compose.yaml

# compose.yaml — podman-first (also works with docker compose).
#
# Three vendor-specific services share one Containerfile, parameterized
# via build args. Pick one based on your GPU; the build context is the
# same so the AdaptiveCpp + xchplot2 build layers cache across services.
#
# Build & run examples:
#
#   # NVIDIA (default sm_89 / RTX 4090; override via $CUDA_ARCH=120 etc.)
#   podman compose build cuda
#   podman compose run --rm cuda test 22 <plot_id_hex> 2 0 0 -G -o /out
#
#   # NVIDIA Pascal/Volta (sm_61 / GTX 10-series, sm_70 / V100): CUDA 13.x
#   # dropped codegen for pre-Turing archs, so pin to a 12.x base image.
#   # scripts/build-container.sh does this automatically when it detects
#   # CUDA_ARCH < 75; if invoking compose directly, set the base manually:
#   CUDA_ARCH=61 \
#     BASE_DEVEL=docker.io/nvidia/cuda:12.9.1-devel-ubuntu24.04 \
#     BASE_RUNTIME=docker.io/nvidia/cuda:12.9.1-devel-ubuntu24.04 \
#     podman compose build cuda
#
#   # AMD ROCm — set $ACPP_GFX to your card's gfx target (rocminfo | grep gfx).
#   #   gfx1031 = Navi 22 (RX 6700/6700 XT/6800M)
#   #   gfx1100 = Navi 31 (RX 7900 XTX/XT)   ← default
#   #   gfx900  = Vega 10 (RX Vega 56/64, MI25)
#   ACPP_GFX=gfx1031 podman compose build rocm
#   podman compose run --rm rocm test 22 <plot_id_hex> 2 0 0 -G -o /out
#
#   # Intel oneAPI (experimental, untested).
#   podman compose build intel
#
# Plot files land in ./plots/ on the host (mounted at /out in the
# container).

services:
  cuda:
    build:
      context: .
      dockerfile: Containerfile
      args:
        # BASE_DEVEL / BASE_RUNTIME default to CUDA 13.x (latest, sm_75+).
        # scripts/build-container.sh overrides both to nvidia/cuda:12.9.1
        # when it detects a pre-Turing GPU (Pascal/Volta, CUDA_ARCH < 75)
        # — CUDA 13.0 dropped codegen for those archs. Set BASE_DEVEL
        # explicitly to bypass the auto-pick (e.g. for cross-targeting an
        # arch the host doesn't have).
        BASE_DEVEL:           "${BASE_DEVEL:-docker.io/nvidia/cuda:13.0.0-devel-ubuntu24.04}"
        BASE_RUNTIME:         "${BASE_RUNTIME:-docker.io/nvidia/cuda:13.0.0-devel-ubuntu24.04}"
        ACPP_TARGETS:         "generic"
        XCHPLOT2_BUILD_CUDA:  "ON"
        INSTALL_CUDA_HEADERS: "0"
        CUDA_ARCH:            "${CUDA_ARCH:-89}"
    image: xchplot2:cuda
    # GPU pass-through. Works on both engines:
    #   - Docker (with nvidia-container-toolkit + `nvidia-ctk runtime
    #     configure --runtime=docker && systemctl restart docker`)
    #   - Podman 5.x (with podman-compose 1.x+; equivalent to
    #     `--device nvidia.com/gpu=all` via CDI)
    # The previous `devices: nvidia.com/gpu=all` shorthand worked on
    # podman but Docker silently ignored it as an unknown device path,
    # leaving the container without libcuda.so.1 and producing a
    # confusing "No matching device" failure mid-plot.
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    volumes:
      - ./plots:/out

  rocm:
    build:
      context: .
      dockerfile: Containerfile
      args:
        # Pinned to ROCm 6.2.x for two reasons:
        #   1. ROCm 7.x's rocm-llvm package no longer ships LLVMConfig.cmake,
        #      so AdaptiveCpp's find_package(LLVM) can't run.
        #   2. ROCm 6.2 ships LLVM 18.0git, matching Ubuntu's llvm-18 so the
        #      device bitcode (ocml.bc, ockl.bc) is readable by AdaptiveCpp
        #      built against Ubuntu's LLVM. No "Unknown attribute kind"
        #      mismatch.
        # AdaptiveCpp is therefore built against Ubuntu's /usr/lib/llvm-18
        # (the Containerfile default), and ROCm provides its own clang +
        # device libs at /opt/rocm/llvm for the HIP backend at runtime.
        BASE_DEVEL:           docker.io/rocm/dev-ubuntu-24.04:6.2-complete
        BASE_RUNTIME:         docker.io/rocm/dev-ubuntu-24.04:6.2-complete
        # IMPORTANT: ACPP_GFX is intentionally *required* — no silent default.
        # If it's unset the SYCL kernels are AOT-compiled for the wrong amdgcn
        # ISA, which HIP loads without error but the kernels execute as silent
        # no-ops at runtime (sort returns input, AES match finds zero results,
        # plot content diverges from the canonical reference). That failure
        # mode is extremely confusing to diagnose — it looks like a correctness
        # bug in the kernels rather than a build-time config error.
        #
        # Set ACPP_GFX explicitly. If you sudo compose, pass the var through
        # (sudo strips env by default):
        #   ACPP_GFX=gfx1031 sudo -E podman compose build rocm
        #   sudo ACPP_GFX=gfx1031 podman compose build rocm
        #
        # Common gfx targets (see `rocminfo | grep gfx`):
        #   gfx1030 = RDNA2 Navi 21 (RX 6800/6800 XT/6900 XT)
        #   gfx1031 = RDNA2 Navi 22 (RX 6700/6700 XT/6800M)
        #   gfx1100 = RDNA3 Navi 31 (RX 7900 XTX/XT)
        #   gfx1101 = RDNA3 Navi 32 (RX 7800 XT/7700 XT)
        #   gfx906  = Vega 20 (Radeon VII, MI50)
        #   gfx900  = Vega 10 (RX Vega 56/64, MI25)
        # Use ${VAR:-default} (NOT ${VAR:?error}) so that building cuda
        # / intel / cpu services without ACPP_GFX set doesn't trip a
        # parse-time error — podman-compose evaluates :? across ALL
        # services during YAML parse, not just the one being built.
        # The placeholder value is intentionally invalid as a gfx
        # target so AdaptiveCpp's HIP backend fails loudly with the
        # placeholder string in its error message — much better than
        # silently building wrong-arch amdgcn ISA from a default like
        # gfx1100 (kernels would then execute as runtime no-ops, see
        # the IMPORTANT block above).
        ACPP_TARGETS:         "hip:${ACPP_GFX:-MISSING-set-ACPP_GFX-or-use-scripts-build-container-sh}"
        XCHPLOT2_BUILD_CUDA:  "OFF"
        # No CUDA headers on the AMD path — they conflict with HIP's
        # uchar1/etc. typedefs. CudaHalfShim.hpp's __has_include guard
        # handles the absence cleanly.
        INSTALL_CUDA_HEADERS: "0"
    image: xchplot2:rocm
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
    # Rootless podman's default seccomp filter + capability set blocks
    # some of the KFD IOCTLs libhsa-runtime64 issues during DMA setup,
    # which surfaces as a segfault inside the HSA runtime on the first
    # host→device copy (rocminfo-level queries still work, so the
    # failure is subtle and confusing). Loosen the sandbox just enough
    # for HSA's DMA path. If rootless still fails on your host, run
    # rootful + privileged instead:
    #   sudo podman run --rm --privileged --device /dev/kfd \
    #        --device /dev/dri -v $PWD/plots:/out xchplot2:rocm \
    #        plot -k 28 -n 10 -f <farmer-pk> -c <pool-contract> -o /out
    security_opt:
      - seccomp=unconfined
    cap_add:
      - SYS_ADMIN
    volumes:
      - ./plots:/out

  intel:
    build:
      context: .
      dockerfile: Containerfile
      args:
        BASE_DEVEL:           docker.io/intel/oneapi-basekit:latest
        BASE_RUNTIME:         docker.io/intel/oneapi-runtime:latest
        ACPP_TARGETS:         "generic"
        XCHPLOT2_BUILD_CUDA:  "OFF"
        INSTALL_CUDA_HEADERS: "1"
    image: xchplot2:intel
    devices:
      - /dev/dri
    volumes:
      - ./plots:/out

  cpu:
    # CPU-only image: AdaptiveCpp's OpenMP backend compiles the SYCL
    # kernels for the host CPU. No GPU runtime needed. Plotting is
    # 1-2 orders of magnitude slower than GPU; useful for headless CI,
    # dev machines without a GPU, or as an extra worker on a
    # heterogeneous `--devices` list. See README's CPU section.
    build:
      context: .
      dockerfile: Containerfile
      args:
        BASE_DEVEL:           docker.io/ubuntu:24.04
        BASE_RUNTIME:         docker.io/ubuntu:24.04
        ACPP_TARGETS:         "omp"
        XCHPLOT2_BUILD_CUDA:  "OFF"
        # AdaptiveCpp's libkernel/half.hpp includes cuda_fp16.h on every
        # build path; pull the headers (no libcudart link, just headers).
        INSTALL_CUDA_HEADERS: "1"
    image: xchplot2:cpu
    volumes:
      - ./plots:/out