@Test
public void writeCopyReadBufferTest() {
out.println(" - - - highLevelTest; copy buffer test - - - ");
final int elements = NUM_ELEMENTS;
final CLContext context = CLContext.create();
// the CL.MEM_* flag is probably completely irrelevant in our case since we do not use a kernel
// in this test
final CLBuffer<ByteBuffer> clBufferA =
context.createByteBuffer(elements * SIZEOF_INT, Mem.READ_ONLY);
final CLBuffer<ByteBuffer> clBufferB =
context.createByteBuffer(elements * SIZEOF_INT, Mem.READ_ONLY);
// fill only first read buffer -> we will copy the payload to the second later.
fillBuffer(clBufferA.buffer, 12345);
final CLCommandQueue queue = context.getDevices()[0].createCommandQueue();
// asynchronous write of data to GPU device, blocking read later to get the computed results
// back.
queue
.putWriteBuffer(clBufferA, false) // write A
.putCopyBuffer(clBufferA, clBufferB, clBufferA.buffer.capacity()) // copy A -> B
.putReadBuffer(clBufferB, true) // read B
.finish();
context.release();
out.println("validating computed results...");
checkIfEqual(clBufferA.buffer, clBufferB.buffer, elements);
out.println("results are valid");
}
@Test
public void copyLimitedSlicedBuffersTest() {
final int size = 4200 * SIZEOF_INT; // Arbitrary number that is a multiple of SIZEOF_INT;
final int padding = 307; // Totally arbitrary number > 0
final CLContext context = CLContext.create();
final CLCommandQueue queue = context.getDevices()[0].createCommandQueue();
// Make a buffer that is offset relative to the originally allocated position and has a limit
// that is
// not equal to the capacity to test whether all these attributes are correctly handled.
ByteBuffer hostBuffer = ByteBuffer.allocateDirect(size + padding);
hostBuffer.position(padding / 2); // Offset the original buffer
hostBuffer = hostBuffer.slice(); // Slice it to have a new buffer that starts at the offset
hostBuffer.limit(size);
hostBuffer.order(ByteOrder.nativeOrder()); // Necessary for comparisons to work later on.
fillBuffer(hostBuffer, 12345);
final CLBuffer<ByteBuffer> bufferA = context.createBuffer(size).cloneWith(hostBuffer);
final CLBuffer<ByteBuffer> bufferB = context.createByteBuffer(size);
queue
.putWriteBuffer(bufferA, false)
.putCopyBuffer(bufferA, bufferB, bufferA.getNIOSize())
.putReadBuffer(bufferB, true)
.finish();
hostBuffer.rewind();
bufferB.buffer.rewind();
checkIfEqual(hostBuffer, bufferB.buffer, size / SIZEOF_INT);
context.release();
}
@Test
public void mapBufferTest() {
out.println(" - - - highLevelTest; map buffer test - - - ");
final int elements = NUM_ELEMENTS;
final int sizeInBytes = elements * SIZEOF_INT;
CLContext context;
CLBuffer<?> clBufferA;
CLBuffer<?> clBufferB;
// We will have to allocate mappable NIO memory on non CPU contexts
// since we can't map e.g GPU memory.
if (CLPlatform.getDefault().listCLDevices(CLDevice.Type.CPU).length > 0) {
context = CLContext.create(CLDevice.Type.CPU);
clBufferA = context.createBuffer(sizeInBytes, Mem.READ_WRITE);
clBufferB = context.createBuffer(sizeInBytes, Mem.READ_WRITE);
} else {
context = CLContext.create();
clBufferA = context.createByteBuffer(sizeInBytes, Mem.READ_WRITE, Mem.USE_BUFFER);
clBufferB = context.createByteBuffer(sizeInBytes, Mem.READ_WRITE, Mem.USE_BUFFER);
}
final CLCommandQueue queue = context.getDevices()[0].createCommandQueue();
// fill only first buffer -> we will copy the payload to the second later.
final ByteBuffer mappedBufferA = queue.putMapBuffer(clBufferA, Map.WRITE, true);
assertEquals(sizeInBytes, mappedBufferA.capacity());
fillBuffer(mappedBufferA, 12345); // write to A
queue
.putUnmapMemory(clBufferA, mappedBufferA) // unmap A
.putCopyBuffer(clBufferA, clBufferB); // copy A -> B
// map B for read operations
final ByteBuffer mappedBufferB = queue.putMapBuffer(clBufferB, Map.READ, true);
assertEquals(sizeInBytes, mappedBufferB.capacity());
out.println("validating computed results...");
checkIfEqual(mappedBufferA, mappedBufferB, elements); // A == B ?
out.println("results are valid");
queue.putUnmapMemory(clBufferB, mappedBufferB); // unmap B
context.release();
}
@Test
public void bufferWithHostPointerTest() {
out.println(" - - - highLevelTest; host pointer test - - - ");
final int elements = NUM_ELEMENTS;
final CLContext context = CLContext.create();
final ByteBuffer buffer = Buffers.newDirectByteBuffer(elements * SIZEOF_INT);
// fill only first read buffer -> we will copy the payload to the second later.
fillBuffer(buffer, 12345);
final CLCommandQueue queue = context.getDevices()[0].createCommandQueue();
final Mem[] bufferConfig = new Mem[] {Mem.COPY_BUFFER, Mem.USE_BUFFER};
for (int i = 0; i < bufferConfig.length; i++) {
out.println("testing with " + bufferConfig[i] + " config");
final CLBuffer<ByteBuffer> clBufferA =
context.createBuffer(buffer, Mem.READ_ONLY, bufferConfig[i]);
final CLBuffer<ByteBuffer> clBufferB =
context.createByteBuffer(elements * SIZEOF_INT, Mem.READ_ONLY);
// asynchronous write of data to GPU device, blocking read later to get the computed results
// back.
queue
.putCopyBuffer(clBufferA, clBufferB, clBufferA.buffer.capacity()) // copy A -> B
.putReadBuffer(clBufferB, true) // read B
.finish();
assertEquals(2, context.getMemoryObjects().size());
clBufferA.release();
assertEquals(1, context.getMemoryObjects().size());
clBufferB.release();
assertEquals(0, context.getMemoryObjects().size());
// uploading worked when a==b.
out.println("validating computed results...");
checkIfEqual(clBufferA.buffer, clBufferB.buffer, elements);
out.println("results are valid");
}
context.release();
}
CLProgram(CLContext context, CLDevice... devices) {
super(null, true);
this.context = context;
this.devices = devices == null || devices.length == 0 ? context.getDevices() : devices;
}