aboutsummaryrefslogtreecommitdiffstats
path: root/test/com/jogamp/opencl
diff options
context:
space:
mode:
Diffstat (limited to 'test/com/jogamp/opencl')
-rw-r--r--test/com/jogamp/opencl/CLBufferTest.java94
-rw-r--r--test/com/jogamp/opencl/CLCommandQueueTest.java6
-rw-r--r--test/com/jogamp/opencl/HighLevelBindingTest.java31
-rw-r--r--test/com/jogamp/opencl/LowLevelBindingTest.java33
4 files changed, 115 insertions, 49 deletions
diff --git a/test/com/jogamp/opencl/CLBufferTest.java b/test/com/jogamp/opencl/CLBufferTest.java
index 3ac1fffe..50820d4e 100644
--- a/test/com/jogamp/opencl/CLBufferTest.java
+++ b/test/com/jogamp/opencl/CLBufferTest.java
@@ -33,6 +33,7 @@ import com.jogamp.opencl.CLMemory.Map;
import com.jogamp.opencl.test.util.MiscUtils;
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;
@@ -256,7 +257,7 @@ public class CLBufferTest extends UITestCase {
}
@Test
- public void subBufferTest() {
+ public void subBufferTest01ByteBuffer() {
out.println(" - - - subBufferTest - - - ");
if(MiscUtils.isOpenCLUnavailable())
@@ -271,48 +272,77 @@ 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/CLCommandQueueTest.java b/test/com/jogamp/opencl/CLCommandQueueTest.java
index 76260289..491eab53 100644
--- a/test/com/jogamp/opencl/CLCommandQueueTest.java
+++ b/test/com/jogamp/opencl/CLCommandQueueTest.java
@@ -487,7 +487,9 @@ public class CLCommandQueueTest extends UITestCase {
@Override
public void run() {
- int groupSize = queue2.getDevice().getMaxWorkItemSizes()[0];
+ int maxWorkItemSize = queue2.getDevice().getMaxWorkItemSizes()[0];
+ int kernelWorkGroupSize = (int)vectorAddKernel2.getWorkGroupSize( queue2.getDevice() );
+ int localWorkSize = Math.min( maxWorkItemSize, kernelWorkGroupSize );
fillBuffer(clBufferA2.buffer, 12345);
fillBuffer(clBufferB2.buffer, 67890);
@@ -501,7 +503,7 @@ public class CLCommandQueueTest extends UITestCase {
// System.out.println("D kernels");
CLEventList events2 = new CLEventList(2);
- queue2.put1DRangeKernel(vectorAddKernel2, 0, elements, groupSize, events2);
+ queue2.put1DRangeKernel(vectorAddKernel2, 0, elements, localWorkSize, events2);
queue2.putReadBuffer(clBufferD, false, events2);
barrier.waitFor(queue2, events2);
diff --git a/test/com/jogamp/opencl/HighLevelBindingTest.java b/test/com/jogamp/opencl/HighLevelBindingTest.java
index 6a003435..a0a67595 100644
--- a/test/com/jogamp/opencl/HighLevelBindingTest.java
+++ b/test/com/jogamp/opencl/HighLevelBindingTest.java
@@ -167,6 +167,7 @@ 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");
@@ -323,13 +324,29 @@ public class HighLevelBindingTest extends UITestCase {
int elementCount = 11444777; // Length of float arrays to process (odd # for illustration)
int localWorkSize = device.getMaxWorkItemSizes()[0]; // set and log Global and Local work size dimensions
- int globalWorkSize = roundUp(localWorkSize, elementCount); // rounded up to the nearest multiple of the LocalWorkSize
-
- out.println("allocateing buffers of size: "+globalWorkSize);
-
- ByteBuffer srcA = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
- ByteBuffer srcB = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
- ByteBuffer dest = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ int globalWorkSize = 0;
+
+ ByteBuffer srcA = null;
+ ByteBuffer srcB = null;
+ ByteBuffer dest = null;
+ boolean allocated = false;
+ int divisor = 1;
+ while( !allocated ) {
+ try {
+ // round up to the nearest multiple of the LocalWorkSize
+ globalWorkSize = roundUp(localWorkSize, elementCount);
+ out.println("allocating three buffers of size: "+globalWorkSize);
+ srcA = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ srcB = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ dest = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ allocated = true;
+ }
+ catch( OutOfMemoryError oome ) {
+ ++divisor;
+ elementCount /= divisor;
+ out.println("not enough direct buffer memory; retrying with smaller buffers");
+ }
+ }
fillBuffer(srcA, 23456);
fillBuffer(srcB, 46987);
diff --git a/test/com/jogamp/opencl/LowLevelBindingTest.java b/test/com/jogamp/opencl/LowLevelBindingTest.java
index 5381cab0..90027e13 100644
--- a/test/com/jogamp/opencl/LowLevelBindingTest.java
+++ b/test/com/jogamp/opencl/LowLevelBindingTest.java
@@ -280,13 +280,30 @@ public class LowLevelBindingTest extends UITestCase {
checkError("on clCreateCommandQueue", intBuffer.get(0));
int localWorkSize = Math.min(128, maxWGS); // set and log Global and Local work size dimensions
- int globalWorkSize = roundUp(localWorkSize, ELEMENT_COUNT); // rounded up to the nearest multiple of the LocalWorkSize
-
- out.println("allocateing buffers of size: "+globalWorkSize);
-
- ByteBuffer srcA = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
- ByteBuffer srcB = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
- ByteBuffer dest = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ int elementCount = ELEMENT_COUNT;
+ int globalWorkSize = 0;
+
+ ByteBuffer srcA = null;
+ ByteBuffer srcB = null;
+ ByteBuffer dest = null;
+ boolean allocated = false;
+ int divisor = 1;
+ while( !allocated ) {
+ try {
+ // round up to the nearest multiple of the LocalWorkSize
+ globalWorkSize = roundUp(localWorkSize, elementCount);
+ out.println("allocating three buffers of size: "+globalWorkSize);
+ srcA = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ srcB = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ dest = newDirectByteBuffer(globalWorkSize*SIZEOF_INT);
+ allocated = true;
+ }
+ catch( OutOfMemoryError oome ) {
+ ++divisor;
+ elementCount /= divisor;
+ out.println("not enough direct buffer memory; retrying with smaller buffers");
+ }
+ }
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
long devSrcA = cl.clCreateBuffer(context, CL.CL_MEM_READ_ONLY, srcA.capacity(), null, intBuffer);
@@ -374,7 +391,7 @@ public class LowLevelBindingTest extends UITestCase {
ret = cl.clSetKernelArg(kernel, 0, is32Bit()?SIZEOF_INT:SIZEOF_LONG, wrap(devSrcA)); checkError("on clSetKernelArg0", ret);
ret = cl.clSetKernelArg(kernel, 1, is32Bit()?SIZEOF_INT:SIZEOF_LONG, wrap(devSrcB)); checkError("on clSetKernelArg1", ret);
ret = cl.clSetKernelArg(kernel, 2, is32Bit()?SIZEOF_INT:SIZEOF_LONG, wrap(devDst)); checkError("on clSetKernelArg2", ret);
- ret = cl.clSetKernelArg(kernel, 3, SIZEOF_INT, wrap(ELEMENT_COUNT)); checkError("on clSetKernelArg3", ret);
+ ret = cl.clSetKernelArg(kernel, 3, SIZEOF_INT, wrap(elementCount)); checkError("on clSetKernelArg3", ret);
out.println("used device memory: "+ (srcA.capacity()+srcB.capacity()+dest.capacity())/1000000 +"MB");
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-28 18:51:37 +0000