Intel EU / XVE Stall Counter Visualizer

Audited tiny examples: documented counter semantics separated from performance interpretation
Pseudo ASM · source-scoped semantics
Fast mental model · audited
ALUWR / SBID

Both are dependency-side waits, but different dependencies: ALUWR is waiting for an ALU write/result; SBID is scoreboard token/dependency state. Neither alone proves DRAM latency or physical write-port contention.

PIPESTALL

The instruction can be ready yet unable to enter the target pipe. Intel VTune gives a GRF bank conflict as a concrete example.

SENDWR

This one must be source-scoped: Arc/ACM = waiting for SEND to write GRF; Lunar Lake-style metric = waiting for SEND dispatch from XVE.

Compact audited reference
CounterDocumented core meaningSource scopeGood tiny mental model
EU_STALL_ALUWRWaiting for ALU to write GRF/ACCArc/ACM PRMproducer → destination not written → dependent consumer waits
EU_STALL_SENDWRArc: SEND writes GRF; newer Xe: SEND dispatch waitGeneration-sensitivedo not use one universal flow
EU_STALL_SBIDScoreboard token/dependency unavailableArc/ACM + VTune + LNLasync tracked work holds scoreboard state → later work waits
EU_STALL_PIPESTALLReady instruction cannot dispatch to target pipeVTune XVE / LNLwon arbitration → GRF/resource conflict → pipe cannot accept
EU_STALL_PS_DEPENDENCYPixel Shader dependencyLunar Lake metricgraphics dependency remains unresolved
EU_STALL_INSTFETCHWaiting for Instruction FetchArc/ACM + VTunenew PC → target instruction not available
EU_STALL_CONTROLWaiting for Branch unit / branch completionVTune XVE / LNLconditional branch → control unit unavailable
EU_STALL_BARRIERWaiting for synchronization dependency; Arc mentions Gateway NotifyArc/ACM + VTunesome work-items arrive → others late → barrier wait
EU_STALL_OTHEROther dependency; Flag/EoT are explicit examplesArc/ACM + LNLflag/predicate dependency not covered elsewhere
Audit sources
Primary: Intel Arc A-Series / ACM PRM Vol. 13 event definitions; Intel VTune — Reasons for XVE Stalls; Intel VTune — XVE Pipe Stall; Intel VTune — XVE SBID Stall.
Cross-check for newer Xe/Lunar Lake metric wording: Mesa Intel performance-metric XML (including PS_DEPENDENCY and newer SENDWR wording).