MultiDeviceFractal demo can now use devices with different FP precision in parallel.

Added support for cl_amd_fp64.

3 files changed, 166 insertions, 116 deletions

diff --git a/nbproject/project.properties b/nbproject/project.properties
index 68fc9bf..62a1870 100644
--- a/nbproject/project.properties
+++ b/nbproject/project.properties
@@ -1,93 +1,98 @@
-

-#default values

-gluegen.root=${basedir}/../gluegen

-jogl.root=${basedir}/../jogl

-jocl.root=${basedir}/../jocl

-rootrel.build=build

-

-application.title=JOCL Demos

-application.vendor=jogamp

-build.classes.excludes=**/*.java,**/*.form

-# This directory is removed when the project is cleaned:

-build.dir=${rootrel.build}

-build.classes.dir=${build.dir}/classes

-build.generated.dir=${build.dir}/generated

-build.generated.sources.dir=${build.dir}/generated-sources

-# Only compile against the classpath explicitly listed here:

-build.sysclasspath=ignore

-build.test.classes.dir=${build.dir}/test/classes

-build.test.results.dir=${build.dir}/test/results

-# Uncomment to specify the preferred debugger connection transport:

-#debug.transport=dt_socket

-debug.classpath=\

-    ${run.classpath}

-debug.test.classpath=\

-    ${run.test.classpath}

-# This directory is removed when the project is cleaned:

-dist.dir=dist

-dist.jar=${dist.dir}/jocl-demos.jar

-dist.javadoc.dir=${dist.dir}/javadoc

-endorsed.classpath=

-excludes=

-includes=**

-jar.compress=false

-javac.classpath=\

-    ${reference.jocl.jar}:\

-    ${file.reference.jogl.all.jar}:\

-    ${file.reference.newt.all.jar}:\

-    ${file.reference.nativewindow.all.jar}:\

-    ${file.reference.gluegen-rt.jar}

-# Space-separated list of extra javac options

-javac.compilerargs=

-javac.deprecation=false

-javac.source=1.5

-javac.target=1.5

-javac.test.classpath=\

-    ${javac.classpath}:\

-    ${build.classes.dir}:\

-    ${libs.junit.classpath}:\

-    ${libs.junit_4.classpath}

-javadoc.additionalparam=

-javadoc.author=false

-javadoc.encoding=${source.encoding}

-javadoc.noindex=false

-javadoc.nonavbar=false

-javadoc.notree=false

-javadoc.private=false

-javadoc.splitindex=true

-javadoc.use=true

-javadoc.version=false

-javadoc.windowtitle=

-jaxbwiz.endorsed.dirs="${netbeans.home}/../ide12/modules/ext/jaxb/api"

-main.class=com.jogamp.opencl.demos.hellojocl.HelloJOCL

-manifest.file=manifest.mf

-meta.inf.dir=${src.dir}/META-INF

-no.dependencies=true

-platform.active=default_platform

-

-project.jocl=${jocl.root}

-reference.jocl.jar=${jocl.root}/dist/jocl.jar

-file.reference.gluegen-rt.jar=${gluegen.root}/${rootrel.build}/gluegen-rt.jar

-file.reference.jogl.all.jar=${jogl.root}/${rootrel.build}/jogl/jogl.all.jar

-file.reference.nativewindow.all.jar=${jogl.root}/${rootrel.build}/nativewindow/nativewindow.all.jar

-file.reference.newt.all.jar=${jogl.root}/${rootrel.build}/newt/newt.all.jar

-

-run.classpath=\

-    ${javac.classpath}:\

-    ${build.classes.dir}

-# Space-separated list of JVM arguments used when running the project

-# (you may also define separate properties like run-sys-prop.name=value instead of -Dname=value

-# or test-sys-prop.name=value to set system properties for unit tests):

-#-verbose:jni

-run.jvmargs=-Djava.library.path="\

-    ${jocl.root}/${rootrel.build}/natives/jocl${path.separator}\

-    ${jogl.root}/${rootrel.build}/jogl/obj${path.separator}\

-    ${jogl.root}/${rootrel.build}/nativewindow/obj${path.separator}\

-    ${jogl.root}/${rootrel.build}/newt/obj${path.separator}\

-    ${gluegen.root}/${rootrel.build}/obj"

-run.test.classpath=\

-    ${javac.test.classpath}:\

-    ${build.test.classes.dir}

-source.encoding=UTF-8

-src.dir=src

-test.src.dir=test

+
+annotation.processing.enabled=true
+annotation.processing.enabled.in.editor=false
+annotation.processing.run.all.processors=true
+#default values
+gluegen.root=${basedir}/../gluegen
+javac.processorpath=\
+    ${javac.classpath}
+jogl.root=${basedir}/../jogl
+jocl.root=${basedir}/../jocl
+rootrel.build=build
+
+application.title=JOCL Demos
+application.vendor=jogamp
+build.classes.excludes=**/*.java,**/*.form
+# This directory is removed when the project is cleaned:
+build.dir=${rootrel.build}
+build.classes.dir=${build.dir}/classes
+build.generated.dir=${build.dir}/generated
+build.generated.sources.dir=${build.dir}/generated-sources
+# Only compile against the classpath explicitly listed here:
+build.sysclasspath=ignore
+build.test.classes.dir=${build.dir}/test/classes
+build.test.results.dir=${build.dir}/test/results
+# Uncomment to specify the preferred debugger connection transport:
+#debug.transport=dt_socket
+debug.classpath=\
+    ${run.classpath}
+debug.test.classpath=\
+    ${run.test.classpath}
+# This directory is removed when the project is cleaned:
+dist.dir=dist
+dist.jar=${dist.dir}/jocl-demos.jar
+dist.javadoc.dir=${dist.dir}/javadoc
+endorsed.classpath=
+excludes=
+includes=**
+jar.compress=false
+javac.classpath=\
+    ${reference.jocl.jar}:\
+    ${file.reference.jogl.all.jar}:\
+    ${file.reference.newt.all.jar}:\
+    ${file.reference.nativewindow.all.jar}:\
+    ${file.reference.gluegen-rt.jar}
+# Space-separated list of extra javac options
+javac.compilerargs=
+javac.deprecation=false
+javac.source=1.5
+javac.target=1.5
+javac.test.classpath=\
+    ${javac.classpath}:\
+    ${build.classes.dir}:\
+    ${libs.junit.classpath}:\
+    ${libs.junit_4.classpath}
+javadoc.additionalparam=
+javadoc.author=false
+javadoc.encoding=${source.encoding}
+javadoc.noindex=false
+javadoc.nonavbar=false
+javadoc.notree=false
+javadoc.private=false
+javadoc.splitindex=true
+javadoc.use=true
+javadoc.version=false
+javadoc.windowtitle=
+jaxbwiz.endorsed.dirs="${netbeans.home}/../ide12/modules/ext/jaxb/api"
+main.class=com.jogamp.opencl.demos.hellojocl.HelloJOCL
+manifest.file=manifest.mf
+meta.inf.dir=${src.dir}/META-INF
+no.dependencies=true
+platform.active=default_platform
+
+project.jocl=${jocl.root}
+reference.jocl.jar=${jocl.root}/dist/jocl.jar
+file.reference.gluegen-rt.jar=${gluegen.root}/${rootrel.build}/gluegen-rt.jar
+file.reference.jogl.all.jar=${jogl.root}/${rootrel.build}/jogl/jogl.all.jar
+file.reference.nativewindow.all.jar=${jogl.root}/${rootrel.build}/nativewindow/nativewindow.all.jar
+file.reference.newt.all.jar=${jogl.root}/${rootrel.build}/newt/newt.all.jar
+
+run.classpath=\
+    ${javac.classpath}:\
+    ${build.classes.dir}
+# Space-separated list of JVM arguments used when running the project
+# (you may also define separate properties like run-sys-prop.name=value instead of -Dname=value
+# or test-sys-prop.name=value to set system properties for unit tests):
+#-verbose:jni
+run.jvmargs=-Xcheck:jni -Djava.library.path="\
+    ${jocl.root}/${rootrel.build}/natives/jocl${path.separator}\
+    ${jogl.root}/${rootrel.build}/jogl/obj${path.separator}\
+    ${jogl.root}/${rootrel.build}/nativewindow/obj${path.separator}\
+    ${jogl.root}/${rootrel.build}/newt/obj${path.separator}\
+    ${gluegen.root}/${rootrel.build}/obj"
+run.test.classpath=\
+    ${javac.test.classpath}:\
+    ${build.test.classes.dir}
+source.encoding=UTF-8
+src.dir=src
+test.src.dir=test
diff --git a/src/com/jogamp/opencl/demos/fractal/Mandelbrot.cl b/src/com/jogamp/opencl/demos/fractal/Mandelbrot.cl
index 640c775..2fc959c 100644
--- a/src/com/jogamp/opencl/demos/fractal/Mandelbrot.cl
+++ b/src/com/jogamp/opencl/demos/fractal/Mandelbrot.cl
@@ -1,5 +1,9 @@
 #ifdef DOUBLE_FP
-    #pragma OPENCL EXTENSION cl_khr_fp64 : enable
+    #ifdef AMD_FP
+        #pragma OPENCL EXTENSION cl_amd_fp64 : enable
+    #else
+        #pragma OPENCL EXTENSION cl_khr_fp64 : enable
+    #endif
     typedef double varfloat;
 #else
     typedef float varfloat;
diff --git a/src/com/jogamp/opencl/demos/fractal/MultiDeviceFractal.java b/src/com/jogamp/opencl/demos/fractal/MultiDeviceFractal.java
index 9420c6d..1b7f19d 100644
--- a/src/com/jogamp/opencl/demos/fractal/MultiDeviceFractal.java
+++ b/src/com/jogamp/opencl/demos/fractal/MultiDeviceFractal.java
@@ -12,6 +12,7 @@ import com.jogamp.opencl.CLKernel;
 import com.jogamp.opencl.CLPlatform;
 import com.jogamp.opencl.CLProgram;
 import com.jogamp.opencl.CLProgram.CompilerOptions;
+import com.jogamp.opencl.util.CLProgramConfiguration;
 import com.jogamp.opengl.util.awt.TextRenderer;
 import java.awt.Color;
 import java.awt.Dimension;
@@ -72,7 +73,7 @@ public class MultiDeviceFractal implements GLEventListener {
     private CLGLContext clContext;
     private CLCommandQueue[] queues;
     private CLKernel[] kernels;
-    private CLProgram program;
+    private CLProgram[] programs;
     private CLEventList probes;
     private CLGLBuffer<?>[] pboBuffers;
     private CLBuffer<IntBuffer>[] colorMap;
@@ -166,8 +167,31 @@ public class MultiDeviceFractal implements GLEventListener {
 
             }
 
-            // load and build program
-            program = clContext.createProgram(getClass().getResourceAsStream("Mandelbrot.cl"));
+            // check if we have 64bit FP support on all devices
+            // if yes we can use only one program for all devices + one kernel per device.
+            // if not we will have to create (at least) one program for 32 and one for 64bit devices.
+            // since there are different vendor extensions for double FP we use one program per device.
+            // (OpenCL spec is not very clear about this usecases)
+            boolean all64bit = true;
+            for (CLDevice device : devices) {
+                if(!isDoubleFPAvailable(device)) {
+                    all64bit = false;
+                    break;
+                }
+            }
+
+            // load program(s)
+            if(all64bit) {
+                programs = new CLProgram[] {
+                    clContext.createProgram(getClass().getResourceAsStream("Mandelbrot.cl"))
+                };
+            }else{
+                programs = new CLProgram[slices];
+                for (int i = 0; i < slices; i++) {
+                    programs[i] = clContext.createProgram(getClass().getResourceAsStream("Mandelbrot.cl"));
+                }
+            }
+
             buildProgram();
 
         } catch (IOException ex) {
@@ -258,33 +282,42 @@ public class MultiDeviceFractal implements GLEventListener {
          * workaround: The driver keeps using the old binaries for some reason.
          * to solve this we simple create a new program and release the old.
          * however rebuilding programs should be possible -> remove when drivers are fixed.
+         * (again: the spec is not very clear about this kind of usages)
          */
-        if(program != null && rebuild) {
-            String source = program.getSource();
-            program.release();
-            program = clContext.createProgram(source);
+        if(programs[0] != null && rebuild) {
+            for(int i = 0; i < programs.length; i++) {
+                String source = programs[i].getSource();
+                programs[i].release();
+                programs[i] = clContext.createProgram(source);
+            }
         }
 
         // disable 64bit floating point math if not available
-        if(doublePrecision) {
-            for (CLDevice device : program.getCLDevices()) {
-                if(!device.isDoubleFPAvailable()) {
-                    doublePrecision = false;
-                    break;
+        for(int i = 0; i < programs.length; i++) {
+            CLDevice device = queues[i].getDevice();
+
+            CLProgramConfiguration configure = programs[i].prepare();
+            if(doublePrecision && isDoubleFPAvailable(device)) {
+                //cl_khr_fp64
+                configure.withDefine("DOUBLE_FP");
+
+                //amd's verson of double precision floating point math
+                if(!device.isDoubleFPAvailable() && device.isExtensionAvailable("cl_amd_fp64")) {
+                    configure.withDefine("AMD_FP");
                 }
             }
-        }
+            if(programs.length > 1) {
+                configure.forDevice(device);
+            }
+            System.out.println(configure);
+            configure.withOption(CompilerOptions.FAST_RELAXED_MATH).build();
+         }
 
-        if(doublePrecision) {
-            program.build(CompilerOptions.FAST_RELAXED_MATH, "-D DOUBLE_FP");
-        }else{
-            program.build(CompilerOptions.FAST_RELAXED_MATH);
-        }
         rebuild = false;
 
         for (int i = 0; i < kernels.length; i++) {
             // init kernel with constants
-            kernels[i] = program.createCLKernel("mandelbrot");
+            kernels[i] = programs[min(i, programs.length)].createCLKernel("mandelbrot");
         }
 
     }
@@ -292,7 +325,7 @@ public class MultiDeviceFractal implements GLEventListener {
     // init kernels with constants
     private void setKernelConstants() {
         for (int i = 0; i < slices; i++) {
-            kernels[i].setForce32BitArgs(!doublePrecision)
+            kernels[i].setForce32BitArgs(!doublePrecision || !isDoubleFPAvailable(queues[i].getDevice()))
                       .setArg(6, pboBuffers[i])
                       .setArg(7, colorMap[i])
                       .setArg(8, colorMap[i].getBuffer().capacity())
@@ -374,14 +407,17 @@ public class MultiDeviceFractal implements GLEventListener {
         //draw info text
         textRenderer.beginRendering(width, height, false);
 
-            textRenderer.draw("precision: "+ (doublePrecision?"64bit":"32bit"), 10, height-15);
+            textRenderer.draw("device/time/precision", 10, height-15);
 
             for (int i = 0; i < slices; i++) {
                 CLDevice device = queues[i].getDevice();
+                boolean doubleFP = doublePrecision && isDoubleFPAvailable(device);
                 CLEvent event = probes.getEvent(i);
                 long start = event.getProfilingInfo(START);
                 long end = event.getProfilingInfo(END);
-                textRenderer.draw(device.getType().toString()+i +" "+(int)((end-start)/1000000.0f)+"ms", 10, height-(20+16*(slices-i)));
+                textRenderer.draw(device.getType().toString()+i +" "
+                               + (int)((end-start)/1000000.0f)+"ms @"
+                               + (doubleFP?"64bit":"32bit"), 10, height-(20+16*(slices-i)));
             }
 
         textRenderer.endRendering();
@@ -474,6 +510,11 @@ public class MultiDeviceFractal implements GLEventListener {
         canvas.addKeyListener(keyAdapter);
     }
 
+
+    private boolean isDoubleFPAvailable(CLDevice device) {
+        return device.isDoubleFPAvailable() || device.isExtensionAvailable("cl_amd_fp64");
+    }
+
     public void dispose(GLAutoDrawable drawable) {
     }