COMPMID-3604: Graph failures during tuning

Update ICLTuner interface to account for the new memory injection
interface.
Redirect to appropriate kernel execution interface depending on if the
kernel supports memory injection or not.

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I8ce29f5c22f1865c9e688d12b65e68ee4486f99c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3588
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 44 insertions, 22 deletions

diff --git a/arm_compute/runtime/CL/CLTuner.h b/arm_compute/runtime/CL/CLTuner.h
index 745d57a959..aa31181d2d 100644
--- a/arm_compute/runtime/CL/CLTuner.h
+++ b/arm_compute/runtime/CL/CLTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 Arm Limited.
+ * Copyright (c) 2017-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -116,6 +116,7 @@ public:
     // Inherited methods overridden:
     void tune_kernel_static(ICLKernel &kernel) override;
     void tune_kernel_dynamic(ICLKernel &kernel) override;
+    void tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs) override;
 
     /** Is the kernel_event set ?
      *
@@ -130,7 +131,7 @@ private:
      *
      * @return The optimal LWS to use
      */
-    cl::NDRange find_optimal_lws(ICLKernel &kernel);
+    cl::NDRange find_optimal_lws(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs);
 
     std::unordered_map<std::string, cl::NDRange> _lws_table;
     cl::Event   _kernel_event;
diff --git a/arm_compute/runtime/CL/ICLTuner.h b/arm_compute/runtime/CL/ICLTuner.h
index 0b238180eb..4bc8ddf632 100644
--- a/arm_compute/runtime/CL/ICLTuner.h
+++ b/arm_compute/runtime/CL/ICLTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 Arm Limited.
+ * Copyright (c) 2017-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,8 @@
 #ifndef ARM_COMPUTE_ICLTUNER_H
 #define ARM_COMPUTE_ICLTUNER_H
 
+#include "arm_compute/core/experimental/Types.h"
+
 namespace arm_compute
 {
 class ICLKernel;
@@ -49,6 +51,13 @@ public:
      * @param[in] kernel Kernel to tune
      */
     virtual void tune_kernel_dynamic(ICLKernel &kernel) = 0;
+    /** Tune OpenCL kernel dynamically
+     *
+     * @param[in]      kernel  Kernel to tune
+     * @param[in]      inputs  Inputs for the kernel to use
+     * @param[in, out] outputs Outputs for the kernel to use
+     */
+    virtual void tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs) = 0;
 };
-}
+} // namespace arm_compute
 #endif /*ARM_COMPUTE_ICLTUNER_H */
diff --git a/arm_compute/runtime/CL/tuners/BifrostTuner.h b/arm_compute/runtime/CL/tuners/BifrostTuner.h
index b7ce6e96f9..830f7d9067 100644
--- a/arm_compute/runtime/CL/tuners/BifrostTuner.h
+++ b/arm_compute/runtime/CL/tuners/BifrostTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,6 +37,7 @@ public:
     // Inherited overriden methods
     void tune_kernel_static(ICLKernel &kernel) override;
     void tune_kernel_dynamic(ICLKernel &kernel) override;
+    void tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs) override;
 };
 } // namespace tuners
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/tuners/MidgardTuner.h b/arm_compute/runtime/CL/tuners/MidgardTuner.h
index 418b80728d..c702e7a2aa 100644
--- a/arm_compute/runtime/CL/tuners/MidgardTuner.h
+++ b/arm_compute/runtime/CL/tuners/MidgardTuner.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,6 +37,7 @@ public:
     // Inherited overriden methods
     void tune_kernel_static(ICLKernel &kernel) override;
     void tune_kernel_dynamic(ICLKernel &kernel) override;
+    void tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs) override;
 };
 } // namespace tuners
 } // namespace arm_compute
diff --git a/src/runtime/CL/CLScheduler.cpp b/src/runtime/CL/CLScheduler.cpp
index 56f5f212a8..5ef66f456a 100644
--- a/src/runtime/CL/CLScheduler.cpp
+++ b/src/runtime/CL/CLScheduler.cpp
@@ -157,22 +157,16 @@ void CLScheduler::enqueue_common(ICLKernel &kernel, const InputTensorMap &inputs
                              "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 = !inputs.empty();
+
     // Tune the kernel if the CLTuner has been provided
     if(_cl_tuner != nullptr)
     {
-        // Tune the OpenCL kernel
-        _cl_tuner->tune_kernel_dynamic(kernel);
+        inject_memory ? _cl_tuner->tune_kernel_dynamic(kernel, inputs, outputs) : _cl_tuner->tune_kernel_dynamic(kernel);
     }
 
     // Run kernel
-    if(inputs.empty())
-    {
-        kernel.run(kernel.window(), _queue);
-    }
-    else
-    {
-        kernel.run_op(inputs, outputs, kernel.window(), _queue);
-    }
+    inject_memory ? kernel.run_op(inputs, outputs, kernel.window(), _queue) : kernel.run(kernel.window(), _queue);
 
     if(flush)
     {
diff --git a/src/runtime/CL/CLTuner.cpp b/src/runtime/CL/CLTuner.cpp
index e3119c1db9..b2e3476e20 100644
--- a/src/runtime/CL/CLTuner.cpp
+++ b/src/runtime/CL/CLTuner.cpp
@@ -77,6 +77,11 @@ void CLTuner::tune_kernel_static(ICLKernel &kernel)
 
 void CLTuner::tune_kernel_dynamic(ICLKernel &kernel)
 {
+    tune_kernel_dynamic(kernel, {}, {});
+}
+
+void CLTuner::tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs)
+{
     // Get the configuration ID from the kernel and append GPU target name and number of available compute units
     const std::string config_id = kernel.config_id() + "_" + string_from_target(kernel.get_target()) + "_MP" + support::cpp11::to_string(CLKernelLibrary::get().get_num_compute_units());
 
@@ -90,7 +95,7 @@ void CLTuner::tune_kernel_dynamic(ICLKernel &kernel)
             if(_tune_new_kernels)
             {
                 // Find the optimal LWS for the kernel
-                cl::NDRange opt_lws = find_optimal_lws(kernel);
+                cl::NDRange opt_lws = find_optimal_lws(kernel, inputs, outputs);
 
                 // Insert the optimal LWS in the table
                 add_lws_to_table(config_id, opt_lws);
@@ -112,7 +117,7 @@ void CLTuner::add_lws_to_table(const std::string &kernel_id, cl::NDRange optimal
     _lws_table.emplace(kernel_id, optimal_lws);
 }
 
-cl::NDRange CLTuner::find_optimal_lws(ICLKernel &kernel)
+cl::NDRange CLTuner::find_optimal_lws(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs)
 {
     // Profiling queue
     cl::CommandQueue queue_profiler;
@@ -167,7 +172,8 @@ cl::NDRange CLTuner::find_optimal_lws(ICLKernel &kernel)
     cl::NDRange gws = ICLKernel::gws_from_window(kernel.window());
 
     // Run the kernel with default lws to be used as baseline
-    kernel.run(kernel.window(), queue_profiler);
+    const bool inject_memory = !inputs.empty();
+    inject_memory ? kernel.run_op(inputs, outputs, kernel.window(), queue_profiler) : kernel.run(kernel.window(), queue_profiler);
 
     queue_profiler.finish();
 
@@ -178,7 +184,7 @@ cl::NDRange CLTuner::find_optimal_lws(ICLKernel &kernel)
 
     cl::NDRange opt_lws = cl::NullRange;
 
-    //Construct the list of LWS values to be tested based on the tuner mode.
+    // Construct the list of LWS values to be tested based on the tuner mode.
     auto lws_list = cl_tuner::CLLWSListFactory::get_lws_list(_tuner_mode, gws);
     for(size_t i = 0; i < lws_list->size(); ++i)
     {
@@ -197,7 +203,7 @@ cl::NDRange CLTuner::find_optimal_lws(ICLKernel &kernel)
         kernel.set_lws_hint(lws_test);
 
         // Run the kernel
-        kernel.run(kernel.window(), queue_profiler);
+        inject_memory ? kernel.run_op(inputs, outputs, kernel.window(), queue_profiler) : kernel.run(kernel.window(), queue_profiler);
 
         queue_profiler.finish();
 
diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp
index 3fecd04455..1797c2ceb1 100644
--- a/src/runtime/CL/tuners/BifrostTuner.cpp
+++ b/src/runtime/CL/tuners/BifrostTuner.cpp
@@ -315,5 +315,10 @@ void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel)
 {
     ARM_COMPUTE_UNUSED(kernel);
 }
+
+void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs)
+{
+    ARM_COMPUTE_UNUSED(kernel, inputs, outputs);
+}
 } // namespace tuners
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CL/tuners/MidgardTuner.cpp b/src/runtime/CL/tuners/MidgardTuner.cpp
index a95ca1998e..68c98cebe7 100644
--- a/src/runtime/CL/tuners/MidgardTuner.cpp
+++ b/src/runtime/CL/tuners/MidgardTuner.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -73,5 +73,10 @@ void MidgardTuner::tune_kernel_dynamic(ICLKernel &kernel)
 {
     ARM_COMPUTE_UNUSED(kernel);
 }
+
+void MidgardTuner::tune_kernel_dynamic(ICLKernel &kernel, const InputTensorMap &inputs, const OutputTensorMap &outputs)
+{
+    ARM_COMPUTE_UNUSED(kernel, inputs, outputs);
+}
 } // namespace tuners
 } // namespace arm_compute