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>

3 files changed, 43 insertions, 5 deletions

diff --git a/arm_compute/runtime/CL/CLScheduler.h b/arm_compute/runtime/CL/CLScheduler.h
index 362d2ba137..5bfaaf4b5d 100644
--- a/arm_compute/runtime/CL/CLScheduler.h
+++ b/arm_compute/runtime/CL/CLScheduler.h
@@ -211,6 +211,11 @@ public:
 
 private:
     void enqueue_common(ICLKernel &kernel, ITensorPack &tensors, bool flush);
+    /** If job chain is disabled, then flush the command queue according to @p flush. Otherwise @p flush is ignored and the queue is only flushed when job chain count exceeds allocated job chain size
+     *
+     * @param[in] flush Flush the command queue. Ignored when job chain is enabled.
+     */
+    void flush_queue(bool flush);
 
 #if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
     void enqueue_common(ICLKernel &kernel, experimental::dynamic_fusion::TensorBinding &tensors, const experimental::dynamic_fusion::ClExecutionDescriptor &exec_desc, bool flush);
diff --git a/arm_compute/runtime/CL/CLTuner.h b/arm_compute/runtime/CL/CLTuner.h
index f96edc962b..e595f8f34b 100644
--- a/arm_compute/runtime/CL/CLTuner.h
+++ b/arm_compute/runtime/CL/CLTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -124,6 +124,9 @@ public:
     void tune_kernel_static(ICLKernel &kernel) override;
     void tune_kernel_dynamic(ICLKernel &kernel) override;
     void tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) override;
+#if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
+    void tune_kernel_dynamic(ICLKernel &kernel, experimental::dynamic_fusion::TensorBinding &tensors, const experimental::dynamic_fusion::ClExecutionDescriptor &exec_desc) override;
+#endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
 
     /** Is the kernel_event set ?
      *
@@ -131,15 +134,26 @@ public:
      */
     bool kernel_event_is_set() const;
 
+    /** A wrapper wrapping tensors and other objects needed for running the kernel
+     */
+    struct IKernelData;
+
 private:
+    /** Perform tune_kernel_dynamic
+     *
+     * @param[in]     kernel OpenCL kernel to be tuned with tuning parameters
+     * @param[in,out] data   IKernelData object wrapping tensors and other objects needed for running the kernel
+     *
+     */
+    void do_tune_kernel_dynamic(ICLKernel &kernel, IKernelData *data);
     /** Find optimal tuning parameters using brute-force approach
      *
-     * @param[in]     kernel  OpenCL kernel to be tuned with tuning parameters
-     * @param[in,out] tensors Tensors for the kernel to operate on
+     * @param[in]     kernel OpenCL kernel to be tuned with tuning parameters
+     * @param[in,out] data   IKernelData object wrapping tensors and other objects needed for running the kernel
      *
      * @return The optimal tuning parameters to use
      */
-    CLTuningParams find_optimal_tuning_params(ICLKernel &kernel, ITensorPack &tensors);
+    CLTuningParams find_optimal_tuning_params(ICLKernel &kernel, IKernelData *data);
 
     std::unordered_map<std::string, CLTuningParams> _tuning_params_table;
     std::unordered_map<std::string, cl::NDRange>    _lws_table;
diff --git a/arm_compute/runtime/CL/ICLTuner.h b/arm_compute/runtime/CL/ICLTuner.h
index 0f951c384e..a327497255 100644
--- a/arm_compute/runtime/CL/ICLTuner.h
+++ b/arm_compute/runtime/CL/ICLTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,6 +30,16 @@ namespace arm_compute
 {
 class ICLKernel;
 
+#if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
+namespace experimental
+{
+namespace dynamic_fusion
+{
+struct TensorBinding;
+struct ClExecutionDescriptor;
+} // namespace dynamic_fusion
+} // namespace experimental
+#endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
 /** Basic interface for tuning the OpenCL kernels */
 class ICLTuner
 {
@@ -57,6 +67,15 @@ public:
      * @param[in, out] tensors Tensors for the kernel to use
      */
     virtual void tune_kernel_dynamic(ICLKernel &kernel, ITensorPack &tensors) = 0;
+#if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
+    /** Tune OpenCL kernel dynamically for dynamic fusion interface
+     *
+     * @param[in]      kernel    Kernel to tune
+     * @param[in, out] tensors   Tensors for the kernel to use
+     * @param[in]      exec_desc Execution descriptor
+     */
+    virtual void tune_kernel_dynamic(ICLKernel &kernel, experimental::dynamic_fusion::TensorBinding &tensors, const experimental::dynamic_fusion::ClExecutionDescriptor &exec_desc) = 0;
+#endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION)
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_ICLTUNER_H */