Introduce roofit and cuda options instead of inheriting from ROOT

.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

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)
 

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)

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)

root/roofit/roofit/RooFitUnBinnedBenchmarks.cxx

@@ -246,23 +246,23 @@ 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
 
 BENCHMARK(BDecayGaussResolution)->Name("BDecayGaussResolution_FitLegacy")->Args({nEvents, Legacy, 1})->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
 
 BENCHMARK(BDecayDoubleGauss)->Name("BDecayDoubleGauss_FitLegacy")->Args({nEvents, Legacy, 1})->ARGS;
 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
 

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.