Bug 979: Add 'CL_DEVICE_MEM_BASE_ADDR_ALIGN' to CLDevice and overal maximum to CLContext - Split CLBufferTest and use alignment.

See Bug 979 for remaining issues and discussion.

6 files changed, 101 insertions, 36 deletions

diff --git a/src/com/jogamp/opencl/CLContext.java b/src/com/jogamp/opencl/CLContext.java
index 163375e..378f42f 100644
--- a/src/com/jogamp/opencl/CLContext.java
+++ b/src/com/jogamp/opencl/CLContext.java
@@ -628,6 +628,17 @@ public class CLContext extends CLObjectResource {
     }
 
     /**
+     * Returns the maximum {@link CLDevice#getMemBaseAddrAlign()} of all devices.
+     */
+    public long getMaxMemBaseAddrAlign() {
+        long maxAlignment = 0;
+        for (CLDevice device : getDevices()) {
+            maxAlignment = Math.max(maxAlignment, device.getMemBaseAddrAlign());
+        }
+        return maxAlignment;
+    }
+
+    /**
      * Returns all devices associated with this CLContext.
      */
     public CLDevice[] getDevices() {
diff --git a/src/com/jogamp/opencl/CLDevice.java b/src/com/jogamp/opencl/CLDevice.java
index 25fb009..860fa1a 100644
--- a/src/com/jogamp/opencl/CLDevice.java
+++ b/src/com/jogamp/opencl/CLDevice.java
@@ -48,7 +48,7 @@ import static com.jogamp.opencl.llb.CL.*;
  * @see CLPlatform#getMaxFlopsDevice(com.jogamp.opencl.CLDevice.Type...)
  * @see CLContext#getDevices()
  * @see CLContext#getMaxFlopsDevice(com.jogamp.opencl.CLDevice.Type)
- * @author Michael Bien
+ * @author Michael Bien, et al.
  */
 public class CLDevice extends CLObject {
 
@@ -356,7 +356,16 @@ public class CLDevice extends CLObject {
      */
     @CLProperty("CL_DEVICE_MAX_MEM_ALLOC_SIZE")
     public long getMaxMemAllocSize() {
-        return deviceInfo.getLong(CL_DEVICE_MAX_MEM_ALLOC_SIZE);
+        return deviceInfo.getLong(CL.CL_DEVICE_MAX_MEM_ALLOC_SIZE);
+    }
+
+    /**
+     * Returns the <code>uint32_t</code> memory base address alignment
+     * value reinterpreted as a <code>long</code> value.
+     */
+    @CLProperty("CL_DEVICE_MEM_BASE_ADDR_ALIGN")
+    public long getMemBaseAddrAlign() {
+        return deviceInfo.getUInt32Long(CL.CL_DEVICE_MEM_BASE_ADDR_ALIGN);
     }
 
     /**
diff --git a/src/com/jogamp/opencl/impl/CLTLInfoAccessor.java b/src/com/jogamp/opencl/impl/CLTLInfoAccessor.java
index c31b22a..2811530 100644
--- a/src/com/jogamp/opencl/impl/CLTLInfoAccessor.java
+++ b/src/com/jogamp/opencl/impl/CLTLInfoAccessor.java
@@ -41,7 +41,7 @@ import static com.jogamp.opencl.CLException.*;
 /**
  * Internal utility for common OpenCL clGetFooInfo calls.
  * Threadsafe, threadlocal implementation.
- * @author Michael Bien
+ * @author Michael Bien, et al.
  */
 public abstract class CLTLInfoAccessor implements CLInfoAccessor {
 
@@ -65,6 +65,14 @@ public abstract class CLTLInfoAccessor implements CLInfoAccessor {
     };
 
     @Override
+    public final long getUInt32Long(int key) {
+        final ByteBuffer buffer = getBB(4).putInt(0, 0);
+        final int ret = getInfo(key, 4, buffer, null);
+        CLException.checkForError(ret, "error while asking for info value");
+        return Bitstream.toUInt32Long(buffer.getInt(0));
+    }
+
+    @Override
     public final long getLong(int key) {
 
         ByteBuffer buffer = getBB(8).putLong(0, 0);
diff --git a/src/com/jogamp/opencl/spi/CLInfoAccessor.java b/src/com/jogamp/opencl/spi/CLInfoAccessor.java
index 0ff0aea..a02a38e 100644
--- a/src/com/jogamp/opencl/spi/CLInfoAccessor.java
+++ b/src/com/jogamp/opencl/spi/CLInfoAccessor.java
@@ -6,13 +6,19 @@ package com.jogamp.opencl.spi;
 /**
  * Internal utility for common OpenCL clGetFooInfo calls.
  * Provides common accessors to CL objects.
- * @author Michael Bien
+ * @author Michael Bien, et al.
  */
 public interface CLInfoAccessor {
 
     int[] getInts(int key, int n);
 
     /**
+     * Returns the <code>uint32_t</code> value for the given key,
+     * reinterpreted as a <code>long</code> value.
+     */
+    long getUInt32Long(int key);
+
+    /**
      * Returns the long value for the given key.
      */
     long getLong(int key);
diff --git a/test/com/jogamp/opencl/CLBufferTest.java b/test/com/jogamp/opencl/CLBufferTest.java
index 0638844..5b5d0e3 100644
--- a/test/com/jogamp/opencl/CLBufferTest.java
+++ b/test/com/jogamp/opencl/CLBufferTest.java
@@ -32,6 +32,7 @@ import com.jogamp.opencl.CLMemory.Mem;
 import com.jogamp.opencl.CLMemory.Map;
 import com.jogamp.opencl.test.util.UITestCase;
 import com.jogamp.common.nio.Buffers;
+import com.jogamp.common.util.Bitstream;
 
 import java.io.IOException;
 import java.nio.Buffer;
@@ -247,7 +248,7 @@ public class CLBufferTest extends UITestCase {
     }
 
     @Test
-    public void subBufferTest() {
+    public void subBufferTest01ByteBuffer() {
 
         out.println(" - - - subBufferTest - - - ");
 
@@ -260,48 +261,77 @@ public class CLBufferTest extends UITestCase {
         CLContext context = CLContext.create(platform);
         try{
             final int subelements = 5;
+            final long lMaxAlignment = context.getMaxMemBaseAddrAlign();
+            final int iMaxAlignment = Bitstream.uint32LongToInt(lMaxAlignment);
+            System.err.println("XXX: maxAlignment "+lMaxAlignment+", 0x"+Long.toHexString(lMaxAlignment)+", (int)"+iMaxAlignment+", (int)0x"+Integer.toHexString(iMaxAlignment));
+            if( -1 == iMaxAlignment ) {
+                throw new RuntimeException("Cannot handle MaxMemBaseAddrAlign > MAX_INT, has 0x"+Long.toHexString(lMaxAlignment));
+            }
             // device only
-            {
-                CLBuffer<?> buffer = context.createBuffer(64);
+            CLBuffer<?> buffer = context.createBuffer(iMaxAlignment+subelements);
 
-                assertFalse(buffer.isSubBuffer());
-                assertNotNull(buffer.getSubBuffers());
-                assertTrue(buffer.getSubBuffers().isEmpty());
+            assertFalse(buffer.isSubBuffer());
+            assertNotNull(buffer.getSubBuffers());
+            assertTrue(buffer.getSubBuffers().isEmpty());
 
-                CLSubBuffer<?> subBuffer = buffer.createSubBuffer(10, subelements);
+            CLSubBuffer<?> subBuffer = buffer.createSubBuffer(iMaxAlignment, subelements);
 
-                assertTrue(subBuffer.isSubBuffer());
-                assertEquals(subelements, subBuffer.getCLSize());
-                assertEquals(10, subBuffer.getOffset());
-                assertEquals(10, subBuffer.getCLOffset());
-                assertEquals(buffer, subBuffer.getParent());
-                assertEquals(1, buffer.getSubBuffers().size());
+            assertTrue(subBuffer.isSubBuffer());
+            assertEquals(subelements, subBuffer.getCLSize());
+            assertEquals(iMaxAlignment, subBuffer.getOffset());
+            assertEquals(iMaxAlignment, subBuffer.getCLOffset());
+            assertEquals(buffer, subBuffer.getParent());
+            assertEquals(1, buffer.getSubBuffers().size());
 
-                subBuffer.release();
-                assertEquals(0, buffer.getSubBuffers().size());
-            }
+            subBuffer.release();
+            assertEquals(0, buffer.getSubBuffers().size());
+        }finally{
+            context.release();
+        }
 
-            // device + direct buffer
-            {
-                CLBuffer<FloatBuffer> buffer = context.createFloatBuffer(64);
-                assertFalse(buffer.isSubBuffer());
-                assertNotNull(buffer.getSubBuffers());
-                assertTrue(buffer.getSubBuffers().isEmpty());
+    }
 
-                CLSubBuffer<FloatBuffer> subBuffer = buffer.createSubBuffer(10, subelements);
+    @Test
+    public void subBufferTest02FloatBuffer() {
 
-                assertTrue(subBuffer.isSubBuffer());
-                assertEquals(subelements, subBuffer.getBuffer().capacity());
-                assertEquals(10, subBuffer.getOffset());
-                assertEquals(40, subBuffer.getCLOffset());
-                assertEquals(buffer, subBuffer.getParent());
-                assertEquals(1, buffer.getSubBuffers().size());
+        out.println(" - - - subBufferTest - - - ");
 
-                assertEquals(subBuffer.getCLCapacity(), subBuffer.getBuffer().capacity());
+        CLPlatform platform = CLPlatform.getDefault(version(CL_1_1));
+        if(platform == null) {
+            out.println("aborting subBufferTest");
+            return;
+        }
 
-                subBuffer.release();
-                assertEquals(0, buffer.getSubBuffers().size());
+        CLContext context = CLContext.create(platform);
+        try{
+            final int subelements = 5;
+            final long lMaxAlignment = context.getMaxMemBaseAddrAlign();
+            final int iMaxAlignment = Bitstream.uint32LongToInt(lMaxAlignment);
+            System.err.println("XXX: maxAlignment "+lMaxAlignment+", 0x"+Long.toHexString(lMaxAlignment)+", (int)"+iMaxAlignment+", (int)0x"+Integer.toHexString(iMaxAlignment));
+            if( -1 == iMaxAlignment ) {
+                throw new RuntimeException("Cannot handle MaxMemBaseAddrAlign > MAX_INT, has 0x"+Long.toHexString(lMaxAlignment));
             }
+            // FIXME: See Bug 979: Offset/Alignment via offset calculation per element-count is faulty!
+            final int floatsPerAlignment = iMaxAlignment / Buffers.SIZEOF_FLOAT;
+            // device + direct buffer
+            CLBuffer<FloatBuffer> buffer = context.createFloatBuffer(floatsPerAlignment+subelements);
+            assertFalse(buffer.isSubBuffer());
+            assertNotNull(buffer.getSubBuffers());
+            assertTrue(buffer.getSubBuffers().isEmpty());
+
+            CLSubBuffer<FloatBuffer> subBuffer = buffer.createSubBuffer(floatsPerAlignment, subelements);
+
+            assertTrue(subBuffer.isSubBuffer());
+            assertEquals(subelements, subBuffer.getBuffer().capacity());
+            assertEquals(floatsPerAlignment, subBuffer.getOffset());
+            assertEquals(iMaxAlignment, subBuffer.getCLOffset());
+            assertEquals(buffer, subBuffer.getParent());
+            assertEquals(1, buffer.getSubBuffers().size());
+
+            assertEquals(subBuffer.getCLCapacity(), subBuffer.getBuffer().capacity());
+
+            subBuffer.release();
+            assertEquals(0, buffer.getSubBuffers().size());
 
         }finally{
             context.release();
diff --git a/test/com/jogamp/opencl/HighLevelBindingTest.java b/test/com/jogamp/opencl/HighLevelBindingTest.java
index 6dd6a73..40d2b9e 100644
--- a/test/com/jogamp/opencl/HighLevelBindingTest.java
+++ b/test/com/jogamp/opencl/HighLevelBindingTest.java
@@ -164,6 +164,7 @@ public class HighLevelBindingTest extends UITestCase {
 //                out.println("    C version: "+device.getCVersion()); //CL 1.1
                 out.println("    driver version: "+device.getDriverVersion());
                 out.println("    type: "+device.getType());
+                out.println("    mem base addr align: "+device.getMemBaseAddrAlign());
                 out.println("    global mem: "+device.getGlobalMemSize()/(1024*1024)+" MB");
                 out.println("    max alloc mem: "+device.getMaxMemAllocSize()/(1024*1024)+" MB");
                 out.println("    max param size: "+device.getMaxParameterSize()+" byte");