Auto-detect CPU capabilities at runtime in Linux

.github/workflows/main.yml

@@ -14,10 +14,6 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest, linux-gpu-cuda]
-        fullopt: ["True", "False"]
-        exclude:
-        - os: linux-gpu-cuda
-          fullopt: "False"
     runs-on: ${{ matrix.os }}
     steps:
     - uses: actions/checkout@v3
@@ -49,23 +45,19 @@ jobs:
           clang -v
         fi
         which h5c++
-        if [[ "$(uname -s)" == "Linux" && "${{ matrix.fullopt }}" != "False" ]];
+        if [[ "$(uname -s)" == "Linux" ]];
         then
-          # always use PGI when fully optimizing
-          # but never when in compatible/low-optimization mode
+          # install PGI but do not source it
+          # the makefile will do it automatically
           ./scripts/install_hpc_sdk.sh </dev/null
-          source setup_nv_h5.sh
         fi
         df -h .
         export PERFORMING_CONDA_BUILD=True
-        export BUILD_FULL_OPTIMIZATION=${{ matrix.fullopt }}
         echo "======= begin env ====="
         env
         echo "=======  end env  ====="
-        make api && \
-        make main && \
-        make install && \
-        make test_binaries
+        # all == build (shlib,bins,tests) and install
+        make all
         df -h .
         pushd src
         if [[ "$(uname -s)" == "Linux" ]]; 

Makefile

@@ -1,24 +1,93 @@
-.PHONY: all main api test_binaries test install clean
+.PHONY: test clean all
 
-all: api main install
+# Note: This Makefile will NOT properly work with the -j option 
 
-main:
-	cd src && make main
+PLATFORM := $(shell uname -s)
+COMPILER := $(shell ($(CXX) -v 2>&1) | tr A-Z a-z )
+
+ifeq ($(PLATFORM),Darwin)
+all: api main install test_binaries
+
+else
+
+# Note: important that all_nv is after all_cpu_basic and all_nv_avx2 for tests to work
+all: all_cpu_basic all_nv_avx2 all_nv all_combined test_binaries_nv
+
+all_cpu_basic: api_cpu_basic main_cpu_basic install_cpu_basic
+
+all_nv: api_nv main_nv install_nv
+
+all_nv_avx2: api_nv_avx2 main_nv_avx2 install_nv_avx2
+
+all_combined: api_combined install_combined
+
+endif
+
+clean:
+	-cd test && make clean
+	-cd src && make clean
+	-cd combined && make clean
+
+########### api
 
 api:
 	cd src && make api
 
+api_cpu_basic:
+	export BUILD_VARIANT=cpu_basic ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make clean && make api
+
+api_nv:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make clean && make api
+
+api_nv_avx2:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv_avx2 ; export BUILD_FULL_OPTIMIZATION=True ; cd src && make clean && make api
+
+api_combined:
+	cd combined && make api
+
+########### main
+
+main:
+	cd src && make main
+
+main_cpu_basic:
+	export BUILD_VARIANT=cpu_basic ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make main
+
+main_nv:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make main
+
+main_nv_avx2:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv_avx2 ; export BUILD_FULL_OPTIMIZATION=True ; cd src && make main
+
+########### install
+
+install:
+	cd src && make install
+
+install_cpu_basic:
+	export BUILD_VARIANT=cpu_basic ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make install
+
+install_nv:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make install
+
+install_nv_avx2:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv_avx2 ; export BUILD_FULL_OPTIMIZATION=True ; cd src && make install
+
+install_combined:
+	cd combined && make install
+
+########### test
+
 test_binaries:
 	cd src && make test_binaries
 	cd test && make test_binaries
 
+test_binaries_nv:
+	source ./setup_nv_h5.sh; export BUILD_VARIANT=nv ; export BUILD_FULL_OPTIMIZATION=False ; cd src && make test_binaries
+	# use the default compiler for the test subdir as it tests the combined shlib
+	cd test && make test_binaries
+
 test:
 	cd src && make test
 	cd test && make test
 
-install:
-	cd src && make install
-
-clean:
-	-cd test && make clean
-	-cd src && make clean

README.md

@@ -90,6 +90,17 @@ To restrict the number of cores used, set:
 
     export OMP_NUM_THREADS=nthreads
 
+## Older CPU support
+
+On Linux platforms, Unifrac will auto-detect the CPU generation, i.e. if it supports avx or avx2 vector instructions.
+To force the most compatible binary variant, one can set:
+
+    export UNIFRAC_MAX_CPU=basic
+
+To check which binary is used (Unifrac will print it to standard output at runtime), set:
+
+    export UNIFRAC_CPU_INFO=Y
+
 ## GPU support
 
 On Linux platforms, Unifrac will run on a GPU, if one is found. 

combined/Makefile

@@ -0,0 +1,27 @@
+.PHONY: all api main install clean
+
+all: api install
+
+api: libssu.so
+
+# no-op, just for completeness
+main:
+
+
+ifeq ($(PREFIX),)
+    PREFIX := $(CONDA_PREFIX)
+endif
+
+libssu.o: libssu.c
+	$(CC) -c libssu.c -fPIC
+
+libssu.so: libssu.o
+	$(CC) -shared -o libssu.so libssu.o -fPIC -ldl
+
+install: libssu.so
+	rm -f ${PREFIX}/lib//libssu.so; cp libssu.so ${PREFIX}/lib/
+	rm -f ${PREFIX}/bin/ssu; cp ssu ${PREFIX}/bin/
+	rm -f ${PREFIX}/bin/faithpd; cp faithpd ${PREFIX}/bin/
+
+clean:
+	rm -f libssu.o libssu.so

combined/faithpd

@@ -0,0 +1,44 @@
+#!/bin/bash
+
+#default
+FPD=faithpd_nv
+
+# Need at least AVX to support GPUs
+if [ "${UNIFRAC_MAX_CPU}" == "basic" ]; then
+   has_no_avx=1
+else
+   cat /proc/cpuinfo |grep flags |head -1 | grep -q avx
+   has_no_avx=$?
+fi
+
+if [ "${has_no_avx}" -eq 1 ]; then
+   FPD=faithpd_cpu_basic
+else
+
+if [ "${UNIFRAC_MAX_CPU}" == "avx" ]; then
+   has_no_avx2=1
+else
+   cat /proc/cpuinfo |grep flags |head -1 | grep -q avx2
+   has_no_avx2=$?
+fi
+
+if [ "${has_no_avx2}" -eq 1 ]; then
+   FPD=faithpd_nv
+else
+   FPD=faithpd_nv_avx2
+fi # if "${has_no_avx2}" -eq 1 
+
+
+fi # if "${has_no_avx}" -eq 1 
+
+if [ "${UNIFRAC_CPU_INFO}" == "Y" ]; then
+  echo "INFO (unifrac): Using executable" ${FPD}
+fi
+
+#
+#
+#
+BASEDIR=$(dirname "$0")
+
+exec ${BASEDIR}/${FPD} "$@"
+