#!/bin/bash
# Extract what the compilers already generated, so you can start from it instead of from scratch.
#   ./dump_ir.sh triton   python3 my_triton.py     # PTX / TTGIR / LLIR / cubin from Triton
#   ./dump_ir.sh inductor python3 my_model.py      # the Triton that torch.compile generated
#   ./dump_ir.sh sass     my_kernel.cubin|.so      # what actually landed on the machine
set -uo pipefail
MODE=${1:-triton}; shift
case "$MODE" in
  triton)
    D=${TRITON_DUMP_DIR:-/tmp/triton_dump}; rm -rf "$D"; mkdir -p "$D"
    TRITON_KERNEL_DUMP=1 TRITON_DUMP_DIR="$D" TRITON_ALWAYS_COMPILE=1 "$@" || exit $?
    echo "== artifacts in $D =="
    find "$D" \( -name "*.ptx" -o -name "*.ttgir" -o -name "*.llir" -o -name "*.cubin" \) | head -24
    echo
    echo "Read the .ttgir for the layouts Triton chose and the .ptx for what it emitted."
    echo "Then: ./dump_ir.sh sass <that .cubin>   -- the SASS that actually ran."
    ;;
  inductor)
    TORCH_LOGS=output_code TORCHINDUCTOR_MAX_AUTOTUNE=1 "$@" 2>&1 | tee /tmp/inductor_out.log
    echo "== generated Triton kernels =="; grep -n "def triton_" /tmp/inductor_out.log | head -20
    echo "Inductor's autotuned choice is a strong starting point -- read it before writing your own."
    ;;
  sass)
    F=${1:-}
    [ -n "$F" ] && [ -f "$F" ] || { echo "usage: dump_ir.sh sass <file.cubin|.so>  (no such file: '${F}')" >&2; exit 1; }
    cuobjdump -sass "$F" > /tmp/kernel.sass || { echo "cuobjdump failed on $F" >&2; exit 1; }
    [ -s /tmp/kernel.sass ] || { echo "no SASS in $F (not a cubin / no device code?)" >&2; exit 1; }
    head -40 /tmp/kernel.sass; echo "..."
    echo "== register spills (LDL/STL) ==";   grep -cE "\bLDL\b|\bSTL\b" /tmp/kernel.sass || true
    echo "== tensor-core instructions ==";    grep -coE "HMMA|IMMA|QGMMA|wgmma" /tmp/kernel.sass || true
    echo "full listing: /tmp/kernel.sass"
    echo "Zero MMA instructions where you expected tensor cores means the MMA never issued."
    ;;
  *) echo "usage: dump_ir.sh {triton|inductor|sass} ..." >&2; exit 1 ;;
esac