)]}' { "log": [ { "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" }, { "commit": "f1ab4551669c5552f095dadf1a46881594ec9589", "tree": "89b9fdf84f6116dad26afcda32a83b003f69391e", "parents": [ "d1da9a531ba33bfda32b5dc558460f42c98ed734" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 27 21:16:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 21:17:28 2026" }, "message": "Implement `static_expand_dims` using `static_transpose`\n\nFewer ops to support is nice, and this op is easier to simplify (see reshape -\u003e transpose rewrite for example).\n\nPiperOrigin-RevId: 922341041\n" }, { "commit": "d1da9a531ba33bfda32b5dc558460f42c98ed734", "tree": "a0929b4398b68668ba6eeed67b8bf951fa22aac5", "parents": [ "38eb8abda8a7a56c7828b51df92329ef06c49050" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 27 15:30:30 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 15:31:38 2026" }, "message": "Add transcendental ops for every x86 architecture\n\nThe current selection of which ISAs have kernels is now pretty stale. The first step in cleaning this up is to just add all of the kernels to every ISA.\n\nPiperOrigin-RevId: 922148772\n" }, { "commit": "38eb8abda8a7a56c7828b51df92329ef06c49050", "tree": "b9f1a282d79635c8c27b4351a0d945189601439c", "parents": [ "ac73c5b87399a1845a43368e109d9e1d4defe5c0" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed May 27 15:27:17 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 15:28:17 2026" }, "message": "Remove RTTI from the Tensor API. Rework the `Quantization` hierarchy.\n\nInstances of `dynamic_cast` should be replaced with calls to\n`Quantization::As\u003cT\u003e()` to safely downcast to a concrete quantization type.\n\n`GetTypeId()` and `IsA()` have been implemented for the existing quantization\ntypes and let us find out the concrete type of a quantization object when it\u0027s\nkept as part of the graph and are used by `Quantization::As\u003cT\u003e()` to safely\ndowncast from a buffer.\n\nPiperOrigin-RevId: 922147064\n" }, { "commit": "ac73c5b87399a1845a43368e109d9e1d4defe5c0", "tree": "505b1b6bb6cdacd95f7eb3603ffb5b2cf9267e35", "parents": [ "43118f4fc34456a392a06a635ac5d85ccca2f874" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed May 27 14:59:19 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 15:00:37 2026" }, "message": "Remove RTTI from the Tensor API. Rework the `Buffer` hierarchy.\n\nThis change removes virtual inheritance from the `Buffer` hierarchy.\n`MutableBuffer` is removed and `LockMutable()` is merged into `Buffer`.\n\nInstances of `dynamic_cast` on buffers should now use `Buffer::As\u003cT\u003e()`. This\nis also safer as it forces users to check that the cast was performed\nsuccessfully.\n\n`GetTypeId()` and `IsA()` have been implemented for the existing buffer types\nand let us retrieve the concrete type of a buffer when it\u0027s kept as part of the\ngraph and are used by `Buffer::As\u003cT\u003e()` to safely downcast from a buffer.\n\nPiperOrigin-RevId: 922132824\n" }, { "commit": "43118f4fc34456a392a06a635ac5d85ccca2f874", "tree": "d9b6a0419604ad8f2cdf182ab9cd4c3b62bc6228", "parents": [ "30e1a98b4c6a32afa28641758a9f53777b2129c5" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Wed May 27 14:36:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 14:36:51 2026" }, "message": "Remove RTTI from the Tensor API. Introduce `TypeId` class.\n\n`TypeId` will be used to identify the concrete type of type erased elements in\nthe Tensor API graph.\n\nThis change is part of a series of changes that aim to make the Tensor API\ncompile when RTTI is disabled.\n\nPiperOrigin-RevId: 922122531\n" }, { "commit": "30e1a98b4c6a32afa28641758a9f53777b2129c5", "tree": "9801c06bfc60efb3e6a6a98a0b0db0ac2a90c562", "parents": [ "f9ee799919bd14b18677dfce901285e3c3b84297" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed May 27 02:58:38 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 02:59:31 2026" }, "message": "f16-vtanh using high-accuracy rational polynomial implementation.\n\nPiperOrigin-RevId: 921833633\n" }, { "commit": "f9ee799919bd14b18677dfce901285e3c3b84297", "tree": "bcd3ec5b804a4ec1266c9e4e7876dbd5ffd9bdfc", "parents": [ "134e8ef6cd903f75f6ce7dcfcfddc8e5c271d82e" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Wed May 27 02:31:06 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 02:31:48 2026" }, "message": "Add indirection_bench to test performance of indirection init\n\nPiperOrigin-RevId: 921822572\n" }, { "commit": "134e8ef6cd903f75f6ce7dcfcfddc8e5c271d82e", "tree": "78f0b9813677c69d4f2fa5d196cb7f0c0b26e740", "parents": [ "dddad07f84631ef6c8fe8ecb5e879c866d3538bb" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 27 01:36:15 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 01:37:11 2026" }, "message": "Fix test timeouts on emulators\n\nThe pi_test does pretty big reductions, not suitable for use on emulators.\n\nPiperOrigin-RevId: 921803185\n" }, { "commit": "dddad07f84631ef6c8fe8ecb5e879c866d3538bb", "tree": "499dd1c886c46ab6ec188c70587e5ccd92c83b1a", "parents": [ "4d837f20dfff9177c579e160cad339172c0843e5" ], "author": { "name": "Marie White", "email": "mariewhite@google.com", "time": "Wed May 27 00:02:54 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 27 00:03:36 2026" }, "message": "Split dot operation on K.\n\nCurrently split_k is the length of k so we don\u0027t expect major performance differences.\n\nSample IR for `f32[110,10240,2560]`.\n\nBefore:\n```\nnull \u003d constant_buffer([], 0, {}) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[0, 10239], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n out2.d0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, out2.d0} in {\n v3.out2.d0 \u003d crop_dim(v3, 3, [(out2.d0 / 16), ((out2.d0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.out2.d0}, {})\n }\n out2.out2.d0 \u003d crop_dim(out2, 0, [out2.d0, (out2.d0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.out2.d0}, {})\n }\n }\n }\n }\n}\n```\n\nAfter where k_split\u003d1024.\n```\nnull \u003d constant_buffer([], 0, {}) {\n reduction \u003d allocate(automatic, 0, {\n {[0, 10239], 0, \u003c\u003e}\n }) {\n k0#0 \u003d loop(serial, [0, 10239], 1024) {\n reduction \u003d crop_dim(reduction, 0, [k0#0, (k0#0 + 1023)]) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[k0#0, buffer_max(reduction, 0)], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n d0#0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, reduction, d0#0} in {\n v3.d0#0 \u003d crop_dim(v3, 3, [(d0#0 / 16), ((d0#0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.d0#0}, {})\n }\n out2.d0#0 \u003d crop_dim(out2, 0, [d0#0, (d0#0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.d0#0, reduction}, {})\n }\n }\n }\n }\n }\n }\n }\n}\n```\nPiperOrigin-RevId: 921767627\n" }, { "commit": "4d837f20dfff9177c579e160cad339172c0843e5", "tree": "16ecfc903e02871f4fd37c79196fd913d05f63dc", "parents": [ "acb00f5ed496bc2b39f3e7918fe0baae32e3d006" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 23:14:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 23:14:52 2026" }, "message": "Fix attempts to use AVX2 instructions on non-AVX2 targets\n\nPiperOrigin-RevId: 921744427\n" }, { "commit": "acb00f5ed496bc2b39f3e7918fe0baae32e3d006", "tree": "f94af980da403919776833fb5cbd6d2f559354e9", "parents": [ "e60eb9dcc68a46684454cbb4116408c63fe8e8d1" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 21:25:41 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 21:28:30 2026" }, "message": "Remove tile size from kernel function name\n\nThis will make it easier to compare performance when the tile size changes\n\nPiperOrigin-RevId: 921682098\n" }, { "commit": "e60eb9dcc68a46684454cbb4116408c63fe8e8d1", "tree": "e095ea2c5aae49c2e4af4a6e21ea5b3acf18c50e", "parents": [ "6cb1a2d35e0f148ff0d9f48ae641af81ec2e6982" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 21:24:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 21:25:05 2026" }, "message": "Add missing build of arm_neonfma benchmarks\n\nPiperOrigin-RevId: 921681198\n" }, { "commit": "6cb1a2d35e0f148ff0d9f48ae641af81ec2e6982", "tree": "6e89c92e4004dd66a59579e13606ab51de68dec0", "parents": [ "8ea1945e5d99cff0b214db889548bffb567606f5" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Tue May 26 20:45:49 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 20:46:40 2026" }, "message": "Add XNN_ENABLE_RNDNU16 build flag and conditionally use rndnu16 kernels\n\nPiperOrigin-RevId: 921659833\n" }, { "commit": "8ea1945e5d99cff0b214db889548bffb567606f5", "tree": "37a383dffdab1febbb4aea445ebd5cc826cbb2e3", "parents": [ "cc4daec108add3fbc2d820160627f687489fbfd4" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 19:02:54 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 19:03:33 2026" }, "message": "`tanh` accuracy improvements\n\n- Use `expm1(2x) / (expm1(2x) + 2)` for better accuracy (2 ULPs vs. 5)\n- Keep the old direct polynomial approximation as `approx_tanh(x)`\n- Add a small x region to `approx_tanh(x)` to avoid higher error for denormals.\n\nThis is a significant performance regression due to the first change:\n```\nname time/op time/op vs base\nbench_reference/tanh_float/m:1/n:4096/real_time 42.66µ ± 10% 43.11µ ± 2% ~ (p\u003d0.818 n\u003d6)\nbench_reference/tanh_double/m:1/n:4096/real_time 110.1µ ± 44% 107.6µ ± 6% ~ (p\u003d0.589 n\u003d6)\nbench/tanh_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.381µ ± 30% 2.925µ ± 4% +111.75% (p\u003d0.002 n\u003d6)\nbench/tanh_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7.696µ ± 13% 7.448µ ± 7% ~ (p\u003d0.485 n\u003d6)\nbench/tanh_fp32_1x32_x86_avx/m:1/n:4096/real_time 2.879µ ± 6% 7.561µ ± 3% +162.59% (p\u003d0.002 n\u003d6)\nbench/tanh_fp64_1x16_x86_avx/m:1/n:4096/real_time 19.26µ ± 14% 19.39µ ± 3% ~ (p\u003d1.000 n\u003d6)\nbench/tanh_fp32_1x16_x86_sse2/m:1/n:4096/real_time 5.813µ ± 8% 10.910µ ± 3% +87.68% (p\u003d0.002 n\u003d6)\nbench/tanh_fp64_1x8_x86_sse2/m:1/n:4096/real_time 28.26µ ± 4% 28.14µ ± 8% ~ (p\u003d0.937 n\u003d6)\nbench/tanh_fp32_1x16_x86_sse2_fma/m:1/n:4096/real_time 86.54µ ± 2% 116.29µ ± 10% +34.38% (p\u003d0.002 n\u003d6)\ngeomean 15.04µ 20.07µ +33.47%\n```\n\nPiperOrigin-RevId: 921603838\n" }, { "commit": "cc4daec108add3fbc2d820160627f687489fbfd4", "tree": "d6f6f88366d3c5bbab06633509c55e160a1add51", "parents": [ "35ba08cb5823236db61f2f326e49e88104349233" ], "author": { "name": "Gerardo Carranza", "email": "gcarranza@google.com", "time": "Tue May 26 17:39:47 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 17:40:45 2026" }, "message": "Migrate LiteRT ATS unary op graph generation to use litert::tensor API.\n\nPiperOrigin-RevId: 921553787\n" }, { "commit": "35ba08cb5823236db61f2f326e49e88104349233", "tree": "cc967b10eda8ee70f8cdbf36fd7b1a4669ad59e1", "parents": [ "f8deca0bb9e9f2312419fab30f601c8f070763d1" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 16:08:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 16:11:13 2026" }, "message": "Add `tanh` SIMD wrappers\n\nTo avoid a significant performance regression, this change includes another change to carefully order the `min` and `max` operands (assuming that the scalar is not NaN), so we don\u0027t need special cases for `NaN`.\n\nBizarrely, the elimination of the `NaN` special case caused a significant regression for the `erf_fp32` kernel only for x86_fma3. This is just a compiler quirk (since this change strictly eliminates instructions vs. baseline in this kernel), I worked around it as much as possible by tweaking the unrolling factor for this kernel.\n\nThis is also the last usage of the polynomial arithmetic in the elementwise op compiler, so the helpers for those is removed.\n\n```\nname time/op time/op vs base\nbench/erf_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.111µ ± 13% 4.669µ ± 3% -8.65% (p\u003d0.002 n\u003d6)\nbench/tanh_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.374µ ± 6% 1.358µ ± 2% -1.17% (p\u003d0.026 n\u003d6)\nbench/erf_fp32_1x16_x86_fma3/m:1/n:4096/real_time 8.789µ ± 2%\nbench/erf_fp64_1x8_x86_fma3/m:1/n:4096/real_time 26.33µ ± 6% 24.40µ ± 3% -7.33% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x64_x86_avx/m:1/n:4096/real_time 13.91µ ± 4% 13.40µ ± 1% -3.62% (p\u003d0.002 n\u003d6)\nbench/erf_fp64_1x8_x86_avx/m:1/n:4096/real_time 40.67µ ± 5% 39.23µ ± 4% -3.54% (p\u003d0.009 n\u003d6)\nbench/tanh_fp32_1x32_x86_avx/m:1/n:4096/real_time 2.736µ ± 2% 2.910µ ± 3% +6.37% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 9.433µ ± 1% 9.862µ ± 3% +4.55% (p\u003d0.002 n\u003d6)\nbench/tanh_fp32_1x16_x86_sse2/m:1/n:4096/real_time 5.433µ ± 4% 5.887µ ± 2% +8.36% (p\u003d0.002 n\u003d6)\nbench/tanh_fp32_1x16_x86_sse2_fma/m:1/n:4096/real_time 66.59µ ± 19% 86.32µ ± 3% +29.63% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x32_x86_fma3/m:1/n:4096/real_time 8.646µ ± 4%\ngeomean 14.67µ 14.68µ +0.14% ¹\n```\n\nPiperOrigin-RevId: 921503300\n" }, { "commit": "f8deca0bb9e9f2312419fab30f601c8f070763d1", "tree": "a04848963110dc3f5f683a0ff16620c58f15b2e4", "parents": [ "f4ed055f9166daad591d82346d5ff9dc53d8c669" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 26 00:31:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 26 00:32:50 2026" }, "message": "Replace rational polynomials for exp with non-rational polynomials\n\nThe rational polynomials provided some ILP, which is now gone. To avoid a performance regression due to this, this change also adds an overload of `eval_polynomial` that uses x^2 to bring back some ILP.\n\nThis change is a bit of a tradeoff:\n- It\u0027s faster on AVX512, by a lot\n- It\u0027s a mixed bag on other ISAs, including a regression on ARM64 for fp32\n\nI think we should avoid using different algorithms on different targets to maintain numerical consistency, so for now, I think this change is the right tradeoff.\n\n```\nname cpu/op cpu/op vs base\nbench_reference/exp_float/m:1/n:4096/real_time 35.25µ ± 5% 35.35µ ± 3% ~ (p\u003d0.818 n\u003d6)\nbench_reference/exp_double/m:1/n:4096/real_time 72.89µ ± 2% 72.47µ ± 1% ~ (p\u003d0.240 n\u003d6)\nbench_reference/expm1_float/m:1/n:4096/real_time 38.08µ ± 1% 38.09µ ± 5% ~ (p\u003d1.000 n\u003d6)\nbench_reference/expm1_double/m:1/n:4096/real_time 86.36µ ± 1% 86.05µ ± 1% -0.35% (p\u003d0.026 n\u003d6)\nbench_reference/erf_float/m:1/n:4096/real_time 36.07µ ± 7% 35.90µ ± 3% ~ (p\u003d0.937 n\u003d6)\nbench_reference/erf_double/m:1/n:4096/real_time 41.36µ ± 3% 41.47µ ± 3% ~ (p\u003d0.937 n\u003d6)\nbench_reference/sigmoid_float/m:1/n:4096/real_time 37.59µ ± 1% 37.63µ ± 3% ~ (p\u003d0.485 n\u003d6)\nbench_reference/sigmoid_double/m:1/n:4096/real_time 73.26µ ± 2% 73.16µ ± 2% ~ (p\u003d0.937 n\u003d6)\nbench/erf_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.131µ ± 4% 4.796µ ± 2% -6.51% (p\u003d0.002 n\u003d6)\nbench/erf_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 25.41µ ± 2% 18.74µ ± 3% -26.22% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2.280µ ± 1% 1.943µ ± 1% -14.79% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.779µ ± 2% 4.592µ ± 1% -20.54% (p\u003d0.002 n\u003d6)\nbench/expm1_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2.389µ ± 1% 2.018µ ± 1% -15.53% (p\u003d0.002 n\u003d6)\nbench/expm1_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.925µ ± 1% 4.828µ ± 1% -18.51% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2.902µ ± 4% 2.397µ ± 2% -17.39% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7.928µ ± 2% 5.794µ ± 3% -26.91% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 3.107µ ± 2% 3.070µ ± 2% ~ (p\u003d0.132 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 9.315µ ± 3% 8.371µ ± 2% -10.13% (p\u003d0.002 n\u003d6)\nbench/expm1_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 3.241µ ± 4% 3.182µ ± 5% ~ (p\u003d0.485 n\u003d6)\nbench/expm1_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 9.491µ ± 4% 9.278µ ± 4% -2.25% (p\u003d0.015 n\u003d6)\nbench/erf_fp32_1x16_x86_fma3/m:1/n:4096/real_time 8.597µ ± 3% 8.737µ ± 5% ~ (p\u003d0.180 n\u003d6)\nbench/erf_fp64_1x8_x86_fma3/m:1/n:4096/real_time 27.59µ ± 3% 25.83µ ± 24% ~ (p\u003d0.065 n\u003d6)\nbench/exp_fp32_1x16_x86_avx2/m:1/n:4096/real_time 5.488µ ± 5% 5.637µ ± 10% ~ (p\u003d0.065 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 16.70µ ± 9% 15.40µ ± 4% -7.80% (p\u003d0.002 n\u003d6)\nbench/expm1_fp32_1x16_x86_avx2/m:1/n:4096/real_time 5.596µ ± 3% 5.655µ ± 2% ~ (p\u003d0.240 n\u003d6)\nbench/expm1_fp64_1x16_x86_avx2/m:1/n:4096/real_time 17.13µ ± 5% 15.98µ ± 5% -6.72% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time 5.623µ ± 5% 5.715µ ± 19% +1.63% (p\u003d0.041 n\u003d6)\nbench/sigmoid_fp64_1x8_x86_avx2/m:1/n:4096/real_time 17.16µ ± 2% 15.28µ ± 3% -10.98% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x64_x86_avx/m:1/n:4096/real_time 13.59µ ± 4% 14.00µ ± 4% ~ (p\u003d0.093 n\u003d6)\nbench/erf_fp64_1x8_x86_avx/m:1/n:4096/real_time 38.19µ ± 6% 40.27µ ± 2% +5.46% (p\u003d0.041 n\u003d6)\nbench/erf_fp32_1x16_x86_sse2/m:1/n:4096/real_time 21.88µ ± 2% 23.86µ ± 2% +9.02% (p\u003d0.002 n\u003d6)\nbench/erf_fp64_1x16_x86_sse2/m:1/n:4096/real_time 54.22µ ± 3% 54.27µ ± 1% ~ (p\u003d0.818 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 7.435µ ± 2% 9.585µ ± 4% +28.92% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 25.48µ ± 1% 25.22µ ± 3% ~ (p\u003d0.093 n\u003d6)\nbench/expm1_fp32_1x16_x86_sse2/m:1/n:4096/real_time 7.561µ ± 1% 9.819µ ± 3% +29.86% (p\u003d0.002 n\u003d6)\nbench/expm1_fp64_1x8_x86_sse2/m:1/n:4096/real_time 25.73µ ± 7% 26.00µ ± 2% ~ (p\u003d0.485 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time 9.060µ ± 4% 10.162µ ± 2% +12.16% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp64_1x8_x86_sse2/m:1/n:4096/real_time 26.36µ ± 2% 26.12µ ± 1% ~ (p\u003d0.310 n\u003d6)\nbench/erf_fp32_1x16_x86_sse2_fma/m:1/n:4096/real_time 271.4µ ± 1% 283.7µ ± 1% +4.54% (p\u003d0.002 n\u003d6)\nbench/expm1_fp32_1x16_x86_sse2_fma/m:1/n:4096/real_time 92.40µ ± 2% 105.19µ ± 2% +13.85% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2_fma/m:1/n:4096/real_time 94.12µ ± 3% 116.66µ ± 2% +23.94% (p\u003d0.002 n\u003d6)\ngeomean 16.25µ 15.90µ -2.15%\n```\n```\nname time/op time/op vs base\nbench_reference/exp_float/m:1/n:4096/real_time 5.818µ ± 8% 5.888µ ± 12% ~ (p\u003d0.485 n\u003d6)\nbench_reference/exp_double/m:1/n:4096/real_time 6.939µ ± 6% 6.881µ ± 10% ~ (p\u003d0.818 n\u003d6)\nbench_reference/expm1_float/m:1/n:4096/real_time 18.98µ ± 2% 18.15µ ± 11% ~ (p\u003d0.065 n\u003d6)\nbench_reference/expm1_double/m:1/n:4096/real_time 19.17µ ± 4% 19.76µ ± 5% +3.09% (p\u003d0.026 n\u003d6)\nbench_reference/erf_float/m:1/n:4096/real_time 15.06µ ± 5% 14.97µ ± 7% ~ (p\u003d0.818 n\u003d6)\nbench_reference/erf_double/m:1/n:4096/real_time 16.75µ ± 6% 16.77µ ± 8% ~ (p\u003d0.818 n\u003d6)\nbench_reference/sigmoid_float/m:1/n:4096/real_time 6.159µ ± 6% 6.239µ ± 10% ~ (p\u003d0.485 n\u003d6)\nbench_reference/sigmoid_double/m:1/n:4096/real_time 7.192µ ± 5% 7.159µ ± 8% ~ (p\u003d0.818 n\u003d6)\nbench/erf_fp64_1x8_neon/m:1/n:4096/real_time 14.53µ ± 5% 13.31µ ± 5% -8.44% (p\u003d0.004 n\u003d6)\nbench/exp_fp64_1x8_neon/m:1/n:4096/real_time 4.920µ ± 3% 4.210µ ± 5% -14.43% (p\u003d0.002 n\u003d6)\nbench/expm1_fp64_1x8_neon/m:1/n:4096/real_time 5.122µ ± 3% 4.290µ ± 5% -16.25% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp64_1x4_neon/m:1/n:4096/real_time 5.474µ ± 3% 4.973µ ± 4% -9.16% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x16_neon/m:1/n:4096/real_time 6.269µ ± 3% 6.470µ ± 8% ~ (p\u003d0.065 n\u003d6)\nbench/exp_fp32_1x16_neon/m:1/n:4096/real_time 2.388µ ± 2% 2.704µ ± 13% +13.23% (p\u003d0.041 n\u003d6)\nbench/expm1_fp32_1x16_neon/m:1/n:4096/real_time 2.454µ ± 3% 2.611µ ± 9% ~ (p\u003d0.132 n\u003d6)\nbench/sigmoid_fp32_1x8_neon/m:1/n:4096/real_time 2.643µ ± 2% 3.473µ ± 58% +31.39% (p\u003d0.002 n\u003d6)\ngeomean 6.983µ 6.968µ -0.21%\n```\n\nPiperOrigin-RevId: 921140178\n" }, { "commit": "f4ed055f9166daad591d82346d5ff9dc53d8c669", "tree": "d2dcd2e92f4c6e542123cac6bb67ddf716470048", "parents": [ "41fde00824707ba4d4507fdbfc3bb70f9862b884" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 25 21:47:23 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 25 21:48:11 2026" }, "message": "Add `erf` SIMD math functions\n\nThese are the toughest functions I\u0027ve tried to approximate so far, and thus are the most complicated.\n\nFor float32, there are two approximations:\n- `fast_erf`, which is a single rational polynomial, accurate to 5 ULPs\n- `erf`, which is 3 piecewise (non-rational) polynomials, the first of which is a direct approximation of erf, and the next two are approximating the \"scaled\" complementary error function, accurate to 2 ULPs, and is ~3x slower.\n\nFor float64, `erf` is two rational polynomials, the second of which uses the scaled complementary error function like the float32 approximation.\n\nThis is a significant regression in performance, but a necessary improvement in correctness:\n```\nname time/op time/op vs base\nbench_reference/erf_float/m:1/n:4096/real_time 35.71µ ± 16% 36.26µ ± 2% ~ (p\u003d0.394 n\u003d6)\nbench_reference/erf_double/m:1/n:4096/real_time 42.67µ ± 40% 42.31µ ± 4% ~ (p\u003d0.310 n\u003d6)\nbench/erf_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.626µ ± 23% 5.230µ ± 8% +221.76% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x16_x86_fma3/m:1/n:4096/real_time 2.429µ ± 3% 8.561µ ± 3% +252.48% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x64_x86_avx/m:1/n:4096/real_time 3.491µ ± 3% 13.446µ ± 3% +285.21% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x16_x86_sse2/m:1/n:4096/real_time 6.997µ ± 4% 22.067µ ± 1% +215.39% (p\u003d0.002 n\u003d6)\nbench/erf_fp32_1x16_x86_sse2_fma/m:1/n:4096/real_time 83.35µ ± 3% 270.31µ ± 1% +224.29% (p\u003d0.002 n\u003d6)\nbench/erf_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 25.49µ ± 1%\nbench/erf_fp64_1x8_x86_fma3/m:1/n:4096/real_time 27.89µ ± 2%\nbench/erf_fp64_1x8_x86_avx/m:1/n:4096/real_time 38.04µ ± 3%\nbench/erf_fp64_1x16_x86_sse2/m:1/n:4096/real_time 55.07µ ± 2%\ngeomean 10.29µ 27.97µ +139.34%\n```\n\nPiperOrigin-RevId: 921097089\n" }, { "commit": "41fde00824707ba4d4507fdbfc3bb70f9862b884", "tree": "db3d2353f9ea2afc14b9a7b62c9adf7ba481bcb9", "parents": [ "2b99af3f109a3802ae63bb3d8883eaccec341c90" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 25 21:26:15 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 25 21:27:07 2026" }, "message": "Improve exp approximation\n\nThe basic idea is to pull a few terms from the Taylor series of `exp(x)` out of the polynomial approximation. This enables tightening the error bounds further, `expm1` tolerances are reduced from 3 to 2 ULPs.\n\nWhile I think `exp` could be reduced from 2 to 1 ULP, it fails because of the one single (output) value that is 2 ULPs away from infinity. We produce infinity, while `std::exp` produces a finite value there. I don\u0027t think this is worth a performance cost to fix.\n\nPiperOrigin-RevId: 921092139\n" }, { "commit": "2b99af3f109a3802ae63bb3d8883eaccec341c90", "tree": "6124513e0746d63a22a448f21c1e18acfa65044a", "parents": [ "894ae65353c98e0256112ca6f0ae69f0d472a697" ], "author": { "name": "Marie White", "email": "mariewhite@google.com", "time": "Mon May 25 07:13:22 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 25 07:14:11 2026" }, "message": "Add clarifying comments in call to define_transpose_a().\n\nPiperOrigin-RevId: 920809733\n" }, { "commit": "894ae65353c98e0256112ca6f0ae69f0d472a697", "tree": "efc028a206c8f1501239620aa7786d76ffc7876a", "parents": [ "eafd6fe46bbb93ed46c71fc56796f1fbb82bf0db" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Sun May 24 23:36:56 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sun May 24 23:38:06 2026" }, "message": "Fix `floor_log2(NaN)` to be `NaN`\n\nThis is a performance regression, but seems necessary for correctness:\n```\nname time/op time/op vs base\nBM_floor_log2_f32x4_sse2 2.290n ± 2% 2.292n ± 1% ~ (p\u003d0.937 n\u003d6)\nBM_floor_log2_f64x2_sse2 1.915n ± 2% 2.303n ± 2% +20.32% (p\u003d0.002 n\u003d6)\nBM_log_f32x4_sse2 6.999n ± 2% 7.303n ± 2% +4.34% (p\u003d0.002 n\u003d6)\nBM_log_f64x2_sse2 11.62n ± 3% 12.47n ± 3% +7.35% (p\u003d0.002 n\u003d6)\nBM_log1p_f32x4_sse2 9.873n ± 3% 10.439n ± 4% +5.73% (p\u003d0.006 n\u003d6)\nBM_log1p_f64x2_sse2 14.63n ± 3% 15.96n ± 4% +9.10% (p\u003d0.002 n\u003d6)\ngeomean 6.099n 6.566n +7.65%\n```\n\nAVX512 is unaffected:\n```\nname time/op time/op vs base\nBM_floor_log2_f32x16_avx512 1.039n ± 3% 1.034n ± 4% ~ (p\u003d1.000 n\u003d6)\nBM_floor_log2_f32x8_avx512 0.6460n ± 2% 0.6450n ± 2% ~ (p\u003d0.907 n\u003d6)\nBM_floor_log2_f32x4_avx512 0.5720n ± 2% 0.5755n ± 2% ~ (p\u003d0.162 n\u003d6)\nBM_floor_log2_f64x8_avx512 1.030n ± 3% 1.043n ± 2% ~ (p\u003d0.329 n\u003d6)\nBM_floor_log2_f64x4_avx512 0.8695n ± 21% 0.8715n ± 3% ~ (p\u003d0.937 n\u003d6)\nBM_floor_log2_f64x2_avx512 0.7570n ± 20% 0.7660n ± 2% ~ (p\u003d0.331 n\u003d6)\nBM_log_f32x16_avx512 6.230n ± 4% 6.301n ± 2% ~ (p\u003d0.589 n\u003d6)\nBM_log_f64x8_avx512 10.172n ± 5% 9.771n ± 5% ~ (p\u003d0.132 n\u003d6)\nBM_log1p_f32x16_avx512 8.073n ± 3% 8.127n ± 2% ~ (p\u003d0.180 n\u003d6)\nBM_log1p_f64x8_avx512 11.53n ± 4% 11.46n ± 2% ~ (p\u003d0.699 n\u003d6)\ngeomean 2.083n 2.082n -0.01%\n```\n\nPiperOrigin-RevId: 920668251\n" }, { "commit": "eafd6fe46bbb93ed46c71fc56796f1fbb82bf0db", "tree": "b2a478f03a359607ebaa471bd0f393b737e9f111", "parents": [ "38631e8f9ec99555c302fe744739b2ce14808da8" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Sun May 24 23:13:26 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sun May 24 23:15:27 2026" }, "message": "Add benchmarks of exp and log for avx and avx2\n\nPiperOrigin-RevId: 920663137\n" }, { "commit": "38631e8f9ec99555c302fe744739b2ce14808da8", "tree": "00da42bd4a83476c3db2a26b6be8461127ec328b", "parents": [ "fda3ca7b87546df6288c9036775c0fc0235e873f" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Sun May 24 23:08:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sun May 24 23:10:19 2026" }, "message": "Refactor `exp` and `expm1` to use the same implementation\n\nThe expm1 polynomial should be better, because it passes through 0, and so it should be easier to fit numerically.\n\nIndeed, this change enables reducing the ULP tolerance of `exp` from 3 to 2.\n\nPiperOrigin-RevId: 920662177\n" }, { "commit": "fda3ca7b87546df6288c9036775c0fc0235e873f", "tree": "bff9cf00228001ce0479b1ed8c2027fd14d0bd44", "parents": [ "52bd8d0a6ef2b6c941973df67864daf33a4e783c" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Sat May 23 01:36:07 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sat May 23 01:37:01 2026" }, "message": "Fix precision issue in rndnu16 requantization for scales near powers of 2.\n\nPiperOrigin-RevId: 919955609\n" }, { "commit": "52bd8d0a6ef2b6c941973df67864daf33a4e783c", "tree": "28a31af436dbd3be92eac38bc3e3e9a2054786dc", "parents": [ "0ccb84e7c33a5527d7ba417057b305debdb2a48f" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 22:33:13 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 22:34:08 2026" }, "message": "Tighten tolerances of `log` from 3 ULPs to 2\n\nAnd add tests/benchmarks for ARM\n\nPiperOrigin-RevId: 919889466\n" }, { "commit": "0ccb84e7c33a5527d7ba417057b305debdb2a48f", "tree": "d825e7724c25efcc262b62874c256865e0fc6985", "parents": [ "c45d6b4166d3e8ac581e6fdabb371fcd33ef065d" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 21:27:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 21:27:54 2026" }, "message": "Implement `YNN_FLAG_CONSISTENT_ARITHMETIC` for unary elementwise kernels\n\nPiperOrigin-RevId: 919861252\n" }, { "commit": "c45d6b4166d3e8ac581e6fdabb371fcd33ef065d", "tree": "9307fdc68ebb5ee21a3d7745753e8ab3ca7aa79f", "parents": [ "5ff101c80c734d97aef504940482350953351be2" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Fri May 22 18:33:11 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 18:34:02 2026" }, "message": "Don\u0027t split the innermost dimension if the type of the input is sub-byte.\n\nPiperOrigin-RevId: 919777496\n" }, { "commit": "5ff101c80c734d97aef504940482350953351be2", "tree": "3523a6cb58e807a0d9444ad0c5af3ae0f347be9c", "parents": [ "3b5dbb202628f7687c374f6c9a2caf9c4c7adce5", "0f6ee412250039c51b002352cd88537b64d856aa" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 18:19:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 18:19:10 2026" }, "message": "Merge pull request #10298 from wangw-1991:fix_LUT_fusion\n\nPiperOrigin-RevId: 919769875\n" }, { "commit": "3b5dbb202628f7687c374f6c9a2caf9c4c7adce5", "tree": "bdd7e9ad70e8695a02dbfa2e95ee05d4ebdf822e", "parents": [ "008036729dd12d3b3d6985008ae258c4078a32e0" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 18:02:30 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 18:03:18 2026" }, "message": "Remove `fma` when not available, and add `multiply_add` which optionally uses `fma` when available.\n\nAnd add the `sse2fma` target which emulates fma (slowly).\n\nPiperOrigin-RevId: 919762437\n" }, { "commit": "008036729dd12d3b3d6985008ae258c4078a32e0", "tree": "e46ee8a6ad087f1e4dd8d8a1634d3ca4e866249f", "parents": [ "c4e49c65c72674932fc60977e2985d6326ad7188" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 17:25:04 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 17:26:07 2026" }, "message": "Combine x86 SIMD wrapper headers\n\nThis is a large change that is mostly code motion, to combine the various microarchitecture specific header files for x86 intrinsics wrappers. The problem this solves is that things don\u0027t really work properly if you include multiple of these headers at once. But, sometimes that is necessary, e.g. you want to use `x86_f16c.h` and `x86_fma3.h` at the same time. I\u0027m currently trying to make an optional fma or multiply-add intrinsic, and that runs into this problem broadly.\n\nPiperOrigin-RevId: 919744261\n" }, { "commit": "c4e49c65c72674932fc60977e2985d6326ad7188", "tree": "d5c132b0ad9ab6ed794c6fc6b991702116ff97b7", "parents": [ "00825fc8ecdf3c3d58a2cae3471c959e28c9b147" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 16:44:03 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 16:44:50 2026" }, "message": "Combine ARM SIMD wrapper headers\n\nThis is a prototype for what I\u0027d like to do to the x86 headers too, but this is an easier place to start.\n\nThe main advantage of this approach is it\u0027s easier to add new things with fallback implementations on older microarchitectures. The problem I\u0027m specifically trying to solve is how to add `multiply_add` to use fma when available, and fall back to multiply + add when not available. Doing this with the header per microarchitecture approach is really hard. Consider AVX2, AVX, and FMA: there are a lot of combinations of headers to support. And then if you are using AVX512, it should use FMA for AVX and SSE vectors too.\n\nThe main disadvantage is that there\u0027s less safety, I think it would be easy to accidentally not use a lower capability microarchitecture (i.e. get worse performance).\n\nPiperOrigin-RevId: 919723374\n" }, { "commit": "00825fc8ecdf3c3d58a2cae3471c959e28c9b147", "tree": "4399116b98e7e571a3aa72325afe1b5a0969b910", "parents": [ "cccad557c84c1e03c1adb595946e79f9a84fafbd" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 22 16:07:20 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 22 16:08:14 2026" }, "message": "Define all architecture flags transitively implied by enabled architectures.\n\nThis means we can rely on these flags instead of compiler specific features to enable or disable intrinsics, which is safer when we have a fallback path (we won\u0027t silently fall back to a lesser implementation than the user was expecting).\n\nPiperOrigin-RevId: 919707309\n" }, { "commit": "cccad557c84c1e03c1adb595946e79f9a84fafbd", "tree": "d35ed483430cd87c44d9bed6dea5674b4c425071", "parents": [ "549deb8dd62218ba926d0b6cd4730c408b4a6718" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 21 22:56:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 22:56:50 2026" }, "message": "Add math helpers to SIMD wrappers\n\nThe main benefit of this change is the ability to test the math functions in isolation from the rest of the kernel infrastructure. This led me to discover an issue with our tolerances that was otherwise difficult to find.\n\nThis change is a significant upgrade in the accuracy of `log` and `log1p` especially, and especially for fp64 types. We now have ~3 ULPs for all exp and log functions (and no more absolute tolerances).\n\nCurrently, I\u0027ve implemented `tanh` and `sigmoid` in terms of these functions. These implementations could likely be optimized and improved further, but I think this change is already big enough by itself.\n\nI intend to further remove the sophisticated approximations from the python code and move them to the C++ SIMD wrappers.\n\nCurrently it is added to `generic.inc`, but this is really not a good place for them. The SIMD wrappers need some restructuring to enable a better implementation for this. However, this is a good improvement that unblocks other work, I\u0027d like to work on this cleanup in parallel with other tasks.\n\nThere are some performance improvements and regressions in this change:\n```\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2.550µ ± 3% 2.309µ ± 3% -9.46% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 6.255µ ± 2% 5.905µ ± 5% -5.59% (p\u003d0.009 n\u003d6)\nbench/expm1_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2.876µ ± 2% 2.713µ ± 4% -5.67% (p\u003d0.002 n\u003d6)\nbench/expm1_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7.094µ ± 2% 6.776µ ± 2% -4.48% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.506µ ± 2% 1.883µ ± 3% +25.07% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.851µ ± 7% 5.473µ ± 3% +12.83% (p\u003d0.002 n\u003d6)\nbench/floor_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 297.6n ± 3% 325.0n ± 4% +9.20% (p\u003d0.002 n\u003d6)\nbench/negate_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 270.8n ± 9% 319.7n ± 11% +18.06% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.837µ ± 1% 3.086µ ± 4% +68.03% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.616µ ± 5% 8.505µ ± 1% +84.23% (p\u003d0.002 n\u003d6)\nbench/tanh_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7.683µ ± 4% 9.246µ ± 2% +20.35% (p\u003d0.002 n\u003d6)\n```\n\nThe sigmoid regressions are significant, but I think we should prioritize correctness for now. A dedicated sigmoid op using this new mechanism would be faster and not too hard to implement.\n\nPiperOrigin-RevId: 919298994\n" }, { "commit": "549deb8dd62218ba926d0b6cd4730c408b4a6718", "tree": "322ca473054e6e5f83a5b9767232decb74436a5b", "parents": [ "58a65db7c2b470e967fca11c3be29689946683c9" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 21 21:14:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 21:15:52 2026" }, "message": "Remove unused `transpose` SIMD wrapper\n\nPiperOrigin-RevId: 919250056\n" }, { "commit": "58a65db7c2b470e967fca11c3be29689946683c9", "tree": "0fd151193cf08db59a591d8b000cc30551ad15e7", "parents": [ "6c9a1ab6259e7a6a8bce27dac6922ecc28d2bbe0" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Thu May 21 17:55:59 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 17:56:49 2026" }, "message": "Mark values as external outputs in constant folding only if they are actually used in the non-constant pipeline.\n\nPiperOrigin-RevId: 919143197\n" }, { "commit": "6c9a1ab6259e7a6a8bce27dac6922ecc28d2bbe0", "tree": "aa89c88e01b27f7dc4239b8e2318a5acad5dedce", "parents": [ "ea77aab35cfd965afa93891af6a2fa249af98164" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 21 17:20:30 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 17:21:12 2026" }, "message": "Consolidate some SIMD wrapper headers\n\nThis is a low hanging fruit part of eliminating micro-architecture specific headers.\n\nPiperOrigin-RevId: 919122512\n" }, { "commit": "ea77aab35cfd965afa93891af6a2fa249af98164", "tree": "1afee7f4392d7058c0dd8dc4d63031e38f29bc5e", "parents": [ "1b849f44fb6ce9eae460586158669461044acc2d" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 21 16:36:52 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 16:37:53 2026" }, "message": "Generalize FMA emulation helper\n\nThis doesn\u0027t need to depend specifically on SSE2, it should work for any SIMD type.\n\nThis eliminates one of the microarchitecture entangled headers in the SIMD wrappers.\n\nPiperOrigin-RevId: 919098203\n" }, { "commit": "1b849f44fb6ce9eae460586158669461044acc2d", "tree": "116aca0116f85a3314d83a2e03cfc1f6a44b1e12", "parents": [ "d4adfcdc4c5c84eff8b8f2d68bf29c0b9d185eea" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 21 16:18:06 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 16:18:37 2026" }, "message": "Tune params for unary kernels to avoid tolerance issues\n\nIn some cases, the absolute tolerances are needed due to the choice of \"params\" that scale/offset the result. This changes the choices of params such that we can remove the absolute tolerances from the tests, by avoiding zero in the output and avoiding \"shrinking\" the output too much.\n\nPiperOrigin-RevId: 919089273\n" }, { "commit": "0f6ee412250039c51b002352cd88537b64d856aa", "tree": "0f0c50b93cb94ea5f6e643c77b63af2e5fbc12b3", "parents": [ "d877e1a189580fb562a66026021f3403cc48bf18" ], "author": { "name": "Wei Wang", "email": "wei4.wang@intel.com", "time": "Thu May 21 11:58:47 2026" }, "committer": { "name": "Wei Wang", "email": "wei4.wang@intel.com", "time": "Thu May 21 11:58:47 2026" }, "message": "Initial upload.\n" }, { "commit": "d4adfcdc4c5c84eff8b8f2d68bf29c0b9d185eea", "tree": "44fff0143cf2e931a161917acdd27211e92caf76", "parents": [ "f56a6c716d6c544b74e8388a68f89a49132375ab" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Thu May 21 01:01:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 21 01:01:50 2026" }, "message": "gemm benchmark documentation fix\n- update names of models to match files\n\nbenchmarks now require 3 int values on command line\neg f16_gemm_minmax_bench 784 512 144\n\nPiperOrigin-RevId: 918743819\n" }, { "commit": "f56a6c716d6c544b74e8388a68f89a49132375ab", "tree": "abb8deec3c40aadc5f08e7bffed4f29e6fc25a00", "parents": [ "29a1c7333c02fedf408ba913f959781058d26843" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 20 23:36:45 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 23:37:37 2026" }, "message": "Add numerically correct `expm1` kernels\n\n- Add a feature to `rational_approximation.py` to include 0 as one of the points that Remez optimizes.\n- Use it to generate polynomials for `exp2m1`\n- Implement `expm1` kernels. These are very similar to `exp`, but rearranging the arithmetic to avoid catastrophic cancellation when 2^x is near 1.\n\n```\n--------------------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n--------------------------------------------------------------------------------------------------------------------------\nbench_reference/exp_float/m:1/n:4096/real_time 36990 ns 36980 ns 18487 Bytes\u003d885.865M/s Op\u003d110.733M/s\nbench_reference/exp_double/m:1/n:4096/real_time 75183 ns 75157 ns 9337 Bytes\u003d871.682M/s Op\u003d54.4801M/s\nbench_reference/expm1_float/m:1/n:4096/real_time 36931 ns 36925 ns 18730 Bytes\u003d887.269M/s Op\u003d110.909M/s\nbench_reference/expm1_double/m:1/n:4096/real_time 80463 ns 80438 ns 8501 Bytes\u003d814.486M/s Op\u003d50.9054M/s\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2576 ns 2576 ns 268977 Bytes\u003d12.7188G/s Op\u003d1.58985G/s\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 6399 ns 6397 ns 109094 Bytes\u003d10.2419G/s Op\u003d640.117M/s\nbench/expm1_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 2938 ns 2938 ns 238502 Bytes\u003d11.1519G/s Op\u003d1.39398G/s\nbench/expm1_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7284 ns 7282 ns 96595 Bytes\u003d8.99685G/s Op\u003d562.303M/s\nbench/exp_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 3397 ns 3397 ns 207260 Bytes\u003d9.64533G/s Op\u003d1.20567G/s\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 8476 ns 8474 ns 81181 Bytes\u003d7.73191G/s Op\u003d483.245M/s\nbench/expm1_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 3761 ns 3761 ns 182875 Bytes\u003d8.71165G/s Op\u003d1.08896G/s\nbench/expm1_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 12944 ns 12941 ns 54861 Bytes\u003d5.06303G/s Op\u003d316.439M/s\nbench/exp_fp32_1x16_x86_avx2/m:1/n:4096/real_time 4995 ns 4994 ns 138480 Bytes\u003d6.56011G/s Op\u003d820.014M/s\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 11544 ns 11541 ns 58414 Bytes\u003d5.67696G/s Op\u003d354.81M/s\nbench/expm1_fp32_1x16_x86_avx2/m:1/n:4096/real_time 6345 ns 6343 ns 110725 Bytes\u003d5.16475G/s Op\u003d645.594M/s\nbench/expm1_fp64_1x16_x86_avx2/m:1/n:4096/real_time 18278 ns 18273 ns 36881 Bytes\u003d3.58542G/s Op\u003d224.089M/s\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 7399 ns 7398 ns 95901 Bytes\u003d4.42847G/s Op\u003d553.558M/s\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 20106 ns 20101 ns 35700 Bytes\u003d3.25956G/s Op\u003d203.722M/s\nbench/expm1_fp32_1x16_x86_sse2/m:1/n:4096/real_time 7980 ns 7979 ns 86063 Bytes\u003d4.1063G/s Op\u003d513.287M/s\nbench/expm1_fp64_1x8_x86_sse2/m:1/n:4096/real_time 24816 ns 24813 ns 27562 Bytes\u003d2.64087G/s Op\u003d165.054M/s\n```\n\nPiperOrigin-RevId: 918713274\n" }, { "commit": "29a1c7333c02fedf408ba913f959781058d26843", "tree": "a36ad44de144b45576f88353048b6bd454d89d91", "parents": [ "b1a0a5d3725f9b75b9ca449e88e0f197fbc3a7db" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 21:33:47 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 21:34:41 2026" }, "message": "Add std::string overloads for tensor::Create.\n\nPiperOrigin-RevId: 918656793\n" }, { "commit": "b1a0a5d3725f9b75b9ca449e88e0f197fbc3a7db", "tree": "0860ed7cb470be7f01d86e23b61fc69dc83fa194", "parents": [ "a0dbef32efcd73eff5d13f6900d93bb58ddd91f4", "4ae8b8f607b2a35a91bba3ef491bda68a405fdc3" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 19:27:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 19:27:08 2026" }, "message": "Merge pull request #10261 from velonica0:f16\n\nPiperOrigin-RevId: 918589980\n" }, { "commit": "a0dbef32efcd73eff5d13f6900d93bb58ddd91f4", "tree": "4e352a23c07a167fa68e0046ff96c3f8e7327d87", "parents": [ "3f33e550b6270dbad0b9549170de4b9f107a4052" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 20 19:21:53 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 19:22:42 2026" }, "message": "Improve `exp` accuracy\n\nCurrently we effectively have an `exp2` kernel, and turn it into an `exp` kernel with an `input_multiplier`.\n\nThis strategy is a bit problematic because a little precision is lost in the `exp` -\u003e `exp2` conversion. This was hidden by the fact that we computed the reference result in the same way.\n\nThis change rewrites the `exp` kernels to compute `exp` directly.\n\nThis is a performance regression, but I think we need to prioritize accuracy for now. Most of the performance regression is due to the more accurate range reduction.\n\n```\nname time/op time/op vs base\nbench_reference/exp_float/m:1/n:4096/real_time 53.98µ ± 2% 35.47µ ± 5% -34.30% (p\u003d0.002 n\u003d6)\nbench_reference/exp_double/m:1/n:4096/real_time 66.48µ ± 2% 72.48µ ± 2% +9.04% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.873µ ± 3% 2.486µ ± 2% +32.70% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.732µ ± 5% 6.318µ ± 2% +33.51% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 2.530µ ± 4% 3.216µ ± 2% +27.10% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 6.717µ ± 3% 8.113µ ± 4% +20.79% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_avx2/m:1/n:4096/real_time 3.799µ ± 1% 4.839µ ± 2% +27.39% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 10.02µ ± 2% 11.17µ ± 1% +11.49% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 5.585µ ± 3% 6.913µ ± 2% +23.78% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 16.33µ ± 7% 19.19µ ± 2% +17.53% (p\u003d0.002 n\u003d6)\ngeomean 8.459µ 9.723µ +14.94%\n```\n\nPiperOrigin-RevId: 918587354\n" }, { "commit": "3f33e550b6270dbad0b9549170de4b9f107a4052", "tree": "ddb29e2b19836f6988cdc13694380c6185d88415", "parents": [ "5a59a54af5e777acbb0c81a8b0d8bff8cc1d3464" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 20 18:50:16 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 18:51:06 2026" }, "message": "Add `select` and conditional operations to SIMD wrappers\n\nTo make this work across the variety of SIMD ISA conventions for this, mask types should just be `auto`. They can\u0027t be stored to memory.\n\nIf we need `select` to work on conditions with a different size type than the values, we should be able to handle that with more overloads (and some conversions), but we don\u0027t need it yet.\n\nThis removes the workarounds that avoided the need for this until now: `select_greater_than`, `copynan`, `kahan_sum`.\n\nThis was partly AI generated, here is what it had to say:\n\n✦ I have completed the implementation of boolean operators and conditional selection across the entire SIMD abstraction layer. Here is a summary of the work:\n\n 1. SIMD Core \u0026 Generic Layer\n * Boolean Operators: Implemented \u003d\u003d, !\u003d, \u003c, \u003c\u003d, \u003e, \u003e\u003d for all vector types. These now use auto return types to accommodate varying architecture-specific mask representations (e.g., integer vectors on SSE/AVX vs. __mmaskN on AVX-512).\n * Conditional Selection: Added a unified select(mask, true_val, false_val) intrinsic.\n * Predicates: Implemented isnan, isinf, and isfinite for floating-point vectors.\n * Recursive Generics: Updated generic.inc to support recursive implementation of these operators for wide vectors (e.g., f32x16 on AVX2) by defining lo()/hi()/concat() helpers for all vector and mask types.\n\n 2. Architecture-Specific Implementations\n * x86 AVX-512: Utilized native mask registers (__mmask16, etc.) and _mm512_mask_blend_* instructions. Implemented mask-specific glue logic for recursive templates.\n * x86 SSE/AVX: Implemented comparisons returning signed integer vectors. Added s64x2 and s64x4 specializations to ensure 64-bit comparisons (for doubles) maintain the correct element count in the resulting mask.\n * ARM NEON: Leveraged vceqq, vcgtq, and vbslq (bitselect) for mask operations. Added 64-bit integer vector support for ARM64.\n * WASM SIMD128: Integrated wasm_v128_bitselect and standard WASM comparison intrinsics.\n\n 3. Kernel Compiler \u0026 Intrinsics\n * Intrinsic Support: Updated the elementwise kernel compiler (compiler.py) to support select, isnan, isinf, and isfinite.\n * Type Deduction: Modified the compiler to emit auto for comparison results, allowing the C++ compiler to handle architecture-specific mask types.\n * Kernel Refactoring: Migrated existing kernels (e.g., exp, sigmoid) away from the legacy copynan helper to the more idiomatic select(isnan(x), ...) pattern.\n\n 4. Verification \u0026 Cleanup\n * Comprehensive Testing: Expanded the SIMD test suite in base/simd/test/ to verify comparisons, predicates, and selection logic across all architectures.\n * Code Quality: Resolved header inclusion ordering issues, fixed redefinition errors between SSE2 and SSE4.1, and performed a significant cleanup of bitwise operator consistency in the AVX headers.\n\n Status: All tasks in PLAN.md are complete, and all SIMD test targets (x86_avx512, x86_avx2, x86_avx, x86_sse2, arm_neon, wasm_simd128) are passing.\n\nI expected this to be a no-op in terms of performance/generated code, but it improves some benchmarks:\n```\nbench/sum_squared_k1_bf16_fp32_avx512/real_time [256x256] 5.520µ ± 4% 5.483µ ± 3% ~ (p\u003d0.937 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_avx512/real_time [256x256] 4.766µ ± 3% 5.053µ ± 11% ~ (p\u003d0.065 n\u003d6)\nbench/sum_squared_k1_fp16_fp32_avx512/real_time [256x256] 4.495µ ± 3% 4.571µ ± 5% ~ (p\u003d0.310 n\u003d6)\nbench/sum_squared_kn_fp16_fp32_avx512/real_time [256x256] 4.157µ ± 1% 4.181µ ± 17% ~ (p\u003d0.937 n\u003d6)\nbench/sum_squared_k1_fp32_avx512/real_time [256x256] 5.506µ ± 15% 5.286µ ± 7% ~ (p\u003d0.310 n\u003d6)\nbench/sum_squared_kn_fp32_avx512/real_time [256x256] 4.770µ ± 9% 4.691µ ± 17% ~ (p\u003d0.937 n\u003d6)\nbench/sum_squared_k1_fp64_avx512/real_time [256x256] 9.632µ ± 10% 8.636µ ± 2% -10.34% (p\u003d0.015 n\u003d6)\nbench/sum_squared_kn_fp64_avx512/real_time [256x256] 10.36µ ± 6% 10.21µ ± 9% ~ (p\u003d0.699 n\u003d6)\nbench/sum_squared_k1_bf16_fp32_avx2/real_time [256x256] 8.651µ ± 8% 7.314µ ± 4% -15.45% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_avx2/real_time [256x256] 7.510µ ± 3% 6.949µ ± 3% -7.47% (p\u003d0.002 n\u003d6)\nbench/sum_squared_k1_fp32_avx/real_time [256x256] 7.076µ ± 3% 6.491µ ± 3% -8.27% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp32_avx/real_time [256x256] 6.368µ ± 4% 5.859µ ± 3% -7.99% (p\u003d0.002 n\u003d6)\nbench/sum_squared_k1_fp64_avx/real_time [256x256] 12.28µ ± 2% 11.31µ ± 4% -7.89% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp64_avx/real_time [256x256] 12.72µ ± 4% 11.61µ ± 9% -8.76% (p\u003d0.009 n\u003d6)\nbench/sum_squared_k1_fp16_fp32_f16c/real_time [256x256] 7.480µ ± 4% 6.799µ ± 2% -9.11% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp16_fp32_f16c/real_time [256x256] 6.445µ ± 2% 6.163µ ± 12% ~ (p\u003d0.065 n\u003d6)\nbench/sum_squared_k1_fp32_sse2/real_time [256x256] 10.97µ ± 4% 10.31µ ± 5% -6.05% (p\u003d0.009 n\u003d6)\nbench/sum_squared_kn_fp32_sse2/real_time [256x256] 10.42µ ± 3% 10.21µ ± 3% ~ (p\u003d0.065 n\u003d6)\nbench/sum_squared_k1_bf16_fp32_sse2/real_time [256x256] 12.85µ ± 5% 12.31µ ± 3% -4.20% (p\u003d0.004 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_sse2/real_time [256x256] 11.56µ ± 2% 11.25µ ± 4% ~ (p\u003d0.240 n\u003d6)\nbench/sum_squared_k1_fp64/real_time [256x256] 24.13µ ± 3% 22.98µ ± 19% ~ (p\u003d0.065 n\u003d6)\nbench/sum_squared_kn_fp64/real_time [256x256] 21.80µ ± 4% 20.33µ ± 16% ~ (p\u003d0.065 n\u003d6)\nbench/sum_squared_k1_fp16_fp32/real_time [256x256] 183.7µ ± 1% 181.1µ ± 2% ~ (p\u003d0.093 n\u003d6)\nbench/sum_squared_kn_fp16_fp32/real_time [256x256] 50.73µ ± 7% 47.96µ ± 2% -5.45% (p\u003d0.002 n\u003d6)\ngeomean 9.443µ 9.233µ -2.22%\n```\nBut some kernels slower:\n```\nname time/op time/op vs base\nbench_reference/exp_float/m:1/n:4096/real_time 53.03µ ± 1% 53.24µ ± 2% ~ (p\u003d0.485 n\u003d6)\nbench_reference/exp_double/m:1/n:4096/real_time 64.92µ ± 1% 65.43µ ± 2% ~ (p\u003d0.180 n\u003d6)\nbench_reference/sigmoid_float/m:1/n:4096/real_time 37.11µ ± 1% 37.41µ ± 2% ~ (p\u003d0.394 n\u003d6)\nbench_reference/sigmoid_double/m:1/n:4096/real_time 72.69µ ± 20% 72.44µ ± 1% ~ (p\u003d0.699 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.844µ ± 1% 1.872µ ± 3% +1.51% (p\u003d0.015 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.816µ ± 6% 4.743µ ± 2% ~ (p\u003d0.394 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.841µ ± 3% 1.827µ ± 1% ~ (p\u003d0.394 n\u003d6)\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.549µ ± 7% 4.619µ ± 3% ~ (p\u003d0.093 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 2.481µ ± 3% 2.575µ ± 3% +3.77% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 6.648µ ± 3% 6.834µ ± 3% +2.80% (p\u003d0.026 n\u003d6)\nbench/exp_fp32_1x16_x86_avx2/m:1/n:4096/real_time 3.777µ ± 2% 3.876µ ± 3% +2.60% (p\u003d0.015 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 9.753µ ± 3% 9.919µ ± 3% ~ (p\u003d0.180 n\u003d6)\nbench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time 4.010µ ± 3% 3.910µ ± 2% ~ (p\u003d0.093 n\u003d6)\nbench/sigmoid_fp64_1x8_x86_avx2/m:1/n:4096/real_time 10.65µ ± 3% 10.66µ ± 2% ~ (p\u003d0.589 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 5.544µ ± 1% 5.770µ ± 1% +4.07% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 16.14µ ± 7% 16.53µ ± 3% ~ (p\u003d0.310 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time 5.792µ ± 4% 5.917µ ± 4% ~ (p\u003d0.394 n\u003d6)\nbench/sigmoid_fp64_1x8_x86_sse2/m:1/n:4096/real_time 17.36µ ± 3% 17.72µ ± 2% ~ (p\u003d0.132 n\u003d6)\ngeomean 9.036µ 9.144µ +1.19%\n```\n\nPiperOrigin-RevId: 918570419\n" }, { "commit": "5a59a54af5e777acbb0c81a8b0d8bff8cc1d3464", "tree": "8b9f3bc749e9a76bc81f7b33546e255f08f560a5", "parents": [ "bcc179a8b7b8c2cadefd52a48d75cb64887228d3" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 00:15:12 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 00:16:30 2026" }, "message": "Properly open source tensor api in github through copybara\n\nPiperOrigin-RevId: 918107140\n" }, { "commit": "bcc179a8b7b8c2cadefd52a48d75cb64887228d3", "tree": "db88c4b261d43d2d71664361dd8105ce2c1d619b", "parents": [ "0547829228c2bab8f58f468b54db0a8ec91ebc62" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Wed May 20 00:01:00 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Wed May 20 00:01:56 2026" }, "message": "Remove fp64 wasm support\n\nWe should never need this\n\nPiperOrigin-RevId: 918101335\n" }, { "commit": "0547829228c2bab8f58f468b54db0a8ec91ebc62", "tree": "72c762fb6fbea18567f12fd0677b3133de2a2aaa", "parents": [ "cc68da84bd4e062a79600b82109ea28bd1553e49" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 19 23:21:38 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 23:22:12 2026" }, "message": "Remove lo/hi as member functions of `vec\u003cT, N\u003e`\n\nI think this is necessary as a step towards attempting to implement conditions and `select`, where the mask types might be things like `__mmask8`.\n\nPiperOrigin-RevId: 918084846\n" }, { "commit": "cc68da84bd4e062a79600b82109ea28bd1553e49", "tree": "cc9b3f6473c89abb9c28b63f93e0b4f288b97b1a", "parents": [ "4b53a435774a5ec388621aa594b7c60c2e129b02" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 19 23:14:22 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 23:15:11 2026" }, "message": "Add sigmoid_fp64 kernels\n\nAnd rewrite the sigmoid_fp32 kernel using the same technique.\n\nIt turns out that this kernel is faster, which is a little surprising. It does have less \"overhead\" (special cases, piecewise branches, etc.) in exchange for more polynomial arithmetic.\n\nChange in performance for `sigmoid_fp32`:\n\n```\nname time/op time/op vs base\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.759µ ± 3% 1.771µ ± 21% ~ (p\u003d0.818 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time 1.740µ ± 3% 1.774µ ± 3% ~ (p\u003d0.065 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time 1.773µ ± 2% 1.783µ ± 1% ~ (p\u003d0.699 n\u003d6)\nbench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time 4.909µ ± 1% 3.670µ ± 2% -25.24% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x16_x86_avx2/m:4/n:1024/real_time 4.830µ ± 2% 3.706µ ± 4% -23.28% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x16_x86_avx2/m:16/n:256/real_time 4.912µ ± 1% 3.740µ ± 2% -23.87% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time 6.632µ ± 3% 5.437µ ± 2% -18.02% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2/m:4/n:1024/real_time 6.637µ ± 3% 5.524µ ± 3% -16.77% (p\u003d0.002 n\u003d6)\nbench/sigmoid_fp32_1x32_x86_sse2/m:16/n:256/real_time 6.692µ ± 4% 5.493µ ± 1% -17.92% (p\u003d0.002 n\u003d6)\ngeomean 3.851µ 3.305µ -14.19%\n```\n\n`sigmoid_fp64` compared to other kernels:\n```\n----------------------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n----------------------------------------------------------------------------------------------------------------------------\nbench_reference/sigmoid_float/m:1/n:4096/real_time 38579 ns 38571 ns 7348 Bytes\u003d849.382M/s Op\u003d106.173M/s\nbench_reference/sigmoid_float/m:4/n:1024/real_time 38345 ns 38338 ns 7440 Bytes\u003d854.556M/s Op\u003d106.819M/s\nbench_reference/sigmoid_float/m:16/n:256/real_time 39192 ns 39187 ns 7190 Bytes\u003d836.095M/s Op\u003d104.512M/s\nbench_reference/sigmoid_double/m:1/n:4096/real_time 91326 ns 91313 ns 3045 Bytes\u003d717.606M/s Op\u003d44.8504M/s\nbench_reference/sigmoid_double/m:4/n:1024/real_time 91307 ns 91290 ns 3043 Bytes\u003d717.757M/s Op\u003d44.8598M/s\nbench_reference/sigmoid_double/m:16/n:256/real_time 93505 ns 93486 ns 3018 Bytes\u003d700.885M/s Op\u003d43.8053M/s\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1786 ns 1786 ns 155658 Bytes\u003d18.3422G/s Op\u003d2.29277G/s\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time 1802 ns 1802 ns 157599 Bytes\u003d18.1847G/s Op\u003d2.27309G/s\nbench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time 1791 ns 1791 ns 156134 Bytes\u003d18.2963G/s Op\u003d2.28704G/s\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4475 ns 4475 ns 60425 Bytes\u003d14.6433G/s Op\u003d915.207M/s\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 4822 ns 4821 ns 59593 Bytes\u003d13.5913G/s Op\u003d849.459M/s\nbench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 4842 ns 4840 ns 56596 Bytes\u003d13.5363G/s Op\u003d846.016M/s\nbench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time 3789 ns 3788 ns 69486 Bytes\u003d8.64752G/s Op\u003d1.08094G/s\nbench/sigmoid_fp32_1x16_x86_avx2/m:4/n:1024/real_time 3892 ns 3892 ns 74142 Bytes\u003d8.41825G/s Op\u003d1.05228G/s\nbench/sigmoid_fp32_1x16_x86_avx2/m:16/n:256/real_time 3757 ns 3756 ns 72827 Bytes\u003d8.72073G/s Op\u003d1.09009G/s\nbench/sigmoid_fp64_1x8_x86_avx2/m:1/n:4096/real_time 10451 ns 10450 ns 26516 Bytes\u003d6.27103G/s Op\u003d391.939M/s\nbench/sigmoid_fp64_1x8_x86_avx2/m:4/n:1024/real_time 11010 ns 11007 ns 24451 Bytes\u003d5.95261G/s Op\u003d372.038M/s\nbench/sigmoid_fp64_1x8_x86_avx2/m:16/n:256/real_time 10475 ns 10472 ns 26374 Bytes\u003d6.2567G/s Op\u003d391.044M/s\nbench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time 5649 ns 5648 ns 49675 Bytes\u003d5.80048G/s Op\u003d725.06M/s\nbench/sigmoid_fp32_1x32_x86_sse2/m:4/n:1024/real_time 5646 ns 5645 ns 50916 Bytes\u003d5.80353G/s Op\u003d725.441M/s\nbench/sigmoid_fp32_1x32_x86_sse2/m:16/n:256/real_time 5571 ns 5571 ns 48792 Bytes\u003d5.88151G/s Op\u003d735.188M/s\nbench/sigmoid_fp64_1x8_x86_sse2/m:1/n:4096/real_time 15957 ns 15952 ns 17116 Bytes\u003d4.10712G/s Op\u003d256.695M/s\nbench/sigmoid_fp64_1x8_x86_sse2/m:4/n:1024/real_time 15657 ns 15654 ns 17451 Bytes\u003d4.18581G/s Op\u003d261.613M/s\nbench/sigmoid_fp64_1x8_x86_sse2/m:16/n:256/real_time 15748 ns 15744 ns 17804 Bytes\u003d4.16163G/s Op\u003d260.102M/s\n```\n\nPiperOrigin-RevId: 918081537\n" }, { "commit": "4b53a435774a5ec388621aa594b7c60c2e129b02", "tree": "766ec8b1ce83eee127182e28288345d92cb56898", "parents": [ "0741ac549c8f6357a8e311ffee2cee234a6da7cf" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Tue May 19 21:41:49 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 21:42:21 2026" }, "message": "Prevent scheduling of ki/ko loops in packing.\n\nIn order to keep rest of the loops fused I added \"identity\" splits where step \u003d\u003d extent, as a side effect now the packed buffer is stored inside of the loop.\n\nPiperOrigin-RevId: 918036935\n" }, { "commit": "0741ac549c8f6357a8e311ffee2cee234a6da7cf", "tree": "14b17258fa9a1501f96207d7afd54d79dd11ec33", "parents": [ "ce14e188cd02a9726c00970babc087a3be6de942" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 21:28:13 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 21:29:03 2026" }, "message": "Open source Tensor API in google-ai-edge/LiteRT\n\nPiperOrigin-RevId: 918030392\n" }, { "commit": "ce14e188cd02a9726c00970babc087a3be6de942", "tree": "a3a36647db738550cf888f0cf91c2d8d59f0ff93", "parents": [ "d0004f80c78fed80c230045ee83ff34dc55be81a" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Tue May 19 18:25:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 18:26:41 2026" }, "message": "Adjust bounds for elementwise unary kernels with sub-byte inputs.\n\nPiperOrigin-RevId: 917935792\n" }, { "commit": "d0004f80c78fed80c230045ee83ff34dc55be81a", "tree": "0d15f93c0ae44f1d4eae1de180fa9d014e4dd832", "parents": [ "adf9795c009f3d3190358c7df773dfa7180f694f" ], "author": { "name": "Pedro Gonnet", "email": "gonnet@google.com", "time": "Tue May 19 13:22:40 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 13:23:28 2026" }, "message": "Add `xnn_datatype_qint2` for tensorwise quantized 2-bit values.\n\nPiperOrigin-RevId: 917797252\n" }, { "commit": "adf9795c009f3d3190358c7df773dfa7180f694f", "tree": "ba4cdaafc4f7f76e925eb328ad5a1404cbdf3173", "parents": [ "ea4434133af5537f58c2616d6d939e82966a7f53" ], "author": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 08:37:33 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 08:38:15 2026" }, "message": "Split dot operation on K.\n\nCurrently split_k is the length of k so we don\u0027t expect major performance differences.\n\nSample IR for `f32[110,10240,2560]`.\n\nBefore:\n```\nnull \u003d constant_buffer([], 0, {}) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[0, 10239], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n out2.d0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, out2.d0} in {\n v3.out2.d0 \u003d crop_dim(v3, 3, [(out2.d0 / 16), ((out2.d0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.out2.d0}, {})\n }\n out2.out2.d0 \u003d crop_dim(out2, 0, [out2.d0, (out2.d0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.out2.d0}, {})\n }\n }\n }\n }\n}\n```\n\nAfter where k_split\u003d1024.\n```\nnull \u003d constant_buffer([], 0, {}) {\n reduction \u003d allocate(automatic, 0, {\n {[0, 10239], 0, \u003c\u003e}\n }) {\n k0#0 \u003d loop(serial, [0, 10239], 1024) {\n reduction \u003d crop_dim(reduction, 0, [k0#0, (k0#0 + 1023)]) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[k0#0, buffer_max(reduction, 0)], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n d0#0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, reduction, d0#0} in {\n v3.d0#0 \u003d crop_dim(v3, 3, [(d0#0 / 16), ((d0#0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.d0#0}, {})\n }\n out2.d0#0 \u003d crop_dim(out2, 0, [d0#0, (d0#0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.d0#0, reduction}, {})\n }\n }\n }\n }\n }\n }\n }\n}\n```\nPiperOrigin-RevId: 917686168\n" }, { "commit": "ea4434133af5537f58c2616d6d939e82966a7f53", "tree": "fa70567e65f794cb3346d4a0880f5e7c418a43f7", "parents": [ "7e3c789b2523ddc0f567019c3fa4c96fe4f6697d" ], "author": { "name": "Marie White", "email": "mariewhite@google.com", "time": "Tue May 19 06:14:39 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 06:15:36 2026" }, "message": "Split dot operation on K.\n\nCurrently split_k is the length of k so we don\u0027t expect major performance differences.\n\nSample IR for `f32[110,10240,2560]`.\n\nBefore:\n```\nnull \u003d constant_buffer([], 0, {}) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[0, 10239], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n out2.d0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, out2.d0} in {\n v3.out2.d0 \u003d crop_dim(v3, 3, [(out2.d0 / 16), ((out2.d0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.out2.d0}, {})\n }\n out2.out2.d0 \u003d crop_dim(out2, 0, [out2.d0, (out2.d0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.out2.d0}, {})\n }\n }\n }\n }\n}\n```\n\nAfter where k_split\u003d1024.\n```\nnull \u003d constant_buffer([], 0, {}) {\n reduction \u003d allocate(automatic, 0, {\n {[0, 10239], 0, \u003c\u003e}\n }) {\n k0#0 \u003d loop(serial, [0, 10239], 1024) {\n reduction \u003d crop_dim(reduction, 0, [k0#0, (k0#0 + 1023)]) {\n v3 \u003d allocate(automatic, 4, {\n {},\n {[0, 15], 4, \u003c\u003e},\n {[k0#0, buffer_max(reduction, 0)], 64, \u003c\u003e},\n {[(buffer_min(out2, 0) / 16), (buffer_max(out2, 0) / 16)], \u003c\u003e, \u003c\u003e}\n }) {\n d0#0 \u003d loop(parallel, [buffer_min(out2, 0), buffer_max(out2, 0)], 32) {\n closure {in0, in1, out2, v3, null, reduction, d0#0} in {\n v3.d0#0 \u003d crop_dim(v3, 3, [(d0#0 / 16), ((d0#0 / 16) + 1)]) {\n call(pack_b, {in1}, {v3.d0#0}, {})\n }\n out2.d0#0 \u003d crop_dim(out2, 0, [d0#0, (d0#0 + 31)]) {\n call(dot num_k_dims\u003d1, {in0, v3, null}, {out2.d0#0, reduction}, {})\n }\n }\n }\n }\n }\n }\n }\n}\n```\nPiperOrigin-RevId: 917625813\n" }, { "commit": "7e3c789b2523ddc0f567019c3fa4c96fe4f6697d", "tree": "ba4cdaafc4f7f76e925eb328ad5a1404cbdf3173", "parents": [ "7333afbd475d2f99c1af84512fff382271758a43" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 19 05:44:27 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 05:45:34 2026" }, "message": "Add tanh_fp64 kernels\n\nThese are not direct approximations like most other transcendental ops. Instead it uses our `exp_fp64` polynomials to implement the textbook definition of `(e^x - e^-x) / (e^x + e^-x)`, with some minor refactoring and optimization to group and reuse like terms.\n\nI also tried to improve the `tanh_fp32` kernels. This new approximation is not clearly better (it has the same error as before), but we should switch to it anyways because it is reproducible (via the new cell in rational_approximation.py).\n\nBenchmarks:\n\n```\n-------------------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n-------------------------------------------------------------------------------------------------------------------------\nbench_reference/tanh_float/m:1/n:4096/real_time 43571 ns 43563 ns 6383 Bytes\u003d752.055M/s Op\u003d94.0068M/s\nbench_reference/tanh_float/m:4/n:1024/real_time 43129 ns 43123 ns 6590 Bytes\u003d759.763M/s Op\u003d94.9704M/s\nbench_reference/tanh_float/m:16/n:256/real_time 43515 ns 43506 ns 6420 Bytes\u003d753.033M/s Op\u003d94.1291M/s\nbench_reference/tanh_double/m:1/n:4096/real_time 119486 ns 119465 ns 2175 Bytes\u003d548.482M/s Op\u003d34.2801M/s\nbench_reference/tanh_double/m:4/n:1024/real_time 117688 ns 117657 ns 2364 Bytes\u003d556.861M/s Op\u003d34.8038M/s\nbench_reference/tanh_double/m:16/n:256/real_time 117342 ns 117319 ns 2364 Bytes\u003d558.503M/s Op\u003d34.9064M/s\nbench/tanh_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1650 ns 1650 ns 153859 Bytes\u003d19.8583G/s Op\u003d2.48229G/s\nbench/tanh_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time 1619 ns 1619 ns 168542 Bytes\u003d20.2341G/s Op\u003d2.52927G/s\nbench/tanh_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time 1623 ns 1622 ns 171671 Bytes\u003d20.1957G/s Op\u003d2.52446G/s\nbench/tanh_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 7737 ns 7734 ns 37410 Bytes\u003d8.47083G/s Op\u003d529.427M/s\nbench/tanh_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 8076 ns 8074 ns 34535 Bytes\u003d8.11451G/s Op\u003d507.157M/s\nbench/tanh_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 7981 ns 7980 ns 36502 Bytes\u003d8.21114G/s Op\u003d513.196M/s\nbench/tanh_fp32_1x32_x86_avx/m:1/n:4096/real_time 3209 ns 3209 ns 85684 Bytes\u003d10.2099G/s Op\u003d1.27624G/s\nbench/tanh_fp32_1x32_x86_avx/m:4/n:1024/real_time 3284 ns 3284 ns 84044 Bytes\u003d9.97765G/s Op\u003d1.24721G/s\nbench/tanh_fp32_1x32_x86_avx/m:16/n:256/real_time 3274 ns 3273 ns 84131 Bytes\u003d10.0089G/s Op\u003d1.25111G/s\nbench/tanh_fp64_1x16_x86_avx/m:1/n:4096/real_time 12061 ns 12059 ns 23596 Bytes\u003d5.43361G/s Op\u003d339.6M/s\nbench/tanh_fp64_1x16_x86_avx/m:4/n:1024/real_time 12361 ns 12358 ns 22641 Bytes\u003d5.30194G/s Op\u003d331.371M/s\nbench/tanh_fp64_1x16_x86_avx/m:16/n:256/real_time 12167 ns 12164 ns 22833 Bytes\u003d5.38646G/s Op\u003d336.654M/s\nbench/tanh_fp32_1x16_x86_sse2/m:1/n:4096/real_time 5526 ns 5525 ns 49327 Bytes\u003d5.92943G/s Op\u003d741.179M/s\nbench/tanh_fp32_1x16_x86_sse2/m:4/n:1024/real_time 5536 ns 5534 ns 48796 Bytes\u003d5.9192G/s Op\u003d739.9M/s\nbench/tanh_fp32_1x16_x86_sse2/m:16/n:256/real_time 5585 ns 5583 ns 50897 Bytes\u003d5.86748G/s Op\u003d733.435M/s\nbench/tanh_fp64_1x8_x86_sse2/m:1/n:4096/real_time 20497 ns 20495 ns 14192 Bytes\u003d3.19736G/s Op\u003d199.835M/s\nbench/tanh_fp64_1x8_x86_sse2/m:4/n:1024/real_time 19400 ns 19397 ns 14340 Bytes\u003d3.37812G/s Op\u003d211.132M/s\nbench/tanh_fp64_1x8_x86_sse2/m:16/n:256/real_time 20904 ns 20902 ns 14394 Bytes\u003d3.13508G/s Op\u003d195.943M/s\n```\n\nPiperOrigin-RevId: 917613592\n" }, { "commit": "7333afbd475d2f99c1af84512fff382271758a43", "tree": "55417041aa29039d79e329e6149c5ecc97ebd648", "parents": [ "5191fee1a7b2aaa6f9f37790321b6902bf5cbdf6" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 19 03:24:51 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 03:25:31 2026" }, "message": "Optimize `floor_log2` for fp64 for non-AVX512 targets\n\nPreviously, `floor_log2` called the scalar `floor_log2`, which was quite slow.\n\n```\nname time/op time/op vs base\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 4.845µ ± 9% 4.776µ ± 7% ~ (p\u003d0.240 n\u003d6)\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 4.867µ ± 4% 4.841µ ± 2% ~ (p\u003d0.240 n\u003d6)\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 4.982µ ± 11% 4.797µ ± 4% -3.70% (p\u003d0.041 n\u003d6)\nbench/log_fp64_1x16_x86_avx2/m:1/n:4096/real_time 45.13µ ± 3% 13.87µ ± 4% -69.27% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x16_x86_avx2/m:4/n:1024/real_time 45.07µ ± 9% 14.00µ ± 7% -68.93% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x16_x86_avx2/m:16/n:256/real_time 45.20µ ± 9% 14.01µ ± 8% -69.00% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x4_x86_sse2/m:1/n:4096/real_time 46.31µ ± 6% 25.67µ ± 6% -44.56% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x4_x86_sse2/m:4/n:1024/real_time 46.15µ ± 2% 25.71µ ± 2% -44.30% (p\u003d0.002 n\u003d6)\nbench/log_fp64_1x4_x86_sse2/m:16/n:256/real_time 46.13µ ± 1% 25.80µ ± 2% -44.07% (p\u003d0.002 n\u003d6)\ngeomean 21.69µ 12.00µ -44.71%\n```\n\nMostly AI generated, here is what it had to say:\n\n✦ I have replaced the scalar floor_log2 implementations for doubles (f64x2, f64x4, etc.) with SIMD versions in the ynnpack library. This was done for ARM NEON (ARM64), x86 SSE2/AVX2, and WASM SIMD128 architectures, using the same bit-manipulation technique as the existing float (f32x4) implementations.\n\n Key changes:\n - base/simd/arm_neon_base.h: Implemented f64x2 floor_log2 using vshrq_n_s64 and bitwise operations.\n - base/simd/x86_sse2.h: Implemented f64x2 floor_log2 using _mm_srai_epi32 on high 32-bit elements to emulate 64-bit arithmetic shift for the exponent bits.\n - base/simd/x86_avx2.h: Updated f64x2 floor_log2 to use the same SIMD implementation as SSE2.\n - base/simd/wasm_simd128.h: Implemented f64x2 floor_log2 using wasm_i64x2_shr and bitwise operations.\n\nPiperOrigin-RevId: 917567607\n" }, { "commit": "5191fee1a7b2aaa6f9f37790321b6902bf5cbdf6", "tree": "2a71eb22e9b5392d8c4b0d3d2fbc9bd0f777daa2", "parents": [ "7ef5fdca247dfc94e9ac8a93fac74e558565efac" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Tue May 19 01:16:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Tue May 19 01:17:51 2026" }, "message": "Change tree reduction factor from 32 to 16, and add another level\n\nThis also fixes a bug where we clocked the accumulation mechanism based on the number of input values instead of accumulators in k1 reductions. This was inconsistent with kn reductions, which always have K \u003d 1, and also was excessive accumulation for bf16 and fp16 (which widen to fp32, which gives a lot of headroom on the precision).\n\nThis should make reductions more accurate. Performance is a mixed bag, it improves in some cases and regresses in others:\n```\nname sec/op sec/op vs base\nbench/sum_k1_bf16_fp32_avx512/real_time [256x256] 5.176µ ± 4% 5.395µ ± 12% +4.23% (p\u003d0.009 n\u003d6)\nbench/sum_kn_bf16_fp32_avx512/real_time [256x256] 4.794µ ± 4% 4.793µ ± 29% ~ (p\u003d0.818 n\u003d6)\nbench/sum_k1_fp16_fp32_avx512/real_time [256x256] 5.593µ ± 5% 4.610µ ± 4% -17.58% (p\u003d0.002 n\u003d6)\nbench/sum_kn_fp16_fp32_avx512/real_time [256x256] 4.141µ ± 3% 4.302µ ± 29% +3.87% (p\u003d0.026 n\u003d6)\nbench/sum_k1_fp32_avx512/real_time [256x256] 5.257µ ± 3% 4.325µ ± 5% -17.73% (p\u003d0.002 n\u003d6)\nbench/sum_kn_fp32_avx512/real_time [256x256] 4.228µ ± 4% 4.624µ ± 22% +9.36% (p\u003d0.002 n\u003d6)\nbench/sum_k1_fp64_avx512/real_time [256x256] 6.964µ ± 3% 6.799µ ± 5% ~ (p\u003d0.065 n\u003d6)\nbench/sum_kn_fp64_avx512/real_time [256x256] 8.662µ ± 20% 9.734µ ± 32% ~ (p\u003d0.132 n\u003d6)\nbench/sum_k1_bf16_fp32_avx2/real_time [256x256] 6.199µ ± 7% 6.726µ ± 2% +8.50% (p\u003d0.009 n\u003d6)\nbench/sum_kn_bf16_fp32_avx2/real_time [256x256] 6.279µ ± 3% 6.304µ ± 3% ~ (p\u003d0.937 n\u003d6)\nbench/sum_k1_fp32_avx/real_time [256x256] 5.431µ ± 10% 4.540µ ± 7% -16.41% (p\u003d0.002 n\u003d6)\nbench/sum_kn_fp32_avx/real_time [256x256] 5.343µ ± 8% 5.474µ ± 6% ~ (p\u003d1.000 n\u003d6)\nbench/sum_k1_fp64_avx/real_time [256x256] 8.136µ ± 4% 8.021µ ± 3% ~ (p\u003d0.589 n\u003d6)\nbench/sum_kn_fp64_avx/real_time [256x256] 10.85µ ± 14% 10.87µ ± 14% ~ (p\u003d0.818 n\u003d6)\nbench/sum_k1_fp16_fp32_f16c/real_time [256x256] 6.826µ ± 2% 7.002µ ± 4% +2.57% (p\u003d0.026 n\u003d6)\nbench/sum_kn_fp16_fp32_f16c/real_time [256x256] 5.118µ ± 2% 5.160µ ± 2% ~ (p\u003d0.240 n\u003d6)\nbench/sum_k1_fp32_sse2/real_time [256x256] 10.113µ ± 2% 8.407µ ± 6% -16.88% (p\u003d0.002 n\u003d6)\nbench/sum_kn_fp32_sse2/real_time [256x256] 7.663µ ± 16% 7.846µ ± 21% ~ (p\u003d0.132 n\u003d6)\nbench/sum_k1_bf16_fp32_sse2/real_time [256x256] 8.770µ ± 10% 11.084µ ± 3% +26.38% (p\u003d0.002 n\u003d6)\nbench/sum_kn_bf16_fp32_sse2/real_time [256x256] 9.860µ ± 1% 9.982µ ± 2% +1.24% (p\u003d0.026 n\u003d6)\nbench/sum_k1_fp64/real_time [256x256] 15.39µ ± 2% 18.08µ ± 1% +17.46% (p\u003d0.002 n\u003d6)\nbench/sum_kn_fp64/real_time [256x256] 16.77µ ± 4% 16.77µ ± 10% ~ (p\u003d0.937 n\u003d6)\nbench/sum_k1_fp16_fp32/real_time [256x256] 174.5µ ± 8% 173.9µ ± 18% ~ (p\u003d0.589 n\u003d6)\nbench/sum_kn_fp16_fp32/real_time [256x256] 45.37µ ± 2% 45.38µ ± 3% ~ (p\u003d0.699 n\u003d6)\nbench/sum_squared_k1_bf16_fp32_avx512/real_time [256x256] 5.076µ ± 4% 5.480µ ± 2% +7.97% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_avx512/real_time [256x256] 4.825µ ± 3% 4.783µ ± 30% ~ (p\u003d0.937 n\u003d6)\nbench/sum_squared_k1_fp16_fp32_avx512/real_time [256x256] 5.660µ ± 10% 4.519µ ± 5% -20.17% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp16_fp32_avx512/real_time [256x256] 4.114µ ± 6% 4.250µ ± 29% ~ (p\u003d0.132 n\u003d6)\nbench/sum_squared_k1_fp32_avx512/real_time [256x256] 6.666µ ± 3% 5.932µ ± 10% -11.01% (p\u003d0.009 n\u003d6)\nbench/sum_squared_kn_fp32_avx512/real_time [256x256] 4.536µ ± 4% 5.101µ ± 30% +12.45% (p\u003d0.009 n\u003d6)\nbench/sum_squared_k1_fp64_avx512/real_time [256x256] 9.204µ ± 5% 9.625µ ± 10% ~ (p\u003d0.485 n\u003d6)\nbench/sum_squared_kn_fp64_avx512/real_time [256x256] 9.599µ ± 11% 11.265µ ± 19% +17.36% (p\u003d0.009 n\u003d6)\nbench/sum_squared_k1_bf16_fp32_avx2/real_time [256x256] 7.023µ ± 5% 8.295µ ± 6% +18.12% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_avx2/real_time [256x256] 7.110µ ± 2% 7.484µ ± 4% +5.26% (p\u003d0.002 n\u003d6)\nbench/sum_squared_k1_fp32_avx/real_time [256x256] 7.916µ ± 5% 7.010µ ± 4% -11.46% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp32_avx/real_time [256x256] 5.849µ ± 5% 6.309µ ± 5% +7.87% (p\u003d0.002 n\u003d6)\nbench/sum_squared_k1_fp64_avx/real_time [256x256] 11.70µ ± 8% 12.27µ ± 4% +4.90% (p\u003d0.041 n\u003d6)\nbench/sum_squared_kn_fp64_avx/real_time [256x256] 11.86µ ± 5% 12.42µ ± 13% +4.69% (p\u003d0.015 n\u003d6)\nbench/sum_squared_k1_fp16_fp32_f16c/real_time [256x256] 8.402µ ± 4% 7.321µ ± 2% -12.86% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp16_fp32_f16c/real_time [256x256] 6.290µ ± 3% 6.410µ ± 3% ~ (p\u003d0.132 n\u003d6)\nbench/sum_squared_k1_fp32_sse2/real_time [256x256] 11.63µ ± 10% 11.16µ ± 2% -4.03% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_fp32_sse2/real_time [256x256] 9.218µ ± 12% 10.755µ ± 9% +16.67% (p\u003d0.002 n\u003d6)\nbench/sum_squared_k1_bf16_fp32_sse2/real_time [256x256] 10.70µ ± 2% 13.09µ ± 3% +22.38% (p\u003d0.002 n\u003d6)\nbench/sum_squared_kn_bf16_fp32_sse2/real_time [256x256] 11.27µ ± 16% 11.57µ ± 2% ~ (p\u003d0.132 n\u003d6)\nbench/sum_squared_k1_fp64/real_time [256x256] 24.54µ ± 4% 24.17µ ± 21% ~ (p\u003d0.485 n\u003d6)\nbench/sum_squared_kn_fp64/real_time [256x256] 21.43µ ± 4% 22.22µ ± 5% +3.66% (p\u003d0.026 n\u003d6)\nbench/sum_squared_k1_fp16_fp32/real_time [256x256] 185.7µ ± 15% 184.9µ ± 3% ~ (p\u003d0.937 n\u003d6)\nbench/sum_squared_kn_fp16_fp32/real_time [256x256] 48.95µ ± 2% 50.72µ ± 3% +3.60% (p\u003d0.009 n\u003d6)\ngeomean 9.426µ 9.563µ +1.44%\n```\n\nPiperOrigin-RevId: 917527134\n" }, { "commit": "7ef5fdca247dfc94e9ac8a93fac74e558565efac", "tree": "6a671dd9f0cd17a5d2a7cea8838863f443d8fb28", "parents": [ "56ac34b3f45fae2eca1f32584f7f0b279be2cf1f" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 18 20:05:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 20:06:19 2026" }, "message": "Add `round_to_bf16`\n\n- Add kernels for `round_to_bf16(x) \u003d convert(x_type, convert(bf16, x))`\n- Instead of abandoning fusion of `convert(x_type, convert(bf16, x))` when we can\u0027t change the numerics, replace the two operations with `round_to_bf16(x)` instead.\n\nPiperOrigin-RevId: 917386285\n" }, { "commit": "56ac34b3f45fae2eca1f32584f7f0b279be2cf1f", "tree": "9320610ae4e923cc0317969a9504e5ea837116fa", "parents": [ "f873466b2dab32b907f912ff7e63d1e77435572b" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Mon May 18 19:38:30 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 19:39:22 2026" }, "message": "Add a graph rewrite to fallback to fp32 when fp16 isn\u0027t supported.\n\nPiperOrigin-RevId: 917373652\n" }, { "commit": "f873466b2dab32b907f912ff7e63d1e77435572b", "tree": "86537c30cfdcb0ee8cdb9120ee3ac7021d1d1061", "parents": [ "3ca1b08328931a3abf0fe5dd2ef392adc3655407" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Mon May 18 19:09:29 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 19:11:31 2026" }, "message": "Align C++ standard to C++17 in CMake builds to be equal to Bazel builds.\n\nPiperOrigin-RevId: 917360156\n" }, { "commit": "3ca1b08328931a3abf0fe5dd2ef392adc3655407", "tree": "b2e5eeac06bd7cd615581b5f05912244ac144af9", "parents": [ "98c8ded4369968fab823ebbef877bfcdd87beb58" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 18 18:26:44 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 18:30:10 2026" }, "message": "Relax tolerance for sum squared kernel test\n\nI saw a rare flake of this test whiel running locally\n\nPiperOrigin-RevId: 917338626\n" }, { "commit": "98c8ded4369968fab823ebbef877bfcdd87beb58", "tree": "a53c57689dc1a8a4323562394168863ce1d76a41", "parents": [ "f1fe9b5c42a97e20b46c0c9f0c2e1bb570a51e70" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 18 18:25:14 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 18:26:33 2026" }, "message": "Polynomial approximation improvements for `exp` and `log`\n\n- Improve rational_approximation.py to use Remez algorithm (thanks AI), improves accuracy of `exp` and `log`.\n- Add `log_fp64` kernels.\n- Reduce `exp_fp64` numerator from degree 8 to degree 6 (while improving accuracy).\n\nRefactoring:\n- Split `exp.py` into `exp.py`, `log.py`, `erf.py`.\n- Add `eval_polynomial` helper.\n\n`exp_fp64` speedup:\n```\nname time/op time/op vs base\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.347µ ± 60% 4.863µ ± 7% -9.05% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 5.296µ ± 59% 4.816µ ± 2% -9.06% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 5.349µ ± 62% 4.751µ ± 2% -11.18% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 7.810µ ± 72% 6.786µ ± 3% -13.11% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:4/n:1024/real_time 8.008µ ± 19% 6.758µ ± 1% -15.61% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:16/n:256/real_time 7.899µ ± 13% 6.763µ ± 3% -14.37% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 11.35µ ± 3% 10.04µ ± 4% -11.56% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:4/n:1024/real_time 11.605µ ± 7% 9.930µ ± 5% -14.44% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:16/n:256/real_time 11.575µ ± 17% 9.985µ ± 14% -13.74% (p\u003d0.004 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 18.59µ ± 2% 16.57µ ± 39% ~ (p\u003d0.065 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:4/n:1024/real_time 18.60µ ± 3% 16.78µ ± 21% ~ (p\u003d0.065 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:16/n:256/real_time 18.44µ ± 3% 16.58µ ± 20% ~ (p\u003d0.065 n\u003d6)\ngeomean 9.738µ 8.576µ -11.93%\n```\n\n`log_fp64` benchmarks:\n```\n------------------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n------------------------------------------------------------------------------------------------------------------------\nbench_reference/log_double/m:1/n:4096/real_time 17434 ns 17429 ns 51083 Bytes\u003d3.75916G/s Op\u003d234.948M/s\nbench_reference/log_double/m:4/n:1024/real_time 20406 ns 20399 ns 35681 Bytes\u003d3.21163G/s Op\u003d200.727M/s\nbench_reference/log_double/m:16/n:256/real_time 21140 ns 21135 ns 31858 Bytes\u003d3.10012G/s Op\u003d193.758M/s\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 5006 ns 5004 ns 131436 Bytes\u003d13.0923G/s Op\u003d818.27M/s\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 5004 ns 5003 ns 100000 Bytes\u003d13.0979G/s Op\u003d818.62M/s\nbench/log_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 4836 ns 4835 ns 134936 Bytes\u003d13.5507G/s Op\u003d846.917M/s\nbench/log_fp64_1x16_x86_avx2/m:1/n:4096/real_time 45733 ns 45723 ns 15445 Bytes\u003d1.43303G/s Op\u003d89.5643M/s\nbench/log_fp64_1x16_x86_avx2/m:4/n:1024/real_time 45963 ns 45953 ns 15467 Bytes\u003d1.42584G/s Op\u003d89.115M/s\nbench/log_fp64_1x16_x86_avx2/m:16/n:256/real_time 45525 ns 45519 ns 15375 Bytes\u003d1.43957G/s Op\u003d89.9728M/s\nbench/log_fp64_1x4_x86_sse2/m:1/n:4096/real_time 46637 ns 46629 ns 14979 Bytes\u003d1.40525G/s Op\u003d87.8279M/s\nbench/log_fp64_1x4_x86_sse2/m:4/n:1024/real_time 49583 ns 49574 ns 14991 Bytes\u003d1.32174G/s Op\u003d82.6087M/s\nbench/log_fp64_1x4_x86_sse2/m:16/n:256/real_time 47609 ns 47594 ns 14585 Bytes\u003d1.37655G/s Op\u003d86.0346M/s\n```\n\nPiperOrigin-RevId: 917337886\n" }, { "commit": "f1fe9b5c42a97e20b46c0c9f0c2e1bb570a51e70", "tree": "da720450a9713ee7b4f879a8bc3c7caa2bd0a871", "parents": [ "01db6e141adf744bb193a3a059c76c6799a37ab6" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 18 16:50:16 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 16:54:05 2026" }, "message": "Only rewrite reduce(convert(x)) if we have a kernel for that reduction type.\n\nPiperOrigin-RevId: 917286989\n" }, { "commit": "01db6e141adf744bb193a3a059c76c6799a37ab6", "tree": "ae1b50afe045bb2ff6d59ea5bfe7ce2050a9abd2", "parents": [ "1052f90b15785d632e2dcbb6e9661c26d98271ed" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Mon May 18 16:49:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Mon May 18 16:50:36 2026" }, "message": "Fix possible infinite recursion in convert\n\nIf we try to convert from fp32 to an unsupported type (e.g. int4), we try to convert to an fp32 intermediate and try again, which leads to infinite recursion.\n\nPiperOrigin-RevId: 917286839\n" }, { "commit": "4ae8b8f607b2a35a91bba3ef491bda68a405fdc3", "tree": "2c7999de3cf7b7a322311e788624119cbed07cfd", "parents": [ "49e266f77cf7930237c9c0bec5a0a1dd2c767ffd" ], "author": { "name": "velonica0", "email": "like@mail.nankai.edu.cn", "time": "Mon May 18 03:15:45 2026" }, "committer": { "name": "velonica0", "email": "like@mail.nankai.edu.cn", "time": "Mon May 18 03:15:45 2026" }, "message": "fix bug\n" }, { "commit": "1052f90b15785d632e2dcbb6e9661c26d98271ed", "tree": "6641dc7784c047ad38950ac2aa6b3439adadb147", "parents": [ "8da42ae2bd2626ae13d39c7c936076acac8b3735" ], "author": { "name": "Richard Townsend", "email": "ritownsend@google.com", "time": "Sat May 16 01:45:32 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sat May 16 01:46:36 2026" }, "message": "[gn] Add support for building/testing AArch32\n\nAlthough AArch32 is not really the current focus for Chrome\nperformance, as we\u0027ve got the kernels around, there\u0027s no harm in\nhaving it.\n\nAlso modifies the GN test script to support testing on x64 hosts.\n\nPiperOrigin-RevId: 916281274\n" }, { "commit": "8da42ae2bd2626ae13d39c7c936076acac8b3735", "tree": "994d4df13615aebe743c24bc00c7b7fbbdf77a37", "parents": [ "1c292bfc98d0bc412721c335e72f7a188e436c8c" ], "author": { "name": "Gerardo Carranza", "email": "gcarranza@google.com", "time": "Sat May 16 00:20:41 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Sat May 16 00:21:17 2026" }, "message": "Add support for log fp16 in XNNPACK.\n\nPiperOrigin-RevId: 916257784\n" }, { "commit": "1c292bfc98d0bc412721c335e72f7a188e436c8c", "tree": "ea53af08a5bfbdf27cb5e6f1242822b05c3c410d", "parents": [ "c3ac56a5bb4eb939370253ad1e4607ad04c803b0" ], "author": { "name": "Richard Townsend", "email": "ritownsend@google.com", "time": "Fri May 15 23:05:50 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 23:06:42 2026" }, "message": "[gn] Test building AVX512\n\nBecause we don\u0027t test it in Github Actions, AVX512 support\nhas bit-rotted in GN a bit. Should take care of a cluster of\nfailures in the latest Chrome integration attempt. [1]\n\nAlso simplify/unify things for 32-bit.\n\n[1] https://chromium-review.googlesource.com/c/chromium/src/+/7615883?tab\u003dchecks\n\nPiperOrigin-RevId: 916227360\n" }, { "commit": "c3ac56a5bb4eb939370253ad1e4607ad04c803b0", "tree": "900eeacda6433e698777b7a21aeb0623bb5684cf", "parents": [ "7bf9c692b18a695b8681139508746b3959f70aad" ], "author": { "name": "Quentin Khan", "email": "qkhan@google.com", "time": "Fri May 15 19:36:08 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 19:37:21 2026" }, "message": "Add subgraph matcher target to `BUILD.gn`.\n\nPiperOrigin-RevId: 916133551\n" }, { "commit": "7bf9c692b18a695b8681139508746b3959f70aad", "tree": "fabd56fbc5ddff906daf6da709a1b664121b6956", "parents": [ "34c80155b875d102e0ffbcbd12a628477ceb610b" ], "author": { "name": "Frank Barchard", "email": "fbarchard@google.com", "time": "Fri May 15 17:57:10 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 17:58:00 2026" }, "message": "Fix ambiguous std::isfinite, std::abs, and std::fpclassify calls for _Float16 in test framework by explicitly casting to float.\n\nPiperOrigin-RevId: 916085457\n" }, { "commit": "34c80155b875d102e0ffbcbd12a628477ceb610b", "tree": "9cce7b1849f42d91229f31a540583f5ff4d9428d", "parents": [ "ace56b6162087f1926d782d39797a00fb56f2a30" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Fri May 15 16:21:58 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 16:22:57 2026" }, "message": "Make sure partial reduction splits match the loop step.\n\nPiperOrigin-RevId: 916039463\n" }, { "commit": "ace56b6162087f1926d782d39797a00fb56f2a30", "tree": "d196206776645ed3668d1fbae7fd568f97c63920", "parents": [ "49e266f77cf7930237c9c0bec5a0a1dd2c767ffd" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 15 06:20:51 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 06:24:36 2026" }, "message": "Improve `exp` kernel accuracy and correctness\n\n- Use a more accurate polynomial approximation for `exp`.\n- reference kernels accidentally used double precision intermediates due to double precision params.\n- `exp` op did not propagate NaN\n- `sign` op did not propagate NaN\n\nThis is more correct, but costs some performance:\n```\nname time/op time/op vs base\nbench_reference/exp_float/m:1/n:4096/real_time 74.80µ ± 1% 37.63µ ± 4% -49.69% (p\u003d0.002 n\u003d6)\nbench_reference/exp_float/m:4/n:1024/real_time 75.09µ ± 1% 37.22µ ± 1% -50.43% (p\u003d0.002 n\u003d6)\nbench_reference/exp_float/m:16/n:256/real_time 74.87µ ± 2% 37.40µ ± 4% -50.05% (p\u003d0.002 n\u003d6)\nbench_reference/exp_double/m:1/n:4096/real_time 67.29µ ± 1% 67.16µ ± 9% ~ (p\u003d0.818 n\u003d6)\nbench_reference/exp_double/m:4/n:1024/real_time 67.26µ ± 7% 66.93µ ± 2% ~ (p\u003d0.699 n\u003d6)\nbench_reference/exp_double/m:16/n:256/real_time 67.35µ ± 2% 67.59µ ± 2% ~ (p\u003d0.589 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.689µ ± 1% 1.857µ ± 2% +9.89% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time 1.711µ ± 4% 1.854µ ± 1% +8.36% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time 1.711µ ± 1% 1.886µ ± 2% +10.20% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time 5.004µ ± 8% 5.251µ ± 5% +4.92% (p\u003d0.041 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time 5.148µ ± 11% 5.289µ ± 5% ~ (p\u003d0.180 n\u003d6)\nbench/exp_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time 5.117µ ± 4% 5.292µ ± 4% +3.42% (p\u003d0.041 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 2.259µ ± 3% 2.552µ ± 1% +12.94% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:4/n:1024/real_time 2.273µ ± 2% 2.559µ ± 4% +12.58% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2_fma3/m:16/n:256/real_time 2.314µ ± 1% 2.600µ ± 2% +12.34% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:1/n:4096/real_time 7.145µ ± 3% 7.941µ ± 7% +11.15% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:4/n:1024/real_time 7.158µ ± 3% 7.803µ ± 5% +9.02% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2_fma3/m:16/n:256/real_time 7.199µ ± 2% 7.805µ ± 5% +8.41% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x32_x86_avx2/m:1/n:4096/real_time 2.948µ ± 3%\nbench/exp_fp32_1x32_x86_avx2/m:4/n:1024/real_time 2.951µ ± 1%\nbench/exp_fp32_1x32_x86_avx2/m:16/n:256/real_time 2.961µ ± 2%\nbench/exp_fp64_1x16_x86_avx2/m:1/n:4096/real_time 11.00µ ± 6% 11.43µ ± 3% ~ (p\u003d0.180 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:4/n:1024/real_time 11.16µ ± 3% 11.34µ ± 2% ~ (p\u003d0.093 n\u003d6)\nbench/exp_fp64_1x16_x86_avx2/m:16/n:256/real_time 11.24µ ± 4% 11.31µ ± 2% ~ (p\u003d0.699 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:1/n:4096/real_time 4.957µ ± 1% 5.654µ ± 4% +14.05% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:4/n:1024/real_time 4.937µ ± 1% 5.591µ ± 1% +13.26% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_sse2/m:16/n:256/real_time 4.919µ ± 1% 5.632µ ± 3% +14.49% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:1/n:4096/real_time 17.42µ ± 3% 18.26µ ± 1% +4.79% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:4/n:1024/real_time 17.53µ ± 3% 18.15µ ± 2% +3.57% (p\u003d0.002 n\u003d6)\nbench/exp_fp64_1x8_x86_sse2/m:16/n:256/real_time 17.51µ ± 1% 18.19µ ± 12% +3.87% (p\u003d0.002 n\u003d6)\nbench/exp_fp32_1x16_x86_avx2/m:1/n:4096/real_time 3.900µ ± 26%\nbench/exp_fp32_1x16_x86_avx2/m:4/n:1024/real_time 3.913µ ± 4%\nbench/exp_fp32_1x16_x86_avx2/m:16/n:256/real_time 3.906µ ± 3%\ngeomean 8.526µ 8.622µ -1.85% ¹\n¹ benchmark set differs from baseline; geomeans may not be comparable\n```\n\nPiperOrigin-RevId: 915813349\n" }, { "commit": "49e266f77cf7930237c9c0bec5a0a1dd2c767ffd", "tree": "1865c3cf46a10d29024cc26e51aca0301c6b3d13", "parents": [ "11fb88590dfa17facb3b7a57e41c5b2b6b1cb1a6" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Fri May 15 01:48:55 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 01:49:29 2026" }, "message": "Add optimized convert int2/int4 to int8 kernels.\n\nAt the moment, I only added AVX2 and AVX512 kernels. I think we can do better if we use AVX512-VBMI or something, but I think that\u0027s lower priority.\n\n```\n------------------------------------------------------------------------------------------------------------------------------------\nBenchmark Time CPU Iterations UserCounters...\n------------------------------------------------------------------------------------------------------------------------------------\nbench_reference_convert/int2x4_int8_t/m:1/n:4096/real_time 1333 ns 1333 ns 290663 Bytes\u003d6.14607G/s Op\u003d3.07303G/s\nbench_reference_convert/int2x4_int8_t/m:4/n:1024/real_time 1361 ns 1361 ns 306814 Bytes\u003d6.01974G/s Op\u003d3.00987G/s\nbench_reference_convert/int2x4_int8_t/m:16/n:256/real_time 1318 ns 1317 ns 319517 Bytes\u003d6.21558G/s Op\u003d3.10779G/s\nbench/convert_int2_to_int8_1x64_x86_avx512f_avx512bw/m:1/n:4096/real_time 63.8 ns 63.8 ns 6551296 Bytes\u003d128.357G/s Op\u003d64.1784G/s\nbench/convert_int2_to_int8_1x64_x86_avx512f_avx512bw/m:4/n:1024/real_time 65.1 ns 65.1 ns 6545915 Bytes\u003d125.786G/s Op\u003d62.8929G/s\nbench/convert_int2_to_int8_1x64_x86_avx512f_avx512bw/m:16/n:256/real_time 71.6 ns 71.6 ns 4778839 Bytes\u003d114.456G/s Op\u003d57.228G/s\nbench/convert_int2_to_int8_1x32_x86_avx2/m:1/n:4096/real_time 95.8 ns 95.8 ns 4580022 Bytes\u003d85.5078G/s Op\u003d42.7539G/s\nbench/convert_int2_to_int8_1x32_x86_avx2/m:4/n:1024/real_time 94.4 ns 94.4 ns 4572399 Bytes\u003d86.7469G/s Op\u003d43.3734G/s\nbench/convert_int2_to_int8_1x32_x86_avx2/m:16/n:256/real_time 96.4 ns 96.4 ns 4480847 Bytes\u003d85.0169G/s Op\u003d42.5084G/s\nbench_reference_convert/int4x2_int8_t/m:1/n:4096/real_time 1675 ns 1674 ns 248804 Bytes\u003d4.89207G/s Op\u003d2.44603G/s\nbench_reference_convert/int4x2_int8_t/m:4/n:1024/real_time 1695 ns 1695 ns 249163 Bytes\u003d4.83283G/s Op\u003d2.41642G/s\nbench_reference_convert/int4x2_int8_t/m:16/n:256/real_time 1806 ns 1806 ns 233733 Bytes\u003d4.53658G/s Op\u003d2.26829G/s\nbench/convert_int4_to_int8_1x64_x86_avx512f_avx512bw/m:1/n:4096/real_time 64.2 ns 64.2 ns 6632649 Bytes\u003d127.538G/s Op\u003d63.7688G/s\nbench/convert_int4_to_int8_1x64_x86_avx512f_avx512bw/m:4/n:1024/real_time 64.2 ns 64.2 ns 6560923 Bytes\u003d127.632G/s Op\u003d63.8158G/s\nbench/convert_int4_to_int8_1x64_x86_avx512f_avx512bw/m:16/n:256/real_time 64.8 ns 64.8 ns 6320290 Bytes\u003d126.449G/s Op\u003d63.2245G/s\nbench/convert_int4_to_int8_1x32_x86_avx2/m:1/n:4096/real_time 116 ns 116 ns 3617014 Bytes\u003d70.7825G/s Op\u003d35.3913G/s\nbench/convert_int4_to_int8_1x32_x86_avx2/m:4/n:1024/real_time 119 ns 119 ns 3538506 Bytes\u003d68.6638G/s Op\u003d34.3319G/s\nbench/convert_int4_to_int8_1x32_x86_avx2/m:16/n:256/real_time 134 ns 134 ns 3077822 Bytes\u003d61.0267G/s Op\u003d30.5134G/s\n\n```\n\nPiperOrigin-RevId: 915727140\n" }, { "commit": "11fb88590dfa17facb3b7a57e41c5b2b6b1cb1a6", "tree": "300e221e67fdc8f42a6fe2c9a2ea2c16b906255c", "parents": [ "9ab80cd66ef6f077e87551424c07e6327a374b3b" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Fri May 15 01:07:27 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Fri May 15 01:10:03 2026" }, "message": "Implement round to nearest even for float -\u003e bf16 conversions\n\nAlso adds float -\u003e bf16 conversions to SIMD benchmarks\n\nThis is a performance regression, but I think it is worth the accuracy improvement.\n\n```\nname sec/op sec/op vs base\nbench/convert_fp32_to_bf16_1x64_x86_avx512f_avx512bw/m:1/n:4096/real_time 188.1n ± 2% 421.3n ± 1% +123.94% (p\u003d0.002 n\u003d6)\nbench/convert_fp32_to_bf16_1x64_x86_avx512f_avx512bw/m:4/n:1024/real_time 192.6n ± 3% 418.3n ± 1% +117.19% (p\u003d0.002 n\u003d6)\nbench/convert_fp32_to_bf16_1x64_x86_avx512f_avx512bw/m:16/n:256/real_time 197.3n ± 3% 423.9n ± 4% +114.86% (p\u003d0.002 n\u003d6)\nbench/convert_fp32_to_bf16_1x16_x86_avx2/m:1/n:4096/real_time 270.8n ± 3% 753.3n ± 1% +178.14% (p\u003d0.002 n\u003d6)\nbench/convert_fp32_to_bf16_1x16_x86_avx2/m:4/n:1024/real_time 277.5n ± 1% 769.6n ± 2% +177.32% (p\u003d0.002 n\u003d6)\nbench/convert_fp32_to_bf16_1x16_x86_avx2/m:16/n:256/real_time 310.5n ± 2% 795.2n ± 1% +156.11% (p\u003d0.002 n\u003d6)\ngeomean 234.6n 570.4n +143.11%\n```\n\nAs a side effect of the new test coverage in this change, I found a lot of bugs, like hitting UB in float -\u003e int conversions, that motivated the following refactoring:\n- We really shouldn\u0027t have a truncating `cast` at all for `float` -\u003e `int` conversions if we can avoid it.\n- Then I realized that some uses of cast really should be `round_float_to_int`, but some uses of this want to do `float` -\u003e `float` (`half` or `bfloat16`) conversions, so we should rename `round_float_to_int` to `cast` to allow for templates that handle both `int` and `float`.\n- Now the old cast is a duplicate symbol, we should just remove it.\n- `saturate_cast` is similarly redundant, we should just never not saturate when casting to any integer.\n- When casting floats to integers, we clamped at a value lower in magnitude than the int max/min. This was a small bug, fixed in this change.\n\nI think that we never need cast with overflow and no rounding, and is in fact a bug prone helper to offer.\n\nPiperOrigin-RevId: 915714096\n" }, { "commit": "9ab80cd66ef6f077e87551424c07e6327a374b3b", "tree": "38a000a7e88f78e19cf68e9b99b061df120ea5ff", "parents": [ "95ee916af734c126e7bf65d83cf0c08673a32f16" ], "author": { "name": "Volodymyr Kysenko", "email": "vksnk@google.com", "time": "Thu May 14 17:48:20 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 14 17:48:57 2026" }, "message": "Allow adding function own loops even if some of its non-trivial loops has been already fused.\n\nThis really was just a heuristic to avoid creating too many loops I added because I was afraid of creating unnecessary overhead.\n\nPiperOrigin-RevId: 915507602\n" }, { "commit": "95ee916af734c126e7bf65d83cf0c08673a32f16", "tree": "31fc7c8c180698ec416fdb0816011e53266df1db", "parents": [ "d72fa85c9bd2ee33365497fa6264ff84f4b2b593" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 14 17:35:22 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 14 17:35:57 2026" }, "message": "Use a better unroll factor for log2_fp32_sse2\n\n```\nname time/op time/op vs base\nbench/log_fp32_1x8_x86_sse2/m:1/n:4096/real_time 6.296µ ± 3%\nbench/log_fp32_1x8_x86_sse2/m:4/n:1024/real_time 6.478µ ± 5%\nbench/log_fp32_1x8_x86_sse2/m:16/n:256/real_time 6.452µ ± 5%\nbench/log_fp32_1x16_x86_sse2/m:1/n:4096/real_time 6.055µ ± 3%\nbench/log_fp32_1x16_x86_sse2/m:4/n:1024/real_time 6.068µ ± 2%\nbench/log_fp32_1x16_x86_sse2/m:16/n:256/real_time 6.049µ ± 2%\ngeomean 6.408µ 6.057µ ? ¹ ²\n¹ benchmark set differs from baseline; geomeans may not be comparable\n² ratios must be \u003e0 to compute geomean\n```\n\nPiperOrigin-RevId: 915500812\n" }, { "commit": "d72fa85c9bd2ee33365497fa6264ff84f4b2b593", "tree": "dbb188bc987d5b187aef97f8e5456b37b9feaf71", "parents": [ "4fad5b3975c43ac74426e07f30af2e6ea419aa3d" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 14 16:41:12 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 14 16:42:11 2026" }, "message": "Improve log_fp32 kernels\n\n- Avoid the need for the sqrt(2) multiplier by making the polynomial approximate log2 on [1, 2).\n- Use more accurate polynomial\n\nThis is both more accurate and faster on x86 (with a change to the kernel unrolling on SSE2).\n```\nname time/op time/op vs base\nbench/log_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time 1.620µ ± 2% 1.547µ ± 3% -4.48% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time 1.624µ ± 2% 1.560µ ± 6% -3.95% (p\u003d0.041 n\u003d6)\nbench/log_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time 1.631µ ± 2% 1.563µ ± 2% -4.17% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx2_fma3/m:1/n:4096/real_time 3.573µ ± 3% 2.888µ ± 3% -19.16% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx2_fma3/m:4/n:1024/real_time 3.605µ ± 3% 2.918µ ± 11% -19.06% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx2_fma3/m:16/n:256/real_time 3.597µ ± 1% 2.890µ ± 5% -19.67% (p\u003d0.002 n\u003d6)\nbench/log_fp32_1x32_x86_avx2/m:1/n:4096/real_time 3.592µ ± 5% 3.432µ ± 3% -4.45% (p\u003d0.041 n\u003d6)\nbench/log_fp32_1x32_x86_avx2/m:4/n:1024/real_time 3.592µ ± 3% 3.492µ ± 6% ~ (p\u003d0.132 n\u003d6)\nbench/log_fp32_1x32_x86_avx2/m:16/n:256/real_time 3.571µ ± 4% 3.471µ ± 3% -2.80% (p\u003d0.026 n\u003d6)\nbench/log_fp32_1x8_x86_sse2/m:1/n:4096/real_time 6.682µ ± 3%\nbench/log_fp32_1x8_x86_sse2/m:4/n:1024/real_time 6.750µ ± 3%\nbench/log_fp32_1x8_x86_sse2/m:16/n:256/real_time 6.831µ ± 4%\nbench/log_fp32_1x16_x86_sse2/m:1/n:4096/real_time 6.405µ ± 4%\nbench/log_fp32_1x16_x86_sse2/m:4/n:1024/real_time 6.466µ ± 4%\nbench/log_fp32_1x16_x86_sse2/m:16/n:256/real_time 6.531µ ± 3%\n```\n\nPiperOrigin-RevId: 915474623\n" }, { "commit": "4fad5b3975c43ac74426e07f30af2e6ea419aa3d", "tree": "aab7c6a6bf02b342ea7e0037556d9823848483cb", "parents": [ "fe166973d8e38aeee4ca9a08bb025eca28821b9a" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 14 05:37:04 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 14 05:39:28 2026" }, "message": "Disable static_slice test until slinky bug is fixed\n\nPiperOrigin-RevId: 915232622\n" }, { "commit": "393da7ddec358bcaa2a1ec1b8e50634a010db4a9", "tree": "46d64a4f008884d3f5e0ed024312430c3ea27fb7", "parents": [ "f351319454aaf6612e4dc67394bbe75eb9601815" ], "author": { "name": "Ken Unger", "email": "ken.j.unger@gmail.com", "time": "Thu May 14 04:51:07 2026" }, "committer": { "name": "Ken Unger", "email": "ken.j.unger@gmail.com", "time": "Thu May 14 04:51:07 2026" }, "message": "add rvv kernel for f16-vlog\n" }, { "commit": "fe166973d8e38aeee4ca9a08bb025eca28821b9a", "tree": "5a9bb65fca3395ba8f8ecfad5dfe3d5fe99ed5c6", "parents": [ "53007d6933e46552850148e90fc49747516fd4cf" ], "author": { "name": "Dillon Sharlet", "email": "dsharlet@google.com", "time": "Thu May 14 03:45:16 2026" }, "committer": { "name": "XNNPACK Team", "email": "xnnpack-github-robot@google.com", "time": "Thu May 14 03:46:27 2026" }, "message": "Disable static_slice test until slinky bug is fixed\n\nPiperOrigin-RevId: 915199617\n" } ], "next": "53007d6933e46552850148e90fc49747516fd4cf" }