Generalize CLKernel.setArg (velocity voodoo) + fix size of vector types of arity 3 (must be treated as size 4 as per section 6.1.5 of spec)

ochafik · ochafik · commit efaa6148a61b · 2015-03-22T22:38:52.000Z
diff --git a/Core/src/main/velocity/com/nativelibs4java/opencl/CLKernel.java b/Core/src/main/velocity/com/nativelibs4java/opencl/CLKernel.java
@@ -231,55 +231,37 @@ public void setArg(int i, Pointer<?> ptr) {
 		setKernelArg(i, ptr.getValidBytes(), ptr);
     }
     
+#foreach ($prim in $primitives)
     /**
+     * Sets a vector type value as the arg at index {@code i}.
+     *
+     * For instance to pass a {@code ${prim.Name}3}, one must pass a {@code ${prim.Name}[]} of size {@code 3}. 
+     *
 #documentCallsFunction("clSetKernelArg")
+     * @param i index of the argument in the kernel's parameters list
+     * @param arg array of ${prim.Name}s (of length 2, 3, 4, 8 or 16, which must match the arity of the OpenCL vector type used by the kernel).
      */
-    public void setArg(int i, float[] arg) {
-        setKernelArg(i, arg.length * 4, arg.length <= MAX_TMP_ITEMS ? localPointer.setFloats(arg) : pointerToFloats(arg));
+    public void setArg(int i, ${prim.Name}[] arg) {
+        int length = normalizeVectorTypeArity(arg.length);
+        setKernelArg(i, length * ${prim.Size}, length <= MAX_TMP_ITEMS ? localPointer.set${prim.CapName}s(arg) : pointerTo${prim.CapName}s(arg));
     }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, int[] arg) {
-        setKernelArg(i, arg.length * 4, arg.length <= MAX_TMP_ITEMS ? localPointer.setInts(arg) : pointerToInts(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, double[] arg) {
-        setKernelArg(i, arg.length * 8, arg.length <= MAX_TMP_ITEMS ? localPointer.setDoubles(arg) : pointerToDoubles(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, long[] arg) {
-        setKernelArg(i, arg.length * 8, arg.length <= MAX_TMP_ITEMS ? localPointer.setLongs(arg) : pointerToLongs(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, short[] arg) {
-        setKernelArg(i, arg.length * 2, arg.length <= MAX_TMP_ITEMS ? localPointer.setShorts(arg) : pointerToShorts(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, byte[] arg) {
-        setKernelArg(i, arg.length, arg.length <= MAX_TMP_ITEMS ? localPointer.setBytes(arg) : pointerToBytes(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, boolean[] arg) {
-        setKernelArg(i, arg.length, arg.length <= MAX_TMP_ITEMS ? localPointer.setBooleans(arg) : pointerToBooleans(arg));
-    }
-    /**
-#documentCallsFunction("clSetKernelArg")
-     */
-    public void setArg(int i, char[] arg) {
-        setKernelArg(i, arg.length * 2, arg.length <= MAX_TMP_ITEMS ? localPointer.setChars(arg) : pointerToChars(arg));
+#end
+
+    private static int normalizeVectorTypeArity(int length) {
+      switch (length) {
+        case 3:
+          // cl_<prim>3 is identical in size, alignment and behavior to cl_<prim>4. See section 6.1.5
+          return 4;
+        case 2:
+        case 4:
+        case 8:
+        case 16:
+          return length;
+        default:
+          throw new IllegalArgumentException("Invalid vector type length: " + length + " (valid vector type sizes are 2, 3, 4, 8, 16)");
+      }
     }
-    
+
     /**
 #documentCallsFunction("clSetKernelArg")
      */
diff --git a/Core/src/main/velocity/com/nativelibs4java/opencl/ReusablePointer.java b/Core/src/main/velocity/com/nativelibs4java/opencl/ReusablePointer.java
@@ -12,8 +12,14 @@ final class ReusablePointer {
 
     public ReusablePointer(long bytesCapacity) {
         this.bytesCapacity = bytesCapacity;
-        this.pointer = Pointer.allocateBytes(bytesCapacity).withoutValidityInformation();
+        this.pointer = allocateAlignedBytes(bytesCapacity).withoutValidityInformation();
     }
+
+    static Pointer<?> allocateAlignedBytes(int count) {
+      // Allocate memory aligned to 128 bytes to match alignment of cl_double16.
+      return Pointer.allocateAlignedBytes(null /* io */, count, 128 /* alignment */, null /* beforeDeallocation */);
+    }
+
     public Pointer<Integer> pointerToInts(int... values) {
         if (values == null)
             return null;
@@ -48,7 +54,7 @@ public <T> Pointer<T> allocatedBytes(int needed) {
         if (needed == 0)
             return null;
         if (needed > bytesCapacity) {
-            return (Pointer)Pointer.allocateBytes(needed);
+            return (Pointer)allocateAlignedBytes(needed);
         } else {
             return (Pointer)pointer;
         }
@@ -59,4 +65,4 @@ public <T> Pointer<T> allocatedSizeTs(int needed) {
     public <T> Pointer<T> allocatedInts(int needed) {
     	return allocatedBytes(needed * 4);
     }
-}
+}
diff --git a/Core/src/test/java/com/nativelibs4java/opencl/KernelTest.java b/Core/src/test/java/com/nativelibs4java/opencl/KernelTest.java
@@ -56,7 +56,7 @@ public <T> Object testArrayArg(String type, Object array, Class<T> targetType) {
 
     @Test
     public void nullArg() {
-        assertArrayEquals(new Pointer[] { null }, testArg("int", CLKernel.NULL_POINTER_KERNEL_ARGUMENT, Pointer.class).getPointers());
+        assertArrayEquals(new Pointer[] { null }, testArg("int*", CLKernel.NULL_POINTER_KERNEL_ARGUMENT, Pointer.class).getPointers());
     }
 
     byte[] byteTup(int n) {

Original file line number	Diff line number	Diff line change
`@@ -56,7 +56,7 @@ public <T> Object testArrayArg(String type, Object array, Class<T> targetType) {`
`56`	`56`
`57`	`57`	`@Test`
`58`	`58`	`public void nullArg() {`
`59`		`- assertArrayEquals(new Pointer[] { null }, testArg("int", CLKernel.NULL_POINTER_KERNEL_ARGUMENT, Pointer.class).getPointers());`
	`59`	`+ assertArrayEquals(new Pointer[] { null }, testArg("int*", CLKernel.NULL_POINTER_KERNEL_ARGUMENT, Pointer.class).getPointers());`
`60`	`60`	`}`
`61`	`61`
`62`	`62`	`byte[] byteTup(int n) {`