Add input arg option to automate block factor choice

hal-uw · Jun 1, 2024 · e2fc08e · e2fc08e
1 parent e6bc6b8
commit e2fc08e
Show file tree

Hide file tree

Showing 15 changed files with 106 additions and 46 deletions.
diff --git a/LONESTAR/apps/bfs/kernel_ccg.cu b/LONESTAR/apps/bfs/kernel_ccg.cu
@@ -12,6 +12,7 @@ void kernel_sizing(CSRGraph &, dim3 &, dim3 &);
 const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_blocking_factor=4 $ parcomb=True $ np_schedulers=set(['fg', 'tb', 'wp']) $ cc_disable=set([]) $ tb_lb=True $ hacks=set([]) $ np_factor=8 $ instrument=set([]) $ unroll=[] $ read_props=None $ outline_iterate=True $ ignore_nested_errors=False $ np=True $ write_props=None $ quiet_cgen=True $ retry_backoff=True $ cuda.graph_type=basic $ cuda.use_worklist_slots=True $ cuda.worklist_type=basic";
 struct ThreadWork t_work;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -407,13 +408,18 @@ void gg_main_pipe_1_wrapper(CSRGraph& gg, int& LEVEL, PipeContextT<Worklist2>& p
     check_cuda(cudaFree(cl_LEVEL));
   }
 }
+
 void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   PipeContextT<Worklist2> wl;

diff --git a/LONESTAR/apps/bfs/kernel_gcpusrb.cu b/LONESTAR/apps/bfs/kernel_gcpusrb.cu
@@ -291,6 +291,7 @@ void kernel_sizing(CSRGraph &, dim3 &, dim3 &);
 const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_blocking_factor=4 $ parcomb=True $ np_schedulers=set(['fg', 'tb', 'wp']) $ cc_disable=set([]) $ tb_lb=True $ hacks=set([]) $ np_factor=8 $ instrument=set([]) $ unroll=[] $ read_props=None $ outline_iterate=True $ ignore_nested_errors=False $ np=True $ write_props=None $ quiet_cgen=True $ retry_backoff=True $ cuda.graph_type=basic $ cuda.use_worklist_slots=True $ cuda.worklist_type=basic";
 struct ThreadWork t_work;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -715,9 +716,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   PipeContextT<Worklist2> wl;

diff --git a/LONESTAR/apps/bfs/kernel_gsrb.cu b/LONESTAR/apps/bfs/kernel_gsrb.cu
@@ -289,6 +289,7 @@ void kernel_sizing(CSRGraph &, dim3 &, dim3 &);
 const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_blocking_factor=4 $ parcomb=True $ np_schedulers=set(['fg', 'tb', 'wp']) $ cc_disable=set([]) $ tb_lb=True $ hacks=set([]) $ np_factor=8 $ instrument=set([]) $ unroll=[] $ read_props=None $ outline_iterate=True $ ignore_nested_errors=False $ np=True $ write_props=None $ quiet_cgen=True $ retry_backoff=True $ cuda.graph_type=basic $ cuda.use_worklist_slots=True $ cuda.worklist_type=basic";
 struct ThreadWork t_work;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -713,9 +714,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   PipeContextT<Worklist2> wl;

diff --git a/LONESTAR/apps/bfs/run.sh b/LONESTAR/apps/bfs/run.sh
@@ -1,8 +1,8 @@
 #!/bin/bash
 
 # Run GPU SRB (sense-reversing barrier) variant
-../../bin/bfs_gsrb -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/bfs_gsrb -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
 # Run GPU SRB variant that uses CPU-style SRBs
-../../bin/bfs_gcpusrb -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/bfs_gcpusrb -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
 # Run CUDA Cooperative Graphs (CCG) variant
-../../bin/bfs_ccg -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/bfs_ccg -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
diff --git a/LONESTAR/apps/bfs/support.cu b/LONESTAR/apps/bfs/support.cu
@@ -2,25 +2,29 @@
 
 #include "gg.h"
 
-const char *prog_opts = "ls:";
-const char *prog_usage = "[-l] [-s startNode]";
+const char *prog_opts = "ls:b:";
+const char *prog_usage = "[-l] [-s startNode] [-b blockFactor]";
 const char *prog_args_usage = "-l: enable thread block load balancer (by default false)";
 
 extern const int INF;
 int start_node = 0;
+int block_factor = 0;
 extern bool enable_lb;
 
 int process_prog_arg(int argc, char *argv[], int arg_start) {
    return 1;
 }
 
 void process_prog_opt(char c, char *optarg) {
-   if(c == 's') {
+   if (c == 's') {
      start_node = atoi(optarg);
      assert(start_node >= 0);
    }
-   if(c == 'l') {
-	   enable_lb = true;
+   if (c == 'l') {
+     enable_lb = true;
+   }
+   if (c == 'b') {
+     block_factor = atoi(optarg);
    }
 }
 

diff --git a/LONESTAR/apps/pr/kernel_ccg.cu b/LONESTAR/apps/pr/kernel_ccg.cu
@@ -21,6 +21,7 @@ float* P_NEXT ;
 extern const float ALPHA = 0.85;
 extern const float EPSILON = 0.000001;
 extern int MAX_ITERATIONS ;
+extern int block_factor;
 static const int __tb_gg_main_pipe_1_gpu_gb = 32;
 static const int __tb_one = 1;
 static const int __tb_pagerank_main = TB_SIZE;
@@ -385,9 +386,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/pr/kernel_gcpusrb.cu b/LONESTAR/apps/pr/kernel_gcpusrb.cu
@@ -291,6 +291,7 @@ float* P_NEXT ;
 extern const float ALPHA = 0.85;
 extern const float EPSILON = 0.000001;
 extern int MAX_ITERATIONS ;
+extern int block_factor;
 static const int __tb_gg_main_pipe_1_gpu_gb = 32;
 static const int __tb_one = 1;
 static const int __tb_pagerank_main = TB_SIZE;
@@ -684,9 +685,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/pr/kernel_gsrb.cu b/LONESTAR/apps/pr/kernel_gsrb.cu
@@ -291,6 +291,7 @@ float* P_NEXT ;
 extern const float ALPHA = 0.85;
 extern const float EPSILON = 0.000001;
 extern int MAX_ITERATIONS ;
+extern int block_factor;
 static const int __tb_gg_main_pipe_1_gpu_gb = 32;
 static const int __tb_one = 1;
 static const int __tb_pagerank_main = TB_SIZE;
@@ -684,9 +685,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/pr/run.sh b/LONESTAR/apps/pr/run.sh
@@ -1,8 +1,8 @@
 #!/bin/bash
 
 # Run GPU SRB (sense-reversing barrier) variant
-../../bin/pr_gsrb -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
+../../bin/pr_gsrb -b 4 -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
 # Run GPU SRB variant that uses CPU-style SRBs
-../../bin/pr_gcpusrb -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
+../../bin/pr_gcpusrb -b 4 -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
 # Run CUDA Cooperative Graphs (CCG) variant
-../../bin/pr_ccg -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
+../../bin/pr_ccg -b 4 -o outfile.txt -x 1000000 ../../inputs/USA-road-d.W.gr
diff --git a/LONESTAR/apps/pr/support.cu b/LONESTAR/apps/pr/support.cu
@@ -13,13 +13,14 @@ struct pr_value {
 };
 
 /* TODO: accept ALPHA and EPSILON */
-const char *prog_opts = "nt:x:";
-const char *prog_usage = "[-n] [-t top_ranks] [-x max_iterations]";
+const char *prog_opts = "nt:x:b:";
+const char *prog_usage = "[-n] [-t top_ranks] [-x max_iterations] [-b blockFactor]";
 const char *prog_args_usage = "";
 
 extern float *P_CURR, *P_NEXT;
 extern const float ALPHA, EPSILON;
 extern int MAX_ITERATIONS;
+int block_factor = 0;
 
 int NO_PRINT_PAGERANK = 0;
 int PRINT_TOP = 0;
@@ -30,8 +31,9 @@ int process_prog_arg(int argc, char *argv[], int arg_start) {
 }
 
 void process_prog_opt(char c, char *optarg) {
-  if(c == 'n')
+  if(c == 'n') {
     NO_PRINT_PAGERANK = 1;
+  }
 
   if(c == 't') {
     PRINT_TOP = atoi(optarg);    
@@ -40,6 +42,10 @@ void process_prog_opt(char c, char *optarg) {
   if(c == 'x') {
     MAX_ITERATIONS = atoi(optarg);
   }
+
+  if (c == 'b') {
+    block_factor = atoi(optarg);
+  }
 }
 
 void output(CSRGraphTy &g, const char *output_file) {

diff --git a/LONESTAR/apps/sssp/kernel_ccg.cu b/LONESTAR/apps/sssp/kernel_ccg.cu
@@ -13,6 +13,7 @@ const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_b
 struct ThreadWork t_work;
 extern int DELTA;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -484,9 +485,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/sssp/kernel_gcpusrb.cu b/LONESTAR/apps/sssp/kernel_gcpusrb.cu
@@ -284,6 +284,7 @@ const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_b
 struct ThreadWork t_work;
 extern int DELTA;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -796,9 +797,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/sssp/kernel_gsrb.cu b/LONESTAR/apps/sssp/kernel_gsrb.cu
@@ -284,6 +284,7 @@ const char *GGC_OPTIONS = "coop_conv=False $ outline_iterate_gb=True $ backoff_b
 struct ThreadWork t_work;
 extern int DELTA;
 extern int start_node;
+extern int block_factor;
 bool enable_lb = false;
 typedef int edge_data_type;
 typedef int node_data_type;
@@ -791,9 +792,13 @@ void gg_main(CSRGraph& hg, CSRGraph& gg)
 {
   dim3 blocks, threads;
   kernel_sizing(gg, blocks, threads);
-  std::cout << " Enter Block factor" << std::endl;
-  int block_factor;
-  std::cin >> block_factor;
+  // if no user input block factor then request
+  if (block_factor == 0) {
+    std::cout << " Enter Block factor" << std::endl;
+    std::cin >> block_factor;
+    assert(block_factor > 0);
+  }
+  fprintf(stdout, "Using block factor: %d\n", block_factor);
   blocks = ggc_get_nSM()*block_factor;
   t_work.init_thread_work(gg.nnodes);
   static GlobalBarrierLifetime remove_dups_barrier;

diff --git a/LONESTAR/apps/sssp/run.sh b/LONESTAR/apps/sssp/run.sh
@@ -1,8 +1,8 @@
 #!/bin/bash
 
 # Run GPU SRB (sense-reversing barrier) variant
-../../bin/sssp_gsrb -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/sssp_gsrb -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
 # Run GPU SRB variant that uses CPU-style SRBs
-../../bin/sssp_gcpusrb -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/sssp_gcpusrb -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
 # Run CUDA Cooperative Graphs (CCG) variant
-../../bin/sssp_ccg -s 0 -o outfile.txt ../../inputs/USA-road-d.W.gr
+../../bin/sssp_ccg -s 0 -b 4 -o outfile.txt ../../inputs/USA-road-d.W.gr
diff --git a/LONESTAR/apps/sssp/support.cu b/LONESTAR/apps/sssp/support.cu
@@ -3,13 +3,14 @@
 #include "gg.h"
 #include <cassert>
 
-const char *prog_opts = "ls:d:";
-const char *prog_usage = "[-l] [-d delta] [-s startNode]";
+const char *prog_opts = "ls:d:b:";
+const char *prog_usage = "[-l] [-d delta] [-s startNode] [-b blockFactor]";
 const char *prog_args_usage = "-l: enable thread block load balancer (by default false)";
 
 int DELTA = 10000;
 extern const int INF;
 int start_node = 0;
+int block_factor = 0;
 extern bool enable_lb;
 
 int process_prog_arg(int argc, char *argv[], int arg_start) {
@@ -28,6 +29,9 @@ void process_prog_opt(char c, char *optarg) {
      start_node = atoi(optarg);
      assert(start_node >= 0);
   }
+  if (c == 'b') {
+    block_factor = atoi(optarg);
+  }
 }
 
 void output(CSRGraphTy &g, const char *output_file) {