Merge branch 'time_init_read' into develop

CHANGELOG.md

@@ -7,6 +7,8 @@ All notable changes to this project will be documented in this file.
 - Thrust managed memory.
 - HIP managed memory.
 - New implementation using SYCL2020 USM (sycl2020-acc) and renamed original `sycl2020` to `sycl2020-acc`.
+- Data initialisation and read-back timing for all models, including Java, Scala, Julia, and Rust
+- Add support for the latest Aparapi (3.0.0) and TornadoVM (0.15.x) for Java   
 
 ### Changed
 - RAJA CUDA CMake build issues resolved.
@@ -17,6 +19,7 @@ All notable changes to this project will be documented in this file.
 - Number of thread-blocks in CUDA dot kernel implementation changed to 1024.
 - Fix compatibility of `sycl2020` (now `sycl2020-acc`) with hipSYCL. 
 - Bumped Julia compat to 1.9
+- Bumped Scala to 3.3.1
 - Bumped Rust to 1.74.0-nightly (13e6f24b9 2023-09-23)
 
 

src/java/java-stream/pom.xml

@@ -12,7 +12,7 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <project.reporting.outputEncoding>UTF-8</project.reporting.outputEncoding>
-        <junit.version>5.7.2</junit.version>
+        <junit.version>5.9.2</junit.version>
     </properties>
 
     <repositories>
@@ -27,19 +27,19 @@
         <dependency>
             <groupId>com.beust</groupId>
             <artifactId>jcommander</artifactId>
-            <version>1.81</version>
+            <version>1.82</version>
         </dependency>
 
         <dependency>
             <groupId>tornado</groupId>
             <artifactId>tornado-api</artifactId>
-            <version>0.9</version>
+            <version>0.15.1</version>
         </dependency>
 
         <dependency>
             <groupId>com.aparapi</groupId>
             <artifactId>aparapi</artifactId>
-            <version>2.0.0</version>
+            <version>3.0.0</version>
             <exclusions>
                 <!-- don't pull in the entire Scala ecosystem! -->
                 <exclusion>

src/java/java-stream/src/main/java/javastream/JavaStream.java

@@ -56,7 +56,7 @@ protected JavaStream(Config<T> config) {
 
   protected abstract T dot();
 
-  protected abstract Data<T> data();
+  protected abstract Data<T> readArrays();
 
   public static class EnumeratedStream<T> extends JavaStream<T> {
 
@@ -113,8 +113,8 @@ public T dot() {
     }
 
     @Override
-    public Data<T> data() {
-      return actual.data();
+    public Data<T> readArrays() {
+      return actual.readArrays();
     }
   }
 
@@ -140,6 +140,14 @@ private static Duration timed(Runnable f) {
     return Duration.ofNanos(end - start);
   }
 
+  final Duration runInitArrays() {
+    return timed(this::initArrays);
+  }
+
+  final SimpleImmutableEntry<Duration, Data<T>> runReadArrays() {
+    return timed(this::readArrays);
+  }
+
   final SimpleImmutableEntry<Timings<Duration>, T> runAll(int times) {
     Timings<Duration> timings = new Timings<>();
     T lastSum = null;

src/java/java-stream/src/main/java/javastream/Main.java

@@ -128,6 +128,40 @@ static final class Implementation {
     }
   }
 
+  @SuppressWarnings("unchecked")
+  static void showInit(
+      int totalBytes, double megaScale, Options opt, Duration init, Duration read) {
+    List<Entry<String, Double>> setup =
+        Arrays.asList(
+            new SimpleImmutableEntry<>("Init", durationToSeconds(init)),
+            new SimpleImmutableEntry<>("Read", durationToSeconds(read)));
+    if (opt.csv) {
+      tabulateCsv(
+          true,
+          setup.stream()
+              .map(
+                  x ->
+                      Arrays.asList(
+                          new SimpleImmutableEntry<>("function", x.getKey()),
+                          new SimpleImmutableEntry<>("n_elements", opt.arraysize + ""),
+                          new SimpleImmutableEntry<>("sizeof", totalBytes + ""),
+                          new SimpleImmutableEntry<>(
+                              "max_m" + (opt.mibibytes ? "i" : "") + "bytes_per_sec",
+                              ((megaScale * (double) totalBytes / x.getValue())) + ""),
+                          new SimpleImmutableEntry<>("runtime", x.getValue() + "")))
+              .toArray(List[]::new));
+    } else {
+      for (Entry<String, Double> e : setup) {
+        System.out.printf(
+            "%s: %.5f s (%.5f M%sBytes/sec)%n",
+            e.getKey(),
+            e.getValue(),
+            megaScale * (double) totalBytes / e.getValue(),
+            opt.mibibytes ? "i" : "");
+      }
+    }
+  }
+
   static <T extends Number> boolean run(
       String name, Config<T> config, Function<Config<T>, JavaStream<T>> mkStream) {
 
@@ -183,35 +217,46 @@ static <T extends Number> boolean run(
 
     JavaStream<T> stream = mkStream.apply(config);
 
-    stream.initArrays();
-
+    Duration init = stream.runInitArrays();
     final boolean ok;
     switch (config.benchmark) {
       case ALL:
-        Entry<Timings<Duration>, T> results = stream.runAll(opt.numtimes);
-        ok = checkSolutions(stream.data(), config, Optional.of(results.getValue()));
-        Timings<Duration> timings = results.getKey();
-        tabulateCsv(
-            opt.csv,
-            mkCsvRow(timings.copy, "Copy", 2 * arrayBytes, megaScale, opt),
-            mkCsvRow(timings.mul, "Mul", 2 * arrayBytes, megaScale, opt),
-            mkCsvRow(timings.add, "Add", 3 * arrayBytes, megaScale, opt),
-            mkCsvRow(timings.triad, "Triad", 3 * arrayBytes, megaScale, opt),
-            mkCsvRow(timings.dot, "Dot", 2 * arrayBytes, megaScale, opt));
-        break;
+        {
+          Entry<Timings<Duration>, T> results = stream.runAll(opt.numtimes);
+          SimpleImmutableEntry<Duration, Data<T>> read = stream.runReadArrays();
+          showInit(totalBytes, megaScale, opt, init, read.getKey());
+          ok = checkSolutions(read.getValue(), config, Optional.of(results.getValue()));
+          Timings<Duration> timings = results.getKey();
+          tabulateCsv(
+              opt.csv,
+              mkCsvRow(timings.copy, "Copy", 2 * arrayBytes, megaScale, opt),
+              mkCsvRow(timings.mul, "Mul", 2 * arrayBytes, megaScale, opt),
+              mkCsvRow(timings.add, "Add", 3 * arrayBytes, megaScale, opt),
+              mkCsvRow(timings.triad, "Triad", 3 * arrayBytes, megaScale, opt),
+              mkCsvRow(timings.dot, "Dot", 2 * arrayBytes, megaScale, opt));
+          break;
+        }
       case NSTREAM:
-        List<Duration> nstreamResults = stream.runNStream(opt.numtimes);
-        ok = checkSolutions(stream.data(), config, Optional.empty());
-        tabulateCsv(opt.csv, mkCsvRow(nstreamResults, "Nstream", 4 * arrayBytes, megaScale, opt));
-        break;
+        {
+          List<Duration> nstreamResults = stream.runNStream(opt.numtimes);
+          SimpleImmutableEntry<Duration, Data<T>> read = stream.runReadArrays();
+          showInit(totalBytes, megaScale, opt, init, read.getKey());
+          ok = checkSolutions(read.getValue(), config, Optional.empty());
+          tabulateCsv(opt.csv, mkCsvRow(nstreamResults, "Nstream", 4 * arrayBytes, megaScale, opt));
+          break;
+        }
       case TRIAD:
-        Duration triadResult = stream.runTriad(opt.numtimes);
-        ok = checkSolutions(stream.data(), config, Optional.empty());
-        int triadTotalBytes = 3 * arrayBytes * opt.numtimes;
-        double bandwidth = megaScale * (triadTotalBytes / durationToSeconds(triadResult));
-        System.out.printf("Runtime (seconds): %.5f", durationToSeconds(triadResult));
-        System.out.printf("Bandwidth (%s/s): %.3f ", gigaSuffix, bandwidth);
-        break;
+        {
+          Duration triadResult = stream.runTriad(opt.numtimes);
+          SimpleImmutableEntry<Duration, Data<T>> read = stream.runReadArrays();
+          showInit(totalBytes, megaScale, opt, init, read.getKey());
+          ok = checkSolutions(read.getValue(), config, Optional.empty());
+          int triadTotalBytes = 3 * arrayBytes * opt.numtimes;
+          double bandwidth = megaScale * (triadTotalBytes / durationToSeconds(triadResult));
+          System.out.printf("Runtime (seconds): %.5f", durationToSeconds(triadResult));
+          System.out.printf("Bandwidth (%s/s): %.3f ", gigaSuffix, bandwidth);
+          break;
+        }
       default:
         throw new AssertionError();
     }

src/java/java-stream/src/main/java/javastream/aparapi/AparapiStreams.java

@@ -122,7 +122,7 @@ public T dot() {
     }
 
     @Override
-    public Data<T> data() {
+    public Data<T> readArrays() {
       return kernels.syncAndDispose();
     }
   }

src/java/java-stream/src/main/java/javastream/jdk/GenericPlainStream.java

@@ -86,7 +86,7 @@ public T dot() {
   }
 
   @Override
-  public Data<T> data() {
+  public Data<T> readArrays() {
     return new Data<>(a, b, c);
   }
 }

src/java/java-stream/src/main/java/javastream/jdk/GenericStream.java

@@ -80,7 +80,7 @@ public T dot() {
   }
 
   @Override
-  public Data<T> data() {
+  public Data<T> readArrays() {
     return new Data<>(a, b, c);
   }
 }

src/java/java-stream/src/main/java/javastream/jdk/SpecialisedDoubleStream.java

@@ -78,7 +78,7 @@ public Double dot() {
   }
 
   @Override
-  public Data<Double> data() {
+  public Data<Double> readArrays() {
     return new Data<>(boxed(a), boxed(b), boxed(c));
   }
 }

src/java/java-stream/src/main/java/javastream/jdk/SpecialisedFloatStream.java

@@ -78,7 +78,7 @@ public Float dot() {
   }
 
   @Override
-  public Data<Float> data() {
+  public Data<Float> readArrays() {
     return new Data<>(boxed(a), boxed(b), boxed(c));
   }
 }

src/java/java-stream/src/main/java/javastream/jdk/SpecialisedPlainDoubleStream.java

@@ -78,7 +78,7 @@ public Double dot() {
   }
 
   @Override
-  public Data<Double> data() {
+  public Data<Double> readArrays() {
     return new Data<>(boxed(a), boxed(b), boxed(c));
   }
 }

src/java/java-stream/src/main/java/javastream/jdk/SpecialisedPlainFloatStream.java

@@ -78,7 +78,7 @@ public Float dot() {
   }
 
   @Override
-  public Data<Float> data() {
+  public Data<Float> readArrays() {
     return new Data<>(boxed(a), boxed(b), boxed(c));
   }
 }

src/java/java-stream/src/main/java/javastream/tornadovm/GenericTornadoVMStream.java

@@ -4,27 +4,27 @@
 import java.util.stream.Collectors;
 import javastream.JavaStream;
 import javastream.Main.Config;
-import uk.ac.manchester.tornado.api.TaskSchedule;
-import uk.ac.manchester.tornado.api.TornadoRuntimeCI;
+import uk.ac.manchester.tornado.api.TornadoExecutionPlan;
+import uk.ac.manchester.tornado.api.TornadoRuntimeInterface;
 import uk.ac.manchester.tornado.api.common.TornadoDevice;
 import uk.ac.manchester.tornado.api.runtime.TornadoRuntime;
 
 abstract class GenericTornadoVMStream<T> extends JavaStream<T> {
 
   protected final TornadoDevice device;
 
-  protected TaskSchedule copyTask;
-  protected TaskSchedule mulTask;
-  protected TaskSchedule addTask;
-  protected TaskSchedule triadTask;
-  protected TaskSchedule nstreamTask;
-  protected TaskSchedule dotTask;
+  protected TornadoExecutionPlan copyTask;
+  protected TornadoExecutionPlan mulTask;
+  protected TornadoExecutionPlan addTask;
+  protected TornadoExecutionPlan triadTask;
+  protected TornadoExecutionPlan nstreamTask;
+  protected TornadoExecutionPlan dotTask;
 
   GenericTornadoVMStream(Config<T> config) {
     super(config);
 
     try {
-      TornadoRuntimeCI runtime = TornadoRuntime.getTornadoRuntime();
+      TornadoRuntimeInterface runtime = TornadoRuntime.getTornadoRuntime();
       List<TornadoDevice> devices = TornadoVMStreams.enumerateDevices(runtime);
       device = devices.get(config.options.device);
 
@@ -42,10 +42,6 @@ abstract class GenericTornadoVMStream<T> extends JavaStream<T> {
     }
   }
 
-  protected static TaskSchedule mkSchedule() {
-    return new TaskSchedule("");
-  }
-
   @Override
   public List<String> listDevices() {
     return TornadoVMStreams.enumerateDevices(TornadoRuntime.getTornadoRuntime()).stream()
@@ -55,12 +51,12 @@ public List<String> listDevices() {
 
   @Override
   public void initArrays() {
-    this.copyTask.warmup();
-    this.mulTask.warmup();
-    this.addTask.warmup();
-    this.triadTask.warmup();
-    this.nstreamTask.warmup();
-    this.dotTask.warmup();
+    this.copyTask.withWarmUp();
+    this.mulTask.withWarmUp();
+    this.addTask.withWarmUp();
+    this.triadTask.withWarmUp();
+    this.nstreamTask.withWarmUp();
+    this.dotTask.withWarmUp();
   }
 
   @Override