add Chapel

.gitignore

@@ -57,6 +57,23 @@ func.c                          # PRK C89 stencil generated code
 
 MPI1/AMR/amr
 MPI1/Branch/branch
+MPI1/Branch/func.c
+
+# auto-generated files
+func.c
+*.incl
+
+# scripts
+common/make.defs
+scripts/small/runfgmpi
+scripts/wide/runfgmpi
+
+# binaries
+CHAPEL/Stencil/stencil
+CHAPEL/Synch_p2p/p2p
+CHAPEL/Transpose/transpose
+MPI1/Branch/branch
+MPI1/Branch/func.c
 MPI1/DGEMM/dgemm
 MPI1/Nstream/nstream
 MPI1/PIC-static/pic
@@ -73,7 +90,10 @@ MPIOPENMP/Synch_p2p/p2p
 MPIOPENMP/Transpose/transpose
 MPIRMA/Stencil/stencil
 MPIRMA/Synch_p2p/p2p
+MPIRMA/Stencil/stencil
 MPIRMA/Transpose/transpose
+MPISHM/Stencil/loop_body_compact.incl
+MPISHM/Stencil/loop_body_star.incl
 MPISHM/Stencil/stencil
 MPISHM/Synch_p2p/p2p
 MPISHM/Transpose/transpose
@@ -91,6 +111,9 @@ OPENMP/Synch_p2p/p2p
 OPENMP/Transpose/transpose
 SERIAL/AMR/amr
 SERIAL/Branch/branch
+SERIAL/Branch/func.c
+SERIAL/Branch/branch
+SERIAL/Branch/func.c
 SERIAL/DGEMM/dgemm
 SERIAL/Nstream/nstream
 SERIAL/PIC/pic

.travis.yml

@@ -28,8 +28,8 @@ env:
   #- PRK_TARGET=allupc UPC_IMPL=bupc GASNET_CONDUIT=mpi PRK_FLAGS="-Wc,-O3"
   #- PRK_TARGET=allupc UPC_IMPL=bupc GASNET_CONDUIT=ofi PRK_FLAGS="-Wc,-O3"
   # Chapel kernels are not merged yet.  Activate these when they are.
-  #- PRK_TARGET=allchapel CHPL_COMM=none
-  #- PRK_TARGET=allchapel CHPL_COMM=gasnet
+  - PRK_TARGET=allchapel CHPL_COMM=none
+  - PRK_TARGET=allchapel CHPL_COMM=gasnet
   # HPX-3 kernels are not merged yet.  Activate these when they are.
   #- PRK_TARGET=allhpx3
   # HPX-5 kernels are not merged yet.  Activate these when they are.
@@ -89,9 +89,18 @@ matrix:
   # Revisit this once Grappa is working with Clang
   - compiler: gcc
     env: PRK_TARGET=allgrappa
-  # Mac + Chapel + GASNet does not work and I do not care
+  # Chapel + GASNet does not work in Travis
   - os: osx
     env: PRK_TARGET=allchapel CHPL_COMM=gasnet
+  #- os: linux
+  #  env: PRK_TARGET=allchapel CHPL_COMM=gasnet
+  # Linux Clang/GCC too old ("no member named 'max_align_t' in the global namespace")
+  - os: linux
+    compiler: clang
+    env: PRK_TARGET=allchapel CHPL_COMM=none
+  - os: linux
+    compiler: clang
+    env: PRK_TARGET=allchapel CHPL_COMM=gasnet
   # UPC GASNet OFI conduit is busted, perhaps due to SSH spawner
   - env: PRK_TARGET=allupc UPC_IMPL=bupc GASNET_CONDUIT=ofi PRK_FLAGS="-Wc,-O3"
   # probably because of old GCC (4.6)

CHAPEL/README.md

@@ -0,0 +1,23 @@
+# Intel's Parallel Research Kernels (PRK)
+
+https://github.com/ParRes/Kernels
+
+This is a set of simple programs that can be used to explore the features of a
+parallel platform.
+
+
+## Implementations:
+
+Each kernel is implemented in a variety of parallel flavors, such as "serial",
+"shared", and "dist"-ributed. There is also two versions for each flavor:
+a "fast" version optimized for performance and a
+"pretty" version optimized for elegance. Ideally, the performance delta
+between these versions will approach zero over time.
+
+The test inputs were taken from the inputs used in the PRK test scripts.
+
+### Stencil
+
+### Synch_p2p
+
+### Transpose

CHAPEL/Stencil/Makefile

@@ -0,0 +1,23 @@
+PROGRAM=stencil
+
+include ../../common/CHAPEL.defs
+
+CHPLFLAGS=-O
+
+.PHONY: all clean run
+
+all: $(PROGRAM) $(PROGRAM)-serial
+
+$(PROGRAM): $(PROGRAM).chpl
+	$(CHPL) $(CHPLFLAGS) $< -o $@
+
+$(PROGRAM)-serial: $(PROGRAM)-serial.chpl
+	$(CHPL) $(CHPLFLAGS) $< -o $@
+
+run: $(PROGRAM)
+	$(PROGRAM) --iterations=10 --order=32000 --tileSize=20
+
+clean:
+	-rm -f $(PROGRAM)
+	-rm -f $(PROGRAM)-serial
+

CHAPEL/Stencil/stencil-serial.chpl

@@ -0,0 +1,189 @@
+//
+// Chapel's stencil implementation
+//
+use Time;
+
+param PRKVERSION = "2.15";
+
+config var tileSize: int = 0;
+
+config const iterations: int = 0,
+             order: int = 0,
+             debug: bool = false,
+             validate: bool = false;
+
+config param R = 2,
+             compact = false;
+
+// Configurable type for array elements
+config type dtype = real;
+
+const activePoints = (order-2*R)*(order-2*R),
+      coefx: dtype = 1.0,
+      coefy: dtype = 1.0;
+
+param stencilSize = 4*R + 1,
+      weightSize = 2*R + 1,
+      Wsize = 2*R + 1,
+      R1 = R+1;
+
+var timer: Timer;
+
+//
+// Process and test input configs
+//
+if ( (iterations == 0) && (order==0) ) {
+  writeln("Usage: ./stencil --iterations=<# iterations> --order=<matrix order> [--tileSize=<tile_size>]");
+  // None of these work...
+  //writeln("          [--compact=<boolean>] [--useBlockDist=<boolean>] [--useStencilDist=<boolean>]");
+  exit(1);
+}
+if (iterations < 1) {
+  writeln("ERROR: iterations must be >= 1: ", iterations);
+  exit(1);
+}
+if (order < 1) {
+  writeln("ERROR: Matrix Order must be greater than 0 : ", order);
+  exit(1);
+}
+if (R < 1) {
+  writeln("ERROR: Stencil radius ", R, " should be positive");
+  exit(1);
+}
+if (2*R + 1 > order) {
+  writeln("ERROR: Stencil radius ", R, " exceeds grid size ", order);
+  exit(1);
+}
+
+// Determine tiling
+var tiling = (tileSize > 0 && tileSize < order);
+
+// Safety check for creation of tiledDom
+if (!tiling) then tileSize = 1;
+
+// Domains
+const    Dom = {0.. # order, 0.. # order},
+    innerDom = Dom.expand(-R),
+   weightDom = {-R..R, -R..R};
+
+var tiledDom = {R.. # order-2*R by tileSize, R.. # order-2*R by tileSize};
+
+// Arrays
+var input, output: [Dom] dtype = 0.0;
+var tmpout : dtype = 0.0;
+
+// Tuple of tuples
+var weight: Wsize*(Wsize*(dtype));
+
+// Set up weight matrix
+for i in 1..R {
+  const element : dtype = 1 / (2*i*R) : dtype;
+  weight[R1][R1+i]  =  element;
+  weight[R1+i][R1]  =  element;
+  weight[R1-i][R1] = -element;
+  weight[R1][R1-i] = -element;
+}
+
+// Initialize the input and output arrays
+[(i, j) in Dom] input[i,j] = coefx*i + coefy*j;
+
+//
+// Print information before main loop
+//
+if (!validate) {
+  writeln("Parallel Research Kernels Version ", PRKVERSION);
+  writeln("Serial stencil execution on 2D grid");
+  writeln("Grid size            = ", order);
+  writeln("Radius of stencil    = ", R);
+  if compact then writeln("Type of stencil      = compact");
+  else              writeln("Type of stencil      = star");
+  writeln("Data type            = ", dtype:string);
+  if tiling then writeln("Tile size            = ", tileSize);
+  else             writeln("Untiled");
+  writeln("Number of iterations = ", iterations);
+}
+
+//
+// Main loop
+//
+for iteration in 0..iterations {
+
+  // Start timer after warmup iteration
+  if (iteration == 1) {
+    timer.start();
+  }
+
+  if (!tiling) {
+    for (i,j) in innerDom {
+      var tmpout: dtype = 0.0;
+      if (!compact) {
+        for param jj in -R..-1 do tmpout += weight[R1][R1+jj] * input[i, j+jj];
+        for param jj in 1..R   do tmpout += weight[R1][R1+jj] * input[i, j+jj];
+        for param ii in -R..-1 do tmpout += weight[R1+ii][R1] * input[i+ii, j];
+        for param ii in 1..R   do tmpout += weight[R1+ii][R1] * input[i+ii, j];
+      } else {
+        for (ii, jj) in weightDom do
+          tmpout += weight[R1+ii][R1+jj] * input[i+ii, j+jj];
+      }
+      output[i, j] += tmpout;
+    }
+  } else {
+    for (it,jt) in tiledDom {
+      for i in it .. # min(order - R - it, tileSize) {
+        for j in jt .. # min(order - R - jt, tileSize) {
+          var tmpout: dtype = 0.0;
+          if (!compact) {
+            for param jj in -R..-1 do tmpout += weight[R1][R1+jj] * input[i, j+jj];
+            for param jj in 1..R   do tmpout += weight[R1][R1+jj] * input[i, j+jj];
+            for param ii in -R..-1 do tmpout += weight[R1+ii][R1] * input[i+ii, j];
+            for param ii in 1..R   do tmpout += weight[R1+ii][R1] * input[i+ii, j];
+          } else {
+            for (ii, jj) in weightDom do
+              tmpout += weight[R1+ii][R1+jj] * input[i+ii, j+jj];
+          }
+          output[i, j] += tmpout;
+        }
+      }
+    }
+  }
+
+  // Add constant to solution to force refresh of neighbor data, if any
+  for (i,j) in Dom do input[i,j] += 1.0;
+
+} // end of iterations
+
+timer.stop();
+
+//
+// Analyze and output results
+//
+
+// Timings
+var stencilTime = timer.elapsed(),
+    flops = (2*stencilSize + 1) * activePoints,
+    avgTime = stencilTime / iterations;
+
+// Compute L1 norm
+var referenceNorm = (iterations + 1) * (coefx + coefy),
+    norm = + reduce abs(output);
+norm /= activePoints;
+
+// Error threshold
+const epsilon = 1.e-8;
+
+// Verify correctness
+if abs(norm-referenceNorm) > epsilon then {
+  writeln("ERROR: L1 norm = ", norm, ", Reference L1 norm = ", referenceNorm);
+  exit(1);
+} else {
+  writeln("Solution validates");
+
+  if debug {
+    writeln("L1 norm = ", norm, ", Reference L1 norm = ", referenceNorm);
+  }
+
+  if (!validate) {
+    writeln("Rate (MFlops/s): ", 1.0E-06 * flops/avgTime, "  Avg time (s): ", 
+            avgTime);
+  }
+}