Enable dynamic cl tuning for dynamically fused kernels

* Add new tune_kernel_dynamic interface
* Add generate_config_id

Resolves: COMPMID-5154

Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: I39870e59fceda875487970061ceb2048995c5a45
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7400
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 23 insertions, 15 deletions

diff --git a/src/runtime/CL/CLScheduler.cpp b/src/runtime/CL/CLScheduler.cpp
index 18fd52232d..4cff707f1a 100644
--- a/src/runtime/CL/CLScheduler.cpp
+++ b/src/runtime/CL/CLScheduler.cpp
@@ -181,19 +181,12 @@ void CLScheduler::enqueue_common(ICLKernel &kernel, ITensorPack &tensors, bool f
 
     // Run kernel
     inject_memory ? kernel.run_op(tensors, kernel.window(), _queue) : kernel.run(kernel.window(), _queue);
-
     if(_job_chaining_enabled)
     {
-        if(++_job_chaining_count >= _job_chaining_size)
-        {
-            _job_chaining_count = 0;
-            _queue.flush();
-        }
-    }
-    else if(flush)
-    {
-        _queue.flush();
+        ++_job_chaining_count;
     }
+
+    flush_queue(flush);
 }
 
 #if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
@@ -204,14 +197,31 @@ void CLScheduler::enqueue_common(ICLKernel &kernel, experimental::dynamic_fusion
                              "The CLScheduler is not initialised yet! Please call the CLScheduler::get().default_init(), \
                              or CLScheduler::get()::init() and CLKernelLibrary::get()::init() function before running functions!");
 
-    const bool inject_memory = !tensors._binding.empty();
+    // ClCompositeKernel is stateless thus alway requires memory injection
+
+    // Tune the kernel if the CLTuner has been provided
+    if(_cl_tuner != nullptr)
+    {
+        _cl_tuner->tune_kernel_dynamic(kernel, tensors, exec_desc);
+    }
 
     // Run kernel
-    inject_memory ? kernel.run_composite_op(tensors, kernel.window(), _queue, exec_desc) : kernel.run(kernel.window(), _queue);
+    kernel.run_composite_op(tensors, kernel.window(), _queue, exec_desc);
+    if(_job_chaining_enabled)
+    {
+        ++_job_chaining_count;
+    }
+
+    flush_queue(flush);
+}
+
+#endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
 
+void CLScheduler::flush_queue(bool flush)
+{
     if(_job_chaining_enabled)
     {
-        if(++_job_chaining_count >= _job_chaining_size)
+        if(_job_chaining_count >= _job_chaining_size)
         {
             _job_chaining_count = 0;
             _queue.flush();
@@ -223,8 +233,6 @@ void CLScheduler::enqueue_common(ICLKernel &kernel, experimental::dynamic_fusion
     }
 }
 
-#endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
-
 void CLScheduler::enqueue(ICLKernel &kernel, bool flush)
 {
     ITensorPack pack;