COMPMID-734: CLTuner rework

Change-Id: I8f20d6ea8a09869d71003e7b08e0d33775282f6c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125802
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

6 files changed, 155 insertions, 76 deletions

diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
index 56ac0c7250..b5526c4fca 100644
--- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
@@ -315,79 +315,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
 
         kernel_name << "_f32_bifrost";
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name.str(), build_options.options()));
-
-        // Through extensive experimentation with over 30 representative tensor
-        // shapes, we found a small number of local work size configurations
-        // that result in nearly optimal execution times. Selecting the right
-        // lws for a given shape, however, required a complex decision tree,
-        // until we constructed a simple feature as described below.
-        //
-        // We started from the number of multiply-accumulate operations for a
-        // convolution layer, which is equal to the product of the input
-        // dimensions 0..2 and the weights dimensions 0..2.  Unfortunately,
-        // this resulted in ties between distinct shapes that required distinct
-        // lws configurations. Replacing the width of the input with the kernel
-        // size, however, resulted in nearly optimal predictions. We use underscores
-        // in variable names to indicate when they are intentionally misleading.
-        const size_t product_of_weights_dimensions = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2);
-        const size_t product_of_input_dimensions_  = input->info()->dimension(0) * weights->info()->dimension(1) * input->info()->dimension(2);
-        const float  mega_ops_                     = 1e-6 * product_of_weights_dimensions * product_of_input_dimensions_;
-
-        switch(kernel_size)
-        {
-            case 1:
-            {
-                if(mega_ops_ < 1.f)
-                {
-                    _lws_hint = cl::NDRange(1, 1, 8);
-                }
-                else if(mega_ops_ < 7.f)
-                {
-                    _lws_hint = cl::NDRange(1, 1, 4);
-                }
-                else
-                {
-                    _lws_hint = cl::NDRange(1, 1, 2);
-                }
-                break;
-            }
-            case 3:
-            {
-                if(mega_ops_ < 1.f)
-                {
-                    _lws_hint = cl::NDRange(1, 1, 8);
-                }
-                else if(mega_ops_ < 13.f)
-                {
-                    _lws_hint = cl::NDRange(2, 1, 4);
-                }
-                else if(mega_ops_ < 50.f)
-                {
-                    _lws_hint = cl::NDRange(3, 1, 4);
-                }
-                else
-                {
-                    _lws_hint = cl::NDRange(2, 1, 6);
-                }
-                break;
-            }
-            case 5:
-            {
-                if(mega_ops_ < 2.f || mega_ops_ > 80.f)
-                {
-                    _lws_hint = cl::NDRange(2, 1, 4);
-                }
-                else
-                {
-                    _lws_hint = cl::NDRange(2, 1, 8);
-                }
-                break;
-            }
-            default:
-            {
-                ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost");
-            }
-        }
     }
     else
     {
diff --git a/src/graph/Graph.cpp b/src/graph/Graph.cpp
index 2fe3a90aef..47bd672114 100644
--- a/src/graph/Graph.cpp
+++ b/src/graph/Graph.cpp
@@ -30,6 +30,7 @@
 #include "arm_compute/graph/Tensor.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTuner.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "support/ToolchainSupport.h"
 
diff --git a/src/runtime/CL/CLScheduler.cpp b/src/runtime/CL/CLScheduler.cpp
index 65292fe837..2a5d836c33 100644
--- a/src/runtime/CL/CLScheduler.cpp
+++ b/src/runtime/CL/CLScheduler.cpp
@@ -52,7 +52,7 @@ void CLScheduler::enqueue(ICLKernel &kernel, bool flush)
     if(_cl_tuner != nullptr)
     {
         // Tune the OpenCL kernel
-        _cl_tuner->tune_kernel(kernel);
+        _cl_tuner->tune_kernel_dynamic(kernel);
     }
 
     // Run kernel
diff --git a/src/runtime/CL/CLTuner.cpp b/src/runtime/CL/CLTuner.cpp
index df8e255356..17a62ab46e 100644
--- a/src/runtime/CL/CLTuner.cpp
+++ b/src/runtime/CL/CLTuner.cpp
@@ -113,7 +113,12 @@ bool CLTuner::tune_new_kernels() const
     return _tune_new_kernels;
 }
 
-void CLTuner::tune_kernel(ICLKernel &kernel)
+void CLTuner::tune_kernel_static(ICLKernel &kernel)
+{
+    ARM_COMPUTE_UNUSED(kernel);
+}
+
+void CLTuner::tune_kernel_dynamic(ICLKernel &kernel)
 {
     // Get the configuration ID from the kernel
     const std::string &config_id = kernel.config_id();
diff --git a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
index d6a335c1ec..c48865a0cc 100644
--- a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,6 +52,9 @@ void CLDirectConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weig
         zero_value = PixelValue(static_cast<uint8_t>(input->info()->quantization_info().offset));
     }
     _input_border_handler.configure(input, _direct_conv_kernel.border_size(), BorderMode::CONSTANT, zero_value);
+
+    // Tune kernels
+    CLScheduler::get().tune_kernel_static(_direct_conv_kernel);
 }
 
 Status CLDirectConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp
new file mode 100644
index 0000000000..c0ebd24afe
--- /dev/null
+++ b/src/runtime/CL/tuners/BifrostTuner.cpp
@@ -0,0 +1,143 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/tuners/BifrostTuner.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernels.h"
+#include "arm_compute/core/utils/misc/Cast.h"
+
+namespace arm_compute
+{
+namespace tuners
+{
+namespace
+{
+/** Tunes a @ref CLDirectConvolutionLayerKernel for a bifrost target
+ *
+ * @param[in] k Kernels to tune
+ */
+void tune_direct_convolution_kernel(CLDirectConvolutionLayerKernel &k)
+{
+    cl::NDRange lws_hint = k.lws_hint();
+
+    const GPUTarget    gpu_target    = k.get_target();
+    const DataType     dt            = k._input->info()->data_type();
+    const TensorShape  weights_shape = k._weights->info()->tensor_shape();
+    const TensorShape  inputs_shape  = k._input->info()->tensor_shape();
+    const size_t       kernel_size   = weights_shape.x();
+    const unsigned int stride_x      = k._conv_stride_x;
+    const unsigned int stride_y      = k._conv_stride_y;
+
+    if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72) && (kernel_size <= 5) && (stride_x == 1) && (stride_y == 1) && (dt == DataType::F32))
+    {
+        // Through extensive experimentation with over 30 representative tensor
+        // shapes, we found a small number of local work size configurations
+        // that result in nearly optimal execution times. Selecting the right
+        // lws for a given shape, however, required a complex decision tree,
+        // until we constructed a simple feature as described below.
+        //
+        // We started from the number of multiply-accumulate operations for a
+        // convolution layer, which is equal to the product of the input
+        // dimensions 0..2 and the weights dimensions 0..2.  Unfortunately,
+        // this resulted in ties between distinct shapes that required distinct
+        // lws configurations. Replacing the width of the input with the kernel
+        // size, however, resulted in nearly optimal predictions. We use underscores
+        // in variable names to indicate when they are intentionally misleading.
+        const size_t product_of_weights_dimensions = weights_shape[0] * weights_shape[1] * weights_shape[2];
+        const size_t product_of_input_dimensions_  = inputs_shape[0] * inputs_shape[1] * inputs_shape[2];
+        const float  mega_ops_                     = 1e-6 * product_of_weights_dimensions * product_of_input_dimensions_;
+
+        switch(kernel_size)
+        {
+            case 1:
+            {
+                if(mega_ops_ < 1.f)
+                {
+                    lws_hint = cl::NDRange(1, 1, 8);
+                }
+                else if(mega_ops_ < 7.f)
+                {
+                    lws_hint = cl::NDRange(1, 1, 4);
+                }
+                else
+                {
+                    lws_hint = cl::NDRange(1, 1, 2);
+                }
+                break;
+            }
+            case 3:
+            {
+                if(mega_ops_ < 1.f)
+                {
+                    lws_hint = cl::NDRange(1, 1, 8);
+                }
+                else if(mega_ops_ < 13.f)
+                {
+                    lws_hint = cl::NDRange(2, 1, 4);
+                }
+                else if(mega_ops_ < 50.f)
+                {
+                    lws_hint = cl::NDRange(3, 1, 4);
+                }
+                else
+                {
+                    lws_hint = cl::NDRange(2, 1, 6);
+                }
+                break;
+            }
+            case 5:
+            {
+                if(mega_ops_ < 2.f || mega_ops_ > 80.f)
+                {
+                    lws_hint = cl::NDRange(2, 1, 4);
+                }
+                else
+                {
+                    lws_hint = cl::NDRange(2, 1, 8);
+                }
+                break;
+            }
+            default:
+                break;
+        }
+        k.set_lws_hint(lws_hint);
+    }
+}
+} // namespace
+
+void BifrostTuner::tune_kernel_static(ICLKernel &kernel)
+{
+    // Continue on tuning if dynamic tuning
+    if(dynamic_cast<CLDirectConvolutionLayerKernel *>(&kernel) != nullptr)
+    {
+        tune_direct_convolution_kernel(*utils::cast::polymorphic_downcast<CLDirectConvolutionLayerKernel *>(&kernel));
+    }
+}
+
+void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel)
+{
+    ARM_COMPUTE_UNUSED(kernel);
+}
+} // namespace tuners
+} // namespace arm_compute
\ No newline at end of file