0000000000000000000000000000000000000000 53ff85d4d36335ccfe6ff6c21233d00743c6bf46 osmarks 1784224365 +0000 commit (initial): Build independent Rust IPU compiler and loader stack 53ff85d4d36335ccfe6ff6c21233d00743c6bf46 4fcf1b6cf5a5d4601337c579b82670e7dd7fc897 osmarks 1784240319 +0000 commit: Add Rust exchange planner and direct runtime 4fcf1b6cf5a5d4601337c579b82670e7dd7fc897 38f223bded0d7662346ad0d168fadeaebbf07192 osmarks 1784240778 +0000 commit: Execute exchange launches with coordinator role 38f223bded0d7662346ad0d168fadeaebbf07192 a799ccffb2cefd77e73cf8465e6f50390fb3d21e osmarks 1784241133 +0000 commit: Relocate exchange fixture memory safely a799ccffb2cefd77e73cf8465e6f50390fb3d21e 442b3477a9c7a9911478701461b8162539b4ee49 osmarks 1784242394 +0000 commit: Run compiler-generated multi-pass reduction 442b3477a9c7a9911478701461b8162539b4ee49 28799f635f8c2cff5dbcd6245759934c6e56bf5b osmarks 1784242627 +0000 commit: Encode IPU host exchange packet headers 28799f635f8c2cff5dbcd6245759934c6e56bf5b 79ab8b4ff7eef82c0be327c8ab5d0202975fc60f osmarks 1784242685 +0000 commit: Plan arbitrary host exchange ranges 79ab8b4ff7eef82c0be327c8ab5d0202975fc60f 0f130473ec152292e45cea1f475276563078be72 osmarks 1784244313 +0000 commit: Add pinned host buffers and TDI memory reads 0f130473ec152292e45cea1f475276563078be72 26c719da478d319ef40129a035d16fe3dbb8665d osmarks 1784244313 +0000 commit: Assemble host exchange device phases 26c719da478d319ef40129a035d16fe3dbb8665d 05aae63f2cb3c36c989ed02877ef51c69bcfad9f osmarks 1784244313 +0000 commit: Prototype native host output scheduling 05aae63f2cb3c36c989ed02877ef51c69bcfad9f f7806be818cdcc4eaaf5895434e3f68dc956af30 osmarks 1784244629 +0000 commit: Document native host output boundary f7806be818cdcc4eaaf5895434e3f68dc956af30 3068e05540436c37c11283d7c112a8441a6aec6c osmarks 1784244708 +0000 commit: Keep native host output explicitly experimental 3068e05540436c37c11283d7c112a8441a6aec6c 5bed6ef069e63fb2fb178f074f8d62728f1741c4 osmarks 1784275264 +0000 commit: Name exchange runtime ABI constants 5bed6ef069e63fb2fb178f074f8d62728f1741c4 cde02fbae66760cc56f8fdf0a456840ef2bb125f osmarks 1784275612 +0000 commit: Validate exchange fixture memory layouts cde02fbae66760cc56f8fdf0a456840ef2bb125f dd21bb77b913f48a0eef1c8b02f64a09a13a8ef2 osmarks 1784276598 +0000 commit: Remove hardcoded reduction runtime patches dd21bb77b913f48a0eef1c8b02f64a09a13a8ef2 4655697ed1939a39067300bf1ec221c69468a9d0 osmarks 1784280938 +0000 commit: Allow sync packet origin to run exchange work 4655697ed1939a39067300bf1ec221c69468a9d0 9a48d77db752e2666f020993b3d41a95d050d7b8 osmarks 1784284514 +0000 commit: Generate global sync configuration in Rust 9a48d77db752e2666f020993b3d41a95d050d7b8 6d51da75fce5799a970b87957832d381b0c9996c osmarks 1784284897 +0000 commit: Remove legacy one-shot exchange path 6d51da75fce5799a970b87957832d381b0c9996c b2b31c5f0e4b1e5eaa80505adff84bce26302080 osmarks 1784285453 +0000 commit: Run all disjoint exchanges concurrently b2b31c5f0e4b1e5eaa80505adff84bce26302080 e119319946132d7a45ee00d43d7f16355f5db9bb osmarks 1784286162 +0000 commit: Run direct arbitrary multicast e119319946132d7a45ee00d43d7f16355f5db9bb e4eb9ab37b70aa32033b446b857e3a2bcf04d6d9 osmarks 1784289167 +0000 commit: Measure exchange plan event timing e4eb9ab37b70aa32033b446b857e3a2bcf04d6d9 4c0a73768d26df5a7e720ef26b8fc930faa22921 osmarks 1784290397 +0000 commit: Schedule conflicting exchange roles in one launch 4c0a73768d26df5a7e720ef26b8fc930faa22921 8207a522255e5af3e6db9bc73c8822973aece829 osmarks 1784292008 +0000 commit: Add graph-driven C600 hardware acceptance tests 8207a522255e5af3e6db9bc73c8822973aece829 7e54b0a268af744f66720a57f7bb070485c7bc7b osmarks 1784292909 +0000 commit: Make host exchange requests self-contained 7e54b0a268af744f66720a57f7bb070485c7bc7b 32cacbc40daf115051981c1bf856d1c70183226b osmarks 1784299289 +0000 commit: Run generated host exchange without debug IO 32cacbc40daf115051981c1bf856d1c70183226b a36142ac508a3a7d58fa57cfd4175056973f29f1 osmarks 1784299321 +0000 commit: Document direct host exchange status a36142ac508a3a7d58fa57cfd4175056973f29f1 48353f05a16e29974b282470309d53d727f3c93c osmarks 1784307767 +0000 commit: Allocate constrained exchange staging 48353f05a16e29974b282470309d53d727f3c93c f9c7d441619511696908222f2543ad173d4248f4 osmarks 1784307767 +0000 commit: Run allocated host exchange round trip f9c7d441619511696908222f2543ad173d4248f4 81089951545f921b3a0f4cb6bb0910c8e3568504 osmarks 1784308146 +0000 commit: Clarify host transfer address limits 81089951545f921b3a0f4cb6bb0910c8e3568504 32cbdd016084d170a65833124000653b6b83ce04 osmarks 1784318944 +0000 commit: Run D2H directly from arbitrary tiles 32cbdd016084d170a65833124000653b6b83ce04 61955fc688cc451c8f52670810f42595c09da2e9 osmarks 1784319798 +0000 commit: Synchronize every exchange epoch 61955fc688cc451c8f52670810f42595c09da2e9 11db53a026ae87c3c1be0fdad764d8ff8e958b41 osmarks 1784322385 +0000 commit: Add randomized exchange hardware tests 11db53a026ae87c3c1be0fdad764d8ff8e958b41 03fc68801c760695e043bde96329c8422ca799de osmarks 1784322692 +0000 commit: Restore host transfer and multicast capabilities 03fc68801c760695e043bde96329c8422ca799de 3a0c2e2884d4b60745373293bb12ad5b829ea5fd osmarks 1784325089 +0000 commit: Investigate split remote host exchange endpoints 3a0c2e2884d4b60745373293bb12ad5b829ea5fd 690c62c543dec5ec1536bdf74055fc8b68a062d2 osmarks 1784325099 +0000 revert: Revert "Investigate split remote host exchange endpoints" 690c62c543dec5ec1536bdf74055fc8b68a062d2 f11076e4b63ad54022d988d64b5e1e0a9a80df7a osmarks 1784325152 +0000 cherry-pick: Investigate split remote host exchange endpoints f11076e4b63ad54022d988d64b5e1e0a9a80df7a 427ed23f7eef690b18dca11c9f4d36d00fe1cd1c osmarks 1784325695 +0000 commit: Test host-assisted inter-command barriers 427ed23f7eef690b18dca11c9f4d36d00fe1cd1c 9327327966ca4060f58b457ffe9c0e24733c0ac0 osmarks 1784325695 +0000 revert: Revert "Test host-assisted inter-command barriers" 9327327966ca4060f58b457ffe9c0e24733c0ac0 8a25c8173d342542961e0f78c09952a73bce843e osmarks 1784325695 +0000 revert: Revert "Investigate split remote host exchange endpoints" 8a25c8173d342542961e0f78c09952a73bce843e 47b034ab83a476817e26cf672a9d9bcc8458c93b osmarks 1784325941 +0000 commit: Separate offline checks from hardware acceptance 47b034ab83a476817e26cf672a9d9bcc8458c93b 3f19a4e7b25e565e83957d8b0c79ff834592eebd osmarks 1784326169 +0000 commit: Report every critical hardware pathway 3f19a4e7b25e565e83957d8b0c79ff834592eebd 3f19a4e7b25e565e83957d8b0c79ff834592eebd osmarks 1784326251 +0000 reset: moving to 3f19a4e7b25e565e83957d8b0c79ff834592eebd 3f19a4e7b25e565e83957d8b0c79ff834592eebd 79d1845a5b9fbbd2eb5801de8440909bd1170f1f osmarks 1784330212 +0000 commit: Compose direct host transfers and harden hardware acceptance 79d1845a5b9fbbd2eb5801de8440909bd1170f1f 979c467210319f4d6955226911a9888c3c1a7f2d osmarks 1784331411 +0000 commit: Serialize D2H commands and derive host endpoints 979c467210319f4d6955226911a9888c3c1a7f2d 73861c5243d07c1ebf9448e50e2f1ad8ab3aeb8f osmarks 1784332591 +0000 commit: Prevent stale host buffers from passing D2H tests 73861c5243d07c1ebf9448e50e2f1ad8ab3aeb8f 5bac4e3cf49b2c27e5bca4a798ee7a48f162af8b osmarks 1784335681 +0000 commit: Make broken D2H a default hardware gate 5bac4e3cf49b2c27e5bca4a798ee7a48f162af8b 72881aaaf2138b0400b281997ae4af7f0d2b4893 osmarks 1784336272 +0000 commit: Make hardware capabilities a truthful serial gate 72881aaaf2138b0400b281997ae4af7f0d2b4893 0fa45b8103fc469b01079427a7d3966457a36b9c osmarks 1784340344 +0000 commit: Make hardware capability tests independently truthful 0fa45b8103fc469b01079427a7d3966457a36b9c f3d73e160a3d1497c9399b8111c9a3a00db9cf3a osmarks 1784340370 +0000 commit (amend): Make hardware capability tests independently truthful f3d73e160a3d1497c9399b8111c9a3a00db9cf3a a1bacefd60776da5e88bc89d0bf3cf1fe1923a4b osmarks 1784343171 +0000 commit: Restore truthful transport gates and static D2D timing a1bacefd60776da5e88bc89d0bf3cf1fe1923a4b 79daa7cf3c5486150f4950ce8b094368b0db8623 osmarks 1784344313 +0000 commit: Document unresolved native D2H protocol roles 79daa7cf3c5486150f4950ce8b094368b0db8623 b115cadfe55d401f6b528952dff1f086d0a87d9e osmarks 1784345388 +0000 commit: Match local host output protocol b115cadfe55d401f6b528952dff1f086d0a87d9e 53ce72b66205c687f219777dc868e4be4afd330b osmarks 1784346627 +0000 commit: Require completed host transfer calls 53ce72b66205c687f219777dc868e4be4afd330b 4dd42cc932afbdda53420e05d7bf1fc15a20cede osmarks 1784347153 +0000 commit: Generate multi-packet host write streams 4dd42cc932afbdda53420e05d7bf1fc15a20cede 1ec44ce3dbc42528b82afbee2f65d27fa1a5077e osmarks 1784349432 +0000 commit: Keep host transport acceptance truthful 1ec44ce3dbc42528b82afbee2f65d27fa1a5077e c2cbd7696e7c4c9e04dbdc2f1b365d2631728f9c osmarks 1784351372 +0000 commit: Stop unit tests from certifying transport guesses c2cbd7696e7c4c9e04dbdc2f1b365d2631728f9c 39f9476cc2e8eb9a9928ea53c34fec45412d3a58 osmarks 1784354520 +0000 commit: Make exchange test claims match their evidence 39f9476cc2e8eb9a9928ea53c34fec45412d3a58 f8f4059b84d543826b8de6f00dab20da81c9715d osmarks 1784354544 +0000 commit: Diagnose host exchange scheduling milestones f8f4059b84d543826b8de6f00dab20da81c9715d 0fd8fb3e4cda6eb2619cc664c4fbaf089356742d osmarks 1784355019 +0000 commit: Separate initialized D2D from H2D composition 0fd8fb3e4cda6eb2619cc664c4fbaf089356742d 64b7f1811140e32df5422ced182af5c7faef6623 osmarks 1784355672 +0000 commit: Expose host exchange boundary diagnostics 64b7f1811140e32df5422ced182af5c7faef6623 08b4d014807e1354b8d376f44d17b3eb2585ec7a osmarks 1784355739 +0000 commit: Name the composed transport gate precisely 08b4d014807e1354b8d376f44d17b3eb2585ec7a b99eeca4f31e19b4519813aac234d48673be3a7e osmarks 1784362579 +0000 commit: Implement command-scoped host and exchange dispatch b99eeca4f31e19b4519813aac234d48673be3a7e fe5bb9d6af4881bc88358450b16ccddabcde507a osmarks 1784365870 +0000 commit: Align host exchange runtime with device protocol fe5bb9d6af4881bc88358450b16ccddabcde507a cc3945c1bf1f74a7eb0cfeb3456322c34fa2740f osmarks 1784369125 +0000 commit: Remove legacy tile command interpreter format cc3945c1bf1f74a7eb0cfeb3456322c34fa2740f 67d41424f39bfd48ee1e799a63a8f5fbc613c2fb osmarks 1784369180 +0000 commit: Define static per-tile execution architecture 67d41424f39bfd48ee1e799a63a8f5fbc613c2fb 3a0b3f843fa4f0996b709b66f8985b8db8f71762 osmarks 1784372538 +0000 commit: Record static lowering hardware boundary 3a0b3f843fa4f0996b709b66f8985b8db8f71762 bdb354f86be0bac5045e33f08ea714072dc72757 osmarks 1784373896 +0000 commit: Replace graph dispatcher with static tile programs bdb354f86be0bac5045e33f08ea714072dc72757 b59a2f6718e233f73f8242a49f59a9a4cbca9dca osmarks 1784374779 +0000 commit: Harden static D2D code generation b59a2f6718e233f73f8242a49f59a9a4cbca9dca d0f84653cb192f614a45f2d70a49359629170841 osmarks 1784376609 +0000 commit: Remove fabricated global sync generator d0f84653cb192f614a45f2d70a49359629170841 94f24bbfaeb30516352f5a4049b4ef7b329fe460 osmarks 1784377184 +0000 commit: Distinguish internal barriers from GSP setup 94f24bbfaeb30516352f5a4049b4ef7b329fe460 cf9c3c89215b070e5ca9f4a0cea573b456379048 osmarks 1784381712 +0000 commit: Generate reusable static D2D phase barriers cf9c3c89215b070e5ca9f4a0cea573b456379048 b94047ad299ca3fd361d7eae2da28aa2e3cbb9c8 osmarks 1784382188 +0000 commit: Restore numeric environment parsing helper b94047ad299ca3fd361d7eae2da28aa2e3cbb9c8 a32ba8b2ee48041d83d6e8d14d54717d5449e87b osmarks 1784383378 +0000 commit: Use internal synchronization for static D2D phases a32ba8b2ee48041d83d6e8d14d54717d5449e87b 427c7766482bf88fac024fb5e438af633ebdf2c3 osmarks 1784389491 +0000 commit: Use point schedules for singleton exchanges 427c7766482bf88fac024fb5e438af633ebdf2c3 adcb5aca456b80870414c58e83083e882ca35ab5 osmarks 1784391808 +0000 commit: Match SDK host target operation roles adcb5aca456b80870414c58e83083e882ca35ab5 87b79edb2e8658bae73842908aa314995eded5f5 osmarks 1784393096 +0000 commit: Encode host hierarchy XREQ role 87b79edb2e8658bae73842908aa314995eded5f5 e83594f5e59ea8b6853e5481772ace025a17ec97 osmarks 1784396376 +0000 commit: Generate static host transfer phases e83594f5e59ea8b6853e5481772ace025a17ec97 28396f24d3dd0d569279995669fd2f7f3bb14ed8 osmarks 1784397336 +0000 commit: Generalize static host exchange routing 28396f24d3dd0d569279995669fd2f7f3bb14ed8 d59105fa67bff38d6d67aea570db84e8ac1b5595 osmarks 1784399905 +0000 commit: Complete arbitrary-tile static host exchange d59105fa67bff38d6d67aea570db84e8ac1b5595 f4b654597d08df263fa2497e0153133f6ce94e02 osmarks 1784400280 +0000 commit: Test complete host compute exchange pipeline f4b654597d08df263fa2497e0153133f6ce94e02 e5bae5f438def98e5e78c2e7b06cbc9e7dbe9edf osmarks 1784401235 +0000 commit: Compress sparse static host schedules e5bae5f438def98e5e78c2e7b06cbc9e7dbe9edf ab11ff544b6818a5b6556b2945599855cc5e82c8 osmarks 1784402313 +0000 commit: Add sharded FP32 block GEMM ab11ff544b6818a5b6556b2945599855cc5e82c8 6295f827226d20b5797e5d51732edb5979062196 osmarks 1784403582 +0000 commit: Add per-tile Cap'n Proto cycle profiles 6295f827226d20b5797e5d51732edb5979062196 01170c57646d6a912de6fdda940eb9890a840836 osmarks 1784404682 +0000 commit: Jitter host synchronization writes in hardware tests 01170c57646d6a912de6fdda940eb9890a840836 292119b9cf896d103b62087b74595f447db18a09 osmarks 1784405431 +0000 commit: Diagnose GEMM output against tile SRAM 292119b9cf896d103b62087b74595f447db18a09 e5f4706c7185ca94c2c1c6cb6dd9685adc4e31a0 osmarks 1784405567 +0000 commit: Repeat GEMM hardware verification in process e5f4706c7185ca94c2c1c6cb6dd9685adc4e31a0 b1cb5338d631b025a05729a2a9d373892709a6fb osmarks 1784406419 +0000 commit: Exclude captured HSP handshakes from device configuration b1cb5338d631b025a05729a2a9d373892709a6fb 79d7dd16abc1b5f00a0862060a5a74dba5c53d28 osmarks 1784406498 +0000 commit: Inspect the actual mismatched GEMM block 79d7dd16abc1b5f00a0862060a5a74dba5c53d28 0463ce81e564a59e60539bf1198900ce8593a411 osmarks 1784406583 +0000 commit: Flush HSP mark writes before polling completion 0463ce81e564a59e60539bf1198900ce8593a411 7d71108c78318854ba3767a763d13ecfd36b05c8 osmarks 1784408267 +0000 commit: Complete host transfers before copying D2H data 7d71108c78318854ba3767a763d13ecfd36b05c8 11c5f1536afacfdfbb5ffa051a3a6453d0892d14 osmarks 1784408365 +0000 commit: Vectorize blocked FP32 GEMM inner products 11c5f1536afacfdfbb5ffa051a3a6453d0892d14 da5614c9d69893c238cbd4932462fccf7ab3ca9d osmarks 1784409910 +0000 commit: Index GEMM allocations during schedule lowering da5614c9d69893c238cbd4932462fccf7ab3ca9d e24570e76ffecbc73e91335587f4379ae7bf1e6b osmarks 1784409958 +0000 commit: Compact static host transfer programs e24570e76ffecbc73e91335587f4379ae7bf1e6b e2277fb7cc465e7bc1f56553fcc95b3e023758f3 osmarks 1784441989 +0000 commit: Accelerate host packet address allocation e2277fb7cc465e7bc1f56553fcc95b3e023758f3 cd744c96354aab082b1daa9569c1184f574b742e osmarks 1784443095 +0000 commit: Use AMP for blocked FP32 GEMM cd744c96354aab082b1daa9569c1184f574b742e 0e79a7ab1d07cdfc1f9539394546f958c367ae93 osmarks 1784445408 +0000 commit: Pipeline FP32 GEMM with PACE 0e79a7ab1d07cdfc1f9539394546f958c367ae93 3aa4f09f1e00b5bb05b5dc09455af7eb7b0bdf30 osmarks 1784445932 +0000 commit: Parallelize GEMM staging copies 3aa4f09f1e00b5bb05b5dc09455af7eb7b0bdf30 29214b97cdbf959331dead38ebbc00c0f7746d64 osmarks 1784446087 +0000 commit: Index package blobs by digest 29214b97cdbf959331dead38ebbc00c0f7746d64 a1915e7c96968cc90d88d24692a81eaa70d0dbe4 osmarks 1784446223 +0000 commit: Parallelize GEMM output initialization a1915e7c96968cc90d88d24692a81eaa70d0dbe4 feeec48998d404467c1b3697001fe8a29678fbd0 osmarks 1784447063 +0000 commit: Reuse host exchange packet storage feeec48998d404467c1b3697001fe8a29678fbd0 e9a9fdf999ac02ac56e64a9cad4b81f0f79cdb38 osmarks 1784447226 +0000 commit: Use both exchange windows for GEMM operands e9a9fdf999ac02ac56e64a9cad4b81f0f79cdb38 e41ff8fa72f81bc4f3bc8d61d10bc7e64bf456d0 osmarks 1784447693 +0000 commit: Schedule paired GEMM exchanges in one launch e41ff8fa72f81bc4f3bc8d61d10bc7e64bf456d0 abcec45575e83dfd19b5a9cac77fd042a9a570e6 osmarks 1784447887 +0000 commit: Remove superseded scalar GEMM kernel abcec45575e83dfd19b5a9cac77fd042a9a570e6 e5f26a58102dd7a0124bfbf2dbbaeb4c95eef4c0 osmarks 1784448051 +0000 commit: Specialize PACE GEMM for shorter row shards e5f26a58102dd7a0124bfbf2dbbaeb4c95eef4c0 9b72f785d594fdd96113e79522c708e386471027 osmarks 1784448447 +0000 commit: Shard GEMM rows across more tiles 9b72f785d594fdd96113e79522c708e386471027 80e64aee01eed7424fd07f13b3736e6141d28c19 osmarks 1784449154 +0000 commit: profiler renderer 80e64aee01eed7424fd07f13b3736e6141d28c19 b7f8d6df7db3bc3232d49b4685d58e4638317c4f osmarks 1784449507 +0000 commit: Parameterize GEMM row sharding b7f8d6df7db3bc3232d49b4685d58e4638317c4f 9eaf0f894643d0d190225e0465e5b10b4093e776 osmarks 1784450795 +0000 commit: Wave GEMM outputs across tiles 9eaf0f894643d0d190225e0465e5b10b4093e776 1b07a75f0c641f81cd3a67b0e0c07f1f2ef63c60 osmarks 1784453904 +0000 commit: Compact detailed cycle profiling 1b07a75f0c641f81cd3a67b0e0c07f1f2ef63c60 b449ee69daea85d1292fd2b83a21cc5e4281508f osmarks 1784456902 +0000 commit: Add semantic and memory profiles b449ee69daea85d1292fd2b83a21cc5e4281508f 3c57d4e2499a62a776cc825b62320f5c5ffa4b47 osmarks 1784457493 +0000 commit: Allow large Cap'n Proto artifacts 3c57d4e2499a62a776cc825b62320f5c5ffa4b47 511287d69dc826c238c5785baf39e3a4b2c825cb osmarks 1784459649 +0000 commit: Make semantic profile reports scalable 511287d69dc826c238c5785baf39e3a4b2c825cb 34fb6641a0c4df32140136dfe2090bab3c5eaef5 osmarks 1784459934 +0000 commit: Restore semantic flamegraph drilldown 34fb6641a0c4df32140136dfe2090bab3c5eaef5 c22fed233b0caab0393cf6dbce5cf25d5a4f2df6 osmarks 1784460156 +0000 commit: Navigate semantic flamegraph by hierarchy c22fed233b0caab0393cf6dbce5cf25d5a4f2df6 dc9019b2ab27e75136e21769414e27179c0a3589 osmarks 1784460378 +0000 commit: Require shift for profile wheel zoom dc9019b2ab27e75136e21769414e27179c0a3589 635d82e3fbe4a808b74dc3ad62ccfbfc8dc5255e osmarks 1784461003 +0000 commit: Add structured profile query interface 635d82e3fbe4a808b74dc3ad62ccfbfc8dc5255e b45ebf06aa34fb987d288c48ecc54b10268b4139 osmarks 1784464595 +0000 commit: Tune GEMM blocking and compact host plans b45ebf06aa34fb987d288c48ecc54b10268b4139 58efedcc7025a779160028cbae9cdd3c22913988 osmarks 1784464666 +0000 commit: Use release builds for large graph tooling 58efedcc7025a779160028cbae9cdd3c22913988 65dab794bed1c368a57ec50723dbe9d4539e2d1d osmarks 1784467373 +0000 commit: Compose blocked GEMMs into an FP32 MLP 65dab794bed1c368a57ec50723dbe9d4539e2d1d 97a62344193df6daa4896a8f4dfff0de08351fae osmarks 1784467373 +0000 commit: Document the composed MLP runner 97a62344193df6daa4896a8f4dfff0de08351fae 57be9c20a8d22ee62051456936d71ace222b51f3 osmarks 1784468751 +0000 commit: Add semantic phase profiling 57be9c20a8d22ee62051456936d71ace222b51f3 2a0b5363270d169ae6bac3787b34f07ed80afbfb osmarks 1784469598 +0000 commit: Split profile synchronization from exchange 2a0b5363270d169ae6bac3787b34f07ed80afbfb 45285cea9bdb148573034331e261d62e71459d87 osmarks 1784471233 +0000 commit: Render inactive profile spans in white 45285cea9bdb148573034331e261d62e71459d87 79fee909ebf4e0903e473a41d17b4adde47936a1 osmarks 1784471233 +0000 commit: Specialize GEMM row partitions statically 79fee909ebf4e0903e473a41d17b4adde47936a1 072de842b6cbbc6ffd06ebfcb01e0883b16c3be4 osmarks 1784471233 +0000 commit: Forward MLP accumulators between layers 072de842b6cbbc6ffd06ebfcb01e0883b16c3be4 42052e23ba33bc64c729d0ced7d02b321d8e6d7e osmarks 1784481773 +0000 commit: Add FP16 inference kernels and planning 42052e23ba33bc64c729d0ced7d02b321d8e6d7e b7776eb0a3d0313728a41ba5f78a5c8e580f1a5d osmarks 1784481788 +0000 commit: Validate randomized FP16 GEMM and MLP b7776eb0a3d0313728a41ba5f78a5c8e580f1a5d 6d83c643717a15b812e00d583922aa9811e97a7a osmarks 1784484235 +0000 commit: Statically specialize FP16 GEMM kernels 6d83c643717a15b812e00d583922aa9811e97a7a a798d18563524206bd1c5312b652e01f406cd97e osmarks 1784488956 +0000 commit: Support compiler-emitted worker kernels a798d18563524206bd1c5312b652e01f406cd97e 375ae7b0bfa6357d3e79f9cdc2804ef437f037bb osmarks 1784488991 +0000 commit: Add non-causal FP16 FlashAttention 375ae7b0bfa6357d3e79f9cdc2804ef437f037bb 2b378b7e8d98a1806442fae2179dc0dbf10c8cdb osmarks 1784489552 +0000 commit: Shard FlashAttention queries across tiles 2b378b7e8d98a1806442fae2179dc0dbf10c8cdb 93addd1da866181929dddefcd68e34879f62e716 osmarks 1784490259 +0000 commit: Profile and vectorize FlashAttention 93addd1da866181929dddefcd68e34879f62e716 baef24b1df13c01b14cb33f6f7dadbaca2d8ae4e osmarks 1784491214 +0000 commit: Run FlashAttention QK on AMP baef24b1df13c01b14cb33f6f7dadbaca2d8ae4e 682e3b7fb0de8b4b91eaf8145fd8d64b267c2033 osmarks 1784491887 +0000 commit: Run FlashAttention value products on AMP 682e3b7fb0de8b4b91eaf8145fd8d64b267c2033 7bd70a8feaa0685aad03a65ffc8d5a237c75e2a9 osmarks 1784492061 +0000 commit: Saturate attention worker contexts by default 7bd70a8feaa0685aad03a65ffc8d5a237c75e2a9 2a5fd7c6f104fdfad5d71f92134509c7a8d5ba43 osmarks 1784492365 +0000 commit: Specialize FlashAttention block shapes statically 2a5fd7c6f104fdfad5d71f92134509c7a8d5ba43 3ff61fefe2bd5e4adabd449523c9e81dde1c7941 osmarks 1784492483 +0000 commit: Specialize FlashAttention merge roles 3ff61fefe2bd5e4adabd449523c9e81dde1c7941 f52976dbdf8716b5e9f15dfa56e14a2072c2e43f osmarks 1784492849 +0000 commit: Choose attention key blocks by static cost f52976dbdf8716b5e9f15dfa56e14a2072c2e43f 4f4024c87a419d8c6f5fdc3475a4356697302d9d osmarks 1784492966 +0000 commit: Document FlashAttention hardware baselines 4f4024c87a419d8c6f5fdc3475a4356697302d9d 002b2e9c05018c110f6d3516e77de9ed28c9a7ba osmarks 1784493266 +0000 commit: Distribute long-sequence attention storage 002b2e9c05018c110f6d3516e77de9ed28c9a7ba e4568a67c235a184e9991c9fde90dbd4d87f8c35 osmarks 1784493770 +0000 commit: Document batched long-sequence attention results e4568a67c235a184e9991c9fde90dbd4d87f8c35 170bf3ddf7bd4804e41bbce15cea0eff9a425f7d osmarks 1784495624 +0000 commit: Vectorize and widen FlashAttention blocks 170bf3ddf7bd4804e41bbce15cea0eff9a425f7d 64f96268e91ba369427dfded220e9ecec87b78ea osmarks 1784495899 +0000 commit: Record optimized FlashAttention baselines 64f96268e91ba369427dfded220e9ecec87b78ea 730fb9819ef45144267dbffb3a3d60d38ee789d8 osmarks 1784538272 +0000 commit: Add SigLIP vision reference generator 730fb9819ef45144267dbffb3a3d60d38ee789d8 d99203065a7a7d28db77ce5b084c4ab00aebf9ef osmarks 1784538782 +0000 commit: Plan resident SigLIP encoder memory d99203065a7a7d28db77ce5b084c4ab00aebf9ef c927c252acf6e44beb611b32775559b8b4c703d9 osmarks 1784539181 +0000 commit: Require contiguous AMP weight storage c927c252acf6e44beb611b32775559b8b4c703d9 a45d2f766502b5b2517193035ccfb4376edb2b5a osmarks 1784539209 +0000 commit: Revert "Plan resident SigLIP encoder memory" a45d2f766502b5b2517193035ccfb4376edb2b5a a45d2f766502b5b2517193035ccfb4376edb2b5a osmarks 1784539233 +0000 reset: moving to a45d2f766502b5b2517193035ccfb4376edb2b5a a45d2f766502b5b2517193035ccfb4376edb2b5a dd2840fbf2df960083bd89d0b2c3599a3f8c9da4 osmarks 1784539377 +0000 commit: Generalize tensor allocation across SRAM arenas dd2840fbf2df960083bd89d0b2c3599a3f8c9da4 51cfc3550932899520aa16b936cca9ef01b8db7a osmarks 1784540097 +0000 commit: Use production startup handshake for diagnostics 51cfc3550932899520aa16b936cca9ef01b8db7a 17eb9db9f3d3915cce0bbc2d9dcceecc93f248db osmarks 1784540509 +0000 commit: Exercise segmented graph allocation on hardware 17eb9db9f3d3915cce0bbc2d9dcceecc93f248db e742275cb656be449e09bd80f58a5c20feb3042e osmarks 1784541230 +0000 commit: Add hardware-tested FP16 affine layer norm e742275cb656be449e09bd80f58a5c20feb3042e 03ecc27cf92241df645684b745e01fa04c9caab2 osmarks 1784542012 +0000 commit: Use AMP-A layout for row-sharded activations 03ecc27cf92241df645684b745e01fa04c9caab2 afd62e042d2681902f20540ad120daba78fe3cad osmarks 1784542334 +0000 commit: Add SafeTensors SigLIP model frontend afd62e042d2681902f20540ad120daba78fe3cad e89ff5fbe26d6cd81192894f1eb08c01ce9af996 osmarks 1784542574 +0000 commit: Run SigLIP patch embedding on hardware e89ff5fbe26d6cd81192894f1eb08c01ce9af996 ac19c5569d76c69bdcb0b3fd730bb7cf06a3beed osmarks 1784542855 +0000 commit: Gather GEMM outputs into row-sharded activations ac19c5569d76c69bdcb0b3fd730bb7cf06a3beed ecdb7696ef4114ced0bae32ff11401524cb53a75 osmarks 1784543143 +0000 commit: Compose SigLIP embedding with encoder layer norm ecdb7696ef4114ced0bae32ff11401524cb53a75 eb54fa2b6fa65cb9bde05afdf7f4d746a1eb00ef osmarks 1784543689 +0000 commit: Compose row-sharded activations into QKV GEMM eb54fa2b6fa65cb9bde05afdf7f4d746a1eb00ef b11d47f041c2aaa66a5d7e61396fc4a98f6ed755 osmarks 1784545227 +0000 commit: Place generated executables in free SRAM elements b11d47f041c2aaa66a5d7e61396fc4a98f6ed755 ba98cc134bcaa5060d6269e98fe1743111b9dc1f osmarks 1784545478 +0000 commit: Compose SigLIP QKV into FlashAttention ba98cc134bcaa5060d6269e98fe1743111b9dc1f 1e8c2b255e6077ed413f0e3c62812b6dd508895f osmarks 1784545936 +0000 commit: Gather attention heads into row-sharded activations 1e8c2b255e6077ed413f0e3c62812b6dd508895f 940a5eba611fb0fb622adbc61aa394fe75d730b8 osmarks 1784546221 +0000 commit: Run SigLIP attention projection and residual 940a5eba611fb0fb622adbc61aa394fe75d730b8 9cab3e5c4a5fb10fd18d6c661306c5265449f734 osmarks 1784547094 +0000 commit: Reuse activation storage across SigLIP phases 9cab3e5c4a5fb10fd18d6c661306c5265449f734 78e30de74df67722efb6170a857872ce308e2abf osmarks 1784549108 +0000 commit: Keep wide SigLIP activations distributed 78e30de74df67722efb6170a857872ce308e2abf 64c85146891524f775d4923d2c0ca9f83ddaca89 osmarks 1784553410 +0000 commit: Run complete SigLIP encoder layer on hardware 64c85146891524f775d4923d2c0ca9f83ddaca89 5f572a64be9fdc8c0b4fc0416d262e93e1f4717b osmarks 1784553764 +0000 commit: Broadcast SigLIP linear biases compactly 5f572a64be9fdc8c0b4fc0416d262e93e1f4717b d1b33ed10b12e22f68fa34059663117f71c79424 osmarks 1784554271 +0000 commit: Share linear bias blocks across row shards d1b33ed10b12e22f68fa34059663117f71c79424 47c51988d484224fb6423364400cf38b33dd1dc4 osmarks 1784554643 +0000 commit: Place appended attention in live SRAM gaps 47c51988d484224fb6423364400cf38b33dd1dc4 f794daa768f23802e526630c188334f4cdb5748a osmarks 1784555663 +0000 commit: Balance appended GEMM placement across tiles f794daa768f23802e526630c188334f4cdb5748a 03073c46ea79bd910dbae474a3b3a45b9ee3e286 osmarks 1784556398 +0000 commit: Compose reusable SigLIP encoder layers 03073c46ea79bd910dbae474a3b3a45b9ee3e286 d163796d5db3fa45d1c578fa5ef4d5b230c7c620 osmarks 1784558158 +0000 commit: Scale SigLIP prefix planning and diagnostics d163796d5db3fa45d1c578fa5ef4d5b230c7c620 2910e0dd9fbc43a0b23603472bd2825981d02d9d osmarks 1784559907 +0000 commit: Reclaim composed activations and distribute norm affine 2910e0dd9fbc43a0b23603472bd2825981d02d9d 4f4a60bf19808e244a80e8f5175650bee7d87984 osmarks 1784560843 +0000 commit: Support short-query FlashAttention plans 4f4a60bf19808e244a80e8f5175650bee7d87984 bc59c4a96f849fceb2e25c310b245c54cb3e2e45 osmarks 1784560931 +0000 commit: Exercise FlashAttention cross-attention on hardware bc59c4a96f849fceb2e25c310b245c54cb3e2e45 ba7d8805bcf79b8796c44d0ecfe1b3b111503277 osmarks 1784561419 +0000 commit: Implement and verify SigLIP MAP pooling head ba7d8805bcf79b8796c44d0ecfe1b3b111503277 e629be2a5f4ad35459d88728b647675c7fbb4d77 osmarks 1784561763 +0000 commit: Verify SigLIP post norm and pooled tail e629be2a5f4ad35459d88728b647675c7fbb4d77 2c141f7f5556c697edf2282bf9e05d282aa46861 osmarks 1784562673 +0000 commit: Place runtime code independently on each tile 2c141f7f5556c697edf2282bf9e05d282aa46861 05c8ff51837159e80fca76903bb552ef8033ddb1 osmarks 1784563532 +0000 commit: Strip runtime kernels per tile 05c8ff51837159e80fca76903bb552ef8033ddb1 0ef76be2fee619b557118fcdf7ee15f8211bef51 osmarks 1784564393 +0000 commit: Deduplicate exchange plans per tile 0ef76be2fee619b557118fcdf7ee15f8211bef51 ad396b1288115bfdc890e13ebd56a6002e6feda9 osmarks 1784566430 +0000 commit: Speed up SigLIP graph lowering ad396b1288115bfdc890e13ebd56a6002e6feda9 dab187b1c3457279078b9284aa17ab118eca8adf osmarks 1784566430 +0000 commit: Report generated code size for packaged graphs dab187b1c3457279078b9284aa17ab118eca8adf a7d04490a6fe5265a0451cfa8a0ca865dddfae02 osmarks 1784568583 +0000 commit: Compact repeated static exchange compute runs a7d04490a6fe5265a0451cfa8a0ca865dddfae02 02ae53679948fc898f903bd33de13a8e97fbb7ff osmarks 1784568840 +0000 commit: Cache per-tile executable placement ranges 02ae53679948fc898f903bd33de13a8e97fbb7ff 70dde994d010733403c1e983e7c1f8197cfd24b0 osmarks 1784570052 +0000 commit: Keep preloaded SigLIP parameters resident 70dde994d010733403c1e983e7c1f8197cfd24b0 c70fb411d89b50d3c7c18b97e39ceb3037b7765a osmarks 1784571681 +0000 commit: Emit repeated SigLIP layers from static templates c70fb411d89b50d3c7c18b97e39ceb3037b7765a 122f71c78a4b88990fd5889bf01db11d8f518409 osmarks 1784572306 +0000 commit: Compact template exchange dispatch 122f71c78a4b88990fd5889bf01db11d8f518409 1cd0277d016da490596d4126774ef36d421bc964 osmarks 1784573416 +0000 commit: Reuse dead allocations in appended GEMMs 1cd0277d016da490596d4126774ef36d421bc964 8f52ba810249463ff7aba7fa1ec954c145cf93e0 osmarks 1784574149 +0000 commit: Deduplicate templated sender exchange plans 8f52ba810249463ff7aba7fa1ec954c145cf93e0 2415a8d2373e4317099cbb8da734d974c7901f97 osmarks 1784574834 +0000 commit: Compact repeated exchange plan records 2415a8d2373e4317099cbb8da734d974c7901f97 abfa098afdc3d819cd195059b5cedf0897e4d21c osmarks 1784575978 +0000 commit: Place template records in segmented SRAM abfa098afdc3d819cd195059b5cedf0897e4d21c 10bbaa03b9b0fe4a1dbdcf7a70fd930643f9e104 osmarks 1784581942 +0000 commit: Allocate resident tensors across ordered SRAM arenas 10bbaa03b9b0fe4a1dbdcf7a70fd930643f9e104 2c96f07b6abee4c23aace928874e836da7f9e5ce osmarks 1784582049 +0000 commit: Document IPU tile memory placement policy 2c96f07b6abee4c23aace928874e836da7f9e5ce 39d65111fcfad1da512bfa7d1b907183d4abe647 osmarks 1784584277 +0000 commit: Balance resident weights by unavailable SRAM 39d65111fcfad1da512bfa7d1b907183d4abe647 ef1f139f296ef70e6eb3526106ecfff34e9742f5 osmarks 1784585022 +0000 commit: Name IPU21 SRAM placement regions ef1f139f296ef70e6eb3526106ecfff34e9742f5 525614f60e582199657600e86318be0b25874faf osmarks 1784586320 +0000 commit: Add compact FP8 GEMM weight storage 525614f60e582199657600e86318be0b25874faf 3232d3b568726a9c6ac19271944af3c7a3ff7970 osmarks 1784587543 +0000 commit: Use compact FP8 weights in SigLIP layers 3232d3b568726a9c6ac19271944af3c7a3ff7970 5db4ec6bcad550a2424085c6e048ec46210a9bcb osmarks 1784588589 +0000 commit: Relocate host runtime around tile allocations 5db4ec6bcad550a2424085c6e048ec46210a9bcb 947ef2cebfb02661e4c7050afbd1ba1173bb59ea osmarks 1784588912 +0000 commit: Avoid F143 NaN on negative underflow 947ef2cebfb02661e4c7050afbd1ba1173bb59ea 5e0b743a36b540e5d5ac912154f23645d617b4b3 osmarks 1784590805 +0000 commit: Pack host runtime across ordinary SRAM gaps 5e0b743a36b540e5d5ac912154f23645d617b4b3 0e7059679b5a6ac6529289eb55764a10478bd9a9 osmarks 1784591902 +0000 commit: Separate executable and data host plan placement 0e7059679b5a6ac6529289eb55764a10478bd9a9 03756973c219ad875c650bae850789dc1d1b76ec osmarks 1784593131 +0000 commit: Compact static host run descriptors 03756973c219ad875c650bae850789dc1d1b76ec dd6eea6877dac731479115880bd9aa0ce2a85153 osmarks 1784594204 +0000 commit: Pack host packet destination into descriptor flags dd6eea6877dac731479115880bd9aa0ce2a85153 68c780b8075fe51b3da8e6daadcfb86de7f79e95 osmarks 1784595569 +0000 commit: Account for per-tile resident memory headroom 68c780b8075fe51b3da8e6daadcfb86de7f79e95 82925d2b93eaecc6b55f69730cdb791bbee09bd2 osmarks 1784598142 +0000 commit: Intern repeated static template operands 82925d2b93eaecc6b55f69730cdb791bbee09bd2 e45f49e41eb2a4e710ccf7c3f47eae9cac6c3838 osmarks 1784599290 +0000 commit: Compact affine static template operands e45f49e41eb2a4e710ccf7c3f47eae9cac6c3838 a02c3d5b5a0ea221714817de9dcb195b2695716e osmarks 1784600526 +0000 commit: Bound affine template compression by SRAM cost a02c3d5b5a0ea221714817de9dcb195b2695716e ed8d3a5f4681ce4a66a3273bb7823eb3991d1817 osmarks 1784601496 +0000 commit: Pack static template records largest first ed8d3a5f4681ce4a66a3273bb7823eb3991d1817 25706a08ca25774f894bd1d52ab6a7f98e02fc44 osmarks 1784603047 +0000 commit: Share layer-constant template operands 25706a08ca25774f894bd1d52ab6a7f98e02fc44 5696ffe62daf59c3a6321f33f634e73c5ce64187 osmarks 1784604295 +0000 commit: Reuse worker guard for runtime completion 5696ffe62daf59c3a6321f33f634e73c5ce64187 8d8f3c311b5ef266573cbcfa9c0c361462185399 osmarks 1784605431 +0000 commit: Spill static host run state on supervisor stack 8d8f3c311b5ef266573cbcfa9c0c361462185399 c2dd49b8106b1d6988083e2ea0676057476fef91 osmarks 1784606569 +0000 commit: Segment static template records into two banks c2dd49b8106b1d6988083e2ea0676057476fef91 3f1275d8f65c4e1c55a271bf089d6386935d9b1b osmarks 1784607959 +0000 commit: Share host staging source across run descriptors 3f1275d8f65c4e1c55a271bf089d6386935d9b1b 26b37fc0a7a8641f883720a9b7a591ffb1e83e3c osmarks 1784609791 +0000 commit: Place support code across executable SRAM gaps 26b37fc0a7a8641f883720a9b7a591ffb1e83e3c e2b86a8dc2aa77a72d85a911d8e41c602d6f6c6f osmarks 1784610320 +0000 commit: Remove per-tile resident headroom override e2b86a8dc2aa77a72d85a911d8e41c602d6f6c6f bf98cbbef1a8e1595ed0eb79c624c47f5d248167 osmarks 1784611093 +0000 commit: Reserve generated code before packing runtime data bf98cbbef1a8e1595ed0eb79c624c47f5d248167 6ce318b70495e232898a8cb121ce20e351e2ffb5 osmarks 1784611356 +0000 commit: Report executable placement capacity 6ce318b70495e232898a8cb121ce20e351e2ffb5 33ce996398b318e042b8baafc7c32d19f1ea9695 osmarks 1784613966 +0000 commit: Reserve only addressable tile context stacks 33ce996398b318e042b8baafc7c32d19f1ea9695 89400fae825816ddfdea00383b51e8d60acd9dc8 osmarks 1784615866 +0000 commit: Offset finite child allocation lifetimes 89400fae825816ddfdea00383b51e8d60acd9dc8 02035b29e767477e56500d89350dc5501eaba88a osmarks 1784616675 +0000 commit: Delta encode repeated layer templates 02035b29e767477e56500d89350dc5501eaba88a e5bc04e6b43a85b5f45d6fbb380adbbaf514b6b5 osmarks 1784618344 +0000 commit: Pack repeated template address deltas e5bc04e6b43a85b5f45d6fbb380adbbaf514b6b5 6824b9eb6708bf30517eb594df8274da05df4f33 osmarks 1784619581 +0000 commit: Reserve support executable capacity before data 6824b9eb6708bf30517eb594df8274da05df4f33 1318bc72df29e30f72bad4ba699d87593b5054c3 osmarks 1784620665 +0000 commit: Derive packed template delta counts 1318bc72df29e30f72bad4ba699d87593b5054c3 10ec52c3ac8d759df23f948413151521e2b2b13d osmarks 1784621918 +0000 commit: Pack template deltas across bitmap groups 10ec52c3ac8d759df23f948413151521e2b2b13d 5e25365aa49817e11d33ab07ac6ac4e80bab13a6 osmarks 1784623146 +0000 commit: Compact repeated host run descriptors 5e25365aa49817e11d33ab07ac6ac4e80bab13a6 5bada042264d2cf5f0618a1d9e816ca9f58c4b9b osmarks 1784623757 +0000 commit: Cache compiled IPU kernel artifacts 5bada042264d2cf5f0618a1d9e816ca9f58c4b9b 03a433d7842cb1009c3f8d37dfe5a9854f17168d osmarks 1784623757 +0000 commit: Reuse allocation index during lowering 03a433d7842cb1009c3f8d37dfe5a9854f17168d 79e33fd4417c4700827c27e72a411ef071314640 osmarks 1784623929 +0000 commit: Compact the lowering allocation index 79e33fd4417c4700827c27e72a411ef071314640 ea9931a9e727d118bd974d975ba2d10c128bec21 osmarks 1784624404 +0000 commit: Avoid duplicate kernel specialization strings ea9931a9e727d118bd974d975ba2d10c128bec21 74ad366abc4731d91b1b5c25cf9fbc406e82d277 osmarks 1784624706 +0000 commit: Skip unused GEMM profile metadata 74ad366abc4731d91b1b5c25cf9fbc406e82d277 d45d171bcc240ec9f4b7e8a31c28c586f82b6617 osmarks 1784624997 +0000 commit: Parallelize independent tile packaging d45d171bcc240ec9f4b7e8a31c28c586f82b6617 406b23ca9a6fbe8d3db72bc3bb4cd0f4577f039f osmarks 1784625243 +0000 commit: Index exchange scheduler conflicts by tile 406b23ca9a6fbe8d3db72bc3bb4cd0f4577f039f 43060fbeda6c2612b8367e337e5a247c4b4b393f osmarks 1784625514 +0000 commit: Reuse occupancy when placing GEMM biases 43060fbeda6c2612b8367e337e5a247c4b4b393f 1deb5ea3ffed79ac54c138f1211896bce5232905 osmarks 1784626753 +0000 commit: Reuse template patch destinations from stack 1deb5ea3ffed79ac54c138f1211896bce5232905 8e3a073911898228bbc206decc940c7a010fa938 osmarks 1784626904 +0000 commit: Reuse template record pointers across instances 8e3a073911898228bbc206decc940c7a010fa938 97c4eae1dde722d6130b1e50d4cb6ee2578dfe7d osmarks 1784627791 +0000 commit: Canonicalize repeated exchange plan rows 97c4eae1dde722d6130b1e50d4cb6ee2578dfe7d 93da7452f41abdf474719db54afce9de1fb4aa6e osmarks 1784628917 +0000 commit: Split repeated template patches for placement 93da7452f41abdf474719db54afce9de1fb4aa6e 5214621c96ad97917043418fe6c2ba4eab6c079f osmarks 1784629525 +0000 commit: Load template patch lengths in runtime 5214621c96ad97917043418fe6c2ba4eab6c079f 1c7ad4b885e993fd02afc2dd8710a7a1a80f54f6 osmarks 1784630153 +0000 commit: Trim empty template patch mask groups 1c7ad4b885e993fd02afc2dd8710a7a1a80f54f6 c18f28c89e33454604941d40525e1ba03acfccb8 osmarks 1784630923 +0000 commit: Pack repeated template patch type bits c18f28c89e33454604941d40525e1ba03acfccb8 2e8c124b26bdd7b2f4c40d4768269ae7c545eb1d osmarks 1784631676 +0000 commit: Place template patches in smaller parts 2e8c124b26bdd7b2f4c40d4768269ae7c545eb1d fd373ae8e005daae04d82776e54554d2ef1acc99 osmarks 1784632580 +0000 commit: Stabilize resident tile assignment for repeated layers fd373ae8e005daae04d82776e54554d2ef1acc99 db8eeb110f76afce7478e9d4a04653c89874f508 osmarks 1784633278 +0000 commit: Balance resident placement within repeated layers db8eeb110f76afce7478e9d4a04653c89874f508 46fce9df4b3710f33cdefd3be997cfaf4804d61c osmarks 1784634691 +0000 commit: Stream large host exchanges through reusable pages 46fce9df4b3710f33cdefd3be997cfaf4804d61c 69c6b367f81e5ec803a3f27bcd6d4d4db3a8b13a osmarks 1784636808 +0000 commit: Reduce repeated-layer compiler scans 69c6b367f81e5ec803a3f27bcd6d4d4db3a8b13a d1f30f1f63e97645f7a055538818f2b8007039c9 osmarks 1784637763 +0000 commit: Automatically stream reused host exchange pages d1f30f1f63e97645f7a055538818f2b8007039c9 589e9861c507e86649c902863a7865e0fd1a3158 osmarks 1784639839 +0000 commit: Scale F143 weights per GEMM block 589e9861c507e86649c902863a7865e0fd1a3158 082e6e3bbc957c903f510a0bcc9020198afec678 osmarks 1784640428 +0000 commit: Add activation-aware F143 reconstruction tool 082e6e3bbc957c903f510a0bcc9020198afec678 840bc68c9c6f91b18a24b8c3a7930fbefe0861d1 osmarks 1784641854 +0000 commit: Calibrate F143 reconstruction on real images 840bc68c9c6f91b18a24b8c3a7930fbefe0861d1 31633d6c4577d4bea881639a7d4142ddc07157db osmarks 1784642787 +0000 commit: Support sequential F143 calibration 31633d6c4577d4bea881639a7d4142ddc07157db fd3f2072a1a5c16635209694d3b1b22af26d84c6 osmarks 1784643433 +0000 commit: Add bounded LayerNorm equalization for F143 fd3f2072a1a5c16635209694d3b1b22af26d84c6 f8226287339b7cdee0d4504be3a4d3e6e8e09f70 osmarks 1784643433 +0000 commit: Compose SigLIP encoder and MAP head f8226287339b7cdee0d4504be3a4d3e6e8e09f70 dbde075e14573c98036dfa603789eba025887d37 osmarks 1784645232 +0000 commit: Place host phase code in executable elements dbde075e14573c98036dfa603789eba025887d37 786c23611dc366044ad10e384a99162ca075b8ac osmarks 1784646554 +0000 commit: Restrict IPU21 code to executable memory 786c23611dc366044ad10e384a99162ca075b8ac daaf10580c69d3f9041dceaca0a28f06170db47d osmarks 1784648118 +0000 commit: Fix composed attention scratch placement daaf10580c69d3f9041dceaca0a28f06170db47d 0c6a7d1c4bfb5cbe17bfa6eec0c9aaf18964497e osmarks 1784648237 +0000 commit: Scale allocation overlap validation 0c6a7d1c4bfb5cbe17bfa6eec0c9aaf18964497e daa34a7edbfcdc0247ea7d1309233405399da046 osmarks 1784648773 +0000 commit: Index host matrix placement occupancy daa34a7edbfcdc0247ea7d1309233405399da046 3540f1c078e8269937e9f2cc0983039b98f36403 osmarks 1784648859 +0000 commit: Limit appended GEMM metadata scans 3540f1c078e8269937e9f2cc0983039b98f36403 3540f1c078e8269937e9f2cc0983039b98f36403 osmarks 1784650175 +0000 reset: moving to HEAD 3540f1c078e8269937e9f2cc0983039b98f36403 ec695a335951d77134c605a79293773b95d24026 osmarks 1784650639 +0000 commit: Validate allocation aliases and patch every FP8 expansion ec695a335951d77134c605a79293773b95d24026 0b9e93feb9de3263f7f3b99ba707f78d9eced22d osmarks 1784650678 +0000 commit: Project F143 weights in FP32 0b9e93feb9de3263f7f3b99ba707f78d9eced22d b00eb227cb9e8e7e828655fa35f30c70f2019e3e osmarks 1784651373 +0000 commit: Use fast deterministic hashes in compiler internals b00eb227cb9e8e7e828655fa35f30c70f2019e3e 0b9e93feb9de3263f7f3b99ba707f78d9eced22d osmarks 1784651497 +0000 checkout: moving from main to 0b9e93f 0b9e93feb9de3263f7f3b99ba707f78d9eced22d b00eb227cb9e8e7e828655fa35f30c70f2019e3e osmarks 1784651625 +0000 checkout: moving from 0b9e93feb9de3263f7f3b99ba707f78d9eced22d to main b00eb227cb9e8e7e828655fa35f30c70f2019e3e 31ed462392236c9323971e4d15395ccc8ccbc4f0 osmarks 1784652373 +0000 commit: Bound SigLIP residency scans to appended allocations 31ed462392236c9323971e4d15395ccc8ccbc4f0 366f919ca5ef731b714712d09f747c443c4447ea osmarks 1784653090 +0000 commit: Verify full SigLIP output by cosine similarity 366f919ca5ef731b714712d09f747c443c4447ea e2e3b6ba9dc15cb6e56b086b99364813f1610aa7 osmarks 1784656326 +0000 commit: Specialize composed attention plans correctly e2e3b6ba9dc15cb6e56b086b99364813f1610aa7 731d465087a407e14a5d7820c7852b4752bbe857 osmarks 1784660497 +0000 commit: Fix generated runtime startup and host completion 731d465087a407e14a5d7820c7852b4752bbe857 09d13538bb0ccec458000565f74f68c7623cf63b osmarks 1784660497 +0000 commit: Add saved-package SigLIP sample validation 09d13538bb0ccec458000565f74f68c7623cf63b 2d53b95e2223e15f27f443572a75d66611e8fe00 osmarks 1784661286 +0000 commit: Separate resident model inputs from inference data 2d53b95e2223e15f27f443572a75d66611e8fe00 e0e053c28300ab75bdf1b24240eb9ee7523c5b2d osmarks 1784661856 +0000 commit: Add static repeated graph execution e0e053c28300ab75bdf1b24240eb9ee7523c5b2d da7d66253453ca622148709e53c790e564966293 osmarks 1784663052 +0000 commit: Make repeated model execution reentrant da7d66253453ca622148709e53c790e564966293 7fad1a32ff1cafc0dfdeb94818e02665345bfdd6 osmarks 1784663947 +0000 commit: Add native FP8 blocked GEMM 7fad1a32ff1cafc0dfdeb94818e02665345bfdd6 2aa1dfd2a12fd7545b581dc7b985c9d81c6ccfdb osmarks 1784664606 +0000 commit: Add mixed-precision SigLIP linear execution 2aa1dfd2a12fd7545b581dc7b985c9d81c6ccfdb f0e925783925d51b1157df8ef8e1da9afb90d474 osmarks 1784664727 +0000 commit: Remove obsolete SigLIP run retries f0e925783925d51b1157df8ef8e1da9afb90d474 63e6bbba6712910498b719331b99a4923cb8593a osmarks 1784666493 +0000 commit: Index resident host binding validation 63e6bbba6712910498b719331b99a4923cb8593a bc4582072503fa36d4531a466115de3d16dcf3e4 osmarks 1784669549 +0000 commit: Add semantic SigLIP stage profiling bc4582072503fa36d4531a466115de3d16dcf3e4 bc1e00506f54f7624866f3f17b38859629f77352 osmarks 1784671060 +0000 commit: Use paired coefficient loads for expanded FP8 GEMM bc1e00506f54f7624866f3f17b38859629f77352 656b357212555f171a5d3e65958c8c5fab641133 osmarks 1784674674 +0000 commit: Profile and rebalance composed attention 656b357212555f171a5d3e65958c8c5fab641133 72b3ac93f6fc9768dbad5f79fb6733c6c43dfc40 osmarks 1784678787 +0000 commit: Generalize segmented SigLIP placement 72b3ac93f6fc9768dbad5f79fb6733c6c43dfc40 f288fdab5ffdbb914bc946cc400d0f5fbd0d7739 osmarks 1784678787 +0000 commit: Reduce LayerNorm worker register pressure f288fdab5ffdbb914bc946cc400d0f5fbd0d7739 2c1b6427961108d7bfec79898def16a93cb2b934 osmarks 1784681028 +0000 commit: Support independently blocked SigLIP GEMMs 2c1b6427961108d7bfec79898def16a93cb2b934 43895864215807b66fbc1381e0bb89720691b508 osmarks 1784682004 +0000 commit: Compact C16 row-reblock transfers 43895864215807b66fbc1381e0bb89720691b508 cfc5c8148cc74d535b3cab2a5408b35a6d5d4853 osmarks 1784682380 +0000 commit: Vectorize FP16 LayerNorm arithmetic cfc5c8148cc74d535b3cab2a5408b35a6d5d4853 24c23cbadd4b3c537ae0527de7f72c62e44c65ec osmarks 1784682844 +0000 commit: Batch GEMM inner-block exchange epochs 24c23cbadd4b3c537ae0527de7f72c62e44c65ec 1e149a5d86664559eb300c55d1cb0797e40fb49c osmarks 1784683419 +0000 commit: Specialize SigLIP GEMMs by inner block width 1e149a5d86664559eb300c55d1cb0797e40fb49c fc7eb2b0bfe9b254fb59adb669988d88d0abaac0 osmarks 1784683653 +0000 commit: Batch FlashAttention key exchange epochs fc7eb2b0bfe9b254fb59adb669988d88d0abaac0 1c32a4a8d1793bea722c5e36e3673f75a2c7831f osmarks 1784684344 +0000 commit: Account for sequential operations in profile queries 1c32a4a8d1793bea722c5e36e3673f75a2c7831f d4757a178806cbfa87b0685322a4c892e8e1cea6 osmarks 1784684348 +0000 commit: Support 32-column FP16 GEMM blocks d4757a178806cbfa87b0685322a4c892e8e1cea6 6f6e6a5ade3f4086d3005ac1cd95d75066ee97da osmarks 1784684740 +0000 commit: Add automatic QKV row blocking mode 6f6e6a5ade3f4086d3005ac1cd95d75066ee97da 4ffa408a903a9d16a3c3c9409441bbabd41bd4f9 osmarks 1784685123 +0000 commit: Preserve automatic QKV layout through attention 4ffa408a903a9d16a3c3c9409441bbabd41bd4f9 cd16bcb4be6e820e4b64e8d9f11d0e97212e1a65 osmarks 1784686267 +0000 commit: Decouple SigLIP MLP row blocking cd16bcb4be6e820e4b64e8d9f11d0e97212e1a65 5b5afdcc81fe8e3f1c026de62123f5338c653a14 osmarks 1784686854 +0000 commit: Vectorize FP16 LayerNorm four lanes wide 5b5afdcc81fe8e3f1c026de62123f5338c653a14 05ddef78dd97474b3e7e5966d26ecc62c7fc533b osmarks 1784687111 +0000 commit: Annotate fine profiles with semantic regions 05ddef78dd97474b3e7e5966d26ecc62c7fc533b ccfe015330268ce86a98750265ad92ab1cd73ac0 osmarks 1784687325 +0000 commit: Pack attention values in row pairs ccfe015330268ce86a98750265ad92ab1cd73ac0 c2c2be88382230ea3377035cae3e21ac649812ed osmarks 1784687417 +0000 commit: Vectorize attention query and key packing c2c2be88382230ea3377035cae3e21ac649812ed 99051189a783c561d27de88b93d9a5fc91eab6c3 osmarks 1784687633 +0000 commit: Vectorize attention output unpacking 99051189a783c561d27de88b93d9a5fc91eab6c3 2949f3f34fbaa3ce7a24a84eab9eedc11c97d979 osmarks 1784688670 +0000 commit: Decouple attention query row blocking 2949f3f34fbaa3ce7a24a84eab9eedc11c97d979 0af82da3701582fc873c5ff65390e33f3167d9a3 osmarks 1784688903 +0000 commit: Model attention blocking from scheduled work 0af82da3701582fc873c5ff65390e33f3167d9a3 7f9db913a89503140082d56c94ac39bbf28afee5 osmarks 1784691041 +0000 commit: Fuse residual adds into affine LayerNorm 7f9db913a89503140082d56c94ac39bbf28afee5 cc2e0ad6cb66a3f4f6d8f10589a30b0c946b0fb3 osmarks 1784692275 +0000 commit: Support batched SigLIP encoder execution cc2e0ad6cb66a3f4f6d8f10589a30b0c946b0fb3 c5bd10f9c622cb6002b701983c8bf5bc17604175 osmarks 1784693131 +0000 commit: Generalize row reblocking across GEMM panels c5bd10f9c622cb6002b701983c8bf5bc17604175 58ca944551e1e61164222af0042faf7746e2514b osmarks 1784695809 +0000 commit: Generalize GEMM blocking and avoid local AMP conflicts 58ca944551e1e61164222af0042faf7746e2514b 8035c572fda421eff22c5be285d46c6442f1fb46 osmarks 1784697741 +0000 commit: Fix uneven FlashAttention block specialization 8035c572fda421eff22c5be285d46c6442f1fb46 3387560828e45b5078bf9d12eb28475a3f23b360 osmarks 1784698803 +0000 commit: Batch static FP8 expansion and GEMM commands 3387560828e45b5078bf9d12eb28475a3f23b360 a895dbf878184f677de4a2125b2c4ba91a734014 osmarks 1784700434 +0000 commit: Tune attention blocking and guard empty worker shards a895dbf878184f677de4a2125b2c4ba91a734014 32c4fbed285c800b6ed36162f7fe941624d86dfb osmarks 1784701617 +0000 commit: Fuse transformer residual adds into following norms 32c4fbed285c800b6ed36162f7fe941624d86dfb b01beed67bb1e85be210a0efa4a1b41d0ee9367d osmarks 1784702125 +0000 commit: Generalize FP16 GEMM hardware harness shapes b01beed67bb1e85be210a0efa4a1b41d0ee9367d 97c72f8fd0713c16c220c053d261f761ceea9aa5 osmarks 1784702391 +0000 commit: Retain GEMM worker row partition across coefficient blocks 97c72f8fd0713c16c220c053d261f761ceea9aa5 b06f352d070ac65caf3f2bd4573ca866b1ec826d osmarks 1784702654 +0000 commit: Retain GEMM output state across inner microblocks b06f352d070ac65caf3f2bd4573ca866b1ec826d 0251da991b11671390a6daee955ba7fa30557470 osmarks 1784703922 +0000 commit: Model GEMM inner-depth costs in row blocking 0251da991b11671390a6daee955ba7fa30557470 33d2b34dc82f7c2e7f295928d1a2bf1caec8eb3e osmarks 1784706178 +0000 commit: Add LayerNorm hardware cycle profiling 33d2b34dc82f7c2e7f295928d1a2bf1caec8eb3e ef75f33c354ccaeed09164ed942cdc7164e0835b osmarks 1784706194 +0000 commit: Balance SigLIP GEMM output blocks across tiles ef75f33c354ccaeed09164ed942cdc7164e0835b 2820782490712469eb2658b3fc0292d3427d0e7b osmarks 1784706518 +0000 commit: Keep automatic GEMM blocking within baseline waves 2820782490712469eb2658b3fc0292d3427d0e7b 0f740f426ee3907a5483fa1c4f5d51bac8279810 osmarks 1784707390 +0000 commit: Partition distributed GEMM input staging 0f740f426ee3907a5483fa1c4f5d51bac8279810 5c1821a775005d1ac2a61f7dcb223101b4121ebe osmarks 1784708694 +0000 commit: Standardize per-tile exchange window packing 5c1821a775005d1ac2a61f7dcb223101b4121ebe 88d6a2bced1eb382cd127f599113d51b8748c5e4 osmarks 1784708803 +0000 commit: Constrain GEMM layout selection by transfer capacity 88d6a2bced1eb382cd127f599113d51b8748c5e4 e5f2057767085c2c23dad4788ea8cae211bb99d3 osmarks 1784709694 +0000 commit: Centralize tile address-space placement e5f2057767085c2c23dad4788ea8cae211bb99d3 1b91d457b7243d9bba68985a9264f380a13cfa9a osmarks 1784709750 +0000 commit: Choose row layouts from live placement constraints 1b91d457b7243d9bba68985a9264f380a13cfa9a 7a824f7b0016ea91fe143e6caf96c5f874463f0c osmarks 1784709903 +0000 commit: Lower exchange role conflicts without a quadratic graph 7a824f7b0016ea91fe143e6caf96c5f874463f0c bdef8c94f1d15fa9ec3e92a22b41186973b41ea1 osmarks 1784710798 +0000 commit: Attach exchange staging placement to transfers bdef8c94f1d15fa9ec3e92a22b41186973b41ea1 4fd78de4d910d3c978b0270e2fbaadec10161075 osmarks 1784711988 +0000 commit: Transition row layouts at constrained residual edges 4fd78de4d910d3c978b0270e2fbaadec10161075 d358428006e0b365e4804f8c0cdfb17215237f49 osmarks 1784713302 +0000 commit: Derive executable placement from measured images d358428006e0b365e4804f8c0cdfb17215237f49 3cf59bc15c1a4869979fc1ea481de1af9a86e422 osmarks 1784715545 +0000 commit: Compact structurally distinct model regions 3cf59bc15c1a4869979fc1ea481de1af9a86e422 2994902707446e179edaa5ec0387dcf21c38a72d osmarks 1784715697 +0000 commit: Avoid repeated template affine-key allocation 2994902707446e179edaa5ec0387dcf21c38a72d c41bb34aad585772319eb19b2474b29d2d636bf4 osmarks 1784716468 +0000 commit: Inline short static template columns c41bb34aad585772319eb19b2474b29d2d636bf4 6aa538259ea6b8196a58d8c5f31c138bbc791307 osmarks 1784717918 +0000 commit: Share compute commands across lowering 6aa538259ea6b8196a58d8c5f31c138bbc791307 a7dcb9552537877b480cfbdf1ce2027770c48022 osmarks 1784718342 +0000 commit: Compact lowered tile programs a7dcb9552537877b480cfbdf1ce2027770c48022 4b356105207b3d1ab0c6fa2bd596aa0d45752d13 osmarks 1784720920 +0000 commit: Stream repeated template lowering by tile 4b356105207b3d1ab0c6fa2bd596aa0d45752d13 afc695d54cb622554c7b60a9265a5a8628fd4cdb osmarks 1784721530 +0000 commit: Avoid quadratic placement diagnostics afc695d54cb622554c7b60a9265a5a8628fd4cdb d94cc6a4ff0674d8855ff54ec9811fdf1b32e5de osmarks 1784722273 +0000 commit: Index executable relocation conflicts by tile d94cc6a4ff0674d8855ff54ec9811fdf1b32e5de 99ab9d8e640f355a93393b678d59e21d72c139b3 osmarks 1784722593 +0000 commit: Co-pack generated and support executables 99ab9d8e640f355a93393b678d59e21d72c139b3 035f3b956ae5933559936886e60004d925bc0fd9 osmarks 1784722966 +0000 commit: Split template patches for fragmented SRAM 035f3b956ae5933559936886e60004d925bc0fd9 436a0108b3b13607bfc7aae873262182412df496 osmarks 1784723293 +0000 commit: Subdivide template patches across SRAM gaps 436a0108b3b13607bfc7aae873262182412df496 6fadd6245c5176245c052b060ed25c7bf163f69a osmarks 1784723567 +0000 commit: Best-fit static template data into SRAM gaps 6fadd6245c5176245c052b060ed25c7bf163f69a b0f09b9e3a7bee1578ea3efde4d59b96b1ed2c01 osmarks 1784723803 +0000 commit: Allow finer static patch placement b0f09b9e3a7bee1578ea3efde4d59b96b1ed2c01 6296ac3379839f17b6413e7a590b5fcd9a14692c osmarks 1784724502 +0000 commit: Place static data around executable elements 6296ac3379839f17b6413e7a590b5fcd9a14692c 553884ec39a29d06a2c19e2ef2ffe26ccaceef1d osmarks 1784724715 +0000 commit: Subdivide template patches for element-aware placement 553884ec39a29d06a2c19e2ef2ffe26ccaceef1d cd0141d7c3644e1ecfaeec4459e1c46f3b14abe8 osmarks 1784725237 +0000 commit: Segment static template patches by encoded size cd0141d7c3644e1ecfaeec4459e1c46f3b14abe8 1497f35e7e7e532f1539354a01aba45d7e646901 osmarks 1784725547 +0000 revert: Revert "Segment static template patches by encoded size" 1497f35e7e7e532f1539354a01aba45d7e646901 83e42d65c2f34e3cab2542c34b8bdc3758ba54e8 osmarks 1784725547 +0000 revert: Revert "Subdivide template patches for element-aware placement" 83e42d65c2f34e3cab2542c34b8bdc3758ba54e8 d616a813d23b9e8faf1d23a81047564a3d654480 osmarks 1784725650 +0000 commit: Reserve relocation target intervals d616a813d23b9e8faf1d23a81047564a3d654480 26529aa041253b1532394b41b8351af60131a646 osmarks 1784725914 +0000 commit: Relocate transient tensors around static runtime data 26529aa041253b1532394b41b8351af60131a646 6a18bc539a469c2a0ca51e4f37bb9cd151dfa485 osmarks 1784726286 +0000 commit: Plan static relocation in one pass 6a18bc539a469c2a0ca51e4f37bb9cd151dfa485 dcb08afabd886ebecb0b149f8ad064bf72401fb1 osmarks 1784726644 +0000 commit: Index fixed allocations for static placement dcb08afabd886ebecb0b149f8ad064bf72401fb1 675c9bbcc8f20119e639ec5199f361035c7748fc osmarks 1784727109 +0000 commit: Preserve memory class for static relocation 675c9bbcc8f20119e639ec5199f361035c7748fc fee08606f4b1eafc4c2f02cd2e79cde4098d93b6 osmarks 1784727481 +0000 commit: Exclude moved allocations from relocation occupancy fee08606f4b1eafc4c2f02cd2e79cde4098d93b6 35b9b466a23e3226a2b1a4a078e6eca0b3dc32c7 osmarks 1784727878 +0000 commit: Minimize static placement relocation cost 35b9b466a23e3226a2b1a4a078e6eca0b3dc32c7 b23263b88cdd927eeadb9e74747ca224ea990fa5 osmarks 1784728313 +0000 commit: Repack transient lifetimes around static runtime b23263b88cdd927eeadb9e74747ca224ea990fa5 d4dac1d91b8498344f962821c47efc9ba7dcece0 osmarks 1784728606 +0000 commit: Allow ordinary transient spill during repack d4dac1d91b8498344f962821c47efc9ba7dcece0 c64115c7d09992f31cb1b7024fef0437916115c2 osmarks 1784728925 +0000 commit: Remove unused local static relocation policy c64115c7d09992f31cb1b7024fef0437916115c2 848ae892fe6dd911f021c735f0c7fc62c80942b2 osmarks 1784729110 +0000 commit: Remove unsuccessful static relocation experiments 848ae892fe6dd911f021c735f0c7fc62c80942b2 9191eebbdcb19a787cab79de3cd25f126a10b142 osmarks 1784729665 +0000 commit: Add compiler repeated-region IR 9191eebbdcb19a787cab79de3cd25f126a10b142 1e2bdf8c6ff12af5d3702fff8e6befb4c2aec8fb osmarks 1784731488 +0000 commit: Unify repeated encoder layer bodies 1e2bdf8c6ff12af5d3702fff8e6befb4c2aec8fb a0755088953af8e1c03b57afb2871cfa9b993920 osmarks 1784737941 +0000 commit: Loop over repeated template instances a0755088953af8e1c03b57afb2871cfa9b993920 958f3782469a8dbf52c24b5bdabbb51b1b7f3cd7 osmarks 1784741028 +0000 commit: Make repeated layer execution preserve tile command order 958f3782469a8dbf52c24b5bdabbb51b1b7f3cd7 ce7efde1407433dacd69d0ce9911a250bf4db156 osmarks 1784741415 +0000 commit: Compress repeated layer records with sparse patches ce7efde1407433dacd69d0ce9911a250bf4db156 79fb1f2e7bb845348615f56509bfacae925c6afc osmarks 1784742007 +0000 commit: Place repeated layer records around live allocations 79fb1f2e7bb845348615f56509bfacae925c6afc 9348c4338a3f5e4b9086fa1ddcf1f929529fe726 osmarks 1784742281 +0000 commit: Minimize transient relocation for template data 9348c4338a3f5e4b9086fa1ddcf1f929529fe726 572d0b74f179b95a68091617f6d7b7dc742743dc osmarks 1784742634 +0000 commit: Segment sparse template patches by mask groups 572d0b74f179b95a68091617f6d7b7dc742743dc e3be2c1cd2512de94f14ff649208514d24b8df7d osmarks 1784743081 +0000 commit: Bound template relocation search cost e3be2c1cd2512de94f14ff649208514d24b8df7d 65006d4c662f7ff8a504c2c54492f20773992c92 osmarks 1784743878 +0000 commit: Precompute template relocation tier per tile 65006d4c662f7ff8a504c2c54492f20773992c92 52b58bbdf2c928b346a93dd9ea673afcde843d1b osmarks 1784744206 +0000 commit: Intern repeated template patch streams 52b58bbdf2c928b346a93dd9ea673afcde843d1b b81d3547abdba636b614cbd9586fad9364c2def5 osmarks 1784744478 +0000 commit: Fit patch streams to fragmented SRAM gaps b81d3547abdba636b614cbd9586fad9364c2def5 f94d3833de3a4735a82dd3bd0cf2c33364bdd309 osmarks 1784744748 +0000 commit: Split dense template patch streams f94d3833de3a4735a82dd3bd0cf2c33364bdd309 295e0fe143ab99e78083fcd0d45b4484b28a6f79 osmarks 1784745026 +0000 commit: Use compact patch mask blocks 295e0fe143ab99e78083fcd0d45b4484b28a6f79 ee67d7f914ba680b5de6bfb8c208440c241068a1 osmarks 1784754403 +0000 commit: Repack transient SRAM around static objects ee67d7f914ba680b5de6bfb8c208440c241068a1 9341253d20e7c2fca3891022ef0f7625533b0e93 osmarks 1784754827 +0000 commit: Sweep active lifetimes during SRAM repacking 9341253d20e7c2fca3891022ef0f7625533b0e93 5b6373aa63854546824c99c01e42152140e1aadb osmarks 1784755091 +0000 commit: Normalize placement inputs and index alias rewrites 5b6373aa63854546824c99c01e42152140e1aadb acf49ca045dbd568982d541ca72f08f3087c9ba8 osmarks 1784755783 +0000 commit: Keep immediate exchange staging on transfers acf49ca045dbd568982d541ca72f08f3087c9ba8 7907e4a1b8f2dad74fdaca8be59831b4aa9a6b77 osmarks 1784755783 +0000 commit: Sweep active SRAM reservations during repacking 7907e4a1b8f2dad74fdaca8be59831b4aa9a6b77 c5f5457b481dc301e4df23a903c7aece5b974137 osmarks 1784756176 +0000 commit: Share kernel specialization metadata c5f5457b481dc301e4df23a903c7aece5b974137 7e917eaa0e69196d8ce43254bd73f4a62ec6c87a osmarks 1784756428 +0000 commit: Omit diagnostic exchange groups during codegen lowering 7e917eaa0e69196d8ce43254bd73f4a62ec6c87a 89202cd63a787ae2ba020f51b29883a49d5a1354 osmarks 1784757740 +0000 commit: Unify and compact static runtime placement 89202cd63a787ae2ba020f51b29883a49d5a1354 5850001903995ed45a2493eabdb02f36ed2727db osmarks 1784758490 +0000 commit: Encode repeated template updates as deltas 5850001903995ed45a2493eabdb02f36ed2727db 6d22b02ef562c33caf6767c960a3e0f287c2eb37 osmarks 1784759328 +0000 commit: Make graph memory placement policy-aware 6d22b02ef562c33caf6767c960a3e0f287c2eb37 d67d2e89e3d5040918b61495fd7a93fc16d6be9d osmarks 1784759637 +0000 commit: Reserve exchange plans during global placement d67d2e89e3d5040918b61495fd7a93fc16d6be9d 3e7473cb4139847c8834dc93c90fa4dd90480fc4 osmarks 1784763138 +0000 commit: Stabilize repeated resident placement 3e7473cb4139847c8834dc93c90fa4dd90480fc4 6e088c327e65ed79076c46430f4dcf4e3505fd27 osmarks 1784767644 +0000 commit: Preserve tile memory classes during runtime placement 6e088c327e65ed79076c46430f4dcf4e3505fd27 990277dabb1cb00df0b6e5af1a4e08b8de25e0ce osmarks 1784768220 +0000 commit: Cache identical resolved exchange phases 990277dabb1cb00df0b6e5af1a4e08b8de25e0ce 4f3cd9689c3eecd91a597231adabb6d0c2e02874 osmarks 1784769162 +0000 commit: Index allocation lifetime queries 4f3cd9689c3eecd91a597231adabb6d0c2e02874 599dfd92c0b89fdc4028cac1415d2269970e8ab3 osmarks 1784770177 +0000 commit: Index active allocation placement queries 599dfd92c0b89fdc4028cac1415d2269970e8ab3 021736798160af431125ae2ae4dab86fa786d9eb osmarks 1784771087 +0000 commit: Index tile program lowering inputs 021736798160af431125ae2ae4dab86fa786d9eb df717f26a6eda69777eaf49905c5a8c8e3a7e262 osmarks 1784771658 +0000 commit: Move occupancy ownership into allocation storage df717f26a6eda69777eaf49905c5a8c8e3a7e262 22ba8f57ce387bba767f543427a3ca25f104dd36 osmarks 1784771936 +0000 commit: Parallelize deterministic host tensor packing 22ba8f57ce387bba767f543427a3ca25f104dd36 b32981a39b866bc26265282e8bb3770ec370432a osmarks 1784773205 +0000 commit: Enable scalable all-tile phase profiling b32981a39b866bc26265282e8bb3770ec370432a de1999f42440db75db70ebd2b736f349133c2f0d osmarks 1784775484 +0000 commit: Compact code and tensors with lifetime-aware placement de1999f42440db75db70ebd2b736f349133c2f0d c0cfbc8b6bf0e985d088c6474a763a26b3b680c2 osmarks 1784777004 +0000 commit: Relocate all owned transient storage c0cfbc8b6bf0e985d088c6474a763a26b3b680c2 71aec9f85097574173af154102d7203cf5d9bebb osmarks 1784778070 +0000 commit: Enforce IPU memory-element placement constraints 71aec9f85097574173af154102d7203cf5d9bebb 8da62604d6371da03cf326b1c5e7ea3a41bfa86f osmarks 1784779126 +0000 commit: Honor kernel memory classes during relocation 8da62604d6371da03cf326b1c5e7ea3a41bfa86f 35c08b73d55c5c161a1a2b74449122cadbd32ad4 osmarks 1784779418 +0000 commit: Allow ordinary data in interleaved address ranges 35c08b73d55c5c161a1a2b74449122cadbd32ad4 bc2e278648390f7db1af5f133fdcaa1dfb27fbb8 osmarks 1784779891 +0000 commit: Preserve constraints across static template compaction bc2e278648390f7db1af5f133fdcaa1dfb27fbb8 f70351159dc2d39e24b46d898bb6f3fb11705065 osmarks 1784780510 +0000 commit: Model kernel memory constraints over access spans f70351159dc2d39e24b46d898bb6f3fb11705065 ed6d5d6499524c10894661a634d7f0d4c425e0ab osmarks 1784781822 +0000 commit: Validate packed template transition streams ed6d5d6499524c10894661a634d7f0d4c425e0ab 947863ff99a7620b542006e5d9c521645458dcdc osmarks 1784783552 +0000 commit: Encode kernel memory ABI explicitly 947863ff99a7620b542006e5d9c521645458dcdc 1717ced43ae1a6a6db58d076d7d28713e8862891 osmarks 1784785373 +0000 commit: Reserve kernel instruction access footprints 1717ced43ae1a6a6db58d076d7d28713e8862891 7a6ec29db049b11fc0f702a6b0670f61aa29b3ca osmarks 1784785386 +0000 commit: Add tile context exception diagnostics 7a6ec29db049b11fc0f702a6b0670f61aa29b3ca 8eeaa8a21b8e039786faa7e88f09898c50bbf109 osmarks 1784786500 +0000 commit: Separate kernel access spans from storage ownership 8eeaa8a21b8e039786faa7e88f09898c50bbf109 d47116b3ab9571fb2b2f2aed68ce522b115e8f35 osmarks 1784787237 +0000 commit: Separate kernel access extents from allocation size d47116b3ab9571fb2b2f2aed68ce522b115e8f35 9a01ab5d7e035fb2f5d3b6542d558370780c56b8 osmarks 1784788439 +0000 commit: Pin executable placement across tensor relayout 9a01ab5d7e035fb2f5d3b6542d558370780c56b8 8f0b8fb7685efa02f7c05f7863e946a8240b4273 osmarks 1784788521 +0000 commit: Use dense tensor allocation indexes 8f0b8fb7685efa02f7c05f7863e946a8240b4273 90d405ef4692e1efc7b61e2689aea03f7478d53a osmarks 1784788883 +0000 commit: Respect static access spans during tensor relayout 90d405ef4692e1efc7b61e2689aea03f7478d53a fa1925325d55c5775253877412c2454e69928231 osmarks 1784790257 +0000 commit: Make static packaging placement iterative fa1925325d55c5775253877412c2454e69928231 b43a789b8042f62e8843e306be5996052387b8e2 osmarks 1784791706 +0000 commit: Resolve compacted tensor addresses by lifetime b43a789b8042f62e8843e306be5996052387b8e2 6aa3560bbf8419b47ace4bbb6e292c29b248ef63 osmarks 1784792107 +0000 commit: Model compute subviews as live aliases 6aa3560bbf8419b47ace4bbb6e292c29b248ef63 51d10de59d70f51c4eb640be37ef4216c665787c osmarks 1784793103 +0000 commit: Propagate relocation through alias chains 51d10de59d70f51c4eb640be37ef4216c665787c abcb547a7c38bca3c2e0305075f6c38c9043631c osmarks 1784793791 +0000 commit: Preserve valid allocations during package relayout abcb547a7c38bca3c2e0305075f6c38c9043631c 139817b73629b8ea5560285ed7c817f53db44084 osmarks 1784794644 +0000 commit: Place support code across executable intervals 139817b73629b8ea5560285ed7c817f53db44084 bedcf1728980230695b9fac33eca2cf0880249df osmarks 1784795495 +0000 commit: Pack executable objects within shared memory elements bedcf1728980230695b9fac33eca2cf0880249df 684d857644356813248b7ec7b0f4478cae4d3374 osmarks 1784795981 +0000 commit: Use architectural base for segmented relocations 684d857644356813248b7ec7b0f4478cae4d3374 1baa724249096ad366da82aefff991bee6b67cd5 osmarks 1784796621 +0000 commit: Separate Colossus relocation and placement bases 1baa724249096ad366da82aefff991bee6b67cd5 5f3ac5caed2823bcdbd95dff956dff5d3eadebc3 osmarks 1784797008 +0000 commit: Index resident allocation occupancy 5f3ac5caed2823bcdbd95dff956dff5d3eadebc3 172db47cf6f3b41a45dff9d4e51fcd269248ea25 osmarks 1784797380 +0000 commit: Remove late repeated-region command scans 172db47cf6f3b41a45dff9d4e51fcd269248ea25 310de983ce69f2373ec82d893436c1a2f8400eb0 osmarks 1784797938 +0000 commit: Repack complete tile memory on global compaction 310de983ce69f2373ec82d893436c1a2f8400eb0 46a7fc418df4ff98f4cc26a5cee81a7e41f54e1c osmarks 1784798250 +0000 commit: Preserve constrained bases during tile repacking 46a7fc418df4ff98f4cc26a5cee81a7e41f54e1c 5da8ad46850640ce792629b553c797a2c9103d81 osmarks 1784798720 +0000 commit: Place constrained allocations before unconstrained storage 5da8ad46850640ce792629b553c797a2c9103d81 54a963f2fab8ef92b0aac955e2df34e8b389a101 osmarks 1784799431 +0000 commit: Allow aggregate profiles across irregular template instances 54a963f2fab8ef92b0aac955e2df34e8b389a101 762e3f7f5c0e4635074f3cb32006bc84064b8bc8 osmarks 1784799631 +0000 commit: Honor graph profiling granularity in SigLIP runner 762e3f7f5c0e4635074f3cb32006bc84064b8bc8 deae85cc044c54450e72b4d27097c8f755eb62f7 osmarks 1784800128 +0000 commit: Defragment tile memory with targeted eviction deae85cc044c54450e72b4d27097c8f755eb62f7 66d71b661245bbca55de3b727e982268170120a1 osmarks 1784801013 +0000 commit: Treat FP8 scales as repeated instance data 66d71b661245bbca55de3b727e982268170120a1 3802f88e872d2c48718a8f2b7350edef051cb721 osmarks 1784801426 +0000 commit: Relocate graph memory through stable allocation references 3802f88e872d2c48718a8f2b7350edef051cb721 ef4d61a3c2e9662283d63da76458feebc0fded45 osmarks 1784801957 +0000 commit: Repack all movable graph allocations globally ef4d61a3c2e9662283d63da76458feebc0fded45 4cc3987aea4e13fb1786621dd8a356b78a1d229f osmarks 1784802876 +0000 commit: Test wide template patches on hardware 4cc3987aea4e13fb1786621dd8a356b78a1d229f e7c39db98c6d0033cf9194575d7b445214c27797 osmarks 1784804102 +0000 commit: Represent composite layer norm storage with aliases e7c39db98c6d0033cf9194575d7b445214c27797 b4e5b44d881a9557df5c79613355795d8fda54df osmarks 1784804287 +0000 commit: Reject unowned graph addresses during relocation b4e5b44d881a9557df5c79613355795d8fda54df 8ef38d6ad3cddbbbd0cde628d3a5e6a11d413afb osmarks 1784804302 +0000 commit: Verify SigLIP encoder boundaries on hardware 8ef38d6ad3cddbbbd0cde628d3a5e6a11d413afb bba3f65637b653bf55b72275bfa110f818fe65b8 osmarks 1784804302 +0000 commit: Stress repeated template patches across model depth bba3f65637b653bf55b72275bfa110f818fe65b8 193f25b90cfa20c680a18cd7698feeefb210f503 osmarks 1784810197 +0000 commit: Support statically specialized batched SigLIP MAP heads 193f25b90cfa20c680a18cd7698feeefb210f503 13eb88d075d676a65abd99eeb4cec15f3f70bf2e osmarks 1784810742 +0000 commit: Support mixed SigLIP precision schedules 13eb88d075d676a65abd99eeb4cec15f3f70bf2e 85d989e8bbe5c14a98c735b3931808164890549a osmarks 1784829247 +0000 commit: Enforce the SDK secondary-loader image limit 85d989e8bbe5c14a98c735b3931808164890549a b13f531624518f9365cd18027e9f4bc46fc26fb6 osmarks 1784831946 +0000 commit: Harden IPU placement and SigLIP tuning b13f531624518f9365cd18027e9f4bc46fc26fb6 7493746a70c4a7e2666546e59a59494d1d73dbc3 osmarks 1784832655 +0000 commit: Report SigLIP accuracy for every run 7493746a70c4a7e2666546e59a59494d1d73dbc3 0f24306c08304725d63435462ab1e82a02141455 osmarks 1784834373 +0000 commit: Respect IPU executable element boundaries 0f24306c08304725d63435462ab1e82a02141455 e6f4b6d2f61d9a7ab8d594471f619fb0012056e7 osmarks 1784835212 +0000 commit: Segment generated tile control programs e6f4b6d2f61d9a7ab8d594471f619fb0012056e7 54ba9a131a391d8727fc64703ed62a31185e9024 osmarks 1784837151 +0000 commit: Synchronize bootloader workers at application entry 54ba9a131a391d8727fc64703ed62a31185e9024 33cc8d3ffcbab4c988589692f8b89fe59e9e6a70 osmarks 1785011196 +0000 commit: Initialize padded attention operands 33cc8d3ffcbab4c988589692f8b89fe59e9e6a70 5e83113ed1006baeabc9b57bf81ff0af8ead7f9a osmarks 1785011243 +0000 commit: Track PACE access spans during placement 5e83113ed1006baeabc9b57bf81ff0af8ead7f9a cdbcd7ced845d9b43d2486c1e4e224705a383e1c osmarks 1785011521 +0000 commit: Validate and segment repeated tile programs cdbcd7ced845d9b43d2486c1e4e224705a383e1c 733fad182f94abb817ab72bf9169910e18665460 osmarks 1785011529 +0000 commit: Batch diagnostic SRAM allocation by phase 733fad182f94abb817ab72bf9169910e18665460 448548a80963a90c06a556b361e9428af8159d1b osmarks 1785012734 +0000 commit: Choose feasible LayerNorm parameter owners 448548a80963a90c06a556b361e9428af8159d1b bd5c04f15a08377450be16c8aae6cb7a1b8d52fd osmarks 1785012988 +0000 commit: Reserve prior transient ranges for LayerNorm constants bd5c04f15a08377450be16c8aae6cb7a1b8d52fd 02e439bd8d9a1d10cf33ec8b1d486bcf25209454 osmarks 1785013270 +0000 commit: Pack large activation spans before resident fragments 02e439bd8d9a1d10cf33ec8b1d486bcf25209454 b3c7ea2aaee7bf4e9e9daff7c7c3d6ed5ac5aa0f osmarks 1785013468 +0000 commit: Prioritize span size over relocatable memory relations b3c7ea2aaee7bf4e9e9daff7c7c3d6ed5ac5aa0f ff05cbabf28e46b1b74ef75ba43875c3b0d87fca osmarks 1785013736 +0000 commit: Make compaction eviction priority monotonic ff05cbabf28e46b1b74ef75ba43875c3b0d87fca cdc5775eda069559c12b271594a1a949118e60a9 osmarks 1785014346 +0000 commit: Relocate self-separated kernel operands cdc5775eda069559c12b271594a1a949118e60a9 25fd87a9a7d05572bb96fad9efdcc012b3a2553f osmarks 1785014580 +0000 commit: Globally recolor transients around executable code 25fd87a9a7d05572bb96fad9efdcc012b3a2553f 87e9f4bd15c332e1168ab4a2c479fd38b6d7a7db osmarks 1785014844 +0000 commit: Fall back to whole-graph memory recoloring 87e9f4bd15c332e1168ab4a2c479fd38b6d7a7db 58a1d30bb153cbb5aeb4e6302fd25c02f5db3acd osmarks 1785015782 +0000 commit: Balance attention tasks against parent tile pressure 58a1d30bb153cbb5aeb4e6302fd25c02f5db3acd 99cf530fc926a5842479960e14f2d958f2cbcb49 osmarks 1785015913 +0000 commit: Keep repeated attention tile maps stable 99cf530fc926a5842479960e14f2d958f2cbcb49 6650aec827234b68da2eeb2af587a10b421a1128 osmarks 1785016482 +0000 commit: Defer resident memory coloring to whole graph 6650aec827234b68da2eeb2af587a10b421a1128 a96f511aa3c2c05c09052a6d12fcf8c0e518f36e osmarks 1785019990 +0000 commit: Generalize transient placement for batched attention a96f511aa3c2c05c09052a6d12fcf8c0e518f36e 38995be031ef1dc8e18efbfe2849b255e411f379 osmarks 1785021682 +0000 commit: Balance resident code pressure and time host calls 38995be031ef1dc8e18efbfe2849b255e411f379 a4a8ed0aef2b717870c0f4eb3e21623def87786d osmarks 1785022708 +0000 commit: Index only moved storage during graph relayout a4a8ed0aef2b717870c0f4eb3e21623def87786d ced30507b09ef254ce92ffa729d99293f200460a osmarks 1785027264 +0000 commit: Converge profiled executable placement ced30507b09ef254ce92ffa729d99293f200460a ba7339c8c72cde2377fd96adaab9640d4f5d1f48 osmarks 1785029211 +0000 commit: Profile semantic regions through static templates ba7339c8c72cde2377fd96adaab9640d4f5d1f48 bfb52a70ae3b7faf29d583172841df3c4fa1314a osmarks 1785029731 +0000 commit: Parallelize generic row reblocking bfb52a70ae3b7faf29d583172841df3c4fa1314a 717186b607ed6a0347204b5959b6a093b2a4c3a0 osmarks 1785030343 +0000 commit: Support full-width static kernel arguments 717186b607ed6a0347204b5959b6a093b2a4c3a0 c55876bb2f20494b4c2db7e86457c892aa79d7bf osmarks 1785030514 +0000 commit: Coalesce contiguous row reblock fragments c55876bb2f20494b4c2db7e86457c892aa79d7bf 92ef9bfca1a3e7ed069211517623fb14e2b260c5 osmarks 1785031202 +0000 commit: Split oversized D2D transfers during lowering 92ef9bfca1a3e7ed069211517623fb14e2b260c5 01491f4a472b13140d4ec6bcd252b38120b2aaa9 osmarks 1785031755 +0000 commit: Reserve repeated-block transient headroom 01491f4a472b13140d4ec6bcd252b38120b2aaa9 e8b1dcd5f93add561283eb7dd8300d59d9c6f984 osmarks 1785032386 +0000 revert: Revert "Reserve repeated-block transient headroom" e8b1dcd5f93add561283eb7dd8300d59d9c6f984 2f3cfa54480924a5ef559aa999048376917b80a5 osmarks 1785036685 +0000 commit: Balance repeated encoder memory placement 2f3cfa54480924a5ef559aa999048376917b80a5 38ace159c54660f2ed182d320a413ee320925336 osmarks 1785037347 +0000 commit: Optimize repeated singleton resident owners 38ace159c54660f2ed182d320a413ee320925336 8e20135dd3f33a6d02b45d36d7b11f253bf6e594 osmarks 1785038996 +0000 commit: Pack permanent SRAM before transient lifetimes 8e20135dd3f33a6d02b45d36d7b11f253bf6e594 c3fa87bf6a2bc92ef8b7e9f72a98181ab11e5bd4 osmarks 1785039345 +0000 commit: Retain size-first ordering for global SRAM packing c3fa87bf6a2bc92ef8b7e9f72a98181ab11e5bd4 0ae76a9c907acecb5bf0968fe57fb3276668ead8 osmarks 1785040135 +0000 commit: Preflight repeated placement against support images 0ae76a9c907acecb5bf0968fe57fb3276668ead8 c30b8e2269f99d20b17556bf9b49d0fa326844f7 osmarks 1785040292 +0000 commit: Recompute executable placement after code shape changes c30b8e2269f99d20b17556bf9b49d0fa326844f7 edde7243aa0b202d98a35a40b07b5cbed3adb368 osmarks 1785041225 +0000 commit: Measure static runtime pressure during placement edde7243aa0b202d98a35a40b07b5cbed3adb368 3b5dfb64e8017d08b0ff45c271e3d9def81f0887 osmarks 1785046730 +0000 commit: Fit repeated programs using measured static placement 3b5dfb64e8017d08b0ff45c271e3d9def81f0887 2a32c7dcf0e2e4e73d0b4b057e427cc8864b5fc4 osmarks 1785047299 +0000 commit: Bound repeated static placement refinement 2a32c7dcf0e2e4e73d0b4b057e427cc8864b5fc4 a19f255a2b824601a8775880b1106307b22c3635 osmarks 1785047902 +0000 commit: Use best-fit for offline multi-arena repacking a19f255a2b824601a8775880b1106307b22c3635 76b63fa9ee5c5f3c59d2df97e6f8dc9da86eb2d2 osmarks 1785048491 +0000 commit: Balance repeated placement against model suffix 76b63fa9ee5c5f3c59d2df97e6f8dc9da86eb2d2 dbfab6b382a03975fe0642940853f8ffbe546c74 osmarks 1785049054 +0000 commit: Preserve resident capacity during role rotation dbfab6b382a03975fe0642940853f8ffbe546c74 df208f86205d2d0d9624956b8f3b94e117c58707 osmarks 1785049524 +0000 commit: Use the loadable SRAM limit for SigLIP planning df208f86205d2d0d9624956b8f3b94e117c58707 367a5fde698578804da34148a30d06afc8e09bff osmarks 1785050202 +0000 commit: Reserve discrete packing slack during role placement 367a5fde698578804da34148a30d06afc8e09bff 000dc32c5e440292dec7359d7b9c2b9990843485 osmarks 1785050645 +0000 commit: Scope role packing slack to repeated residents 000dc32c5e440292dec7359d7b9c2b9990843485 534d3a3199dae4cd8d0c514da771ea8a9ea986a8 osmarks 1785051287 +0000 commit: Place resident storage before transient lifetimes 534d3a3199dae4cd8d0c514da771ea8a9ea986a8 4a145b63622516b6d2992dd647e5a318574b5e1f osmarks 1785051747 +0000 commit: Measure steady-state repeated transient pressure 4a145b63622516b6d2992dd647e5a318574b5e1f 57a4ef629ff0cc6d1f1335bb5dfc7f8f6fe8742c osmarks 1785052290 +0000 commit: Converge repeated placement with steady-state pressure 57a4ef629ff0cc6d1f1335bb5dfc7f8f6fe8742c cf88e805a6efcd5dbcafe0b2ba9a3009bd2f9aca osmarks 1785052962 +0000 commit: Rotate repeated transient pressure with tile roles cf88e805a6efcd5dbcafe0b2ba9a3009bd2f9aca 528d8a3a24695fb7549b0f94efced9ce45cbe7f5 osmarks 1785053434 +0000 commit: Respect transient arena order during global repacking