COMPMID-1004 GLES: Add memory manager to GLES functions

Change-Id: I80fc9c0dd02afd79b501abde751036f9599b7bf2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125103
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 39 insertions, 14 deletions

diff --git a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
index 5689722340..f4c073668a 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
@@ -91,9 +91,9 @@ void GCConvolutionLayerReshapeWeights::run()
     }
 }
 
-GCConvolutionLayer::GCConvolutionLayer()
-    : _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _fill_border(), _input_im2col_reshaped(), _input_interleaved_reshaped(),
-      _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
+GCConvolutionLayer::GCConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _fill_border(), _input_im2col_reshaped(),
+      _input_interleaved_reshaped(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
 {
 }
 
@@ -196,6 +196,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
     TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->info()->fixed_point_position());
     _input_im2col_reshaped.allocator()->init(im2col_reshaped_info);
+    _memory_group.manage(&_input_im2col_reshaped);
 
     // Create tensor (interleave) to prepare input tensor for GEMM
     if(run_interleaved)
@@ -207,6 +208,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
         // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
         TensorInfo interleaved_info(shape_interleaved, 1, dt, input->info()->fixed_point_position());
         _input_interleaved_reshaped.allocator()->init(interleaved_info);
+        _memory_group.manage(&_input_interleaved_reshaped);
     }
 
     // Create GEMM output tensor
@@ -218,6 +220,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
     TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
     _gemm_output.allocator()->init(info_gemm);
+    _memory_group.manage(&_gemm_output);
 
     // Configure kernels
     if(dt == DataType::F16)
@@ -263,6 +266,8 @@ void GCConvolutionLayer::run()
         _reshape_weights.run();
     }
 
+    _memory_group.acquire();
+
     // Run im2col
     GCScheduler::get().dispatch(_fill_border);
     GCScheduler::get().memory_barrier();
@@ -282,4 +287,6 @@ void GCConvolutionLayer::run()
     GCScheduler::get().memory_barrier();
     // Reshape output matrix
     GCScheduler::get().dispatch(_output_col2im_kernel, false);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
index 9e4f0f6c95..0f8f8e6c94 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
@@ -38,9 +38,9 @@ void GCFullyConnectedLayerReshapeWeights::configure(const IGCTensor *input, IGCT
     _kernel = std::move(k);
 }
 
-GCFullyConnectedLayer::GCFullyConnectedLayer()
-    : _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true),
-      _accumulate_biases(false)
+GCFullyConnectedLayer::GCFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _reshape_weights_output(),
+      _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false)
 {
 }
 
@@ -61,6 +61,7 @@ void GCFullyConnectedLayer::configure_conv_fc(const IGCTensor *input, const IGCT
     _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt));
 
     // Configure im2col kernel
+    _memory_group.manage(&_im2col_output);
     _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
 
     // Configure matrix multiply kernel
@@ -156,6 +157,7 @@ void GCFullyConnectedLayer::run()
         _reshape_weights_kernel.run();
     }
 
+    _memory_group.acquire();
     // Linearize input if it comes from a convolutional layer
     if(_is_fc_after_conv)
     {
@@ -177,4 +179,5 @@ void GCFullyConnectedLayer::run()
 
         GCScheduler::get().dispatch(_accumulate_biases_kernel);
     }
+    _memory_group.release();
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index 5122c20504..46424a59f5 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -40,8 +40,8 @@
 using namespace arm_compute;
 using namespace arm_compute::gles_compute;
 
-GCGEMM::GCGEMM()
-    : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false)
+GCGEMM::GCGEMM(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false)
 {
 }
 
@@ -88,6 +88,7 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
 
         TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
         _tmp_a.allocator()->init(info_a);
+        _memory_group.manage(&_tmp_a);
 
         TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position());
         _tmp_b.allocator()->init(info_b);
@@ -118,6 +119,7 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
 
 void GCGEMM::run()
 {
+    _memory_group.acquire();
     if(_is_interleaved_transposed)
     {
         // Run interleave kernel
@@ -137,4 +139,5 @@ void GCGEMM::run()
         GCScheduler::get().memory_barrier();
         GCScheduler::get().dispatch(_ma_kernel);
     }
+    _memory_group.release();
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
index fc3882dbda..13213d2b54 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -33,8 +33,8 @@
 
 using namespace arm_compute;
 
-GCNormalizationLayer::GCNormalizationLayer()
-    : _squared_input(), _norm_kernel(), _multiply_kernel(), _border_handler()
+GCNormalizationLayer::GCNormalizationLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _squared_input(), _norm_kernel(), _multiply_kernel(), _border_handler()
 {
 }
 
@@ -43,6 +43,7 @@ void GCNormalizationLayer::configure(const IGCTensor *input, IGCTensor *output,
     ARM_COMPUTE_ERROR_ON(input == nullptr);
 
     _squared_input.allocator()->init(TensorInfo(input->info()->tensor_shape(), 1, input->info()->data_type()));
+    _memory_group.manage(&_squared_input);
 
     _norm_kernel.configure(input, &_squared_input, output, norm_info);
     _multiply_kernel.configure(input, input, &_squared_input, 1.0f);
@@ -55,9 +56,11 @@ void GCNormalizationLayer::configure(const IGCTensor *input, IGCTensor *output,
 
 void GCNormalizationLayer::run()
 {
+    _memory_group.acquire();
     GCScheduler::get().dispatch(_multiply_kernel, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_border_handler, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_norm_kernel, true);
+    _memory_group.release();
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
index 5221c5cc5d..1748a5952b 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,8 +29,8 @@
 
 using namespace arm_compute;
 
-GCSoftmaxLayer::GCSoftmaxLayer()
-    : _max_kernel(), _shift_exp_sum_kernel(), _norm_kernel(), _max(), _sum(), _tmp()
+GCSoftmaxLayer::GCSoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _max_kernel(), _shift_exp_sum_kernel(), _norm_kernel(), _max(), _sum(), _tmp()
 {
 }
 
@@ -50,6 +50,11 @@ void GCSoftmaxLayer::configure(const IGCTensor *input, IGCTensor *output, float
     _max.allocator()->init(tensor_info_max_sum);
     _sum.allocator()->init(tensor_info_max_sum);
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_tmp);
+    _memory_group.manage(&_max);
+    _memory_group.manage(&_sum);
+
     // Configure Kernels
     _max_kernel.configure(input, &_max);
     _shift_exp_sum_kernel.configure(input, &_max, &_tmp, &_sum);
@@ -63,9 +68,13 @@ void GCSoftmaxLayer::configure(const IGCTensor *input, IGCTensor *output, float
 
 void GCSoftmaxLayer::run()
 {
+    _memory_group.acquire();
+
     GCScheduler::get().dispatch(_max_kernel, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_shift_exp_sum_kernel, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_norm_kernel);
+
+    _memory_group.release();
 }