cms-patatrack · tonydp03 · May 13, 2022
diff --git a/Makefile b/Makefile
@@ -138,8 +138,8 @@ export EIGEN_LDFLAGS :=
 export EIGEN_NVCC_CXXFLAGS := --diag-suppress 20014
 
 BOOST_BASE := /usr
-# Minimum required version of Boost, e.g. 1.78.0
-BOOST_MIN_VERSION := 107800
+# Minimum required version of Boost, e.g. 1.79.0
+BOOST_MIN_VERSION := 107900
 # Check if an external version of Boost is present and recent enough
 ifeq ($(wildcard $(BOOST_BASE)/include/boost/version.hpp),)
 NEED_BOOST := true
@@ -151,7 +151,7 @@ BOOST_BASE := $(EXTERNAL_BASE)/boost
 endif
 export BOOST_DEPS := $(BOOST_BASE)
 export BOOST_CXXFLAGS := -isystem $(BOOST_BASE)/include
-export BOOST_LDFLAGS := -L$(BOOST_BASE)/lib
+export BOOST_LDFLAGS := -L$(BOOST_BASE)/lib -Wl,-rpath,$(BOOST_BASE)/lib -lboost_fiber -lboost_context -lboost_filesystem
 export BOOST_NVCC_CXXFLAGS :=
 
 BACKTRACE_BASE := $(EXTERNAL_BASE)/libbacktrace
@@ -542,8 +542,8 @@ external_boost: $(BOOST_BASE)
 $(BOOST_BASE): CXXFLAGS:=
 $(BOOST_BASE):
  $(eval BOOST_TMP := $(shell mktemp -d))
- curl -L -s -S https://boostorg.jfrog.io/artifactory/main/release/1.78.0/source/boost_1_78_0.tar.bz2 | tar xj -C $(BOOST_TMP)
- cd $(BOOST_TMP)/boost_1_78_0 && ./bootstrap.sh && ./b2 install --prefix=$@ --without-graph_parallel --without-mpi --without-python
+ curl -L -s -S https://boostorg.jfrog.io/artifactory/main/release/1.79.0/source/boost_1_79_0.tar.bz2 | tar xj -C $(BOOST_TMP)
+ cd $(BOOST_TMP)/boost_1_79_0 && ./bootstrap.sh && ./b2 install --prefix=$@ --without-graph_parallel --without-mpi --without-python
  @rm -rf $(BOOST_TMP)
  $(eval undefine BOOST_TMP)
 

diff --git a/src/alpaka/AlpakaCore/AllocatorPolicy.h b/src/alpaka/AlpakaCore/AllocatorPolicy.h
@@ -14,15 +14,15 @@ namespace cms::alpakatools {
  template <typename TDev>
  constexpr inline AllocatorPolicy allocator_policy = AllocatorPolicy::Synchronous;
 
-#if defined ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED
+#if defined ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_ENABLED
  template <>
  constexpr inline AllocatorPolicy allocator_policy<alpaka::DevCpu> =
 #if !defined ALPAKA_DISABLE_CACHING_ALLOCATOR
  AllocatorPolicy::Caching;
 #else
  AllocatorPolicy::Synchronous;
 #endif
-#endif // defined ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED
+#endif // defined ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED || defined ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_ENABLED
 
 #if defined ALPAKA_ACC_GPU_CUDA_ENABLED
  template <>

diff --git a/src/alpaka/AlpakaCore/alpakaConfig.h b/src/alpaka/AlpakaCore/alpakaConfig.h
@@ -111,6 +111,29 @@ namespace alpaka_serial_sync {
 #define ALPAKA_ACCELERATOR_NAMESPACE alpaka_serial_sync
 #endif // ALPAKA_ACC_CPU_B_SEQ_T_SEQ_SYNC_BACKEND
 
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+namespace alpaka_fibers_sync {
+ using namespace alpaka_common;
+
+ using Platform = alpaka::PltfCpu;
+ using Device = alpaka::DevCpu;
+ using Queue = alpaka::QueueCpuBlocking;
+ using Event = alpaka::EventCpu;
+
+ template <typename TDim>
+ using Acc = alpaka::AccCpuFibers<TDim, Idx>;
+ using Acc1D = Acc<Dim1D>;
+ using Acc2D = Acc<Dim2D>;
+ using Acc3D = Acc<Dim3D>;
+
+} // namespace alpaka_fibers_sync
+
+#endif // ALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT
+
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND
+#define ALPAKA_ACCELERATOR_NAMESPACE alpaka_fibers_sync
+#endif // ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND
+
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
 namespace alpaka_tbb_async {
  using namespace alpaka_common;

diff --git a/src/alpaka/AlpakaCore/alpakaFwd.h b/src/alpaka/AlpakaCore/alpakaFwd.h
@@ -85,6 +85,9 @@ namespace alpaka {
  template <typename TDim, typename TIdx>
  class AccCpuSerial;
 
+ template <typename TDim, typename TIdx>
+ class AccCpuFibers;
+
  template <typename TDim, typename TIdx>
  class AccCpuTbbBlocks;
 

diff --git a/src/alpaka/AlpakaCore/backend.h b/src/alpaka/AlpakaCore/backend.h
@@ -1,10 +1,10 @@
 #ifndef AlpakaCore_backend_h
 #define AlpakaCore_backend_h
 
-enum class Backend { SERIAL, TBB, CUDA, HIP };
+enum class Backend { SERIAL, FIBERS, TBB, CUDA, HIP };
 
 inline std::string const& name(Backend backend) {
- static const std::string names[] = {"serial_sync", "tbb_async", "cuda_async", "rocm_async"};
+ static const std::string names[] = {"serial_sync", "fibers_sync", "tbb_async", "cuda_async", "rocm_async"};
  return names[static_cast<int>(backend)];
 }
 

diff --git a/src/alpaka/AlpakaCore/initialise.h b/src/alpaka/AlpakaCore/initialise.h
@@ -12,6 +12,9 @@ namespace cms::alpakatools {
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT
  extern template void initialise<alpaka_serial_sync::Platform>();
 #endif
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+ extern template void initialise<alpaka_fibers_sync::Platform>();
+#endif
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
  extern template void initialise<alpaka_tbb_async::Platform>();
 #endif

diff --git a/src/alpaka/AlpakaDataFormats/TrackingRecHit2DSoAView.h b/src/alpaka/AlpakaDataFormats/TrackingRecHit2DSoAView.h
@@ -21,6 +21,10 @@ namespace alpaka_tbb_async {
  class TrackingRecHit2DAlpaka;
 }
 
+namespace alpaka_fibers_sync {
+ class TrackingRecHit2DAlpaka;
+}
+
 namespace alpaka_serial_sync {
  class TrackingRecHit2DAlpaka;
 }
@@ -38,6 +42,7 @@ class TrackingRecHit2DSoAView {
  friend class alpaka_cuda_async::TrackingRecHit2DAlpaka;
  friend class alpaka_rocm_async::TrackingRecHit2DAlpaka;
  friend class alpaka_tbb_async::TrackingRecHit2DAlpaka;
+ friend class alpaka_fibers_sync::TrackingRecHit2DAlpaka;
  friend class alpaka_serial_sync::TrackingRecHit2DAlpaka;
 
  ALPAKA_FN_ACC ALPAKA_FN_INLINE uint32_t nHits() const { return m_nHits; }

diff --git a/src/alpaka/Makefile b/src/alpaka/Makefile
@@ -9,6 +9,7 @@ test_cpu: $(TARGET)
  @echo
  @echo "Testing $(TARGET)"
  $(TARGET) --maxEvents 2 --serial
+ $(TARGET) --maxEvents 2 --fibers
  $(TARGET) --maxEvents 2 --tbb
  @echo "Succeeded"
 test_nvidiagpu: $(TARGET)
@@ -31,7 +32,7 @@ EXE_DEP := $(EXE_OBJ:$.o=$.d)
 
 LIBNAMES := $(filter-out plugin-% bin test Makefile% plugins.txt%,$(wildcard *))
 PLUGINNAMES := $(patsubst plugin-%,%,$(filter plugin-%,$(wildcard *)))
-MY_CXXFLAGS := -I$(TARGET_DIR) -DLIB_DIR=$(LIB_DIR)/$(TARGET_NAME) -DALPAKA_HOST_ONLY -DALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT -DALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
+MY_CXXFLAGS := -I$(TARGET_DIR) -DLIB_DIR=$(LIB_DIR)/$(TARGET_NAME) -DALPAKA_HOST_ONLY -DALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT -DALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
 ifdef CUDA_BASE
 MY_CXXFLAGS += -DALPAKA_ACC_GPU_CUDA_PRESENT -DALPAKA_ACC_GPU_CUDA_ONLY_MODE
 endif
@@ -60,6 +61,12 @@ $(1)_SERIAL_DEP := $$($(1)_SERIAL_OBJ:$.o=$.d)
 $(1)_SERIAL_LIB := $(LIB_DIR)/$(TARGET_NAME)/lib$(1)_serial.so
 LIBS += $$($(1)_SERIAL_LIB)
 $(1)_SERIAL_LDFLAGS := -l$(1)_serial
+# fibers backend
+$(1)_FIBERS_OBJ := $$(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$$($(1)_PORTABLE_SRC:%=%.fibers.o))
+$(1)_FIBERS_DEP := $$($(1)_FIBERS_OBJ:$.o=$.d)
+$(1)_FIBERS_LIB := $(LIB_DIR)/$(TARGET_NAME)/lib$(1)_fibers.so
+LIBS += $$($(1)_FIBERS_LIB)
+$(1)_FIBERS_LDFLAGS := -l$(1)_fibers
 # TBB backend
 $(1)_TBB_OBJ := $$(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$$($(1)_PORTABLE_SRC:%=%.tbb.o))
 $(1)_TBB_DEP := $$($(1)_TBB_OBJ:$.o=$.d)
@@ -86,7 +93,7 @@ LIBS += $$($(1)_ROCM_LIB)
 $(1)_ROCM_LDFLAGS := -l$(1)_rocm
 endif
 endif # if PORTABLE_SRC is not empty
-ALL_DEPENDS += $$($(1)_DEP) $$($(1)_SERIAL_DEP) $$($(1)_TBB_DEP) $$($(1)_CUDA_DEP) $$($(1)_ROCM_DEP)
+ALL_DEPENDS += $$($(1)_DEP) $$($(1)_SERIAL_DEP) $$($(1)_FIBERS_DEP) $$($(1)_TBB_DEP) $$($(1)_CUDA_DEP) $$($(1)_ROCM_DEP)
 endef
 $(foreach lib,$(LIBNAMES),$(eval $(call LIB_template,$(lib))))
 
@@ -107,6 +114,12 @@ $(1)_SERIAL_DEP := $$($(1)_SERIAL_OBJ:$.o=$.d)
 $(1)_SERIAL_LIB := $(LIB_DIR)/$(TARGET_NAME)/plugin$(1)_serial.so
 PLUGINS += $$($(1)_SERIAL_LIB)
 PLUGINNAMES += $(1)_serial
+# fibers backend
+$(1)_FIBERS_OBJ := $$(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$$($(1)_PORTABLE_SRC:%=%.fibers.o))
+$(1)_FIBERS_DEP := $$($(1)_FIBERS_OBJ:$.o=$.d)
+$(1)_FIBERS_LIB := $(LIB_DIR)/$(TARGET_NAME)/plugin$(1)_fibers.so
+PLUGINS += $$($(1)_FIBERS_LIB)
+PLUGINNAMES += $(1)_fibers
 # TBB backend
 $(1)_TBB_OBJ := $$(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$$($(1)_PORTABLE_SRC:%=%.tbb.o))
 $(1)_TBB_DEP := $$($(1)_TBB_OBJ:$.o=$.d)
@@ -133,7 +146,7 @@ PLUGINS += $$($(1)_ROCM_LIB)
 PLUGINNAMES += $(1)_rocm
 endif
 endif # if PORTABLE_SRC is not empty
-ALL_DEPENDS += $$($(1)_DEP) $$($(1)_SERIAL_DEP) $$($(1)_TBB_DEP) $$($(1)_CUDA_DEP) $$($(1)_ROCM_DEP)
+ALL_DEPENDS += $$($(1)_DEP) $$($(1)_SERIAL_DEP) $$($(1)_FIBERS_DEP) $$($(1)_TBB_DEP) $$($(1)_CUDA_DEP) $$($(1)_ROCM_DEP)
 endef
 $(foreach lib,$(PLUGINNAMES),$(eval $(call PLUGIN_template,$(lib))))
 
@@ -143,9 +156,13 @@ TESTS_PORTABLE_SRC := $(wildcard $(TARGET_DIR)/test/alpaka/*.cc)
 TESTS_SERIAL_OBJ := $(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$(TESTS_PORTABLE_SRC:%=%.serial.o))
 TESTS_SERIAL_DEP := $(TESTS_SERIAL_OBJ:$.o=$.d)
 TESTS_SERIAL_EXE := $(patsubst $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc,$(TEST_DIR)/$(TARGET_NAME)/%.serial,$(TESTS_PORTABLE_SRC))
+# fibers backend
+TESTS_FIBERS_OBJ := $(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$(TESTS_PORTABLE_SRC:%=%.fibers.o))
+TESTS_FIBERS_DEP := $(TESTS_FIBERS_OBJ:$.o=$.d)
+TESTS_FIBERS_EXE := $(patsubst $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc,$(TEST_DIR)/$(TARGET_NAME)/%.fibers,$(TESTS_PORTABLE_SRC))
 # TBB backend
-TESTS_TBB_OBJ := $(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$(TESTS_PORTABLE_SRC:%=tbb.serial.o))
-TESTS_TBB_DEP := $(TESTS_SERIAL_OBJ:$.o=$.d)
+TESTS_TBB_OBJ := $(patsubst $(SRC_DIR)%,$(OBJ_DIR)%,$(TESTS_PORTABLE_SRC:%=%.tbb.o))
+TESTS_TBB_DEP := $(TESTS_TBB_OBJ:$.o=$.d)
 TESTS_TBB_EXE := $(patsubst $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc,$(TEST_DIR)/$(TARGET_NAME)/%.tbb,$(TESTS_PORTABLE_SRC))
 # CUDA backend
 ifdef CUDA_BASE
@@ -161,10 +178,10 @@ TESTS_ROCM_DEP := $(TESTS_ROCM_OBJ:$.o=$.d)
 TESTS_ROCM_EXE := $(patsubst $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc,$(TEST_DIR)/$(TARGET_NAME)/%.rocm,$(TESTS_PORTABLE_SRC))
 endif
 #
-TESTS_EXE := $(TESTS_SERIAL_EXE) $(TESTS_TBB_EXE) $(TESTS_CUDA_EXE) $(TESTS_ROCM_EXE)
-ALL_DEPENDS += $(TESTS_SERIAL_DEP) $(TESTS_TBB_DEP) $(TESTS_CUDA_DEP) $(TESTS_ROCM_DEP)
+TESTS_EXE := $(TESTS_SERIAL_EXE) $(TESTS_FIBERS_EXE) $(TESTS_TBB_EXE) $(TESTS_CUDA_EXE) $(TESTS_ROCM_EXE)
+ALL_DEPENDS += $(TESTS_SERIAL_DEP) $(TESTS_FIBERS_DEP) $(TESTS_TBB_DEP) $(TESTS_CUDA_DEP) $(TESTS_ROCM_DEP)
 # Needed to keep the unit test object files after building $(TARGET)
-.SECONDARY: $(TESTS_SERIAL_OBJ) $(TESTS_TBB_OBJ) $(TESTS_CUDA_OBJ) $(TESTS_CUDADLINK) $(TESTS_ROCM_OBJ)
+.SECONDARY: $(TESTS_SERIAL_OBJ) $(TESTS_FIBERS_OBJ) $(TESTS_TBB_OBJ) $(TESTS_CUDA_OBJ) $(TESTS_CUDADLINK) $(TESTS_ROCM_OBJ)
 
 define RUNTEST_template
 run_$(1): $(1)
@@ -175,6 +192,7 @@ run_$(1): $(1)
 test_$(2): run_$(1)
 endef
 $(foreach test,$(TESTS_SERIAL_EXE),$(eval $(call RUNTEST_template,$(test),cpu)))
+$(foreach test,$(TESTS_FIBERS_EXE),$(eval $(call RUNTEST_template,$(test),cpu)))
 $(foreach test,$(TESTS_TBB_EXE),$(eval $(call RUNTEST_template,$(test),cpu)))
 $(foreach test,$(TESTS_CUDA_EXE),$(eval $(call RUNTEST_template,$(test),nvidiagpu)))
 $(foreach test,$(TESTS_ROCM_EXE),$(eval $(call RUNTEST_template,$(test),amdgpu)))
@@ -187,7 +205,7 @@ $(LIB_DIR)/$(TARGET_NAME)/plugins.txt: $(PLUGINS)
 
 $(TARGET): $(EXE_OBJ) $(LIBS) $(PLUGINS) $(LIB_DIR)/$(TARGET_NAME)/plugins.txt | $(TESTS_EXE)
  # Link all libraries, also the "portable" ones
- $(CXX) $(EXE_OBJ) $(LDFLAGS) $(MY_LDFLAGS) -o $@ -L$(LIB_DIR)/$(TARGET_NAME) $(foreach lib,$(LIBNAMES),$($(lib)_LDFLAGS) $($(lib)_SERIAL_LDFLAGS) $($(lib)_TBB_LDFLAGS) $($(lib)_CUDA_LDFLAGS) $($(lib)_ROCM_LDFLAGS)) $(foreach dep,$(EXTERNAL_DEPENDS),$($(dep)_LDFLAGS))
+ $(CXX) $(EXE_OBJ) $(LDFLAGS) $(MY_LDFLAGS) -o $@ -L$(LIB_DIR)/$(TARGET_NAME) $(foreach lib,$(LIBNAMES),$($(lib)_LDFLAGS) $($(lib)_SERIAL_LDFLAGS) $($(lib)_FIBERS_LDFLAGS) $($(lib)_TBB_LDFLAGS) $($(lib)_CUDA_LDFLAGS) $($(lib)_ROCM_LDFLAGS)) $(foreach dep,$(EXTERNAL_DEPENDS),$($(dep)_LDFLAGS))
 
 define BUILD_template
 $(OBJ_DIR)/$(2)/%.cc.o: $(SRC_DIR)/$(2)/%.cc
@@ -207,6 +225,10 @@ $$($(1)_SERIAL_LIB): $$($(1)_SERIAL_OBJ) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($
  @[ -d $$(@D) ] || mkdir -p $$(@D)
  $(CXX) $$($(1)_SERIAL_OBJ) $(LDFLAGS) -shared $(SO_LDFLAGS) $(LIB_LDFLAGS) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_LDFLAGS)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_SERIAL_LDFLAGS)) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_LDFLAGS)) -o $$@
 
+$$($(1)_FIBERS_LIB): $$($(1)_FIBERS_OBJ) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_DEPS)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_LIB)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_FIBERS_LIB))
+ @[ -d $$(@D) ] || mkdir -p $$(@D)
+ $(CXX) $$($(1)_FIBERS_OBJ) $(LDFLAGS) -shared $(SO_LDFLAGS) $(LIB_LDFLAGS) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_LDFLAGS)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_FIBERS_LDFLAGS)) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_LDFLAGS)) -o $$@
+
 $$($(1)_TBB_LIB): $$($(1)_TBB_OBJ) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_DEPS)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_LIB)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_TBB_LIB))
  @[ -d $$(@D) ] || mkdir -p $$(@D)
  $(CXX) $$($(1)_TBB_OBJ) $(LDFLAGS) -shared $(SO_LDFLAGS) $(LIB_LDFLAGS) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_LDFLAGS)) $$(foreach lib,$$($(1)_DEPENDS),$$($$(lib)_TBB_LDFLAGS)) $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_LDFLAGS)) -o $$@
@@ -230,6 +252,16 @@ $(OBJ_DIR)/$(2)/alpaka/%.cc.serial.o: $(SRC_DIR)/$(2)/alpaka/%.cc
  -e '/^$$$$/ d' -e 's/$$$$/ :/' -e 's/ *//' < $(OBJ_DIR)/$(2)/alpaka/$$*.cc.serial.d.tmp >> $(OBJ_DIR)/$(2)/alpaka/$$*.cc.serial.d; \
  rm $(OBJ_DIR)/$(2)/alpaka/$$*.cc.serial.d.tmp
 
+# Portable code, for fibers backend
+$(OBJ_DIR)/$(2)/alpaka/%.cc.fibers.o: $(SRC_DIR)/$(2)/alpaka/%.cc
+ @[ -d $$(@D) ] || mkdir -p $$(@D)
+ $(CXX) $(CXXFLAGS) $(MY_CXXFLAGS) -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND $$(foreach dep,$(EXTERNAL_DEPENDS),$$($$(dep)_CXXFLAGS)) -c $$< -o $$@ -MMD
+ @cp $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d.tmp; \
+ sed 's#\($(2)/alpaka/$$*\)\.o[ :]*#\1.o \1.d : #g' < $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d.tmp > $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d; \
+ sed -e 's/#.*//' -e 's/^[^:]*: *//' -e 's/ *\\$$$$//' \
+ -e '/^$$$$/ d' -e 's/$$$$/ :/' -e 's/ *//' < $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d.tmp >> $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d; \
+ rm $(OBJ_DIR)/$(2)/alpaka/$$*.cc.fibers.d.tmp
+
 # Portable code, for TBB backend
 $(OBJ_DIR)/$(2)/alpaka/%.cc.tbb.o: $(SRC_DIR)/$(2)/alpaka/%.cc
  @[ -d $$(@D) ] || mkdir -p $$(@D)
@@ -286,6 +318,20 @@ $(TEST_DIR)/$(TARGET_NAME)/%.serial: $(OBJ_DIR)/$(TARGET_NAME)/test/alpaka/%.cc.
  @[ -d $(@D) ] || mkdir -p $(@D)
  $(CXX) $^ $(LDFLAGS) $(MY_LDFLAGS) -o $@ -L$(LIB_DIR)/$(TARGET_NAME) $(foreach lib,$(LIBNAMES),$($(lib)_LDFLAGS) $($(lib)_SERIAL_LDFLAGS)) $(foreach dep,$(EXTERNAL_DEPENDS),$($(dep)_LDFLAGS))
 
+# Fibers backend
+$(OBJ_DIR)/$(TARGET_NAME)/test/alpaka/%.cc.fibers.o: $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc
+ @[ -d $(@D) ] || mkdir -p $(@D)
+ $(CXX) $(CXXFLAGS) $(MY_CXXFLAGS) -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_ENABLED -DALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND $(foreach dep,$(EXTERNAL_DEPENDS),$($(dep)_CXXFLAGS)) -c $< -o $@ -MMD
+ @cp $(@D)/$*.cc.fibers.d $(@D)/$*.cc.fibers.d.tmp; \
+ sed 's#\($(TARGET_NAME)/$*\)\.o[ :]*#\1.o \1.d : #g' < $(@D)/$*.cc.fibers.d.tmp > $(@D)/$*.cc.fibers.d; \
+ sed -e 's/#.*//' -e 's/^[^:]*: *//' -e 's/ *\\$$//' \
+ -e '/^$$/ d' -e 's/$$/ :/' -e 's/ *//' < $(@D)/$*.cc.fibers.d.tmp >> $(@D)/$*.cc.fibers.d; \
+ rm $(@D)/$*.cc.fibers.d.tmp
+
+$(TEST_DIR)/$(TARGET_NAME)/%.fibers: $(OBJ_DIR)/$(TARGET_NAME)/test/alpaka/%.cc.fibers.o | $(LIBS)
+ @[ -d $(@D) ] || mkdir -p $(@D)
+ $(CXX) $^ $(LDFLAGS) $(MY_LDFLAGS) -o $@ -L$(LIB_DIR)/$(TARGET_NAME) $(foreach lib,$(LIBNAMES),$($(lib)_LDFLAGS) $($(lib)_FIBERS_LDFLAGS)) $(foreach dep,$(EXTERNAL_DEPENDS),$($(dep)_LDFLAGS))
+
 # TBB backend
 $(OBJ_DIR)/$(TARGET_NAME)/test/alpaka/%.cc.tbb.o: $(SRC_DIR)/$(TARGET_NAME)/test/alpaka/%.cc
  @[ -d $(@D) ] || mkdir -p $(@D)

diff --git a/src/alpaka/bin/main.cc b/src/alpaka/bin/main.cc
@@ -29,6 +29,9 @@ namespace {
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT
  << "[--serial] "
 #endif
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+ << "[--fibers] "
+#endif
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
  << "[--tbb] "
 #endif
@@ -44,6 +47,9 @@ namespace {
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_SEQ_PRESENT
  << " --serial Use CPU Serial backend\n"
 #endif
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+ << " --fibers Use CPU Fibers backend\n"
+#endif
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
  << " --tbb Use CPU TBB backend\n"
 #endif
@@ -140,6 +146,12 @@ int main(int argc, char** argv) {
  getOptionalArgument(args, i, weight);
  backends.insert_or_assign(Backend::SERIAL, weight);
 #endif
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+ } else if (*i == "--fibers") {
+ float weight = 1.;
+ getOptionalArgument(args, i, weight);
+ backends.insert_or_assign(Backend::FIBERS, weight);
+#endif
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
  } else if (*i == "--tbb") {
  float weight = 1.;
@@ -208,6 +220,11 @@ int main(int argc, char** argv) {
  cms::alpakatools::initialise<alpaka_serial_sync::Platform>();
  }
 #endif
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_PRESENT
+ if (backends.find(Backend::FIBERS) != backends.end()) {
+ cms::alpakatools::initialise<alpaka_fibers_sync::Platform>();
+ }
+#endif
 #ifdef ALPAKA_ACC_CPU_B_TBB_T_SEQ_PRESENT
  if (backends.find(Backend::TBB) != backends.end()) {
  cms::alpakatools::initialise<alpaka_tbb_async::Platform>();

diff --git a/src/alpaka/plugin-SiPixelRecHits/alpaka/PixelRecHits.cc b/src/alpaka/plugin-SiPixelRecHits/alpaka/PixelRecHits.cc
@@ -81,7 +81,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
  alpaka::wait(queue);
 #endif
 
-#if defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ASYNC_BACKEND && defined ALPAKA_DISABLE_CACHING_ALLOCATOR
+#if (defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ASYNC_BACKEND || defined ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND) && defined ALPAKA_DISABLE_CACHING_ALLOCATOR
  // FIXME this is required to keep the host buffer inside hits_d alive; it could be removed once the host buffers are also stream-ordered
  alpaka::wait(queue);
 #endif

diff --git a/src/alpaka/plugin-Validation/alpaka/HistoValidator.cc b/src/alpaka/plugin-Validation/alpaka/HistoValidator.cc
@@ -218,6 +218,8 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
  void HistoValidator::endJob() {
 #if defined ALPAKA_ACC_CPU_B_SEQ_T_SEQ_SYNC_BACKEND
  std::ofstream out("histograms_alpaka_serial.txt");
+#elif defined ALPAKA_ACC_CPU_B_SEQ_T_FIBERS_SYNC_BACKEND
+ std::ofstream out("histograms_alpaka_fibers.txt");
 #elif defined ALPAKA_ACC_CPU_B_TBB_T_SEQ_ASYNC_BACKEND
  std::ofstream out("histograms_alpaka_tbb.txt");
 #elif defined ALPAKA_ACC_GPU_CUDA_ASYNC_BACKEND