Verified Commit 641ad37a authored by Michael R. Crusoe's avatar Michael R. Crusoe 🏳️‍🌈
Browse files

any-{amd64,i386}: multi-SIMD compile with dispatcher

parent e67c31e3

debian/bin/simd-dispatch

0 → 100755
+18 −0
Original line number Diff line number Diff line 
#!/bin/bash



PKG=last-align

DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"

BASE=${DIR}/../lib/$PKG/$(basename "$0")



function test_and_run () {

	if grep -q "$1" /proc/cpuinfo && [ -x "${BASE}-$1" ]; then

		# echo "${BASE}-$1" "$@"

		"${BASE}-$1" "$@"

		exit

	fi

}



for SIMD in avx2 avx sse4.1 ssse3 sse3 sse2 sse mmx ; do test_and_run ${SIMD} "$@" ; done



# fallback to plain option

$BASE-plain "$@"

debian/changelog

+3 −7
Original line number Diff line number Diff line
@@ -2,13 +2,9 @@ last-align (1021-3) UNRELEASED; urgency=medium 


  * Team upload.

  * Enable building on any architecture via simde. (Closes: #946850)

  * TODO

    any-amd64: build multiple times with decreasing levels of

               -m{avx2,avx,sse4.1,ssse3,sse3,sse2}

    any-i386: build multiple times with decreasing levels of

               -m{sse3,sse2,sse,mmx}

    all of the above: add wrapper script to detect at runtime the highest

                      level of SIMD available, and launch the relevant binary

  * For any-amd64 and any-i386: compile multiple times with reducing levels of

    SIMD instruction support and ship a dispatching script to pick the best

    one for the user at run-time.



 -- Michael R. Crusoe <michael.crusoe@gmail.com>  Mon, 16 Dec 2019 19:02:20 +0100

debian/patches/simde

+87 −0
Original line number Diff line number Diff line
@@ -186,3 +186,90 @@ 
 			    s2[-0][s1[0]]);

 	SimdInt x = simdLoad(x2+i);

 	SimdInt y = simdSub(simdLoad(y1+i), mDelGrowCost);

--- last-align.orig/makefile

+++ last-align/makefile

@@ -2,15 +2,24 @@

 all:

 	@cd src && $(MAKE) CXXFLAGS="$(CXXFLAGS)"

 

-progs = src/lastdb src/lastal src/last-split src/last-merge-batches	\

-src/last-pair-probs src/lastdb8 src/lastal8 src/last-split8

+SFX :=

+progs = src/lastdb$(SFX) src/lastal$(SFX) src/last-split$(SFX) \

+	src/last-merge-batches$(SFX) src/last-pair-probs$(SFX) src/lastdb8$(SFX) \

+	src/lastal8$(SFX) src/last-split8$(SFX)

 

 prefix = /usr/local

 exec_prefix = $(prefix)

 bindir = $(exec_prefix)/bin

+libexecdir = $(exec_prefix)/libexec

+

 install: all

 	mkdir -p $(bindir)

-	cp $(progs) scripts/* $(bindir)

+	cp scripts/* $(bindir)

+ifeq (,$(SFX))

+	cp $(progs) $(bindir)

+else

+	cp $(progs) $(libexecdir)

+endif

 

 clean:

 	@cd src && $(MAKE) clean

--- last-align.orig/src/makefile

+++ last-align/src/makefile

@@ -56,8 +56,10 @@

 

 MBOBJ = last-merge-batches.o

 

-ALL = lastdb lastal last-split last-merge-batches last-pair-probs	\

-lastdb8 lastal8 last-split8

+SFX :=

+

+ALL = lastdb$(SFX) lastal$(SFX) last-split$(SFX) last-merge-batches$(SFX) \

+      last-pair-probs$(SFX) lastdb8$(SFX) lastal8$(SFX) last-split8$(SFX)

 

 indexObj8 = $(indexObj4:.o=.o8)

 alignObj8 = $(alignObj4:.o=.o8)

@@ -66,33 +68,33 @@

 all: $(ALL)

 

 indexAllObj4 = $(indexObj0) $(indexObj4)

-lastdb: $(indexAllObj4)

+lastdb$(SFX): $(indexAllObj4)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(indexAllObj4) -lz

 

 indexAllObj8 = $(indexObj0) $(indexObj8)

-lastdb8: $(indexAllObj8)

+lastdb8$(SFX): $(indexAllObj8)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(indexAllObj8) -lz

 

 alignAllObj4 = $(alignObj0) $(alignObj4)

-lastal: $(alignAllObj4)

+lastal$(SFX): $(alignAllObj4)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(alignAllObj4) -lz

 

 alignAllObj8 = $(alignObj0) $(alignObj8)

-lastal8: $(alignAllObj8)

+lastal8$(SFX): $(alignAllObj8)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(alignAllObj8) -lz

 

 splitAllObj4 = $(splitObj0) $(splitObj4)

-last-split: $(splitAllObj4)

+last-split$(SFX): $(splitAllObj4)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(splitAllObj4)

 

 splitAllObj8 = $(splitObj0) $(splitObj8)

-last-split8: $(splitAllObj8)

+last-split8$(SFX): $(splitAllObj8)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(splitAllObj8)

 

-last-pair-probs: $(PPOBJ)

+last-pair-probs$(SFX): $(PPOBJ)

 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(LDFLAGS) -o $@ $(PPOBJ) -lz

 

-last-merge-batches: $(MBOBJ)

+last-merge-batches$(SFX): $(MBOBJ)

 	$(CC) $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) -o $@ $(MBOBJ)

 

 .SUFFIXES:

debian/rules

+66 −10
Original line number Diff line number Diff line
@@ -8,7 +8,9 @@ 


include /usr/share/dpkg/default.mk



mandir=$(CURDIR)/debian/$(DEB_SOURCE)/usr/share/man/man1/

prefix=$(CURDIR)/debian/$(DEB_SOURCE)/usr

mandir=$(prefix)/share/man/man1

libexecdir=$(prefix)/lib/$(DEB_SOURCE)



# Copy upstream CXXFLAGS here because makefile enables only overriding them

CXXFLAGS += -Wall -Wextra -Wcast-qual -Wswitch-enum -Wundef -Wcast-align -Wno-long-long -ansi -pedantic -std=c++11 -fopenmp-simd -O3
@@ -18,18 +20,72 @@ CPPFLAGS += -DHAS_CXX_THREADS -DSIMDE_ENABLE_OPENMP 


LDFLAGS += -pthread



NUMJOBS = 1

ifneq (,$(filter parallel=%,$(DEB_BUILD_OPTIONS)))

  NUMJOBS = $(patsubst parallel=%,%,$(filter parallel=%,$(DEB_BUILD_OPTIONS)))

  MAKEFLAGS += -j$(NUMJOBS)

endif



export DEB_BUILD_MAINT_OPTIONS = hardening=+all



%:

	dh  $@



override_dh_auto_build:

	make all CXXFLAGS="$(CXXFLAGS)" CFLAGS="$(CFLAGS)" CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)"

ifeq (amd64,$(DEB_HOST_ARCH))

	mkdir -p $(prefix)

	mkdir -p $(libexecdir)

	for SIMD in avx2 avx sse4.1 ssse3 sse3 sse2 ; do \

		$(MAKE) -j$(NUMJOBS) all SFX=-$${SIMD} CXXFLAGS="$(CXXFLAGS) -m$$SIMD" CFLAGS="$(CFLAGS) -m$$SIMD" CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)" ; \

		rm src/*.o? ; \

	done

else ifeq (i386,$(DEB_HOST_ARCH))

	mkdir -p $(prefix)

	mkdir -p $(libexecdir)

	for SIMD in ssse3 sse3 sse2 sse mmx; do \

		$(MAKE) -j$(NUMJOBS) all SFX=-$${SIMD} CXXFLAGS="$(CXXFLAGS) -m$$SIMD" CFLAGS="$(CFLAGS) -m$$SIMD" CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)" ; \

		rm src/*.o? ; \

	done

	$(MAKE) -j$(NUMJOBS) all SFX=-plain CXXFLAGS="$(CXXFLAGS)" CFLAGS="$(CFLAGS)" CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)"

else

	$(MAKE) -j$(NUMJOBS) all CXXFLAGS="$(CXXFLAGS)" CFLAGS="$(CFLAGS)" CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)"

endif



HELP2MAN = help2man --no-info --version-string="$(DEB_VERSION_UPSTREAM)"



override_dh_auto_install:

	make install prefix=$(CURDIR)/debian/$(DEB_SOURCE)/usr

ifeq (amd64,$(DEB_HOST_ARCH))

	dh_install debian/bin/simd-dispatch /usr/lib/$(DEB_SOURCE)/

	for SIMD in avx2 avx sse4.1 ssse3 sse3 sse2 ; do \

		$(MAKE) -j$(NUMJOBS) install SFX=-$${SIMD} prefix=$(prefix) libexecdir=$(libexecdir) ; \

	done

	mkdir -p $(prefix)/bin

	cd $(prefix)/bin \

		&& for prog in lastdb lastal last-split last-merge-batches last-pair-probs lastdb8 lastal8 last-split8 ; do \

		ln -s ../lib/$(DEB_SOURCE)/simd-dispatch $${prog} ; done

else ifeq (i386,$(DEB_HOST_ARCH))

	dh_install debian/bin/simd-dispatch /usr/lib/$(DEB_SOURCE)/

	for SIMD in ssse3 sse3 sse2 sse mmx plain; do \

		$(MAKE) -j$(NUMJOBS) install SFX=-$${SIMD} prefix=$(prefix) libexecdir=$(libexecdir) ; \

	done

	mkdir -p $(prefix)/bin

	cd $(prefix)/bin \

		&& for prog in lastdb lastal last-split last-merge-batches last-pair-probs lastdb8 lastal8 last-split8 ; do \

		ln -s ../lib/$(DEB_SOURCE)/simd-dispatch $${prog} ; done

else

	$(MAKE) -j$(NUMJOBS) install prefix=$(prefix)

endif



override_dh_auto_clean:

ifeq (amd64,$(DEB_HOST_ARCH))

	for SIMD in avx2 avx sse4.1 ssse3 sse3 sse2 ; do \

		dh_auto_clean -- SFX=-$${SIMD} ; done

else ifeq (i386,$(DEB_HOST_ARCH))

	for SIMD in ssse3 sse3 sse2 sse mmx plain; do \

		dh_auto_clean -- SFX=-$${SIMD} ; done

else

	dh_auto_clean

endif



override_dh_installman:

	dh_installman
@@ -39,16 +95,16 @@ override_dh_installman: 
	$(HELP2MAN) \

	         --name="genome-scale comparison of biological sequences" \

	         --help-option="-h" \

                 $(CURDIR)/src/lastal > $(mandir)/lastal.1

                 $(prefix)/bin/lastal > $(mandir)/lastal.1

	$(HELP2MAN) \

	         --name="genome-scale comparison of biological sequences" \

	         --help-option="-h" \

                 $(CURDIR)/src/lastal8 > $(mandir)/lastal8.1

                 $(prefix)/bin/lastal8 > $(mandir)/lastal8.1

        # help2man output needs manual postprocessing - see debian/mans

	#$(HELP2MAN) \

	#         --name="genome-scale comparison of biological sequences" \

	#         --help-option="-h" \

        #         $(CURDIR)/src/lastdb > $(mandir)/lastdb.1

        #         $(prefix)/bin/lastdb > $(mandir)/lastdb.1

	$(HELP2MAN) \

                 --name="Sort MAF-format alignments by sequence name" \

                 $(CURDIR)/scripts/maf-sort > $(mandir)/maf-sort.1
@@ -72,10 +128,10 @@ override_dh_installman: 
        #         $(CURDIR)/scripts/last-remove-dominated > $(mandir)/last-remove-dominated.1

	$(HELP2MAN) \

                 --name='Estimates "split" or "spliced" alignments.' \

                 $(CURDIR)/src/last-split > $(mandir)/last-split.1

                 $(prefix)/bin/last-split > $(mandir)/last-split.1

	$(HELP2MAN) \

                 --name='Estimates "split" or "spliced" alignments.' \

                 $(CURDIR)/src/last-split8 > $(mandir)/last-split8.1

                 $(prefix)/bin/last-split8 > $(mandir)/last-split8.1

	$(HELP2MAN) \

                 --name="Join two or more sets of MAF-format multiple alignments" \

                 $(CURDIR)/scripts/maf-join > $(mandir)/maf-join.1
@@ -87,11 +143,11 @@ override_dh_installman: 
                 $(CURDIR)/scripts/fastq-interleave > $(mandir)/fastq-interleave.1

	$(HELP2MAN) \

                 --name="Read files of lastal output, merge corresponding batches, and write them" \

                 $(CURDIR)/src/last-merge-batches > $(mandir)/last-merge-batches.1

                 $(prefix)/bin/last-merge-batches > $(mandir)/last-merge-batches.1

        # help2man output needs manual postprocessing - see debian/mans

	#$(HELP2MAN) \

        #         --name="Read alignments of paired DNA reads to a genome, and do estimations" \

        #         $(CURDIR)/src/last-pair-probs > $(mandir)/last-pair-probs.1

        #         $(prefix)/bin/last-pair-probs > $(mandir)/last-pair-probs.1

	$(HELP2MAN) \

                 --name="Get alignments that have a segment with score >= threshold, with gentle masking of lowercase letters" \

                 $(CURDIR)/scripts/last-postmask > $(mandir)/last-postmask.1