Compile bug: SYCL AOT build for mtl-m (Xe-LPG) fails: `dpas.nosrc0` intrinsic not supported, introduced by d81d0d8fb

Aperta
  • Qualcosa non va?
#31,033 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
1-2 giorni
Idoneità per principianti
48/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
cmake, cpp, linux
Ambito
backend, build-system

Direzione di ricerca

Start with the regression commit d81d0d8fb95e52a66992bcfc212b1891b4d2f94a and the XMX implementation at ggml/src/ggml-sycl/esimd.hpp:1036, where xmx::dpas is called. Reproduce using the provided CMake configuration for GGML_SYCL on mtl-m and inspect how AOT code generation handles that kernel. Done means AOT compilation succeeds on Xe-LPG using a supported DPAS form or by excluding the unsupported kernels for that target.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

bug-unconfirmed
Git commit
1ab09d632ebc30f171f58547a1110f8d6e9546c0 (b11834)
Operating systems

Linux

GGML backends

SYCL

Problem description & steps to reproduce

Compiling llama.cpp with the SYCL backend using AOT compilation for Meteor Lake iGPU (-device mtl-m, Xe-LPG) fails since d81d0d8fb (b11821, PR #29864 "Multi-column on Intel XMX matrix engines refactored"). The new XMX kernels call xmx::dpas (ggml/src/ggml-sycl/esimd.hpp:1036), which the compiler lowers to the llvm.genx.dpas.nosrc0 intrinsic. This intrinsic variant is not supported on Xe-LPG — the DPAS unit on Meteor Lake does not support the no-src0 form (it is a Xe2+ feature), so the AOT backend pass (GenXCisaBuilder) aborts.

The previous build tree at 39da130fcb (b11770, oneAPI 2026.2) compiled fine with the same flags, so this is a regression introduced in this batch. Presumably PR #29864 was validated on Xe2 hardware (B70), where dpas.nosrc0 exists. JIT builds may not notice because native code generation happens lazily at kernel launch; AOT codegen for Xe-LPG fails at build time for any kernel containing the intrinsic.

Expected: AOT compilation for Xe-LPG targets should either succeed (using a DPAS form supported on Xe-LPG) or the XMX multi-column kernels should be excluded from codegen for targets that lack dpas.nosrc0.

First Bad Commit
d81d0d8fb95e52a66992bcfc212b1891b4d2f94a (b11821, PR #29864)

Last known good: 39da130fcb (b11770) built successfully with identical flags and toolchain.

Compile command

cmake -B build-aot \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DEVICE_ARCH=mtl-m \
  -DGGML_SYCL_DNN=OFF
cmake --build build-aot -j16


Toolchain: Intel oneAPI DPC++/C++ Compiler 2026.2.0, Debian 13, kernel 6.12.113.
Relevant log output

[ 77%] Linking CXX shared library ../../../bin/libggml-sycl.so.0.27.0
error: LLVM ERROR: GenXCisaBuilder failed for: <  %.esimd28 = tail call <16 x i32> @llvm.genx.dpas.nosrc0.v16i32.v128i32.v8i32(<128 x i32> %.esimd20.join48, <8 x i32> %.esimd11, i32 17303560)>: Intrinsic is not supported by the <XeLPG> platform
error: LLVM ERROR: GenXCisaBuilder failed for: <  %.esimd30 = tail call <16 x i32> @llvm.genx.dpas.nosrc0.v16i32.v128i32.v8i32(<128 x i32> %.esimd24.join48, <8 x i32> %.esimd111997, i32 17303560)>: Intrinsic is not supported by the <XeLPG> platform
error: LLVM ERROR: GenXCisaBuilder failed for: <  %.esimd28.1 = tail call <16 x i32> @llvm.genx.dpas.nosrc0.v16i32.v128i32.v8i32(<128 x i32> %.esimd20.1.join48, <8 x i32> %.esimd11.1, i32 17303560)>: Intrinsic is not supported by the <XeLPG> platform
gmake[2]: *** [ggml/src/ggml-sycl/CMakeFiles/ggml-sycl.dir/build.make:1772：bin/libggml-sycl.so.0.27.0] 错误 1
gmake[1]: *** [CMakeFiles/Makefile2:2617：ggml/src/ggml-sycl/CMakeFiles/ggml-sycl.dir/all] 错误 2
gmake: *** [Makefile:146：all] 错误 2
Lingua principale
C++
Stelle
133k
Fork
24.9k
Merge medio
2g 1h
PR unite (30g)
475

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di ggml-org/llama.cpp

Tutte le issue di ggml-org/llama.cpp

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.