Merge branch 'main' into ghi-643-fix

Author: vedika-saravanan

.github/actions/build-lib/build_qec.sh

@@ -2,6 +2,7 @@
 set -e
 
 . "$(dirname "$0")/setup_custabilizer.sh"
+. "$(dirname "$0")/../../../scripts/cudaq_realtime_cmake_flags.sh"
 
 build_dir=$1
 install_prefix=$2
@@ -100,7 +101,9 @@ if [ -z "$CUDAQ_REALTIME_ROOT" ]; then
   # which produces libcudaq-realtime-bridge-hololink.so needed by the bridge.
   cd /tmp/cudaq-realtime-src/realtime
   mkdir -p build && cd build
+
   cmake -G Ninja -DCMAKE_INSTALL_PREFIX="$CUDAQ_REALTIME_ROOT" \
+    -DCMAKE_CUDA_FLAGS="$(cudaq_realtime_cmake_cuda_flags)" \
     -DCUDAQ_REALTIME_ENABLE_HOLOLINK_TOOLS=ON \
     -DHOLOSCAN_SENSOR_BRIDGE_SOURCE_DIR=$HSB_ROOT \
     -DHOLOSCAN_SENSOR_BRIDGE_BUILD_DIR=$HSB_BUILD \

.github/actions/get-cudaq-build/action.yaml

@@ -51,6 +51,7 @@ runs:
         to_hash: |
           .github/actions/get-cudaq-build/**
           scripts/build_cudaq_with_realtime.sh
+          scripts/cudaq_realtime_cmake_flags.sh
           .cudaq_version
       run: |
         hash=${{ hashFiles(format('{0}', env.to_hash)) }}

.github/workflows/all_libs_release.yaml

@@ -108,6 +108,7 @@ jobs:
 
       - name: Build cuda-quantum realtime
         run: |
+          source "$GITHUB_WORKSPACE/scripts/cudaq_realtime_cmake_flags.sh"
           CUDAQ_REPO="${{ steps.get-cudaq-version.outputs.repo }}"
           CUDAQ_REF="${{ steps.get-cudaq-version.outputs.ref }}"
           CUDAQ_REALTIME_ROOT=/tmp/cudaq-realtime
@@ -119,7 +120,9 @@ jobs:
           git sparse-checkout set realtime cmake
           git checkout "$CUDAQ_REF"
           cd realtime && mkdir -p build && cd build
-          cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$CUDAQ_REALTIME_ROOT ..
+          cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$CUDAQ_REALTIME_ROOT \
+            -DCMAKE_CUDA_FLAGS="$(cudaq_realtime_cmake_cuda_flags)" \
+            ..
           ninja && ninja install
           echo "CUDAQ_REALTIME_ROOT=$CUDAQ_REALTIME_ROOT" >> $GITHUB_ENV
         shell: bash

docker/build_env/cudaqx.dev.Dockerfile

@@ -23,6 +23,7 @@ RUN apt-get update && CUDA_DASH=$(echo $cuda_version | tr '.' '-') \
 
 COPY .cudaq_version /cudaq_version
 COPY scripts/build_cudaq_with_realtime.sh /usr/local/bin/build_cudaq_with_realtime.sh
+COPY scripts/cudaq_realtime_cmake_flags.sh /usr/local/bin/cudaq_realtime_cmake_flags.sh
 
 ENV CUDAQ_INSTALL_PREFIX=/usr/local/cudaq
 

libs/qec/unittests/realtime/app_examples/surface_code-4-yaml-test.sh

@@ -76,6 +76,18 @@ if [[ $# -ge 7 ]]; then
   EXTRA_APP_ARGS=("${@:7}")
 fi
 
+# The app defaults to one logical patch. The aggregate result sums residual
+# errors across patches, so retain the per-patch ceiling when --num_logical
+# requests multiple patches.
+NUM_LOGICAL=1
+for ((i = 0; i < ${#EXTRA_APP_ARGS[@]}; i++)); do
+  if [[ "${EXTRA_APP_ARGS[$i]}" == "--num_logical" ]] &&
+    ((i + 1 < ${#EXTRA_APP_ARGS[@]})); then
+    i=$((i + 1))
+    NUM_LOGICAL=${EXTRA_APP_ARGS[$i]}
+  fi
+done
+
 export CUDAQ_DEFAULT_SIMULATOR=stim
 if [[ -n "${QEC_DECODING_SERVER:-}" ]]; then
   export CUDAQ_QEC_REALTIME_MODE=external_server
@@ -90,10 +102,11 @@ P_SPAM=0.01
 # d5/T6). A decoder that loads but never corrects reliably exceeds it; any
 # working decoder sits well below it. This is a wiring/correctness check, not
 # a performance target -- it must never be tightened toward a particular
-# decoder's measured rate. Floored at 1 so small shot counts do not truncate
-# the ceiling to 0.
-MAX_NON_ZERO=$((NUM_SHOTS / 50))
-if [[ $MAX_NON_ZERO -lt 1 ]]; then MAX_NON_ZERO=1; fi
+# decoder's measured rate. Floored at 1 per logical patch so small shot counts
+# do not truncate the ceiling to 0, then scaled for the aggregate result.
+MAX_NON_ZERO_PER_LOGICAL=$((NUM_SHOTS / 50))
+if [[ $MAX_NON_ZERO_PER_LOGICAL -lt 1 ]]; then MAX_NON_ZERO_PER_LOGICAL=1; fi
+MAX_NON_ZERO=$((MAX_NON_ZERO_PER_LOGICAL * NUM_LOGICAL))
 
 # Multi-type mode: a comma list binds one decoder type per patch and the
 # per-decoder ceilings below replace the aggregate ceiling (which is
@@ -174,6 +187,7 @@ echo "  distance       = $DISTANCE"
 echo "  num_rounds     = $NUM_ROUNDS"
 echo "  decoder_type   = $DECODER_TYPE"
 echo "  num_shots      = $NUM_SHOTS"
+echo "  num_logical    = $NUM_LOGICAL"
 echo "  realtime mode  = $CUDAQ_QEC_REALTIME_MODE"
 if [[ -n "$ONNX_PATH" ]]; then
   echo "  onnx_path      = $ONNX_PATH"

scripts/build_cudaq_with_realtime.sh

@@ -23,6 +23,9 @@ set -euo pipefail
 
 log() { printf '\033[1;34m==>\033[0m %s\n' "$*"; }
 
+script_dir=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+. "$script_dir/cudaq_realtime_cmake_flags.sh"
+
 BUILD_TYPE=${1:-"Release"}
 CC=${2:-${CC:-"gcc"}}
 CXX=${3:-${CXX:-"g++"}}
@@ -39,6 +42,7 @@ log "Building cudaq-realtime from $CUDAQ_SRC/realtime"
 cmake -G Ninja -S "$CUDAQ_SRC/realtime" -B "$CUDAQ_SRC/realtime/build" \
   -DCMAKE_BUILD_TYPE="$BUILD_TYPE" \
   -DCMAKE_INSTALL_PREFIX="$CUDAQ_INSTALL_PREFIX" \
+  -DCMAKE_CUDA_FLAGS="$(cudaq_realtime_cmake_cuda_flags)" \
   -DCUDAQ_REALTIME_BUILD_TESTS=OFF \
   -DCUDAQ_REALTIME_BUILD_EXAMPLES=OFF \
   -DCUDAQ_REALTIME_ENABLE_HOLOLINK_TOOLS=OFF

scripts/cudaq_realtime_cmake_flags.sh

@@ -0,0 +1,20 @@
+#!/bin/sh
+
+# CMake flags required when CUDA-QX builds CUDA-Q's realtime dependency.
+cudaq_realtime_cmake_cuda_flags() {
+  _cudaq_realtime_cuda_flags="${CMAKE_CUDA_FLAGS:-}"
+  _cudaq_realtime_host_cxx="${CXX:-c++}"
+  _cudaq_realtime_host_cxx_major=$("$_cudaq_realtime_host_cxx" -dumpversion 2>/dev/null | cut -d. -f1)
+  _cudaq_realtime_cuda_major=$(nvcc --version 2>/dev/null |
+    sed -n 's/^.*release \([0-9][0-9]*\)\..*$/\1/p')
+
+  # CUDA 12's nvcc frontend cannot parse GCC 12's AVX-512 BF16 intrinsic
+  # headers. CUDA-Q realtime does not use those intrinsics, so defining the
+  # headers' include guards keeps them out of the nvcc parse.
+  if [ "$(uname -m)" = x86_64 ] && [ "$_cudaq_realtime_cuda_major" = 12 ] && \
+    [ "$_cudaq_realtime_host_cxx_major" = 12 ]; then
+    _cudaq_realtime_cuda_flags="${_cudaq_realtime_cuda_flags} -D_AVX512BF16INTRIN_H_INCLUDED -D_AVX512BF16VLINTRIN_H_INCLUDED"
+  fi
+
+  printf '%s' "$_cudaq_realtime_cuda_flags"
+}