Initial commit

Author: fmorac

Makefile

@@ -0,0 +1,111 @@
+threads?=4
+ver?=omp
+args?=
+BDX?=
+BDY?=
+BDZ?=
+BDO?=
+arch?=cc70
+dvo?=
+dv?=
+av?=
+
+all: preprocessor/preprocessor omp serial
+
+omp: wave-dag-omp/bin
+wave-dag-omp/bin: preprocessor/preprocessor
+	cp -R wavebench-dag wave-dag-omp
+	cd wave-dag-omp; ./cmake.sh
+	
+serial: wave-dag-serial/bin
+wave-dag-serial/bin: preprocessor/preprocessor
+	cp -R wavebench-dag wave-dag-serial
+	cd wave-dag-serial; opts="-s" ./cmake.sh
+	
+graph: wave-dag-graph/bin
+wave-dag-graph/bin: preprocessor/preprocessor
+	cp -R wavebench-dag wave-dag-graph
+	cd wave-dag-graph;  opts="-p" ./cmake.sh
+	
+timing: wave-dag-timing/bin 
+wave-dag-timing/bin: preprocessor/preprocessor
+	cp -R wavebench-dag wave-dag-timing
+	cd wave-dag-timing;  opts="-t"  ./cmake.sh
+
+graph-timing: wave-dag-graph-timing/bin
+wave-dag-graph-timing/bin: preprocessor/preprocessor
+	cp -R wavebench-dag wave-dag-graph-timing
+	cd wave-dag-graph-timing;  opts="-t -p" ./cmake.sh
+
+preprocessor/preprocessor :
+	make -C preprocessor
+
+acc-gpu: wave-acc-gpu/bin
+wave-acc-gpu/bin: 
+	cp -R wavebench wave-acc-gpu
+	cd wave-acc-gpu;  TARCH=${arch} ./cmake_gpu.sh
+	
+omp-block: wavebench-omp-block/bin
+wavebench-omp-block/bin: 
+	cd wavebench-omp-block; ./cmake.sh
+
+omp-naive: wavebench-omp-naive/bin
+wavebench-omp-naive/bin: 
+	cd wavebench-omp-naive; ./cmake.sh
+
+acc-cpu: wave-acc-cpu/bin
+wave-acc-cpu/bin: 
+	cp -R wavebench wave-acc-cpu
+	cd wave-acc-cpu;  TARCH=${arch} ./cmake_cpu.sh
+
+run: wave-dag-${ver}/bin
+	OMP_NUM_THREADS=${threads} \
+	OMP_BLOCK_DIMX=${BDX} \
+	OMP_BLOCK_DIMY=${BDY} \
+	OMP_BLOCK_DIMZ=${BDZ} \
+	OMP_BLOCK_DIMO=${BDO} \
+	wave-dag-${ver}/bin/wavebench ${args}
+
+run-omp: wavebench-omp-${ver}/bin
+	OMP_NUM_THREADS=${threads} \
+	OMP_BLOCK_DIMX=${BDX} \
+	OMP_BLOCK_DIMY=${BDY} \
+	OMP_BLOCK_DIMZ=${BDZ} \
+	OMP_BLOCK_DIMO=${BDO} \
+	wavebench-omp-${ver}/bin/wavebench ${args}
+
+run-acc: wave-acc-${ver}/bin
+	ACC_NUM_CORES=${threads} \
+	wave-acc-${ver}/bin/wavebench ${args}
+
+compare-acc: wave-acc-${av}/bin wave-dag-${dv}/bin
+	wave-acc-${av}/bin/wavebench ${args} > acc.txt
+	OMP_NUM_THREADS=${threads} \
+	OMP_BLOCK_DIMX=${BDX} \
+	OMP_BLOCK_DIMY=${BDY} \
+	OMP_BLOCK_DIMZ=${BDZ} \
+	OMP_BLOCK_DIMO=${BDO} \
+	wave-dag-${dv}/bin/wavebench ${args} > dag.txt
+	diff -U 0 acc.txt dag.txt | grep -v ^@ | wc -l
+	
+compare-dag: wave-dag-${dvo}/bin wave-dag-${dv}/bin
+	OMP_NUM_THREADS=${threads} \
+	OMP_BLOCK_DIMX=${BDX} \
+	OMP_BLOCK_DIMY=${BDY} \
+	OMP_BLOCK_DIMZ=${BDZ} \
+	OMP_BLOCK_DIMO=${BDO} \
+	wave-dag-${dv}/bin/wavebench ${args} > dag1.txt
+	OMP_NUM_THREADS=${threads} \
+	OMP_BLOCK_DIMX=${BDX} \
+	OMP_BLOCK_DIMY=${BDY} \
+	OMP_BLOCK_DIMZ=${BDZ} \
+	OMP_BLOCK_DIMO=${BDO} \
+	wave-dag-${dvo}/bin/wavebench ${args} > dag2.txt
+	diff -U 0 dag1.txt dag2.txt | grep -v ^@ | wc -l
+
+clean:
+	rm -Rf wave-dag-omp wave-dag-serial wave-dag-graph wave-dag-timing wave-dag-graph-timing wave-acc-cpu wave-acc-gpu wavebench-omp-naive/build wavebench-omp-naive/bin wavebench-omp-block/build wavebench-omp-block/bin
+
+purge:
+	rm -Rf wave-dag-omp wave-dag-serial wave-dag-graph wave-dag-timing wave-dag-graph-timing wave-acc-cpu wave-acc-gpu wavebench-omp-naive/build wavebench-omp-naive/bin wavebench-omp-block/build wavebench-omp-block/bin
+	make -C preprocessor purge

README.md

@@ -1 +1,53 @@
-# wavedag
+# Wave-DAG
+
+Wave-DAG is a preprocessor for accelerating structured wavefront patterns on multi core architectures built on top of OpenMP 5.0, ensuing future portability across platforms.
+
+The test cases used in Wave-DAG were taken from the Wavebench project. For more information about Wavebench see [wavebench/README.md](wavebench/README.md).
+
+For more information about the underlying library see [wavebench-dag/include/Readme.md](wavebench-dag/include/Readme.md).
+
+
+## Example
+
+Basic Smith-Waterman serial code.
+```
+  for(int i = 1; i < n; ++i)
+    for(int j = ; j < m; ++j) {
+        int score = (A[i - 1] == B[j - 1])? match : miss;
+        M[i * m + j] = max(M[(i - 1) * m + (j - 1)] + score,
+                       max(M[i * m +(j - 1)] + gap, M[(i - 1) * m + j] + gap));
+    }
+```
+
+Basic Smith-Waterman with Wave-DAG pragmas added.
+```
+#pragma omp dag coarsening(block, 512, 512)
+  for(int i = 1; i < n; ++i)
+    for(int j = ; j < m; ++j) {
+#pragma omp dag task depend({(i + 1) * m + j + 1,((i + 1) < n) && ((j + 1) < m)}, \
+                            {(i + 1) * m + j,(i + 1) < n},                \
+                            {i * m + j + 1,(j + 1) < m})
+      {
+        int score = (A[i - 1] == B[j - 1]) ? match : miss;
+        M[i * m + j] = max(M[(i - 1) * m + (j - 1)] + score,
+                       max(M[i * m +(j - 1)] + gap, M[(i - 1) * m + j] + gap));
+      }
+    }
+```
+The resulting code then will be preprocessod by the Wave-DAG preprocessor and output a C++ OpenMP 5.0 compliant parallel code.
+
+## Building and runninng
+To build Wave-DAG you need GCC 9.2 and in order to run the original wavebench code you need a working PGI compiler.
+To compile the tests use:
+```
+make omp     # Compile the Wave-DAG version of the test cases
+make serial  # Compile a serial version of the test cases
+make acc-gpu # Compile the Wavebench GPU version of the test cases
+```
+To run the samples use:
+```
+bash run.sh
+```
+
+## License
+For more information see [LICENSE](LICENSE).

preprocessor/Makefile

@@ -0,0 +1,26 @@
+BASE = preprocessor
+BISON = bison
+CXX = g++-9
+FLEX = flex
+CXXFLAGS ?= -std=c++17
+
+all: $(BASE)
+
+$(BASE):
+	mkdir -v build
+	$(MAKE) -C parser-lib CXX=$(CXX) CXXFLAGS=$(CXXFLAGS) BISON=$(BISON) FLEX=$(FLEX) static
+	mv -v parser-lib/include build/include
+	mv -v parser-lib/lib build/lib
+	$(MAKE) -C parser-lib purge
+	cp -Rv preprocessor-src/* build/
+	$(MAKE) -C build CXX=$(CXX) CXXFLAGS=$(CXXFLAGS)
+	mv -v build/preprocessor preprocessor
+	$(MAKE) clean
+
+clean:
+	rm -Rf build
+	$(MAKE) -C parser-lib purge
+	
+purge:
+	rm -Rf build preprocessor
+	$(MAKE) -C parser-lib purge

preprocessor/parser-lib/Makefile

@@ -0,0 +1,67 @@
+# This Makefile is designed to be simple and readable.  It does not
+# aim at portability.  It requires GNU Make.
+
+BASE = pragma-parser
+BISON = bison
+CXX = g++-9
+FLEX = flex
+XSLTPROC = xsltproc
+CXXFLAGS ?= -std=c++17
+
+all: static
+
+%.cc %.hh %.xml %.gv: %.yy
+	$(BISON) $(BISONFLAGS) --xml --graph=$*.gv -o $*.cc $<
+
+%.cc: %.ll
+	$(FLEX) $(FLEXFLAGS) -o$@ $<
+
+%.o: %.cc
+	$(CXX) $(CXXFLAGS) -c -o$@ $<
+
+shared: CXXFLAGS = -std=c++17 -fPIC 
+shared: $(BASE).o parser.o driver.o scanner.o
+	$(CXX) -shared astextent.o parser.o driver.o scanner.o $(BASE).o -o libpragma-parser.so
+	mkdir -v lib
+	mv -v libpragma-parser.so lib/
+	mkdir -v include
+	cp -v pragma-parser.hh include/
+	cp -v parser.hh include/
+	cp -v location.hh include/
+	cp -v driver.hh include/
+	ln -sv ${PWD}/ast include/
+	$(MAKE) clean
+
+static: astextent.o parser.o driver.o scanner.o $(BASE).o
+	ar rcs libpragma-parser.a astextent.o parser.o driver.o scanner.o $(BASE).o
+	mkdir -v lib
+	mv -v libpragma-parser.a lib/
+	mkdir -v include
+	cp -v pragma-parser.hh include/
+	cp -v parser.hh include/
+	cp -v location.hh include/
+	cp -v driver.hh include/
+	cp -Rv ast include/
+	$(MAKE) clean
+	
+astextent.o: ast/astextent.cc
+	$(CXX) $(CXXFLAGS) -c -o$@ $<
+$(BASE).o: parser.hh
+parser.o: parser.hh
+scanner.o: parser.hh
+
+html: parser.html
+%.html: %.xml
+	$(XSLTPROC) $(XSLTPROCFLAGS) -o $@ $$($(BISON) --print-datadir)/xslt/xml2xhtml.xsl $<
+
+CLEANFILES =										\
+  $(BASE) *.o										\
+  parser.hh parser.cc parser.output parser.xml parser.html parser.gv location.hh	\
+  scanner.cc position.hh stack.hh
+
+clean:
+	rm -f $(CLEANFILES)
+purge:
+	rm -f $(CLEANFILES)
+	rm -rf ./lib
+	rm -rf ./include