[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85571-en":3,"doc-seo-85571-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},85571,1649267921044,"Ava Thompson","https://us-avatar.wpscdn.com/avatar/1800007509477c92dfb?_k=1782875107921204101",8,"Research & Report","TensorLift Automatic Extraction of Tensor Level ISA Semantics from Accelerator RTL via MLIR Semantic Lifting","TensorLift delivers an end-to-end MLIR pipeline that lifts RTL-extracted accelerator semantics into TAIDL-like tensor ISA specifications. Unlike manual tensor-ISA authoring, it introduces an 8-pass semantic lifting workflow that progressively reconstructs high-level tensor structure, including MAC idioms, saturation behavior, multi-dimensional buffer organization, and data-layout transformations. Generated TAIDL specifications directly enable automatic software-stack creation in the ACT ecosystem. Evaluations on Gemmini and TVM VTA validate correctness with SMT equivalence proofs.","TensorLift: Automatic Extraction of Tensor-Level ISA Semantics from Accelerator RTL via MLIR Semantic Lifting  \nRuijie Gao  \n[ruijieg@umich.edu](ruijieg@umich.edu)[ ](ruijieg@umich.edu)University of Michigan Computer Science & Engineering Ann Arbor, Michigan, USA  \nHaoran Jin  \n[allenjin@umich.edu](allenjin@umich.edu)[ ](allenjin@umich.edu)University of Michigan Computer Science & Engineering Ann Arbor, Michigan, USA  \nJirong Yang∗ [polarisyjr@utexas.edu](polarisyjr@utexas.edu)[ ](polarisyjr@utexas.edu)University of Texas at Austin Computer Science Austin, Texas, USA  \nNathaniel Bleier  \n[nbleier@umich.edu](nbleier@umich.edu)[ ](nbleier@umich.edu)University of Michigan Computer Science & Engineering Ann Arbor, Michigan, USA  \narXiv :2604 . 13523v2 [ cs .AR] 12 Jul 2026  \nAbstract  \nNumerous tensor accelerator designs have been proposed in academia and industry, yet most lack well-documented ISAs and compiler backends. As a consequence, the majority of new designs are only evaluated on a handful of operators or synthetic kernels. Recent work (e.g., TAIDL and the ACT ecosystem) has shown that given a tensor-level ISA specification, complete software stacks including test oracles and compiler backends can be automatically generated. However, writing such specifications remains a manual, expertdriven process.  \nWe present TensorLift, the first end-to-end MLIR-based pipeline that lifts RTL-extracted accelerator semantics to TAIDL-like tensor ISA specifications. Building on prior architecture-level model extraction that produces bit-level LLVM IR, TensorLift introduces an 8-pass MLIR semantic lifting pipeline that progressively recovers high-level tensor structure, including MAC idioms, saturation semantics, multi-dimensional buffer organizations, and data layout transformations. The pipeline emits specifications in the TAIDL formalism, immediately enabling automatic software stack generation through the ACT ecosystem.  \nWe evaluate TensorLift on Gemmini, a systolic-array accelerator, extracting all hardware instructions semantics across 127 MLIR files. The lifting pipeline collapses the extracted bit-level MLIR by up to 92.9% on processing elements (686 → 49 lines, of which only 17 encode the tensor computation core) and by 24–34% on controller modules, with the residue being irreducible control logic (FSMs, address computation, and instruction muxing) . TensorLift discovers hardware features omitted from the hand-written reference, including multi-bank DMA configuration, pooling engine semantics, and im2col hardware support. Correctness is validated through Z3 SMT equivalence proofs. We further confirm generality on TVM’s VTA tensor processor, where the same pipeline lifts all four datapath modules without accelerator-specific changes, with extracted semantics formally verified against the underlying RTL. By feeding the extracted specification into the ACT compiler framework, TensorLift enables an automated path from RTL to a  \n∗ This work was done while Jirong Yang was at the University of Michigan, Computer Science & Engineering.  \nperformance-competitive compiler backend, eliminating the need to manually write tensor-level ISA semantics.  \nKeywords  \nTensor Accelerators, ISA Extraction, RTL Abstraction, MLIR, Semantic Lifting, Hardware-Software Interface  \n1 Introduction  \nThe rapid growth of deep learning workloads has driven a proliferation of specialized tensor accelerators in both academia and industry. Designs such as Gemmini [13], MAERI [24], SIGMA [38], Eyeriss [6], NVDLA [8], and FEATHER [43] explore diverse microarchitectural innovations—systolic arrays, reconfigurable interconnects, flexible dataflow engines—to deliver orders-of-magnitude improvements in performance and energy efficiency over general-purpose processors. Meanwhile, a small number of platforms backed by major vendors, notably Google’s TPU [23] with XLA [39], Intel AMX [19] with oneDNN [35], and NVIDIA GPUs with CUDA/cuDNN [7, 32], enjoy mature compiler ","cbCaiqz8vsM92f2t","https://ap.wps.com/l/cbCaiqz8vsM92f2t","pdf",1102824,2,1,10,"English","en",105,"# Abstract\n# Introduction\n# Existing Landscape (The Critical Gap)\n# TensorLift: Bridging the Gap","[{\"question\":\"What problem does TensorLift address?\",\"answer\":\"Most tensor accelerators lack formally documented ISA semantics, forcing software to rely on limited hand-written kernel libraries. TensorLift provides an automated path from accelerator RTL to tensor-level ISA specifications.\"},{\"question\":\"How does TensorLift work at a high level?\",\"answer\":\"TensorLift uses an end-to-end MLIR pipeline with eight passes to lift RTL-extracted semantics into TAIDL-like tensor ISA descriptions, progressively recovering tensor structure and data-layout transformations.\"},{\"question\":\"How is correctness verified in TensorLift?\",\"answer\":\"Correctness is validated through Z3 SMT equivalence proofs by comparing lifted specifications against the underlying RTL semantics.\"}]",1784204673,25,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"tensorlift-automatic-extraction-of-tensor-level-isa-semantics-from-accelerator-rtl-via-mlir-semantic-lifting","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,47,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":20},"https://docshare.wps.com/document/","Document",{"item":48,"name":12,"@type":43,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/tensorlift-automatic-extraction-of-tensor-level-isa-semantics-from-accelerator-rtl-via-mlir-semantic-lifting/85571/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-23","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does TensorLift address?","Question",{"text":75,"@type":76},"Most tensor accelerators lack formally documented ISA semantics, forcing software to rely on limited hand-written kernel libraries. TensorLift provides an automated path from accelerator RTL to tensor-level ISA specifications.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does TensorLift work at a high level?",{"text":80,"@type":76},"TensorLift uses an end-to-end MLIR pipeline with eight passes to lift RTL-extracted semantics into TAIDL-like tensor ISA descriptions, progressively recovering tensor structure and data-layout transformations.",{"name":82,"@type":73,"acceptedAnswer":83},"How is correctness verified in TensorLift?",{"text":84,"@type":76},"Correctness is validated through Z3 SMT equivalence proofs by comparing lifted specifications against the underlying RTL semantics.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":22,"slug":133},"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]