4 files changed, 76 insertions, 1 deletions

diff --git a/src/opencl/qclkernel.cpp b/src/opencl/qclkernel.cpp
index 58a0db8..e23d820 100644
--- a/src/opencl/qclkernel.cpp
+++ b/src/opencl/qclkernel.cpp
@@ -418,7 +418,7 @@ QCLWorkSize QCLKernel::globalWorkSize() const
 /*!
     Sets the global work size for this instance of the kernel to \a size.
 
-    \sa globalWorkSize(), setLocalWorkSize()
+    \sa globalWorkSize(), setLocalWorkSize(), setRoundedGlobalWorkSize()
 */
 void QCLKernel::setGlobalWorkSize(const QCLWorkSize &size)
 {
@@ -443,6 +443,33 @@ void QCLKernel::setGlobalWorkSize(const QCLWorkSize &size)
 */
 
 /*!
+    \fn void QCLKernel::setRoundedGlobalWorkSize(const QCLWorkSize &size)
+
+    Sets the global work size for this instance of the kernel to \a size,
+    after rounding it up to the next multiple of localWorkSize().
+
+    \sa globalWorkSize(), QCLWorkSize::roundTo()
+*/
+
+/*!
+    \fn void QCLKernel::setRoundedGlobalWorkSize(size_t width, size_t height)
+    \overload
+
+    Sets the global work size for this instance of the kernel to
+    \a width x \a height, after rounding it up to the next multiple
+    of localWorkSize().
+*/
+
+/*!
+    \fn void QCLKernel::setRoundedGlobalWorkSize(size_t width, size_t height, size_t depth)
+    \overload
+
+    Sets the global work size for this instance of the kernel to
+    \a width x \a height x \a depth, after rounding it up to the
+    next multiple of localWorkSize().
+*/
+
+/*!
     \fn void QCLKernel::setLocalWorkSize(size_t width, size_t height)
     \overload
 
diff --git a/src/opencl/qclkernel.h b/src/opencl/qclkernel.h
index b03676b..e886aa6 100644
--- a/src/opencl/qclkernel.h
+++ b/src/opencl/qclkernel.h
@@ -101,6 +101,10 @@ public:
     void setGlobalWorkSize(size_t width, size_t height);
     void setGlobalWorkSize(size_t width, size_t height, size_t depth);
 
+    void setRoundedGlobalWorkSize(const QCLWorkSize &size);
+    void setRoundedGlobalWorkSize(size_t width, size_t height);
+    void setRoundedGlobalWorkSize(size_t width, size_t height, size_t depth);
+
     QCLWorkSize localWorkSize() const;
     void setLocalWorkSize(const QCLWorkSize &size);
     void setLocalWorkSize(size_t width, size_t height);
@@ -298,6 +302,21 @@ inline void QCLKernel::setGlobalWorkSize(size_t width, size_t height, size_t dep
     setGlobalWorkSize(QCLWorkSize(width, height, depth));
 }
 
+inline void QCLKernel::setRoundedGlobalWorkSize(const QCLWorkSize &size)
+{
+    setGlobalWorkSize(size.roundTo(localWorkSize()));
+}
+
+inline void QCLKernel::setRoundedGlobalWorkSize(size_t width, size_t height)
+{
+    setRoundedGlobalWorkSize(QCLWorkSize(width, height));
+}
+
+inline void QCLKernel::setRoundedGlobalWorkSize(size_t width, size_t height, size_t depth)
+{
+    setRoundedGlobalWorkSize(QCLWorkSize(width, height, depth));
+}
+
 inline void QCLKernel::setLocalWorkSize(size_t width, size_t height)
 {
     setLocalWorkSize(QCLWorkSize(width, height));
diff --git a/src/opencl/qclworksize.cpp b/src/opencl/qclworksize.cpp
index 840dc39..eb47213 100644
--- a/src/opencl/qclworksize.cpp
+++ b/src/opencl/qclworksize.cpp
@@ -222,6 +222,33 @@ QCLWorkSize QCLWorkSize::toLocalWorkSize(const QCLDevice &device) const
                            device.maximumWorkItemsPerGroup());
 }
 
+static inline size_t qt_cl_round_to(size_t value, size_t multiple)
+{
+    if (multiple <= 1)
+        return value;
+    size_t remainder = value % multiple;
+    if (!remainder)
+        return value;
+    else
+        return value + multiple - remainder;
+}
+
+/*!
+    Returns the result of rounding this work size up to a multiple of \a size.
+*/
+QCLWorkSize QCLWorkSize::roundTo(const QCLWorkSize &size) const
+{
+    if (m_dim == 1)
+        return QCLWorkSize(qt_cl_round_to(m_sizes[0], size.m_sizes[0]));
+    else if (m_dim == 2)
+        return QCLWorkSize(qt_cl_round_to(m_sizes[0], size.m_sizes[0]),
+                           qt_cl_round_to(m_sizes[1], size.m_sizes[1]));
+    else
+        return QCLWorkSize(qt_cl_round_to(m_sizes[0], size.m_sizes[0]),
+                           qt_cl_round_to(m_sizes[1], size.m_sizes[1]),
+                           qt_cl_round_to(m_sizes[2], size.m_sizes[2]));
+}
+
 /*!
     Returns the string form of this work size, with components
     separated by 'x'.
diff --git a/src/opencl/qclworksize.h b/src/opencl/qclworksize.h
index 290197c..d13bfd5 100644
--- a/src/opencl/qclworksize.h
+++ b/src/opencl/qclworksize.h
@@ -82,6 +82,8 @@ public:
         (const QCLWorkSize &maxWorkItemSize, size_t maxItemsPerGroup) const;
     QCLWorkSize toLocalWorkSize(const QCLDevice &device) const;
 
+    QCLWorkSize roundTo(const QCLWorkSize &size) const;
+
     QString toString() const;
     static QCLWorkSize fromString(const QString &str);