Merge pull request #717 from jszuppe/pr_subgroup_info

Add get_sub_group_info() to kernel class

Author: jszuppe

.appveyor.yml

@@ -18,11 +18,19 @@ environment:
   global:
     INTELOCLSDKROOT: C:\Program Files (x86)\Intel\OpenCL SDK\
     BOOST_COMPUTE_DEFAULT_PLATFORM: Intel(R) OpenCL
+    OPENCL_INCLUDE_DIR: C:\opencl\include
+    # Downloaded OpenCL headers version (2.1)
+    OPENCL_HEADERS_VER: 21
+    # OpenCL version used in tests (2.0)
+    OPENCL_VERSION: 200
+    CXXFLAGS: -DBOOST_COMPUTE_MAX_CL_VERSION=%OPENCL_VERSION%
   matrix:
      - VS_VER: 2015
+       CMAKE_GENERATOR: Visual Studio 14 2015 Win64
        BOOST_ROOT: C:\Libraries\boost_1_59_0
        BOOST_LIBRARYDIR: C:\Libraries\boost_1_59_0\lib64-msvc-14.0
      - VS_VER: 2013
+       CMAKE_GENERATOR: Visual Studio 12 2013 Win64
        BOOST_ROOT: C:\Libraries\boost_1_58_0
        BOOST_LIBRARYDIR: C:\Libraries\boost_1_58_0\lib64-msvc-12.0
 
@@ -42,11 +50,11 @@ before_build:
   - appveyor DownloadFile "http://registrationcenter-download.intel.com/akdlm/irc_nas/9022/opencl_runtime_16.1.1_x64_setup.msi"
   - start /wait msiexec /i opencl_runtime_16.1.1_x64_setup.msi /qn  /l*v msiexec2.log
   #- type msiexec2.log
+  - git clone --depth 1 https://github.com/KhronosGroup/OpenCL-Headers.git -b opencl%OPENCL_HEADERS_VER% %OPENCL_INCLUDE_DIR%\CL
 
 build_script:
   - mkdir build && cd build
-  - if "%VS_VER%" == "2015" cmake -G"Visual Studio 14 2015 Win64" -DBOOST_COMPUTE_BUILD_TESTS=ON -DBOOST_COMPUTE_BUILD_EXAMPLES=ON ..
-  - if "%VS_VER%" == "2013" cmake -G"Visual Studio 12 2013 Win64" -DBOOST_COMPUTE_BUILD_TESTS=ON -DBOOST_COMPUTE_BUILD_EXAMPLES=ON ..
+  - cmake -G"%CMAKE_GENERATOR%" -DBOOST_COMPUTE_BUILD_TESTS=ON -DBOOST_COMPUTE_BUILD_EXAMPLES=ON -DOpenCL_INCLUDE_DIR=%OPENCL_INCLUDE_DIR%  ..
   - cmake --build . --config Debug
 
 test_script:

example/hello_world.cpp

@@ -21,8 +21,9 @@ int main()
     // get the default device
     compute::device device = compute::system::default_device();
 
-    // print the device's name
-    std::cout << "hello from " << device.name() << std::endl;
+    // print the device's name and platform
+    std::cout << "hello from " << device.name();
+    std::cout << " (platform: " << device.platform().name() << ")" << std::endl;
 
     return 0;
 }

include/boost/compute/cl_ext.hpp

@@ -11,6 +11,24 @@
 #ifndef BOOST_COMPUTE_CL_EXT_HPP
 #define BOOST_COMPUTE_CL_EXT_HPP
 
+#if defined(BOOST_COMPUTE_MAX_CL_VERSION)
+#  if !defined(CL_USE_DEPRECATED_OPENCL_2_1_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 202
+#    define CL_USE_DEPRECATED_OPENCL_2_1_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_2_0_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 201
+#    define CL_USE_DEPRECATED_OPENCL_2_0_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_2_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 200
+#    define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_1_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 102
+#    define CL_USE_DEPRECATED_OPENCL_1_1_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_0_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 101
+#    define CL_USE_DEPRECATED_OPENCL_1_0_APIS
+#  endif
+#endif
+
 #if defined(__APPLE__)
 #include <OpenCL/cl_ext.h>
 #else

include/boost/compute/detail/get_object_info.hpp

@@ -36,6 +36,16 @@ struct bound_info_function
     {
     }
 
+    template<class Info>
+    cl_int operator()(Info info, size_t input_size, const void *input,
+                      size_t size, void *value, size_t *size_ret) const
+    {
+        return m_function(
+            m_object, m_aux_info, info,
+            input_size, input, size, value, size_ret
+        );
+    }
+
     template<class Info>
     cl_int operator()(Info info, size_t size, void *value, size_t *size_ret) const
     {
@@ -96,6 +106,20 @@ struct get_object_info_impl
 
         return value;
     }
+
+    template<class Function, class Info>
+    T operator()(Function function, Info info,
+                 const size_t input_size, const void* input) const
+    {
+        T value;
+
+        cl_int ret = function(info, input_size, input, sizeof(T), &value, 0);
+        if(ret != CL_SUCCESS){
+            BOOST_THROW_EXCEPTION(opencl_error(ret));
+        }
+
+        return value;
+    }
 };
 
 // specialization for bool
@@ -167,6 +191,42 @@ struct get_object_info_impl<std::vector<T> >
 
         return vector;
     }
+
+    template<class Function, class Info>
+    std::vector<T> operator()(Function function, Info info,
+                              const size_t input_size, const void* input) const
+    {
+        #ifdef BOOST_COMPUTE_CL_VERSION_2_1
+        // For CL_KERNEL_LOCAL_SIZE_FOR_SUB_GROUP_COUNT in clGetKernelSubGroupInfo
+        // we can't get param_value_size using param_value_size_ret
+        if(info == CL_KERNEL_LOCAL_SIZE_FOR_SUB_GROUP_COUNT)
+        {
+            std::vector<T> vector(3);
+            cl_int ret = function(
+                info, input_size, input,
+                sizeof(T) * vector.size(), &vector[0], 0
+            );
+            if(ret != CL_SUCCESS){
+                BOOST_THROW_EXCEPTION(opencl_error(ret));
+            }
+            return vector;
+        }
+        #endif
+        size_t size = 0;
+
+        cl_int ret = function(info, input_size, input, 0, 0, &size);
+        if(ret != CL_SUCCESS){
+            BOOST_THROW_EXCEPTION(opencl_error(ret));
+        }
+
+        std::vector<T> vector(size / sizeof(T));
+        ret = function(info, input_size, input, size, &vector[0], 0);
+        if(ret != CL_SUCCESS){
+            BOOST_THROW_EXCEPTION(opencl_error(ret));
+        }
+
+        return vector;
+    }
 };
 
 // returns the value (of type T) from the given clGet*Info() function call.
@@ -182,6 +242,12 @@ inline T get_object_info(Function f, Object o, Info i, AuxInfo j)
     return get_object_info_impl<T>()(bind_info_function(f, o, j), i);
 }
 
+template<class T, class Function, class Object, class Info, class AuxInfo>
+inline T get_object_info(Function f, Object o, Info i, AuxInfo j, const size_t k, const void * l)
+{
+    return get_object_info_impl<T>()(bind_info_function(f, o, j), i, k, l);
+}
+
 // returns the value type for the clGet*Info() call on Object with Enum.
 template<class Object, int Enum>
 struct get_object_info_type;

include/boost/compute/interop/opengl/cl_gl_ext.hpp

@@ -11,6 +11,24 @@
 #ifndef BOOST_COMPUTE_INTEROP_OPENGL_CL_GL_EXT_HPP
 #define BOOST_COMPUTE_INTEROP_OPENGL_CL_GL_EXT_HPP
 
+#if defined(BOOST_COMPUTE_MAX_CL_VERSION)
+#  if !defined(CL_USE_DEPRECATED_OPENCL_2_1_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 202
+#    define CL_USE_DEPRECATED_OPENCL_2_1_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_2_0_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 201
+#    define CL_USE_DEPRECATED_OPENCL_2_0_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_2_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 200
+#    define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_1_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 102
+#    define CL_USE_DEPRECATED_OPENCL_1_1_APIS
+#  endif
+#  if !defined(CL_USE_DEPRECATED_OPENCL_1_0_APIS) && BOOST_COMPUTE_MAX_CL_VERSION < 101
+#    define CL_USE_DEPRECATED_OPENCL_1_0_APIS
+#  endif
+#endif
+
 #if defined(__APPLE__)
 #include <OpenCL/cl_gl_ext.h>
 #else

include/boost/compute/interop/opengl/context.hpp

@@ -76,7 +76,7 @@ inline context opengl_create_shared_context()
         const platform &platform = platforms[i];
 
         // check whether this platform supports OpenCL/OpenGL sharing
-        if (!platform.supports_extension(cl_gl_sharing_extension))
+        if (!platform.supports_extension("cl_gl_sharing_extension"))
           continue;
 
         // load clGetGLContextInfoKHR() extension function
@@ -98,7 +98,7 @@ inline context opengl_create_shared_context()
             CL_GLX_DISPLAY_KHR, (cl_context_properties) glXGetCurrentDisplay(),
         #elif defined(_WIN32)
             CL_GL_CONTEXT_KHR, (cl_context_properties) wglGetCurrentContext(),
-            CL_WGL_HDC_KHR, (cl_context_properties) wglGetCurrentDC(), 
+            CL_WGL_HDC_KHR, (cl_context_properties) wglGetCurrentDC(),
         #endif
             0
         };
@@ -118,7 +118,7 @@ inline context opengl_create_shared_context()
 
         // create device object for the GPU and ensure it supports CL-GL sharing
         device gpu(gpu_id, false);
-        if(!gpu.supports_extension(cl_gl_sharing_extension)){
+        if(!gpu.supports_extension("cl_gl_sharing_extension")){
             continue;
         }
 
@@ -129,7 +129,7 @@ inline context opengl_create_shared_context()
 
     // no CL-GL sharing capable devices found
     BOOST_THROW_EXCEPTION(
-        unsupported_extension_error(cl_gl_sharing_extension)
+        unsupported_extension_error("cl_gl_sharing_extension")
     );
 }
 

include/boost/compute/kernel.hpp

@@ -15,10 +15,14 @@
 
 #include <boost/assert.hpp>
 #include <boost/utility/enable_if.hpp>
+#include <boost/optional.hpp>
+
+#include <boost/compute/cl_ext.hpp> // cl_khr_subgroups
 
 #include <boost/compute/config.hpp>
-#include <boost/compute/program.hpp>
 #include <boost/compute/exception.hpp>
+#include <boost/compute/program.hpp>
+#include <boost/compute/platform.hpp>
 #include <boost/compute/type_traits/is_fundamental.hpp>
 #include <boost/compute/detail/get_object_info.hpp>
 #include <boost/compute/detail/assert_cl_success.hpp>
@@ -208,6 +212,101 @@ class kernel
         return detail::get_object_info<T>(clGetKernelWorkGroupInfo, m_kernel, info, device.id());
     }
 
+    #if defined(BOOST_COMPUTE_CL_VERSION_2_1) || defined(BOOST_COMPUTE_DOXYGEN_INVOKED)
+    /// Returns sub-group information for the kernel with \p device. Returns a null
+    /// optional if \p device is not 2.1 device, or is not 2.0 device with support
+    /// for cl_khr_subgroups extension.
+    ///
+    /// \opencl_version_warning{2,1}
+    /// \see_opencl_ref{clGetKernelSubGroupInfo}
+    template<class T>
+    boost::optional<T> get_sub_group_info(const device &device, cl_kernel_sub_group_info info,
+                                          const size_t input_size, const void * input) const
+    {
+        if(device.check_version(2, 1))
+        {
+            return detail::get_object_info<T>(
+                clGetKernelSubGroupInfo, m_kernel, info, device.id(), input_size, input
+            );
+        }
+        else if(!device.check_version(2, 0) || !device.supports_extension("cl_khr_subgroups"))
+        {
+            return boost::optional<T>();
+        }
+        // Only CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE and CL_KERNEL_SUB_GROUP_COUNT_FOR_NDRANGE
+        // are supported in cl_khr_subgroups extension for 2.0 devices.
+        else if(info != CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE && info != CL_KERNEL_SUB_GROUP_COUNT_FOR_NDRANGE)
+        {
+            return boost::optional<T>();
+        }
+
+        clGetKernelSubGroupInfoKHR_fn clGetKernelSubGroupInfoKHR_fptr =
+            reinterpret_cast<clGetKernelSubGroupInfoKHR_fn>(
+                reinterpret_cast<size_t>(
+                    device.platform().get_extension_function_address("clGetKernelSubGroupInfoKHR")
+                )
+            );
+
+        return detail::get_object_info<T>(
+            clGetKernelSubGroupInfoKHR_fptr, m_kernel, info, device.id(), input_size, input
+        );
+    }
+
+    /// \overload
+    template<class T>
+    boost::optional<T> get_sub_group_info(const device &device, cl_kernel_sub_group_info info) const
+    {
+        return get_sub_group_info<T>(device, info, 0, 0);
+    }
+
+    /// \overload
+    template<class T>
+    boost::optional<T> get_sub_group_info(const device &device, cl_kernel_sub_group_info info,
+                                          const size_t input) const
+    {
+        return get_sub_group_info<T>(device, info, sizeof(size_t), &input);
+    }
+    #endif // BOOST_COMPUTE_CL_VERSION_2_1
+
+    #if defined(BOOST_COMPUTE_CL_VERSION_2_0) && !defined(BOOST_COMPUTE_CL_VERSION_2_1)
+    /// Returns sub-group information for the kernel with \p device. Returns a null
+    /// optional if cl_khr_subgroups extension is not supported by \p device.
+    ///
+    /// \opencl_version_warning{2,0}
+    /// \see_opencl_ref{clGetKernelSubGroupInfoKHR}
+    template<class T>
+    boost::optional<T> get_sub_group_info(const device &device, cl_kernel_sub_group_info info,
+                                          const size_t input_size, const void * input) const
+    {
+        if(!device.check_version(2, 0) || !device.supports_extension("cl_khr_subgroups"))
+        {
+            return boost::optional<T>();
+        }
+
+        clGetKernelSubGroupInfoKHR_fn clGetKernelSubGroupInfoKHR_fptr =
+            reinterpret_cast<clGetKernelSubGroupInfoKHR_fn>(
+                reinterpret_cast<size_t>(
+                    device.platform().get_extension_function_address("clGetKernelSubGroupInfoKHR")
+                )
+            );
+
+        return detail::get_object_info<T>(
+            clGetKernelSubGroupInfoKHR_fptr, m_kernel, info, device.id(), input_size, input
+        );
+    }
+    #endif // defined(BOOST_COMPUTE_CL_VERSION_2_0) && !defined(BOOST_COMPUTE_CL_VERSION_2_1)
+
+    #if defined(BOOST_COMPUTE_CL_VERSION_2_0) || defined(BOOST_COMPUTE_DOXYGEN_INVOKED)
+    /// \overload
+    template<class T>
+    boost::optional<T> get_sub_group_info(const device &device, cl_kernel_sub_group_info info,
+                                          const std::vector<size_t> input) const
+    {
+        BOOST_ASSERT(input.size() > 0);
+        return get_sub_group_info<T>(device, info, input.size() * sizeof(size_t), &input[0]);
+    }
+    #endif // BOOST_COMPUTE_CL_VERSION_2_0
+
     /// Sets the argument at \p index to \p value with \p size.
     ///
     /// \see_opencl_ref{clSetKernelArg}

test/test_for_each.cpp

@@ -29,6 +29,7 @@ BOOST_AUTO_TEST_CASE(for_each_nop)
     BOOST_COMPUTE_FUNCTION(void, nop, (int ignored), {});
 
     bc::for_each(vector.begin(), vector.end(), nop, queue);
+    queue.finish();
 }
 
 BOOST_AUTO_TEST_CASE(for_each_n_nop)
@@ -39,6 +40,7 @@ BOOST_AUTO_TEST_CASE(for_each_n_nop)
     BOOST_COMPUTE_FUNCTION(void, nop, (int ignored), {});
 
     bc::for_each_n(vector.begin(), vector.size(), nop, queue);
+    queue.finish();
 }
 
 BOOST_AUTO_TEST_SUITE_END()