636\n" }, { "commit": "494ddd0620a88b2b8787822d1dde3f091be476c8", "tree": "1315c06fa9a06e264b5705dcd7f196bf63c358d6", "parents": [ "4bd2e73bc65224c71edb18e4d0fad0b2b7442f29" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Tue Jun 09 21:04:37 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 21:05:40 2026" }, "message": "Fixed underallocating memory for qint8-\u003eqcint8 conversion which only could have been seen in combination with weight_cache.\n\nPiperOrigin-RevId: 929392409\n" }, { "commit": "4bd2e73bc65224c71edb18e4d0fad0b2b7442f29", "tree": "d8575ab621df012ac1e4cd0fdaa10f348efa2977", "parents": [ "0e469abaec8870178dde9684ffef176f0591b465", "8d8fa1feeb07649acc5982d7aad59b1c45b98b85" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 20:57:28 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 20:57:28 2026" }, "message": "Merge pull request #10357 from aizu-m:constant-pad-dim-overflow\n\nPiperOrigin-RevId: 929387475\n" }, { "commit": "0e469abaec8870178dde9684ffef176f0591b465", "tree": "1f4ef193e9d82b5ea3eecff4ffd454fed5b85168", "parents": [ "a505fe2a9eb4038b9fe159f9be173467b102ddbe" ], "author": { "name": "Byungchul Kim", "email": "byungchul@google.com", "time": "Tue Jun 09 18:17:47 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 18:18:30 2026" }, "message": "Internal changes only\n\nPiperOrigin-RevId: 929303073\n" }, { "commit": "a505fe2a9eb4038b9fe159f9be173467b102ddbe", "tree": "d49340fafe32b8f70d7176275c0da3a191f43e5b", "parents": [ "f5d613307ef88ef784395640f4d48ac1a1e42b21" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue Jun 09 15:54:17 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 15:55:08 2026" }, "message": "Only define `isinf`, `isnan`, `isfinite` if not on MSVC\n\nPiperOrigin-RevId: 929223059\n" }, { "commit": "463c3e99ff7c9d6402f0fe7b90dfdd56ff5da8d8", "tree": "1ef6ce0dcb4f145973b6ee4dfaaf6e0be8e64c33", "parents": [ "f5d613307ef88ef784395640f4d48ac1a1e42b21" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Tue Jun 09 06:59:47 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Tue Jun 09 06:59:47 2026" }, "message": "validate input rank in softmax and qp8 convert reshape\n" }, { "commit": "f5d613307ef88ef784395640f4d48ac1a1e42b21", "tree": "998aeb72769469978706e0d46b115a9702f957a5", "parents": [ "d15762162cb53e41ec7f5405057618c2a9bb8c31", "ab8714dd06f570ec5afe4a1e374c961d725a4d3f" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 02:32:20 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 02:32:20 2026" }, "message": "Merge pull request #10230 from qualcomm:sme1/qp8-f32-qc4w-gemm\n\nPiperOrigin-RevId: 928896476\n" }, { "commit": "d15762162cb53e41ec7f5405057618c2a9bb8c31", "tree": "78e5fc9b93bc4a4e8b7e7c9979aca57f016def42", "parents": [ "95a93bc1a7e20799cae21521d6019a6e709eee01" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue Jun 09 00:02:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 09 00:03:10 2026" }, "message": "Refactor precision for unary ops\n\nI\u0027d like to add a lot more precision variations for unary kernels. The current approach of adding a new operator type for each one is impractical.\n\nThis change changes this to use flags indicating precision requirements, instead of using new operator types. In order to do this, a new parameter `flags` is added to the operator decorator in the elementwise op compiler. However, the `consistent_arithmetic` flag needs to remain as part of the generator...\n\nPiperOrigin-RevId: 928847650\n" }, { "commit": "95a93bc1a7e20799cae21521d6019a6e709eee01", "tree": "e82fb807438be50fc4f53e51a8498cf35d4125a1", "parents": [ "00f16e113df9fb66dbfc09214d1f80a1822ac6dd" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 22:20:34 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 22:21:20 2026" }, "message": "Add `isnan`, `isinf`, `isfinite` for YNNPACK float types\n\nThis avoids overload resolution problems for some configs.\n\nPiperOrigin-RevId: 928798461\n" }, { "commit": "00f16e113df9fb66dbfc09214d1f80a1822ac6dd", "tree": "0fc014edaee999e9be64a27f22855d06a2599f09", "parents": [ "bd205c6baae1ca6e6683a5c1588ed471a55694c4" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 21:59:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 22:00:00 2026" }, "message": "Clean up target suffix for elementwise ops\n\n- avx512bw is not a separate target any more.\n- x86 prefix should not be there (especially on non-x86 targets)\n\nPiperOrigin-RevId: 928787637\n" }, { "commit": "bd205c6baae1ca6e6683a5c1588ed471a55694c4", "tree": "2a315ab6ee4002f8f7322e3ebf21c1c03a68419a", "parents": [ "f6486e3e1dbba03be9a493c8d6ead20c787c9886" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 21:36:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 21:36:50 2026" }, "message": "Refactor precision for unary ops\n\nI\u0027d like to add a lot more precision variations for unary kernels. The current approach of adding a new operator type for each one is impractical.\n\nThis change changes this to use flags indicating precision requirements, instead of using new operator types. In order to do this, a new parameter `flags` is added to the operator decorator in the elementwise op compiler. However, the `consistent_arithmetic` flag needs to remain as part of the generator, because the same kernel may be consistent or not (depending on whether FMA is used).\n\nPiperOrigin-RevId: 928776072\n" }, { "commit": "f6486e3e1dbba03be9a493c8d6ead20c787c9886", "tree": "43bbb8ee72eb8fd45f9da32f84c177f01c996b86", "parents": [ "f4271c2abec491cfd7c52bced38c5f6958ac463d" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 17:53:06 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 17:53:48 2026" }, "message": "Explicitly number enumerations in ynnpack.h for ABI stability\n\nAlso, remove unused `ynn_type_opaque`. This was only ever usable internally, so it should not have ABI stability implications.\n\nPiperOrigin-RevId: 928656162\n" }, { "commit": "f4271c2abec491cfd7c52bced38c5f6958ac463d", "tree": "8081247fcb45ca79f0bc116020ac2a39659f837b", "parents": [ "32df615f7abce44e889afa7862207834a7e3c611" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 17:01:24 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 17:02:11 2026" }, "message": "Remove concatenating constructor from `vec` types\n\nAnd use `concat` free functions instead. This will make it easier to add a clang/GCC vector type implementation of the SIMD wrappers.\n\nPiperOrigin-RevId: 928624690\n" }, { "commit": "32df615f7abce44e889afa7862207834a7e3c611", "tree": "e3e71a6a27b75b984b016da28c355214b02a1114", "parents": [ "e504f4e0ba0c2bf992ee0686175844e5c86443f4" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 08 15:48:50 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 08 15:49:53 2026" }, "message": "Move `sum_kn` kernels from `neondot` to `neon`\n\nThese kernels don\u0027t require `neondot`.\n\nAlso, fix the target requirement of `neonfma` kernels.\n\nPiperOrigin-RevId: 928587822\n" }, { "commit": "337236cbc3cbb0e5ed0173f10b908539065f9f79", "tree": "82d6f745a5e22a4d599d2cd1755c0c78ff934835", "parents": [ "e504f4e0ba0c2bf992ee0686175844e5c86443f4" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Sat Jun 06 08:49:56 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Sat Jun 06 08:49:56 2026" }, "message": "bound rope token count to weights tensor size in reshape\n" }, { "commit": "e504f4e0ba0c2bf992ee0686175844e5c86443f4", "tree": "801022335ce7fd0beaecbca0dccb6f7123a8f71f", "parents": [ "5be8df86dacdebbc3702438ac4adf4ee4a35a960" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri Jun 05 22:42:04 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 05 22:46:12 2026" }, "message": "Add rewrite for add(square(x), y) \u003d\u003e multiply_add(x, x, y)\n\nThe elementwise benchmark hits this case:\n```\nname time/op time/op vs base\nFP32Elementwise/B:1024/N:1024/D:6/process_time/real_time 3.059m ± 5% 2.378m ± 5% -22.26% (p\u003d0.002 n\u003d6)\nFP32Elementwise/B:1024/N:1024/D:10/process_time/real_time 4.330m ± 6% 3.067m ± 2% -29.16% (p\u003d0.002 n\u003d6)\nFP32Elementwise/B:1024/N:1024/D:18/process_time/real_time 6.900m ± 4% 4.404m ± 8% -36.17% (p\u003d0.002 n\u003d6)\nFP32Elementwise/B:1024/N:1024/D:34/process_time/real_time 11.573m ± 4% 7.026m ± 4% -39.29% (p\u003d0.002 n\u003d6)\ngeomean 5.703m 3.876m -32.03%\n```\n\nPiperOrigin-RevId: 927514222\n" }, { "commit": "5be8df86dacdebbc3702438ac4adf4ee4a35a960", "tree": "a1493c052872e5a07d240b56a32678b0002500c9", "parents": [ "3e5f1c7e200d72aeb63b09e4b73d38d9ac2470db" ], "author": { "name": "Alexander Shaposhnikov", "email": "ashaposhnikov@google.com", "time": "Fri Jun 05 22:40:38 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 05 22:41:32 2026" }, "message": "Adjust define_transpose_a declaration.\n\nPiperOrigin-RevId: 927513560\n" }, { "commit": "3e5f1c7e200d72aeb63b09e4b73d38d9ac2470db", "tree": "a6f7c63f93004b8291e66dda335eafaa9eac0dcb", "parents": [ "ce684eb2e29bc825c5d60b64eba8ddac1d1e3db3" ], "author": { "name": "Byungchul Kim", "email": "byungchul@google.com", "time": "Fri Jun 05 16:56:33 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 05 16:57:25 2026" }, "message": "Internal changes only\n\nPiperOrigin-RevId: 927352414\n" }, { "commit": "a9c19098a1ec01747938276bfab046574f282320", "tree": "2235f7445ef576eb3a5b8581c510fdc753178cff", "parents": [ "ce684eb2e29bc825c5d60b64eba8ddac1d1e3db3" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Fri Jun 05 13:15:59 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Fri Jun 05 13:15:59 2026" }, "message": "Validate expand-dims axes in resize_expand_dims_output_tensor\n" }, { "commit": "ce684eb2e29bc825c5d60b64eba8ddac1d1e3db3", "tree": "32e1f53ef4ffe8a00082ac6c20d8dd3131680511", "parents": [ "dc0d8821383b258f31b432f22e5824b7d0d70f23" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Fri Jun 05 12:46:57 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri Jun 05 12:47:31 2026" }, "message": "Rewrote bmm(f32, dequant(qint8)) -\u003e f32 to bmm(f32, qint8 -\u003e qcint8) -\u003e f32.\n\nPiperOrigin-RevId: 927245953\n" }, { "commit": "dc0d8821383b258f31b432f22e5824b7d0d70f23", "tree": "ded308913a6a0922e92e31088002636ae2ad1dc4", "parents": [ "12f3b6e9c6abe4d5ebc9c51bea47237c463d806a" ], "author": { "name": "Alexander Shaposhnikov", "email": "ashaposhnikov@google.com", "time": "Thu Jun 04 23:32:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 23:33:43 2026" }, "message": "Fix wrong types in hexagon_hvx.h\n\nPiperOrigin-RevId: 926951112\n" }, { "commit": "12f3b6e9c6abe4d5ebc9c51bea47237c463d806a", "tree": "60344d5e012e757aa7980eee7f5cf535e0bd949d", "parents": [ "3ed73f57b5ecaa99967bb21862b0c62f47b9bf7c" ], "author": { "name": "Alexander Shaposhnikov", "email": "ashaposhnikov@google.com", "time": "Thu Jun 04 23:16:33 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 23:17:23 2026" }, "message": "On Hexagon std::int32_t is defined as long, which is distinct from int. Because slinky::thread_pool uses int in its base class signature, using int32_t in the derived class causes a type mismatch.\nThis prevented the method from overriding the virtual base class method, leaving slinky_thread_pool as an abstract class.\n\nPiperOrigin-RevId: 926943867\n" }, { "commit": "3ed73f57b5ecaa99967bb21862b0c62f47b9bf7c", "tree": "f9da694295f75d622b83826eebe9e90a1d3c89c3", "parents": [ "43c4f7b3dd728e86fc2ae88e19e60af064345f8e" ], "author": { "name": "Mohammadreza Heydary", "email": "mheydary@google.com", "time": "Thu Jun 04 17:22:37 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 17:23:19 2026" }, "message": "Rewrote bmm(f32, dequant(qint8)) -\u003e f32 to bmm(f32, qint8 -\u003e qcint8) -\u003e f32.\n\nPiperOrigin-RevId: 926757604\n" }, { "commit": "43c4f7b3dd728e86fc2ae88e19e60af064345f8e", "tree": "bfacc5b8c2e07034ed3ae9512ad6b608a98fc866", "parents": [ "1f4e631e0025c17238a6e73d7e2d13e5c8a16ee0" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Thu Jun 04 12:06:45 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 12:07:27 2026" }, "message": "Add fine-grain detection of unsupported fp16 ops when falling back to fp32.\n\nPiperOrigin-RevId: 926624045\n" }, { "commit": "1f4e631e0025c17238a6e73d7e2d13e5c8a16ee0", "tree": "83bf0ed8ff9e26ca023c31e6a63146e78718bd89", "parents": [ "ce3b7c1b4da1911244db7499553ae23d0a3c2434" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Thu Jun 04 08:26:41 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 08:27:22 2026" }, "message": "Rewrote bmm(f32, dequant(qint8)) -\u003e f32 to bmm(f32, qint8 -\u003e qcint8) -\u003e f32.\n\nPiperOrigin-RevId: 926540409\n" }, { "commit": "a9a5a66201285e18b87d80ccd2c54710681ac273", "tree": "8ac4636531ac34805cb8d921fc5dbcafaea32223", "parents": [ "ce3b7c1b4da1911244db7499553ae23d0a3c2434" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Thu Jun 04 06:54:28 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Thu Jun 04 06:54:28 2026" }, "message": "Guard indirection buffer size against overflow in igemm reshape\n" }, { "commit": "ce3b7c1b4da1911244db7499553ae23d0a3c2434", "tree": "0e2e909c61a24cff426ee39a10026c0fdcfba9d0", "parents": [ "68373b6aa3382dcb95bf61f09732a1cc4d70084b" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Thu Jun 04 02:57:51 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 02:58:42 2026" }, "message": "Fix XNNPACK compilation failure on Windows ARM64.\n\nPiperOrigin-RevId: 926414824\n" }, { "commit": "68373b6aa3382dcb95bf61f09732a1cc4d70084b", "tree": "bc7a0573e36c53e9cf2404aa1e6bd2e2e6e03530", "parents": [ "cf112cb82af2cb66c47ddef0ce85177c4682a5ac" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu Jun 04 00:30:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 00:31:39 2026" }, "message": "Disable F32QC8W tests when using YNNPACK\n\nPiperOrigin-RevId: 926364534\n" }, { "commit": "cf112cb82af2cb66c47ddef0ce85177c4682a5ac", "tree": "081750a67e6183a6b4b5a90557575a3b02ea6477", "parents": [ "98930bab768034e2ee87e9936e01bed68961d705" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu Jun 04 00:14:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu Jun 04 00:15:07 2026" }, "message": "Minor loop fusion improvements\n\n- Don\u0027t require splits of extent 1 to match\n- If extents don\u0027t match, stop trying to fuse\n\nPiperOrigin-RevId: 926358417\n" }, { "commit": "98930bab768034e2ee87e9936e01bed68961d705", "tree": "c2fab60bd6aeb1a6304d619b97183028b7c45653", "parents": [ "eb93f24dc52165a803783a974d25c49b9ee4d54e" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 03 23:32:41 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 23:33:58 2026" }, "message": "Test reduce kernels with pi summation in both ascending and descending order\n\nPiperOrigin-RevId: 926340371\n" }, { "commit": "eb93f24dc52165a803783a974d25c49b9ee4d54e", "tree": "f0a4aee1cc81b9bf5dee9da6225ba6e399689279", "parents": [ "5fa627d5a0aaadfe7986807643084ac82d8f774a" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed Jun 03 20:26:46 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 20:27:35 2026" }, "message": "Correctly zero-initialize `xnn_unary_params`.\n\nPiperOrigin-RevId: 926230630\n" }, { "commit": "5fa627d5a0aaadfe7986807643084ac82d8f774a", "tree": "d39cb032f02ddca068430c5218f20fdd8ddb749d", "parents": [ "51af7f59cbf73931df0a324e29f645786386bcdf" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed Jun 03 19:57:20 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 19:58:01 2026" }, "message": "f16 vlog switch from rational-3-3 to rational-1-3\n\n- add f16-f32acc that uses f16 rational-1-3 but with a convert to float\n- add nr as alternative to div\n\nPiperOrigin-RevId: 926213330\n" }, { "commit": "51af7f59cbf73931df0a324e29f645786386bcdf", "tree": "0b968696f6b8b904872364987527ba79caa8abc5", "parents": [ "6f2d84ec43703cfc6374ac24756357555b593b1b" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed Jun 03 19:52:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 19:52:45 2026" }, "message": "Make external tensor order follow topological traversal of the graph.\n\nPiperOrigin-RevId: 926209945\n" }, { "commit": "6f2d84ec43703cfc6374ac24756357555b593b1b", "tree": "c9daaa202ef77c270ded0e2639037e2b7213594e", "parents": [ "859e1a51afac9feeca89e7842852a409cb27d9df" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed Jun 03 19:37:01 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 19:37:46 2026" }, "message": "Fix isomorphic matcher.\n\nUnary and binary nodes would not be stably sorted (all unary ops would modify\nthe hash in the same way, likewise for binary ops).\n\nSort level 0. Graphs that would be identical until level 1 would otherwise not\nbe stable.\n\nPiperOrigin-RevId: 926200411\n" }, { "commit": "859e1a51afac9feeca89e7842852a409cb27d9df", "tree": "7f47be4ec3f4f8de3b01f156c85059ca1a16d5c9", "parents": [ "799de5c9b38344791884ed96040d1c2f4f7b0478", "8649fdd1a12c8a195da8f2c6aad84b495da791f9" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 17:39:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 17:39:55 2026" }, "message": "Merge pull request #10369 from aizu-m:reduce-axis-bounds\n\nPiperOrigin-RevId: 926131208\n" }, { "commit": "799de5c9b38344791884ed96040d1c2f4f7b0478", "tree": "3e13cfe830c229bdd2e2f2b86e3ea7d1e1d65b3c", "parents": [ "4bc0e6ab80e6f5172a32c9a3551e96467d1e2e0a", "6062dc0841c6ccdf5eba87423a72ae27fac60b8f" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 17:31:29 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 17:31:29 2026" }, "message": "Merge pull request #10373 from sin99xx:master\n\nPiperOrigin-RevId: 926125906\n" }, { "commit": "4bc0e6ab80e6f5172a32c9a3551e96467d1e2e0a", "tree": "90bbb1a975383b5e763ab40b21dbe698ce61e5d6", "parents": [ "207dab62c9b4f50ff2a7da31ec89f1a3224524ca" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed Jun 03 11:19:54 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 11:20:56 2026" }, "message": "Remove useless test main function.\n\nPiperOrigin-RevId: 925931669\n" }, { "commit": "207dab62c9b4f50ff2a7da31ec89f1a3224524ca", "tree": "24d99daf7584d2c675d7aff3deba4eaf7067275f", "parents": [ "932ea64e32ee29518e601b8b1dcc4f40d2ca77ad" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Wed Jun 03 10:49:01 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 10:49:51 2026" }, "message": "Enabled f32_qc8w bmm on the subgraph level.\n\nPiperOrigin-RevId: 925919282\n" }, { "commit": "932ea64e32ee29518e601b8b1dcc4f40d2ca77ad", "tree": "9a971dba21172828584cce2a62224812f5bbcee6", "parents": [ "85a6530d7ee905d0fdf2e0ec9eb44f0cafbe5f1a" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed Jun 03 10:13:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 10:14:53 2026" }, "message": "Make hardware configuration and initialization guards (re)setable.\n\nPiperOrigin-RevId: 925905501\n" }, { "commit": "85a6530d7ee905d0fdf2e0ec9eb44f0cafbe5f1a", "tree": "313e785ffb74555bd827ccffa996284b3102a133", "parents": [ "be7bb974f7f5852deefa6a91aca8cc3692613236" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Wed Jun 03 08:58:22 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 08:59:03 2026" }, "message": "Added f32_qc8w operator level support for batch matrix multiply.\n\nPiperOrigin-RevId: 925872216\n" }, { "commit": "be7bb974f7f5852deefa6a91aca8cc3692613236", "tree": "80818f6c74218639f9f1b17c219ac9f02da04638", "parents": [ "a1b0202e619ed69d8f24f1fba9a8558f4ae60577" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed Jun 03 08:30:40 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 08:31:21 2026" }, "message": "Don\u0027t rewrite sum(a * b) \u003d\u003e dot(a, b) if the dot would be a vector-vector multiply\n\nReduce is faster in this case.\n\nPiperOrigin-RevId: 925858104\n" }, { "commit": "a1b0202e619ed69d8f24f1fba9a8558f4ae60577", "tree": "08b45c5d2199fc3e00d23f944c782f16428951b0", "parents": [ "2fef1fcf819bee6990dc95ce52aa1ca2f35aa99f" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed Jun 03 02:34:21 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 02:34:56 2026" }, "message": "Add 2 bit SSE GEMM microkernels\n\nThese updates enable 2-bit quantization support for both QS8-QC2W and QD8-F32-QC2W using SSSE3 instructions with MADD optimization.\n 1. src/qs8-gemm/MRx4c8-ssevnni.c.in:\n * Added support for QS8_QC2, QC2_F32, and QC2_F16 datatypes.\n * Introduced the _MM_SET1_EPI8 macro for consistent constant generation.\n * Updated the ISA and instruction selection logic to support MADD variants (specifically _mm_dpbusd_epi32_madd_kzp2 for 2-bit variants).\n * Updated the function signature to include the row_sum parameter for QD8 variants.\n 2. scripts/generate-qs8-gemm.sh:\n * Added generation rules for SSSE3 with MADD\u003d1 for both QS8_QC2 and QC2_F32 variants.\n 3. src/xnnpack/gemm.h:\n * Added microkernel declarations for xnn_qd8_f32_qc2w_gemm_minmax_ukernel_*x4c8__ssse3_madd.\n * Added microkernel declarations for xnn_qs8_qc2w_gemm_minmax_fp32_ukernel_*x4c8__ssse3_madd.\n\nPiperOrigin-RevId: 925718953\n" }, { "commit": "2fef1fcf819bee6990dc95ce52aa1ca2f35aa99f", "tree": "3600debea44d8bb66c4b10547fcc868ce8d7e4b9", "parents": [ "d43126037f9dd8cc2d182ad658237f72af9e0b7d" ], "author": { "name": "Alexander Shaposhnikov", "email": "ashaposhnikov@google.com", "time": "Wed Jun 03 02:04:34 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 02:05:13 2026" }, "message": "Change cache key schema for bazel builds.\n\nPiperOrigin-RevId: 925708474\n" }, { "commit": "d43126037f9dd8cc2d182ad658237f72af9e0b7d", "tree": "e0001979fee34c11cb5cb83fae350f2428ca3d3b", "parents": [ "cced44b514117bce11f1a65c858b778b3367ee89" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed Jun 03 01:05:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed Jun 03 01:06:50 2026" }, "message": "Fix undeclared identifier XNN_SIMD_NUM_RCP_ITER_F16 in wasm relaxed simd fp16\n\nPiperOrigin-RevId: 925688061\n" }, { "commit": "6062dc0841c6ccdf5eba87423a72ae27fac60b8f", "tree": "5a654f720ca10565eaf6122a3e402f505ac39cce", "parents": [ "cced44b514117bce11f1a65c858b778b3367ee89" ], "author": { "name": "waris )", "email": "sinxx198@gmail.com", "time": "Tue Jun 02 22:15:33 2026" }, "committer": { "name": "GitHub", "email": "noreply@github.com", "time": "Tue Jun 02 22:15:33 2026" }, "message": " fix-concat-oob-write\n\nAdded checks to ensure all inputs have the same rank and dimensions except along the concatenation axis, preventing out-of-bounds writes." }, { "commit": "cced44b514117bce11f1a65c858b778b3367ee89", "tree": "5887afcbf4d276cb91df87a0313e3c6fc02c5016", "parents": [ "383840f80b3f8d96a34f0d59a232f64f1172d0e9" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue Jun 02 20:14:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 02 20:15:41 2026" }, "message": "Add fp16 and bf16 implementations of `exp`, `expm1`, `log`, `log1p`, `erf`, `tanh`\n\n- Add overloads of the fp32 versions of these ops that take a `precision12` tag, indicating we only want 12 bits of precision.\n- Add fp16 and bf16 versions of these ops that cast to float and use the `precision12` tag.\n- Also add `approx_log`, `approx_log1p`, `approx_erf`, `approx_tanh` overloads that actually use fp16 arithmetic. These have a few more ULP of error (despite using the same polynomials, due to intermediates not having extra precision).\n\nPiperOrigin-RevId: 925538751\n" }, { "commit": "383840f80b3f8d96a34f0d59a232f64f1172d0e9", "tree": "88c279be28f118656dc8ee5d1bde77c1384e5c4e", "parents": [ "76de13802d1c1b286b21694734d87f1683767b8f" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Tue Jun 02 18:52:40 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue Jun 02 18:53:17 2026" }, "message": "Update XNNPACK elementwise benchmarks to use consistent N elements.\n\nPiperOrigin-RevId: 925495509\n" }, { "commit": "8649fdd1a12c8a195da8f2c6aad84b495da791f9", "tree": "16de6b6bab290f54c02f7c9403d94b792dc90877", "parents": [ "76de13802d1c1b286b21694734d87f1683767b8f" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Tue Jun 02 06:52:58 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Tue Jun 02 06:52:58 2026" }, "message": "validate reduction axes before nchw remap in static reduce\n" }, { "commit": "76de13802d1c1b286b21694734d87f1683767b8f", "tree": "7d1f2be764717271f04b6ded52721c82fbac3ed2", "parents": [ "6c4f444bcaa25fa9af7f018ad4e125177942bb5c" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 01 19:35:39 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 19:36:16 2026" }, "message": "Add fp16 arithmetic to SIMD wrappers\n\nPiperOrigin-RevId: 924853891\n" }, { "commit": "6c4f444bcaa25fa9af7f018ad4e125177942bb5c", "tree": "b8265c1113b0340119371792543fdf8f82d1db4e", "parents": [ "fa2641200c9aae744215fa7718fb9362e10cca10" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 01 17:51:45 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 17:52:40 2026" }, "message": "Use `topological_sort` instead of `swap`\n\nUsing `swap` is a premature optimization. One of these required this fix, and I\u0027m doing the other one preemptively just in case.\n\nIf this proves to be a performance issue later, there are other ways to fix this problem.\n\nPiperOrigin-RevId: 924796866\n" }, { "commit": "fa2641200c9aae744215fa7718fb9362e10cca10", "tree": "fd110162fece33a213761056fde5b86cec3c23bc", "parents": [ "06540664d20383800f1b5a58ad05312c9018c67c" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 17:01:15 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 17:01:54 2026" }, "message": "Fix NaN handling\n\nPiperOrigin-RevId: 924768660\n" }, { "commit": "06540664d20383800f1b5a58ad05312c9018c67c", "tree": "b015bb576845ddca586d26a95ec4d4870db45b65", "parents": [ "1aa63db55937ee5599264654fb5678d825abc6d0" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon Jun 01 16:23:15 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 16:23:51 2026" }, "message": "Fix bf16 cast\n\nIt was incorrectly disabled when we thought it was available in avx, but it is only available in avx2.\n\nPiperOrigin-RevId: 924749949\n" }, { "commit": "1aa63db55937ee5599264654fb5678d825abc6d0", "tree": "144581249c7e9c11ce08fd5fd26d1d9713a1cc88", "parents": [ "46394a438c056d90f8f18c3f8775febd8a3b6407" ], "author": { "name": "Misha Gutman", "email": "aelphy@google.com", "time": "Mon Jun 01 15:11:39 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon Jun 01 15:12:23 2026" }, "message": "Fixed flakiness of qd8_f16_qc2w operator level tests.\n\nPiperOrigin-RevId: 924716420\n" }, { "commit": "ab8714dd06f570ec5afe4a1e374c961d725a4d3f", "tree": "e2950d5dfbc393e45a0ed59653e2fe1a80247298", "parents": [ "70889f15d5988a271b1d4d830e83bf92fefc7e0e", "46394a438c056d90f8f18c3f8775febd8a3b6407" ], "author": { "name": "Vaisakh K V", "email": "vaiskv@qti.qualcomm.com", "time": "Mon Jun 01 06:08:53 2026" }, "committer": { "name": "Vaisakh K V", "email": "vaiskv@qti.qualcomm.com", "time": "Mon Jun 01 06:08:53 2026" }, "message": "Merge remote-tracking branch \u0027google/master\u0027 into sme1/qp8-f32-qc4w-gemm\n" }, { "commit": "8d8fa1feeb07649acc5982d7aad59b1c45b98b85", "tree": "106b386e973a0cc3237e349c3d102661904aa79e", "parents": [ "46394a438c056d90f8f18c3f8775febd8a3b6407" ], "author": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Sun May 31 15:45:28 2026" }, "committer": { "name": "aizu-m", "email": "aizumusheer2@gmail.com", "time": "Sun May 31 15:45:28 2026" }, "message": "reject overflowing output dimensions in constant pad reshape\n" }, { "commit": "46394a438c056d90f8f18c3f8775febd8a3b6407", "tree": "c4ef4791760cecac4abba26d0a5bd52d0921a660", "parents": [ "078291a7dc5858fb80ad9f5f7fad721f52b95bd1" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Sat May 30 01:31:51 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sat May 30 01:32:31 2026" }, "message": "Simplified fix for warnings in update-microkernels.py by recursively ignoring subdirectories of ignored roots.\n\nPiperOrigin-RevId: 923698983\n" }, { "commit": "078291a7dc5858fb80ad9f5f7fad721f52b95bd1", "tree": "5dd328c7c3e7931a1430f019a596bf487a08e958", "parents": [ "d34f52c6c7962687092c1ad4ccc7a93693510a92" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 29 23:58:03 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 23:58:47 2026" }, "message": "Fix overzealous assert\n\nPiperOrigin-RevId: 923668444\n" }, { "commit": "d34f52c6c7962687092c1ad4ccc7a93693510a92", "tree": "3cfd679823063e8f77cf06892bd3ebee0cabad8e", "parents": [ "5d756cd83d5462805f5a78436271a18ecf2f7af5" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 29 19:32:42 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 19:33:28 2026" }, "message": "Update KleidiAI in XNNPACK\n\nPiperOrigin-RevId: 923547940\n" }, { "commit": "5d756cd83d5462805f5a78436271a18ecf2f7af5", "tree": "48ec49ebbc944ebae553b8fce64d1fa5910cbe59", "parents": [ "76228ba65ee27cb9bb39805e3acfa8ad413ed4e4" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 19:08:44 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 19:10:36 2026" }, "message": "Add approx_tanh operator support behind YNN_FLAG_FAST_MATH.\n\nPiperOrigin-RevId: 923536167\n" }, { "commit": "76228ba65ee27cb9bb39805e3acfa8ad413ed4e4", "tree": "07b0e7bc8a97f2a490dab608e4dc0a95757ef5b9", "parents": [ "58c0a526396203d347fbebec7983f4121d788104" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 17:17:49 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 17:18:29 2026" }, "message": "Fix NaN handling\n\nPiperOrigin-RevId: 923479160\n" }, { "commit": "58c0a526396203d347fbebec7983f4121d788104", "tree": "d83137853f88bf7b4b2b824f4c8fc8a7196a9de0", "parents": [ "d89bf2f383ea5005b4dee58d45079ae5c9a22ef8" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Fri May 29 15:45:24 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 15:46:56 2026" }, "message": "Remove RTTI from the Tensor API. Refactor operation handling in the graph.\n\nRemove virtual inheritance from the operation implementations and stop using\nside-casting using `dynamic_cast` to find out if an operation is implemented by\na backend.\n\nAdd a list of \"graph::BackendExtension\" elements that can be grafted onto an\noperation. Backends implementation of every operation are now implemented by\nderiving from this new class. When lowering to a specific backend, we go\nthrough the list of extensions attached to an operation to find a matching\nbackend.\n\nBackend identification is done by calling `BackendExtension::GetTypeId()`.\n\nAs a result of the above changes, operation node implementations don\u0027t need to\nbe defined as templates anymore.\n\nPiperOrigin-RevId: 923435766\n" }, { "commit": "d89bf2f383ea5005b4dee58d45079ae5c9a22ef8", "tree": "acad2c17b5203f41ab9b26f672fa4a7a00451958", "parents": [ "a74b04841bd6b751833e0eab805b079b757d1a07" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 29 00:33:25 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 29 00:34:07 2026" }, "message": "Don\u0027t allow broadcasts as the first dimension of dot inputs\n\nThis change requires that the stride of the first dimension of dot inputs is 1 element, which forces broadcasts to be realized into memory before running the dot.\n\nBroadcasting is a weird thing to do to a dot input, but some tests do it, and it would be easiest to just handle it than find a workaround.\n\nPiperOrigin-RevId: 923066551\n" }, { "commit": "a74b04841bd6b751833e0eab805b079b757d1a07", "tree": "f021cd041631b5f6875b2b7baceeca4b26ec808e", "parents": [ "23ba0fbd097cafee7aa569b8daaa64f5918bdbf8" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 22:05:46 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 22:06:31 2026" }, "message": "Relax tolerance of sum reduce test\n\nShould address this rare flaky failure: https://github.com/google/XNNPACK/actions/runs/26599479177/job/78379097239\n\nPiperOrigin-RevId: 922996166\n" }, { "commit": "23ba0fbd097cafee7aa569b8daaa64f5918bdbf8", "tree": "faaf91cbb240e6c491d44748d0e27d4dd9467d5a", "parents": [ "0702200b644be75cd569ec5a750c93934f6f39db" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 21:03:38 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 21:04:31 2026" }, "message": "Remove `force_root` from `static_transpose` scheduling\n\nPiperOrigin-RevId: 922963746\n" }, { "commit": "0702200b644be75cd569ec5a750c93934f6f39db", "tree": "9c253947b1516efd4550b5bbc0ec4c1eca7260be", "parents": [ "bf8d96be1971a3a8a7b3c103d32098682ab20919" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 20:35:42 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 20:36:36 2026" }, "message": "Enable `YNN_FLAG_FAST_MATH` in XNNPACK compatibility layer\n\nPiperOrigin-RevId: 922948315\n" }, { "commit": "bf8d96be1971a3a8a7b3c103d32098682ab20919", "tree": "87a51896b7b0c4ccfce8aa166b7d63633ae19c3e", "parents": [ "1eb730046b14f68b8a51cf0f6e3c9d11964ca869" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 20:02:42 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 20:03:34 2026" }, "message": "Add YNN_FLAG_FAST_MATH and approx_erf operator support behind this flag.\n\nThis change introduces the `YNN_FLAG_FAST_MATH` flag and the corresponding `approx_erf` operator to restore performance for classic use cases that do not require high accuracy.\n\nPiperOrigin-RevId: 922930124\n" }, { "commit": "1eb730046b14f68b8a51cf0f6e3c9d11964ca869", "tree": "81111a1245a982abbae70761b3fceeb2b2632256", "parents": [ "9b4a49f0f0943f821bfe30ab9dfb91b10b1454f1" ], "author": { "name": "Marie White", "email": "mariewhite@google.com", "time": "Thu May 28 14:56:19 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 14:57:30 2026" }, "message": "Do not create serial loops for k2, k3, ...\n\nIR before:\n```\nk2#0 \u003d loop(serial, [0, 268], 1) {\n reduction \u003d crop_dim(reduction, 2, [k2#0, k2#0]) {\n k1#0 \u003d loop(serial, [0, 3], 1) {\n reduction \u003d crop_dim(reduction, 1, [k1#0, k1#0]) {\n d0#0 \u003d loop(parallel, [buffer_min(out3, 0), buffer_max(out3, 0)], 128) {\n closure {in0, in2, out3, v4, reduction, d0#0} in {\n out3.d0#0 \u003d crop_dim(out3, 0, [d0#0, (d0#0 + 127)]) {\n d1#0 \u003d loop(parallel, [buffer_min(out3.d0#0, 1), buffer_max(out3.d0#0, 1)], 32) {\n closure {in0, in2, v4, reduction, d1#0, out3.d0#0} in {\n out3.d1#0 \u003d crop_dim(out3.d0#0, 1, [d1#0, (d1#0 + 31)]) {\n call(dot num_k_dims\u003d3, {in0, v4, in2}, {out3.d1#0, reduction}, {})\n```\n\nIR After:\n```\n call(pack_b, {in1}, {v4}, {})\n reduction \u003d allocate(automatic, 0, {\n {[0, 0], 0, \u003c\u003e},\n {[0, 1], 0, \u003c\u003e},\n {[0, 266], 0, \u003c\u003e}\n }) {\n call(dot num_k_dims\u003d3, {in0, v4, const2}, {out3, reduction}, {})\n }\n```\n\nPiperOrigin-RevId: 922773683\n" }, { "commit": "9b4a49f0f0943f821bfe30ab9dfb91b10b1454f1", "tree": "ac8da299b337af2be513e60d8abd13b93af13b41", "parents": [ "05c9b9157a078f13cd6980a067cc6c7a2d72756f" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Thu May 28 10:55:20 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 10:56:01 2026" }, "message": "Fix warnings in update-microkernels.py by recursively ignoring subdirectories of ignored roots.\n\nPiperOrigin-RevId: 922678462\n" }, { "commit": "05c9b9157a078f13cd6980a067cc6c7a2d72756f", "tree": "0acec9c7231b5c54ca2184abeaf96be99cbee685", "parents": [ "e2ab35abf6a5107fa7603184d5f93eb806281742" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Thu May 28 09:13:48 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 09:18:49 2026" }, "message": "Remove fxdiv usages from XNNPACK, keeping it only for pthreadpool\n\nPiperOrigin-RevId: 922638743\n" }, { "commit": "e2ab35abf6a5107fa7603184d5f93eb806281742", "tree": "48ed517b9bfb177dc4b97b39c2ea7801670da40b", "parents": [ "5004f85c86eb6097b2650afabc6b59aa8d8e19be", "393da7ddec358bcaa2a1ec1b8e50634a010db4a9" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 09:14:31 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 09:14:31 2026" }, "message": "Merge pull request #10242 from ken-unger:f16-vlog-rvv\n\nPiperOrigin-RevId: 922638609\n" }, { "commit": "5004f85c86eb6097b2650afabc6b59aa8d8e19be", "tree": "caec45224d7abe30adc69c02a8451c86a8515c88", "parents": [ "01d254db8ac85612afe8c6721ab7a32b1488b6ab" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 02:07:36 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 02:09:42 2026" }, "message": "Rewrite `transpose(static_broadcast(x))` \u003d\u003e `static_broadcast(transpose(x))`\n\nPiperOrigin-RevId: 922472494\n" }, { "commit": "01d254db8ac85612afe8c6721ab7a32b1488b6ab", "tree": "a92b40b638f046161812531462abb79047fda529", "parents": [ "5fc47cc7a4fd0b699c8f1da58a4eb8ff7358e58f" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 01:46:42 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 01:47:19 2026" }, "message": "Remove `broadcast` op implementation\n\n`broadcast` can be represented by a transpose that removes a dimension, and adds a new dimension in its place.\n\nPiperOrigin-RevId: 922464606\n" }, { "commit": "5fc47cc7a4fd0b699c8f1da58a4eb8ff7358e58f", "tree": "dbe26d20282264fc4825e1a2817ad11e169cf361", "parents": [ "f9f2c229991dfaf187606ca5657e7ea3d0838947" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 28 01:31:22 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 28 01:32:55 2026" }, "message": "Fuse sequences of transpose(transpose(x)) into one transpose(x)\n\nPiperOrigin-RevId: 922459271\n" }, { "commit": "f9f2c229991dfaf187606ca5657e7ea3d0838947", "tree": "684b72fd149e13ff82f6e1fcaf7fcea7494a59af", "parents": [ "f1ab4551669c5552f095dadf1a46881594ec9589" ], "author": { "name": "Marie White", "email": "mariewhite@google.com", "time": "Wed May 27 22:35:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 22:36:07 2026" }, "message": "Do not rely on tile_k when aligning split_k\n\ntile_k is dependent on the microkernel which depends on the CPU, which breaks numerical consistency across CPUs.\n\nPiperOrigin-RevId: 922383499\n" } ], "next": "f1ab4551669c5552f095dadf1a46881594ec9589" }