GPU backend branch merge

.gitignore

@@ -5,4 +5,7 @@
 .settings/
 .idea/
 build/
+autotune/*.json
+*.graphit_bin
+*.graphit_sbin
 cmake-build-debug/

CMakeLists.txt

@@ -83,9 +83,19 @@ add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/graphit.py
 	VERBATIM
 )
 
+find_package(CUDA QUIET)
+
+add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/gpu_tests/all_gpu_tests.py
+	COMMAND ${CMAKE_COMMAND} -E make_directory ${CMAKE_BINARY_DIR}/gpu_tests
+	COMMAND sed -e s?\$\{NVCC_COMPILER\}?${CUDA_NVCC_EXECUTABLE}?g -e s?\$\{GRAPHIT_SOURCE_DIRECTORY\}?${CMAKE_SOURCE_DIR}?g -e s?\$\{CXX_COMPILER\}?${CMAKE_CXX_COMPILER}?g -e s?\$\{GRAPHIT_BUILD_DIRECTORY\}?${CMAKE_BINARY_DIR}?g ${CMAKE_SOURCE_DIR}/test/gpu_tests/all_gpu_tests.py > ${CMAKE_BINARY_DIR}/gpu_tests/all_gpu_tests.py
+	DEPENDS ${CMAKE_SOURCE_DIR}/test/gpu_tests/all_gpu_tests.py
+	VERBATIM
+)
+
 add_custom_target(copy_graphitc_py ALL DEPENDS ${GRAPHITC_PY})
 add_custom_target(copy_python_tests ALL DEPENDS ${CMAKE_BINARY_DIR}/python_tests/test_with_schedules.py ${CMAKE_BINARY_DIR}/python_tests/test.py ${CMAKE_BINARY_DIR}/python_tests/pybind_test.py)
 add_custom_target(copy_graphit_py ALL DEPENDS ${CMAKE_BINARY_DIR}/graphit.py)
+add_custom_target(copy_all_gpu_tests_py ALL DEPENDS ${CMAKE_BINARY_DIR}/gpu_tests/all_gpu_tests.py)
 
 configure_file(src/main.cpp ${CMAKE_BINARY_DIR}/bin/main.cpp COPYONLY)
 configure_file(test/library_test_drivers/library_test_driver_cpp.txt ${CMAKE_BINARY_DIR}/bin/library_test_driver.cpp COPYONLY)

apps/cc_pjump.gt

@@ -0,0 +1,55 @@
+element Vertex end
+element Edge end
+
+const edges : edgeset{Edge}(Vertex,Vertex) = load (argv[1]);
+
+const vertices : vertexset{Vertex} = edges.getVertices();
+const IDs : vector{Vertex}(int) = 1;
+
+const update: vector[1](int);
+
+func updateEdge(src : Vertex, dst : Vertex)
+    var src_id: Vertex = IDs[src];
+    var dst_id: Vertex = IDs[dst];
+
+    IDs[dst_id] min= IDs[src_id];
+    IDs[src_id] min= IDs[dst_id];
+end
+
+func init(v : Vertex)
+     IDs[v] = v;
+end
+
+func pjump(v: Vertex) 
+    var y: Vertex = IDs[v];
+    var x: Vertex = IDs[y];
+    if x != y
+        IDs[v] = x;
+        update[0] = 1;
+    end
+end
+
+func main()
+    var n : int = edges.getVertices();
+    for trail in 0:10
+        var frontier : vertexset{Vertex} = new vertexset{Vertex}(n);
+        startTimer();
+        vertices.apply(init);
+        while (frontier.getVertexSetSize() != 0)
+            #s1# var output: vertexset{Vertex} = edges.from(frontier).applyModified(updateEdge,IDs);
+	    delete frontier;
+	    frontier = output;
+            update[0] = 1;
+            while update[0] != 0
+		update[0] = 0;
+		vertices.apply(pjump);
+            end
+        end
+        var elapsed_time : float = stopTimer();
+	delete frontier;
+        print "elapsed time: ";
+        print elapsed_time;
+    end
+end
+
+

autotune/cmd.sh

@@ -0,0 +1 @@
+python3 graphit_gpu_autotuner.py --graph /local/ajaybr/graph-dataset/clean_general/soc-LiveJournal1.mtx --algo_file gpu_apps/sssp_delta_stepping.gt --killed_process_report_runtime_limit 1 --max_delta 100 --runtime_limit 10 --stop-after 900

autotune/compile_gpu.sh

@@ -0,0 +1,3 @@
+python ../build/bin/graphitc.py -a algotorun.gt -f schedule_0 -o test.cu
+/usr/local/cuda/bin/nvcc  -ccbin /usr/bin/c++ -std=c++11 -I ../src/runtime_lib/ -o test -Xcompiler "-w" -O3 test.cu -DNUM_CTA=80 -DCTA_SIZE=512 -Wno-deprecated-gpu-targets -gencode arch=compute_70,code=sm_70 --use_fast_math -Xptxas "-v -dlcm=ca --maxrregcount=64" -rdc=true -DFRONTIER_MULTIPLIER=3
+#/usr/local/cuda/bin/nvcc  -ccbin /usr/bin/c++ -std=c++11 -I ../src/runtime_lib/ -o test -Xcompiler "-w" -O3 test.cu -DNUM_CTA=60 -DCTA_SIZE=512 -Wno-deprecated-gpu-targets -gencode arch=compute_61,code=sm_61 --use_fast_math -Xptxas "-v -dlcm=ca --maxrregcount=64" -rdc=true -DFRONTIER_MULTIPLIER=2

autotune/gpu_apps/bfs.gt

@@ -0,0 +1,41 @@
+element Vertex end
+element Edge end
+
+const edges : edgeset{Edge}(Vertex,Vertex) = load (argv[1]);
+const vertices : vertexset{Vertex} = edges.getVertices();
+const parent : vector{Vertex}(int) = -1;
+
+
+func updateEdge(src : Vertex, dst : Vertex)
+    parent[dst] = src;
+end
+
+func toFilter(v : Vertex) -> output : bool
+    output =  parent[v] == -1;
+end
+
+func reset(v: Vertex)
+    parent[v] = -1;
+end
+
+func main()
+    for trail in 0:10
+    	var frontier : vertexset{Vertex} = new vertexset{Vertex}(0);
+	startTimer();
+        vertices.apply(reset);
+	var start_vertex : int = atoi(argv[2]);
+    	frontier.addVertex(start_vertex);
+    	parent[start_vertex] = start_vertex;
+
+    	#s0# while (frontier.getVertexSetSize() != 0)
+            #s1# var output : vertexset{Vertex} = edges.from(frontier).to(toFilter).applyModified(updateEdge,parent, true);
+	    delete frontier;
+	    frontier = output;
+    	end
+        var elapsed_time : float = stopTimer();
+	delete frontier;
+    	print "elapsed time: ";
+    	print elapsed_time;
+    end
+end
+

autotune/gpu_apps/cc.gt

@@ -0,0 +1,55 @@
+element Vertex end
+element Edge end
+
+const edges : edgeset{Edge}(Vertex,Vertex) = load (argv[1]);
+
+const vertices : vertexset{Vertex} = edges.getVertices();
+const IDs : vector{Vertex}(int) = 1;
+
+const update: vector[1](int);
+
+func updateEdge(src : Vertex, dst : Vertex)
+    var src_id: Vertex = IDs[src];
+    var dst_id: Vertex = IDs[dst];
+
+    IDs[dst_id] min= IDs[src_id];
+    IDs[src_id] min= IDs[dst_id];
+end
+
+func init(v : Vertex)
+     IDs[v] = v;
+end
+
+func pjump(v: Vertex) 
+    var y: Vertex = IDs[v];
+    var x: Vertex = IDs[y];
+    if x != y
+        IDs[v] = x;
+        update[0] = 1;
+    end
+end
+
+func main()
+    var n : int = edges.getVertices();
+    for trail in 0:10
+        var frontier : vertexset{Vertex} = new vertexset{Vertex}(n);
+        startTimer();
+        vertices.apply(init);
+        #s0# while (frontier.getVertexSetSize() != 0)
+            #s1# var output: vertexset{Vertex} = edges.from(frontier).applyModified(updateEdge,IDs);
+	    delete frontier;
+	    frontier = output;
+            update[0] = 1;
+            while update[0] != 0
+		update[0] = 0;
+		vertices.apply(pjump);
+            end
+        end
+        var elapsed_time : float = stopTimer();
+	delete frontier;
+        print "elapsed time: ";
+        print elapsed_time;
+    end
+end
+
+

autotune/gpu_apps/pagerank.gt

@@ -0,0 +1,53 @@
+element Vertex end
+element Edge end
+const edges : edgeset{Edge}(Vertex,Vertex) = load (argv[1]);
+const vertices : vertexset{Vertex} = edges.getVertices();
+const old_rank : vector{Vertex}(float) = 1.0/vertices.size();
+const new_rank : vector{Vertex}(float) = 0.0;
+const out_degree : vector {Vertex}(int) = edges.getOutDegrees();
+const contrib : vector{Vertex}(float) = 0.0;
+const error : vector{Vertex}(float) = 0.0;
+const damp : float = 0.85;
+const beta_score : float = (1.0 - damp) / vertices.size();
+
+func computeContrib(v : Vertex)
+    contrib[v] = old_rank[v] / out_degree[v];
+end
+
+func updateEdge(src : Vertex, dst : Vertex)
+    new_rank[dst] += contrib[src];
+end
+
+func updateVertex(v : Vertex)
+    var old_score : float = old_rank[v];
+    new_rank[v] = beta_score + damp*(new_rank[v]);
+    error[v] = fabs(new_rank[v] - old_rank[v]);
+    old_rank[v] = new_rank[v];
+    new_rank[v] = 0.0;
+end
+
+func printRank(v : Vertex)
+    print old_rank[v];
+end
+
+func reset(v: Vertex)
+    old_rank[v] = 1.0/vertices.size();
+    new_rank[v] = 0.0;
+end
+
+func main()
+    for trail in 0:10
+    	startTimer();
+        vertices.apply(reset);
+    	#s0# for i in 0:20
+    	    vertices.apply(computeContrib);
+            #s1# edges.apply(updateEdge);
+            vertices.apply(updateVertex);
+    	end
+
+    	var elapsed_time : float = stopTimer();
+    	print "elapsed time: ";
+    	print elapsed_time;
+    end
+end
+

autotune/gpu_apps/sssp_delta_stepping.gt

@@ -0,0 +1,38 @@
+element Vertex end
+element Edge end
+const edges : edgeset{Edge}(Vertex,Vertex, int) = load (argv[1]);
+const vertices : vertexset{Vertex} = edges.getVertices();
+const dist : vector{Vertex}(int) = 2147483647; %should be INT_MAX
+const pq: priority_queue{Vertex}(int);
+
+func updateEdge(src : Vertex, dst : Vertex, weight : int)
+    var new_dist : int = dist[src] + weight;
+    pq.updatePriorityMin(dst, dist[dst], new_dist);
+end
+
+func printDist(v : Vertex)
+    print dist[v];
+end
+
+func reset(v: Vertex)
+    dist[v] = 2147483647;
+end
+
+func main()
+    for trail in 0:10
+        var start_vertex : int = atoi(argv[2]);
+        pq = new priority_queue{Vertex}(int)(false, false, dist, 1, 2, false, start_vertex);
+        startTimer();
+        vertices.apply(reset);
+        dist[start_vertex] = 0;
+        #s0# while (pq.finished() == false)
+            var frontier : vertexset{Vertex} = pq.dequeue_ready_set(); % dequeue lowest priority nodes
+            #s1# edges.from(frontier).applyUpdatePriority(updateEdge);
+            delete frontier;
+        end
+        var elapsed_time : float = stopTimer();
+        print "elapsed time: ";
+        print elapsed_time;
+	delete pq;
+    end
+end