Thread: [X10-commits] SF.net SVN: x10:[12125] trunk (Page 10)

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 422-6466

Revision: 12125
          http://x10.svn.sourceforge.net/x10/?rev=12125&view=rev
Author:   sparksparkspark
Date:     2009-11-18 18:25:26 +0000 (Wed, 18 Nov 2009)

Log Message:
-----------
Add autoconfiguration utility for CUDA and use it in KMeansCUDA

Modified Paths:
--------------
    trunk/x10.compiler/src/x10cuda/types/X10CUDAContext_c.java
    trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java
    trunk/x10.dist/samples/KMeansCUDA.x10

Added Paths:
-----------
    trunk/x10.runtime/src-x10/x10/compiler/CUDAUtilities.x10

Modified: trunk/x10.compiler/src/x10cuda/types/X10CUDAContext_c.java
===================================================================

--- trunk/x10.compiler/src/x10cuda/types/X10CUDAContext_c.java	2009-11-18 18:13:39 UTC (rev 12124)
+++ trunk/x10.compiler/src/x10cuda/types/X10CUDAContext_c.java	2009-11-18 18:25:26 UTC (rev 12125)
@@ -18,6 +18,7 @@
 
 import polyglot.ast.Expr;
 import polyglot.ast.Formal;
+import polyglot.ast.LocalDecl;
 import polyglot.frontend.Job;
 import x10.ast.Closure_c;
 import x10cpp.types.X10CPPContext_c;
@@ -58,6 +59,10 @@
         this.threadsVar = threadsVar;
         this.shm = shm;
         this.kernelParams = variables();
+        if (autoBlocks!=null)
+            this.kernelParams.add(autoBlocks.localDef().asInstance());
+        if (autoThreads!=null)
+            this.kernelParams.add(autoThreads.localDef().asInstance());
     }
     public boolean isKernelParam(Name n) {
         for (VarInstance i : kernelParams) {
@@ -85,7 +90,15 @@
     X10CUDAContext_c established;
     public void establishClosure() { established = this; }
     public X10CUDAContext_c established() { return established; }
-    
+
+    LocalDecl autoBlocks;
+    public void autoBlocks(LocalDecl v) { this.autoBlocks = v; }
+    public LocalDecl autoBlocks() { return this.autoBlocks; }
+
+    LocalDecl autoThreads;
+    public void autoThreads(LocalDecl v) { this.autoThreads = v; }
+    public LocalDecl autoThreads() { return this.autoThreads; }
+
 }
 
 //vim:tabstop=4:shiftwidth=4:expandtab

Modified: trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java
===================================================================
--- trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java	2009-11-18 18:13:39 UTC (rev 12124)
+++ trunk/x10.compiler/src/x10cuda/visit/CUDACodeGenerator.java	2009-11-18 18:25:26 UTC (rev 12125)
@@ -405,6 +405,30 @@
         r.var = real_var.name().id();
         return r;
     }
+    
+    public void checkAutoVar(Stmt s) {
+        assert s instanceof LocalDecl : s.getClass(); // FIXME: proper error
+        LocalDecl s_ = (LocalDecl)s;
+        Expr init_expr = s_.init();
+        assert init_expr instanceof X10Call_c : init_expr.getClass(); // FIXME: proper error
+        X10Call_c init_call = (X10Call_c) init_expr;
+        Receiver init_call_target = init_call.target();
+        assert init_call_target instanceof CanonicalTypeNode : init_call_target.getClass(); // FIXME: proper error
+        CanonicalTypeNode init_call_target_node = (CanonicalTypeNode) init_call_target;
+        assert init_call_target_node.nameString().equals("CUDAUtilities") : init_call_target_node.nameString();
+        assert init_call.typeArguments().size() == 0 : init_call.typeArguments();
+        if (init_call.name().toString().equals("autoBlocks")) {
+            assert context().autoBlocks()==null : "Already have autoBlocks: "+context().autoBlocks();
+            context().autoBlocks(s_);
+            context().established().autoBlocks(s_);
+        } else if (init_call.name().toString().equals("autoThreads")) {
+            assert context().autoThreads()==null : "Already have autoThreads: "+context().autoThreads();
+            context().autoThreads(s_);
+            context().established().autoThreads(s_);
+        } else {    
+            assert false : init_call.name();
+        }
+    }
 
     public void visit(Block_c b) {
         super.visit(b);
@@ -412,8 +436,12 @@
             assert !generatingKernel() : "Nesting of cuda annotation makes no sense.";
             // TODO: assert the block is the body of an async
 
-            assert b.statements().size() == 1; // FIXME: proper error
-            Stmt loop_ = b.statements().get(0);
+            assert b.statements().size()==1 || b.statements().size()==3 : b.statements();
+            if (b.statements().size()==3) {
+                checkAutoVar(b.statements().get(0));
+                checkAutoVar(b.statements().get(1));
+            }
+            Stmt loop_ = b.statements().get(b.statements().size()-1);
             // test that it is of the form for (blah in region)
             assert loop_ instanceof For : loop_.getClass(); // FIXME: proper
                                                             // error
@@ -546,31 +574,45 @@
     
             inc.write(make_ref(cnamet)+" __this = "+cnamet+"::"+DESERIALIZE_METHOD+"<"+cnamet+">(__buf);");
             inc.newline();
+            
+            ArrayList<VarInstance> env = context().kernelParams();
 
-            for (int i = 0; i < c.variables.size(); i++) {
-                VarInstance var = (VarInstance) c.variables.get(i);
+            for (VarInstance var : env) {
                 Type t = var.type();
                 String name = var.name().toString();
-                inc.write(Emitter.translateType(t, true)+" "+name+" = __this->"+name+";"); inc.newline();
+                inc.write(Emitter.translateType(t, true)+" "+name);
+                if (var == context().autoBlocks().localDef().asInstance()) {
+                    inc.write(";");
+                } else if (var == context().autoThreads().localDef().asInstance()) {
+                    inc.write(";");
+                } else {
+                    inc.write(" = __this->"+name+";");
+                }
+                inc.newline();
             }
             
-            inc.write("__blocks = ("); inc.begin(0);
-            tr.print(null, context().blocks(), inc);
-            inc.write(")+1;"); inc.end(); inc.newline();
-            
-            inc.write("__threads = ("); inc.begin(0);
-            tr.print(null, context().threads(), inc);
-            inc.write(")+1;"); inc.end(); inc.newline();
-            
             inc.write("__shm = "); inc.begin(0);
             context().shm().generateSize(inc, tr);
             inc.write(";"); inc.end(); inc.newline();
-    
-            generateStruct("__", inc, c.variables);
+            
+            if (context().autoBlocks()!=null && context().autoThreads()!=null) {
+                String bname = context().autoBlocks().name().id().toString();
+                String tname = context().autoThreads().name().id().toString();;
+                inc.write("x10aux::blocks_threads(__gpu, x10aux::DeserializationDispatcher::getMsgType(_serialization_id), __shm, "+bname+", "+tname+");"); inc.newline();
+            }// else {
+                inc.write("__blocks = ("); inc.begin(0);
+                tr.print(null, context().blocks(), inc);
+                inc.write(")+1;"); inc.end(); inc.newline();
+
+                inc.write("__threads = ("); inc.begin(0);
+                tr.print(null, context().threads(), inc);
+                inc.write(")+1;"); inc.end(); inc.newline();
+            //}
+            
+            generateStruct("__", inc, context().kernelParams());
             inc.write("___env __env;"); inc.newline();
                         
-            for (int i = 0; i < c.variables.size(); i++) {
-                VarInstance var = (VarInstance) c.variables.get(i);
+            for (VarInstance var : env) {
                 Type t = var.type();
                 String name = var.name().toString();
                 inc.write("__env."+name+" = ");
@@ -812,7 +854,15 @@
                 super.visit(n);
             }
         } else {
-            super.visit(n);
+            // we end up here in the _deserialize_cuda function because generatingKernel() is false
+            Name ln = n.name().id();
+            if (context().autoBlocks()!=null && ln == context().autoBlocks().name().id()) {
+                sw.write(context().autoBlocks().name().id().toString());
+            } else if (context().autoThreads()!=null && ln == context().autoThreads().name().id()) {
+                    sw.write(context().autoThreads().name().id().toString());
+            } else {
+                super.visit(n);
+            }
         }
     }
 

Modified: trunk/x10.dist/samples/KMeansCUDA.x10
===================================================================
--- trunk/x10.dist/samples/KMeansCUDA.x10	2009-11-18 18:13:39 UTC (rev 12124)
+++ trunk/x10.dist/samples/KMeansCUDA.x10	2009-11-18 18:25:26 UTC (rev 12125)
@@ -10,6 +10,7 @@
 import x10.util.Option;
 
 import x10.compiler.CUDA;
+import x10.compiler.CUDAUtilities;
 
 public class KMeansCUDA {
 
@@ -78,10 +79,10 @@
         
                         val clusters_copy = clusters as ValRail[Float];
 
-                        val kernel_start_time = System.currentTimeMillis();
                         // classify kernel
-                        val blocks = 4, threads = 256;
                         at (gpu) @CUDA {
+                            val blocks = CUDAUtilities.autoBlocks(),
+                                threads = CUDAUtilities.autoThreads();
                             for ((block) in 0..blocks-1) {
                                 val clustercache = Rail.make[Float](num_clusters*4, clusters_copy);
                                 for ((thread) in 0..threads-1) async {

Added: trunk/x10.runtime/src-x10/x10/compiler/CUDAUtilities.x10
===================================================================
--- trunk/x10.runtime/src-x10/x10/compiler/CUDAUtilities.x10	                        (rev 0)
+++ trunk/x10.runtime/src-x10/x10/compiler/CUDAUtilities.x10	2009-11-18 18:25:26 UTC (rev 12125)
@@ -0,0 +1,9 @@
+package x10.compiler;
+
+public class CUDAUtilities {
+    public static def autoBlocks() = 8;
+    public static def autoThreads() = 1;
+}
+
+// vim: shiftwidth=8:tabstop=8:expandtab
+


This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.




Thread: [X10-commits] SF.net SVN: x10:[12125] trunk (Page 10)

Performance and Productivity at Scale

x10-commits