add tracing for cl_intel_concurrent_dispatch

Author: bashbaug

intercept/src/cli_ext.h

@@ -1250,6 +1250,25 @@ typedef struct _cl_queue_family_properties_intel {
 #define CL_QUEUE_CAPABILITY_BARRIER_INTEL                   (1 << 25)
 #define CL_QUEUE_CAPABILITY_KERNEL_INTEL                    (1 << 26)
 
+///////////////////////////////////////////////////////////////////////////////
+// cl_intel_concurrent_dispatch
+
+#define CL_KERNEL_EXEC_INFO_DISPATCH_TYPE_INTEL             0x4257
+
+typedef cl_uint             cl_kernel_exec_info_dispatch_type_intel;
+
+#define CL_KERNEL_EXEC_INFO_DISPATCH_TYPE_DEFAULT_INTEL     0
+#define CL_KERNEL_EXEC_INFO_DISPATCH_TYPE_CONCURRENT_INTEL  1
+
+extern CL_API_ENTRY
+cl_int CL_API_CALL clGetKernelMaxConcurrentWorkGroupCountINTEL(
+    cl_command_queue command_queue,
+    cl_kernel kernel,
+    cl_uint work_dim,
+    const size_t* global_work_offset,
+    const size_t* local_work_size,
+    size_t* max_work_group_count);
+
 ///////////////////////////////////////////////////////////////////////////////
 // cl_intel_create_buffer_with_properties
 

intercept/src/dispatch.cpp

@@ -9055,6 +9055,51 @@ CL_API_ENTRY cl_int CL_API_CALL clGetKernelSuggestedLocalWorkSizeINTEL(
     NULL_FUNCTION_POINTER_RETURN_ERROR(CL_INVALID_COMMAND_QUEUE);
 }
 
+///////////////////////////////////////////////////////////////////////////////
+//
+// cl_intel_concurrent_dispatch
+CL_API_ENTRY cl_int CL_API_CALL clGetKernelMaxConcurrentWorkGroupCountINTEL(
+    cl_command_queue commandQueue,
+    cl_kernel kernel,
+    cl_uint workDim,
+    const size_t *globalWorkOffset,
+    const size_t *localWorkSize,
+    size_t *maxWorkGroupCount)
+{
+    CLIntercept*    pIntercept = GetIntercept();
+
+    if( pIntercept )
+    {
+        const auto& dispatchX = pIntercept->dispatchX(commandQueue);
+        if( dispatchX.clGetKernelMaxConcurrentWorkGroupCountINTEL )
+        {
+            GET_ENQUEUE_COUNTER();
+            CALL_LOGGING_ENTER_KERNEL(
+                kernel,
+                "queue = %p, kernel = %p",
+                commandQueue,
+                kernel );
+            HOST_PERFORMANCE_TIMING_START();
+
+            cl_int retVal = dispatchX.clGetKernelMaxConcurrentWorkGroupCountINTEL(
+                commandQueue,
+                kernel,
+                workDim,
+                globalWorkOffset,
+                localWorkSize,
+                maxWorkGroupCount );
+
+            HOST_PERFORMANCE_TIMING_END();
+            CHECK_ERROR( retVal );
+            CALL_LOGGING_EXIT( retVal );
+
+            return retVal;
+        }
+    }
+
+    NULL_FUNCTION_POINTER_RETURN_ERROR(CL_INVALID_COMMAND_QUEUE);
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 // cl_intel_accelerator

intercept/src/dispatch.h

@@ -547,6 +547,15 @@ struct CLdispatchX
     cl_int (CL_API_CALL *clReleaseAcceleratorINTEL) (
         cl_accelerator_intel accelerator );
 
+    // cl_intel_concurrent_dispatch
+    cl_int (CL_API_CALL *clGetKernelMaxConcurrentWorkGroupCountINTEL) (
+        cl_command_queue command_queue,
+        cl_kernel kernel,
+        cl_uint work_dim,
+        const size_t* global_work_offset,
+        const size_t* local_work_size,
+        size_t* max_work_group_count);
+
     // cl_intel_create_buffer_with_properties
     cl_mem (CL_API_CALL *clCreateBufferWithPropertiesINTEL) (
         cl_context context,

intercept/src/enummap.cpp

@@ -1024,6 +1024,9 @@ CEnumNameMap::CEnumNameMap()
     ADD_ENUM_NAME( m_cl_command_queue_capabilities_intel, CL_QUEUE_CAPABILITY_BARRIER_INTEL );
     ADD_ENUM_NAME( m_cl_command_queue_capabilities_intel, CL_QUEUE_CAPABILITY_KERNEL_INTEL );
 
+    // cl_intel_concurrent_dispatch
+    ADD_ENUM_NAME( m_cl_int, CL_KERNEL_EXEC_INFO_DISPATCH_TYPE_INTEL );
+
     // cl_intel_device_attribute_query
     ADD_ENUM_NAME( m_cl_int, CL_DEVICE_IP_VERSION_INTEL );
     ADD_ENUM_NAME( m_cl_int, CL_DEVICE_ID_INTEL );

intercept/src/intercept.cpp

@@ -13184,6 +13184,9 @@ void* CLIntercept::getExtensionFunctionAddress(
     CHECK_RETURN_EXTENSION_FUNCTION( clRetainAcceleratorINTEL );
     CHECK_RETURN_EXTENSION_FUNCTION( clReleaseAcceleratorINTEL );
 
+    // cl_intel_concurrent_dispatch
+    CHECK_RETURN_EXTENSION_FUNCTION( clGetKernelMaxConcurrentWorkGroupCountINTEL );
+
     // cl_intel_create_buffer_with_properties
     CHECK_RETURN_EXTENSION_FUNCTION( clCreateBufferWithPropertiesINTEL );