added support for cl_khr_command_buffer (#9)

* added support for cl_khr_command_buffer
implements the provisional spec v0.9.0

* update to the latest version of cl_khr_command_buffer

* fix infinite recursion problem

* fix explicit template specialization cannot have a storage class

Author: bashbaug

scripts/call_all.c.mako

@@ -15,6 +15,7 @@ skipExtensions = {
 defaultValueForType = {
     # Handle Types
     'cl_accelerator_intel'              : 'NULL',
+    'cl_command_buffer_khr'             : 'NULL',
     'cl_command_queue'                  : 'NULL',
     'cl_context'                        : 'NULL',
     'cl_device_id'                      : 'NULL',
@@ -38,6 +39,7 @@ defaultValueForType = {
     'cl_accelerator_info_intel'         : 'CL_ACCELERATOR_DESCRIPTOR_INTEL',
     'cl_accelerator_type_intel'         : '0',
     'cl_bool'                           : 'CL_FALSE',
+    'cl_command_buffer_info_khr'        : 'CL_COMMAND_BUFFER_REFERENCE_COUNT_KHR',
     'cl_d3d10_device_source_khr'        : 'CL_D3D10_DEVICE_KHR',
     'cl_d3d10_device_set_khr'           : 'CL_ALL_DEVICES_FOR_D3D10_KHR',
     'cl_d3d11_device_source_khr'        : 'CL_D3D11_DEVICE_KHR',

scripts/openclext.cpp.mako

@@ -273,42 +273,6 @@ static inline cl_platform_id _get_platform(cl_mem memobj)
     return _get_platform(context);
 }
 
-#if defined(cl_khr_semaphore)
-
-static inline cl_platform_id _get_platform(cl_semaphore_khr semaphore)
-{
-    if (semaphore == NULL) return NULL;
-
-    cl_context context = NULL;
-    clGetSemaphoreInfoKHR(
-        semaphore,
-        CL_SEMAPHORE_CONTEXT_KHR,
-        sizeof(context),
-        &context,
-        NULL);
-    return _get_platform(context);
-}
-
-#endif // defined(cl_khr_semaphore)
-
-#if defined(cl_intel_accelerator)
-
-static inline cl_platform_id _get_platform(cl_accelerator_intel accelerator)
-{
-    if (accelerator == NULL) return NULL;
-
-    cl_context context = NULL;
-    clGetAcceleratorInfoINTEL(
-        accelerator,
-        CL_ACCELERATOR_CONTEXT_INTEL,
-        sizeof(context),
-        &context,
-        NULL);
-    return _get_platform(context);
-}
-
-#endif // defined(cl_intel_accelerator)
-
 /***************************************************************
 * Function Pointer Typedefs
 ***************************************************************/
@@ -437,20 +401,54 @@ static void _init(cl_platform_id platform, openclext_dispatch_table* dispatch_pt
 }
 
 #if defined(CLEXT_SINGLE_PLATFORM_ONLY)
+
 static openclext_dispatch_table _dispatch = {0};
 static openclext_dispatch_table* _dispatch_ptr = NULL;
 
 template<typename T>
 static inline openclext_dispatch_table* _get_dispatch(T object)
 {
+    if (object == NULL) return NULL;
+
     if (_dispatch_ptr == NULL) {
         cl_platform_id platform = _get_platform(object);
         _init(platform, &_dispatch);
         _dispatch_ptr = &_dispatch;
     }
+
+    return _dispatch_ptr;
+}
+
+// For some extension objects we cannot reliably query a platform ID without
+// infinitely recursing.  For these objects we cannot initialize the dispatch
+// table if it is not already initialized.
+
+#if defined(cl_khr_semaphore)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_semaphore_khr>(cl_semaphore_khr)
+{
+    return _dispatch_ptr;
+}
+#endif // defined(cl_khr_semaphore)
+
+#if defined(cl_khr_command_buffer)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_command_buffer_khr>(cl_command_buffer_khr)
+{
     return _dispatch_ptr;
 }
+#endif // defined(cl_khr_command_buffer)
+
+#if defined(cl_intel_accelerator)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_accelerator_intel>(cl_accelerator_intel)
+{
+    return _dispatch_ptr;
+}
+#endif // defined(cl_intel_accelerator)
+
 #else // defined(CLEXT_SINGLE_PLATFORM_ONLY)
+
 static size_t _num_platforms = 0;
 static openclext_dispatch_table* _dispatch_array = NULL;
 
@@ -460,6 +458,9 @@ static openclext_dispatch_table* _get_dispatch(T object)
     if (_num_platforms == 0 && _dispatch_array == NULL) {
         cl_uint numPlatforms = 0;
         clGetPlatformIDs(0, NULL, &numPlatforms);
+        if (numPlatforms == 0) {
+            return NULL;
+        }
 
         openclext_dispatch_table* dispatch =
             (openclext_dispatch_table*)malloc(
@@ -490,6 +491,96 @@ static openclext_dispatch_table* _get_dispatch(T object)
 
     return NULL;
 }
+
+// For some extension objects we cannot reliably query a platform ID without
+// infinitely recursing.  For these objects we cannot initialize the dispatch
+// table if it is not already initialized, and we need to use other methods
+// to find the right dispatch table.
+
+#if defined(cl_khr_semaphore)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_semaphore_khr>(cl_semaphore_khr semaphore)
+{
+    if (semaphore == NULL) return NULL;
+    if (_num_platforms <= 1) return _dispatch_array;
+
+    for (size_t i = 0; i < _num_platforms; i++) {
+        openclext_dispatch_table* dispatch_ptr =
+            _dispatch_array + i;
+        if (dispatch_ptr->clGetSemaphoreInfoKHR) {
+            cl_uint refCount = 0;
+            cl_int errorCode = dispatch_ptr->clGetSemaphoreInfoKHR(
+                semaphore,
+                CL_SEMAPHORE_REFERENCE_COUNT_KHR,
+                sizeof(refCount),
+                &refCount,
+                NULL);
+            if (errorCode == CL_SUCCESS) {
+                return dispatch_ptr;
+            }
+        }
+    }
+
+    return NULL;
+}
+#endif // defined(cl_khr_semaphore)
+
+#if defined(cl_khr_command_buffer)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_command_buffer_khr>(cl_command_buffer_khr cmdbuf)
+{
+    if (cmdbuf == NULL) return NULL;
+    if (_num_platforms <= 1) return _dispatch_array;
+
+    for (size_t i = 0; i < _num_platforms; i++) {
+        openclext_dispatch_table* dispatch_ptr =
+            _dispatch_array + i;
+        if (dispatch_ptr->clGetCommandBufferInfoKHR) {
+            cl_uint refCount = 0;
+            cl_int errorCode = dispatch_ptr->clGetCommandBufferInfoKHR(
+                cmdbuf,
+                CL_COMMAND_BUFFER_REFERENCE_COUNT_KHR,
+                sizeof(refCount),
+                &refCount,
+                NULL);
+            if (errorCode == CL_SUCCESS) {
+                return dispatch_ptr;
+            }
+        }
+    }
+
+    return NULL;
+}
+#endif // defined(cl_khr_command_buffer)
+
+#if defined(cl_intel_accelerator)
+template<>
+inline openclext_dispatch_table* _get_dispatch<cl_accelerator_intel>(cl_accelerator_intel accelerator)
+{
+    if (accelerator == NULL) return NULL;
+    if (_num_platforms <= 1) return _dispatch_array;
+
+    for (size_t i = 0; i < _num_platforms; i++) {
+        openclext_dispatch_table* dispatch_ptr =
+            _dispatch_array + i;
+        if (dispatch_ptr->clGetAcceleratorInfoINTEL) {
+            cl_uint refCount = 0;
+            cl_int errorCode = dispatch_ptr->clGetAcceleratorInfoINTEL(
+                accelerator,
+                CL_ACCELERATOR_REFERENCE_COUNT_INTEL,
+                sizeof(refCount),
+                &refCount,
+                NULL);
+            if (errorCode == CL_SUCCESS) {
+                return dispatch_ptr;
+            }
+        }
+    }
+
+    return NULL;
+}
+#endif // defined(cl_intel_accelerator)
+
 #endif // defined(CLEXT_SINGLE_PLATFORM_ONLY)
 
 #ifdef __cplusplus
@@ -525,7 +616,13 @@ ${api.RetType} CL_API_CALL ${api.Name}(
 %        endif
 %      endfor
 {
+%      if api.Name == "clCreateCommandBufferKHR":
+    struct openclext_dispatch_table* dispatch_ptr = _get_dispatch(${api.Params[0].Name} > 0 && ${api.Params[1].Name} ? ${api.Params[1].Name}[0] : NULL);
+%      elif api.Name == "clEnqueueCommandBufferKHR":
+    struct openclext_dispatch_table* dispatch_ptr = _get_dispatch(${api.Params[2].Name});
+%      else:
     struct openclext_dispatch_table* dispatch_ptr = _get_dispatch(${api.Params[0].Name});
+%      endif
     if (dispatch_ptr == NULL || dispatch_ptr->${api.Name} == NULL) {
 %      if api.RetType == "cl_int":
         return CL_INVALID_OPERATION;