Merge pull request #1 from Iainmon/no-remote-vars

Remove and reimplement remote variables to support CPU and GPU operations.

.timedata/f32_fast_kernel.json

@@ -0,0 +1 @@
+{"avg": 1.203106, "stddev": 0.008578116576498592, "times": [1.21923, 1.19135, 1.20671, 1.19214, 1.20526, 1.20393, 1.2009, 1.2122, 1.2063, 1.19304]}

.timedata/f64_fast_kernel.json

@@ -0,0 +1 @@
+{"avg": 1.2119130000000002, "stddev": 0.005646131507501363, "times": [1.22538, 1.21134, 1.21124, 1.20609, 1.2136, 1.21643, 1.2142, 1.20709, 1.20729, 1.20647]}

.timedata/f64_old_kernel.json

@@ -0,0 +1 @@
+{"avg": 1.9952400000000001, "stddev": 0.005532747961004563, "times": [2.00349, 2.00492, 1.99984, 1.99339, 1.98599, 1.99134, 1.99441, 1.99439, 1.9928, 1.99183]}

.timedata/old.json

@@ -0,0 +1 @@
+{"avg": 2.000039, "stddev": 0.024113097042893505, "times": [1.99821, 2.0665, 1.9849, 1.98276, 1.98553, 1.99033, 1.99006, 2.01721, 1.99753, 1.98736]}

MNISTNet.chpl

@@ -9,6 +9,8 @@ import Time;
 
 config param layerDebug = false;
 
+type dtype = real(32);
+
 class CNN : Module(?) {
     var conv1: owned Conv2D(eltType);
     var conv2: owned Conv2D(eltType);
@@ -18,7 +20,7 @@ class CNN : Module(?) {
     var fc1: owned Linear(eltType);
     var fc2: owned Linear(eltType);
 
-    proc init(type eltType = real) {
+    proc init(type eltType = dtype) {
         super.init(eltType);
         // (1,3,3) x 32
         this.conv1 = new Conv2D(eltType,channels=1,features=32,kernel=3,stride=1); // (1,X,X) -> (32,Y,Y)
@@ -142,13 +144,17 @@ if diag {
 }
 
 
-var cnn = new CNN(real);
+var cnn = new CNN(dtype);
 
 
 for (n,m) in cnn.moduleFields() {
     writeln(n);
 }
 
+var model = Network.loadModel(specFile="scripts/models/cnn/specification.json",
+              weightsFolder="scripts/models/cnn/",
+              dtype=dtype);
+
 
 config const testImgSize = 28;
 
@@ -167,7 +173,7 @@ cnn.loadPyTorchDump(modelPath);
 
 config const imageCount = 0;
 
-var images = forall i in 0..<imageCount do Tensor.load("data/datasets/mnist/image_idx_" + i:string + ".chdata");
+var images = forall i in 0..<imageCount do Tensor.load("data/datasets/mnist/image_idx_" + i:string + ".chdata") : dtype;
 var preds: [images.domain] int;
 
 config const numTimes = 1;
@@ -186,7 +192,7 @@ for i in 0..<numTimes {
         // x = conv2(x);
         // var output = x;
         // pred = output.runtimeRank;
-        var output: Tensor(real) = cnn(img);
+        var output: Tensor(dtype) = model(img);
         pred = output.argmax();
         // writeln((i, pred));
     }

MultiLocaleInference.chpl

@@ -0,0 +1,76 @@
+use Tensor;
+
+use Network;
+
+use BlockDist;
+
+import Time;
+
+config const detach = true;
+
+Tensor.detachMode(detach);
+
+type dtype = real(32);
+
+// Load an array of images. 
+config const numImages = 1;
+
+// Create distributed domain for images. 
+const imagesD = blockDist.createDomain({0..<numImages});
+
+// Load distributed array of images.
+var images = forall i in imagesD do 
+    Tensor.load("data/datasets/mnist/image_idx_" + i:string + ".chdata") : dtype;
+
+// Create distributed domain for models. 
+const localeModelsD = blockDist.createDomain(Locales.domain);
+
+// Load distributed array of models. 
+var localeModels = forall li in localeModelsD do 
+    loadModel(specFile="scripts/models/cnn/specification.json",
+              weightsFolder="scripts/models/cnn/",
+              dtype=dtype);
+
+// Create distributed array of output results. 
+var preds: [imagesD] int;
+
+
+config const numTries = 1;
+
+var totalTime: real;
+
+for i in 0..<numTries {
+
+    var st = new Time.stopwatch();
+    st.start();
+
+    // coforall loc in Locales {
+    //     on loc {
+    //         const myAD = A.domain.localIndices();
+    //         forall i in myAD;
+    //     }
+    // }
+    forall (image,pred) in zip(images,preds) {
+        var model = localeModels[here.id].borrow();
+        pred = model(image).argmax();
+    }
+
+    st.stop();
+    const tm = st.elapsed();
+    totalTime += tm;
+
+    writeln("Trial ", i + 1, " of ", numTries," took ", tm, " seconds for ", numImages, " images on ", Locales.size, " nodes.");
+}
+
+const averageTime = totalTime / numTries;
+
+
+config const printResults = false;
+if printResults {
+    for i in images.domain {
+        writeln((i, preds[i]));
+    }
+}
+
+writeln("The average inference time for batch of size ", numImages, " was ", averageTime, " seconds on ", Locales.size, " nodes.");
+writeln("The total inference time for batch of size ", numImages, " was ", totalTime, " seconds on ", Locales.size, " nodes.");

cw.chpl

@@ -0,0 +1,11 @@
+// proc myWarning(param s, type t = string) {
+//   compilerWarning(s);
+// }
+
+proc somethingElse(x1) {
+  compilerWarning("idx");
+}
+// somethingElse(1);
+somethingElse(true);
+
+somethingElse(true);

definit.chpl

@@ -0,0 +1,26 @@
+
+use Reflection;
+
+
+proc (class).defaultInit(args...?nargs) {
+    param numFields = getNumFields(this.type);
+    if numFields != nargs then compilerError("Fields must match args");
+
+    for param i in 0..<numFields {
+        getField(this, i) = args(i);
+    }
+}
+
+
+
+class C {
+    var x: int;
+    var y: bool;
+
+    proc init() {
+        this.defaultInit(1,true);
+    }
+}
+
+var c = new owned C();
+writeln(c);