Add OpenCL 1.1 preferred work size query to QCLKernel

3 files changed, 25 insertions, 0 deletions

diff --git a/src/opencl/qclext_p.h b/src/opencl/qclext_p.h
index 68a17b7..a50df50 100644
--- a/src/opencl/qclext_p.h
+++ b/src/opencl/qclext_p.h
@@ -90,6 +90,9 @@
 #ifndef CL_MEM_OFFSET
 #define CL_MEM_OFFSET 0x1108
 #endif
+#ifndef CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE
+#define CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE 0x11B3
+#endif
 
 // OpenCL-OpenGL sharing.
 #ifndef CL_INVALID_CL_SHAREGROUP_REFERENCE_KHR
diff --git a/src/opencl/qclkernel.cpp b/src/opencl/qclkernel.cpp
index b63cc02..aceaa9f 100644
--- a/src/opencl/qclkernel.cpp
+++ b/src/opencl/qclkernel.cpp
@@ -43,6 +43,7 @@
 #include "qclprogram.h"
 #include "qclbuffer.h"
 #include "qclcontext.h"
+#include "qclext_p.h"
 #include <QtCore/qvarlengtharray.h>
 #include <QtCore/qpoint.h>
 #include <QtGui/qvector2d.h>
@@ -510,6 +511,25 @@ QCLWorkSize QCLKernel::bestLocalWorkSizeImage3D() const
 }
 
 /*!
+    Returns the preferred work group size multiple, which is a
+    performance hint for the local work group size on OpenCL 1.1
+    systems.  Returns zero if the system is OpenCL 1.0, or a
+    preferred work group size multiple is not available.
+*/
+size_t QCLKernel::preferredWorkSizeMultiple() const
+{
+    Q_D(const QCLKernel);
+    size_t size;
+    if (clGetKernelWorkGroupInfo
+            (d->id, d->context->defaultDevice().deviceId(),
+             CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
+             sizeof(size), &size, 0) != CL_SUCCESS)
+        return 0;
+    else
+        return size;
+}
+
+/*!
     Sets argument \a index for this kernel to \a value.
 
     The argument is assumed to have been declared with the
diff --git a/src/opencl/qclkernel.h b/src/opencl/qclkernel.h
index 9ad1d1e..f0721c0 100644
--- a/src/opencl/qclkernel.h
+++ b/src/opencl/qclkernel.h
@@ -109,6 +109,8 @@ public:
     QCLWorkSize bestLocalWorkSizeImage2D() const;
     QCLWorkSize bestLocalWorkSizeImage3D() const;
 
+    size_t preferredWorkSizeMultiple() const;
+
     void setArg(int index, cl_int value);
     void setArg(int index, cl_uint value);
     void setArg(int index, cl_long value);