From 2e18f2fbfb5034553d75d8af5c75b3ae57295c81 Mon Sep 17 00:00:00 2001
From: Jonas Rembser <jonas.rembser@cern.ch>
Date: Tue, 5 Sep 2023 19:41:58 +0200
Subject: [PATCH] Introduce `roofit` and `cuda` options instead of inheriting
 from ROOT

---
 .github/workflows/ci.yaml                       |  3 ++-
 CMakeLists.txt                                  |  3 +++
 root/CMakeLists.txt                             |  2 +-
 root/roofit/roofit/CMakeLists.txt               | 15 +++++++++------
 root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx |  6 +++---
 root/roofit/roofit/benchRooFitBackends.cxx      |  2 +-
 6 files changed, 19 insertions(+), 12 deletions(-)

diff --git a/.github/workflows/ci.yaml b/.github/workflows/ci.yaml
index 29b61f50..e4ba5272 100644
--- a/.github/workflows/ci.yaml
+++ b/.github/workflows/ci.yaml
@@ -46,7 +46,8 @@ jobs:
       run: |
            conda activate root-nightly
            mkdir -p build && cd build
-           cmake ..
+           # The conda nightlies are not built with a cutting-edge ROOT version that the RooFit benchmarks require
+           cmake -Dcuda=OFF -Droofit=OFF ..
            cmake --build .
 
     - name: Run benchmarks
diff --git a/CMakeLists.txt b/CMakeLists.txt
index ae143cb0..9a01ba6e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -9,6 +9,9 @@ endif()
 
 project(rootbench)
 
+option(cuda "Build benchmarks that run on CUDA GPUs" OFF)
+option(roofit "Build RooFit benchmarks" ON)
+
 include(CMakeToolsHelpers OPTIONAL)
 include(ExternalProject)
 
diff --git a/root/CMakeLists.txt b/root/CMakeLists.txt
index f5213558..a25eb8f3 100644
--- a/root/CMakeLists.txt
+++ b/root/CMakeLists.txt
@@ -3,7 +3,7 @@ add_subdirectory(io)
 add_subdirectory(hist)
 add_subdirectory(math)
 add_subdirectory(pyroot)
-if (ROOT_roofit_FOUND AND ROOT_xml_FOUND)
+if (roofit)
   add_subdirectory(roofit)
 endif()
 add_subdirectory(tree)
diff --git a/root/roofit/roofit/CMakeLists.txt b/root/roofit/roofit/CMakeLists.txt
index 75662037..d9d94ff0 100644
--- a/root/roofit/roofit/CMakeLists.txt
+++ b/root/roofit/roofit/CMakeLists.txt
@@ -13,11 +13,14 @@ RB_ADD_GBENCHMARK(benchRooFitBackends
   LABEL long
   LIBRARIES Core MathCore RooFitCore RooFit)
 
+if(cuda)
+  target_compile_definitions(benchRooFitBinned PRIVATE DO_BENCH_ROOFIT_CUDA)
+  target_compile_definitions(benchRooFitBackends PRIVATE DO_BENCH_ROOFIT_CUDA)
+endif()
+
 file(COPY ${CMAKE_CURRENT_SOURCE_DIR}/benchRooFitBackends_make_plot.py DESTINATION ${CMAKE_CURRENT_BINARY_DIR}/)
 
-if(ROOT_VERSION VERSION_GREATER_EQUAL 6.29)
-  RB_ADD_GBENCHMARK(benchCodeSquashAD
-    benchCodeSquashAD.cxx
-    LABEL long
-    LIBRARIES Core MathCore RooFit RooStats RooFitCore)
-endif(ROOT_VERSION VERSION_GREATER_EQUAL 6.29)
+RB_ADD_GBENCHMARK(benchCodeSquashAD
+  benchCodeSquashAD.cxx
+  LABEL long
+  LIBRARIES Core MathCore RooFit RooStats RooFitCore)
diff --git a/root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx b/root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx
index c77fdf8b..58f8ecf5 100644
--- a/root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx
+++ b/root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx
@@ -246,7 +246,7 @@ BENCHMARK(BDecayWithMixing)->Name("BDecayWithMixing_FitLegacy")->Args({nEvents,
 BENCHMARK(BDecayWithMixing)->Name("BDecayWithMixing_FitLegacyNumCPU2")->Args({nEvents, Legacy, 2})->ARGS;
 BENCHMARK(BDecayWithMixing)->Name("BDecayWithMixing_FitLegacyNumCPU4")->Args({nEvents, Legacy, 4})->ARGS;
 BENCHMARK(BDecayWithMixing)->Name("BDecayWithMixing_FitCPU")->Args({nEvents, Cpu, 1})->ARGS;
-#ifdef R__HAS_CUDA
+#ifdef DO_BENCH_ROOFIT_CUDA
 BENCHMARK(BDecayWithMixing)->Name("BDecayWithMixing_FitCUDA")->Args({nEvents, Cuda, 1})->ARGS;
 #endif
 
@@ -254,7 +254,7 @@ BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitLegacy")->Args(
 BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitLegacyNumCPU2")->Args({nEvents, Legacy, 2})->ARGS;
 BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitLegacyNumCPU4")->Args({nEvents, Legacy, 4})->ARGS;
 BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitCPU")->Args({nEvents, Cpu, 1})->ARGS;
-#ifdef R__HAS_CUDA
+#ifdef DO_BENCH_ROOFIT_CUDA
 BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitCUDA")->Args({nEvents, Cuda, 1})->ARGS;
 #endif
 
@@ -262,7 +262,7 @@ BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitLegacy")->Args({nEvents
 BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitLegacyNumCPU2")->Args({nEvents, Legacy, 2})->ARGS;
 BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitLegacyNumCPU4")->Args({nEvents, Legacy, 4})->ARGS;
 BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitCPU")->Args({nEvents, Cpu, 1})->ARGS;
-#ifdef R__HAS_CUDA
+#ifdef DO_BENCH_ROOFIT_CUDA
 BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitCUDA")->Args({nEvents, Cuda, 1})->ARGS;
 #endif
 
diff --git a/root/roofit/roofit/benchRooFitBackends.cxx b/root/roofit/roofit/benchRooFitBackends.cxx
index 0699f0dd..4513baee 100644
--- a/root/roofit/roofit/benchRooFitBackends.cxx
+++ b/root/roofit/roofit/benchRooFitBackends.cxx
@@ -320,7 +320,7 @@ static void benchModel(benchmark::State &state)
 
 auto const unit = benchmark::kMillisecond;
 
-#ifdef R__HAS_CUDA
+#ifdef DO_BENCH_ROOFIT_CUDA
 // Only one iteration for CUDA benchmarks. Otherwise, the benchmark suit will
 // repeat them many times are they appear to be very fast, consuming almost no
 // CPU time.