extract-asm-onednn

name	extract-asm-onednn
description	Extract GPU ISA from oneDNN ngen-JIT kernels. This is the ONLY codegen path that bypasses the standard SYCL/SPIR-V/IGC stack. oneDNN uses its own native code generator (ngen) that directly emits GPU ISA bytes — no SPIR-V, no IGC, no zebin ELF, no .debug_line. Use when extracting ASM from oneDNN kernels (gemm_kernel, gen_conv_kernel), matmul, linear, conv, or SDPA-graph ops dispatched via mkldnn.

Extract ASM from oneDNN ngen-JIT Kernels

This is the ONLY path that does NOT use the standard compilation stack. All other scenarios (SYCL AOT/JIT, Triton) go through SPIR-V → IGC → zebin. oneDNN ngen bypasses all of that.

How this differs from the standard stack

Standard stack (SYCL/Triton):
  Source → LLVM IR → SPIR-V → IGC → zebin ELF (.text + .debug_line)
                                         ↓
                                    ocloc disasm → .asm

oneDNN ngen (THIS skill):
  oneDNN C++ templates → ngen JIT → RAW ISA BYTES (no ELF, no DWARF)
                                         ↓
                                    IGA ctypes → .asm

Key consequences:

• No zebin — output is raw bytes, not ELF
• No .debug_line — source mapping can only use pattern recognition
• No IGC — IGC_ShaderDumpEnable is useless
• No SPIR-V — ngen directly emits machine instructions
• Dump mechanism: ONEDNN_JIT_DUMP=1 (writes .bin files)
• Disassembly: IGA library via ctypes (not ocloc disasm)

When to use

• Op dispatched through oneDNN (mkldnn::*): linear / matmul / mm / bmm / conv* / _scaled_dot_product_attention (oneDNN-graph)

When NOT to use

• Kernel is triton_* → use extract-asm-triton
• Kernel is _ZTS… (SYCL) → use extract-asm-syclkernel-{aot,jit}
• You only need primitive-level timing → use benchdnn directly

Steps

Step 0: Locate tools

# libiga64.so: shipped with oneAPI debugger component
# Detect oneAPI root: check env vars first, then common install locations
ONEAPI=${ONEAPI_ROOT:-${CMPLR_ROOT:+${CMPLR_ROOT%/*}}}
if [ -z "$ONEAPI" ]; then
  for d in /opt/intel/oneapi ~/intel/oneapi /usr/local/oneapi; do
    [ -d "$d" ] && ONEAPI="$d" && break
  done
fi
IGA_LIB=$(find ${ONEAPI:?"oneAPI not found; set ONEAPI_ROOT"} -name 'libiga64.so' 2>/dev/null | head -1)
test -n "$IGA_LIB" || { echo "libiga64.so not found under $ONEAPI; install oneAPI debugger"; exit 1; }
export IGA_LIB

1. Dump raw ISA via ONEDNN_JIT_DUMP.

   OUT="<workdir>/onednn_$(date +%Y%m%d_%H%M%S)"
   mkdir -p "$OUT" && cd "$OUT"
   
   ONEDNN_JIT_DUMP=1 \
   ONEAPI_DEVICE_SELECTOR=level_zero:0 \
     <repro_cmd> 2>&1 | tee run.log
   
   ls dnnl_dump_gpu_*.bin
   

   These are raw ISA bytes (not zebin ELF). ocloc disasm cannot read them.

2. Disassemble with IGA ctypes.

   oneDNN .bin files are raw ISA bytes (not ELF). Use libiga64.so via Python ctypes to disassemble. No separate script needed — run inline:

   for bin in dnnl_dump_gpu_*.bin; do
     name=$(basename "$bin" .bin)
     python3 -c "
   

import ctypes, sys, pathlib iga = ctypes.CDLL('$IGA_LIB') raw = pathlib.Path('$bin').read_bytes() buf = ctypes.create_string_buffer(1 << 20) # 1MB output buffer

Platform IDs: BMG/Xe2=0x2000000, PVC=0x30000, DG2=0x30004

iga.iga_disassemble(0x2000000, raw, len(raw), buf, len(buf)) sys.stdout.write(buf.value.decode()) " > "${name}.asm" done

**Platform ID selection:**
- BMG / Xe2 / LNL: `0x2000000`
- PVC (Ponte Vecchio): `0x30000`
- DG2 (Alchemist): `0x30004`

If `iga_disassemble` returns 0 bytes, the platform ID likely doesn't
match. Try the next one in the list above.

3. **Pin the actually-invoked kernel.**

```bash
ONEDNN_VERBOSE=1 <repro_cmd> 2>&1 | grep -E '^onednn_verbose.*exec' \
  | nl -ba | tee dnnl_verbose.log
# Nth (1-indexed) exec line → dnnl_dump_gpu_*_kernel.<N-1>.bin

The largest .bin by size is typically the GEMM kernel. Cross-check with grep -c dpas <asm> (GEMM has high dpas density).