COMPMID-2069: Rework CL ML layers to run exclusively on CL.

Change-Id: If6cbf7a2e013d264e5d7f7cb54143ce32ba2687b
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/934
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 51 insertions, 55 deletions

diff --git a/arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h b/arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h
index d2f8a78f87..3751178703 100644
--- a/arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h
+++ b/arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -27,6 +27,7 @@
 #include "arm_compute/runtime/IFunction.h"
 
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
+#include "arm_compute/core/CL/kernels/CLMemsetKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/IFunction.h"
@@ -34,9 +35,14 @@
 
 namespace arm_compute
 {
+// Forward declarations
 class ICLTensor;
 
-/** Basic function to run @ref CLDeconvolutionLayerUpsampleKernel */
+/** Basic function to execute deconvolution upsample on OpenCL. This function calls the following OpenCL kernels and functions:
+ *
+ * -# @ref CLMemsetKernel
+ * -# @ref CLDeconvolutionLayerUpsampleKernel
+ */
 class CLDeconvolutionLayerUpsample : public IFunction
 {
 public:
@@ -79,7 +85,8 @@ public:
 
 private:
     CLDeconvolutionLayerUpsampleKernel _upsample;
+    CLMemsetKernel                     _memset;
     ICLTensor                         *_output;
 };
-}
+} // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLDECONVOLUTIONLAYERUPSAMPLE_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h
index 936263d635..b9a435abb2 100644
--- a/arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h
@@ -26,10 +26,9 @@
 
 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h"
+#include "arm_compute/runtime/CL/functions/CLReverse.h"
 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
 
-#include "arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h"
-
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
@@ -70,7 +69,7 @@ class ICLTensor;
  * -# @ref CLConvolutionLayer
  *
  * And the following CPP kernels:
- * -# @ref CPPFlipWeightsKernel
+ * -# @ref CLReverse
  *
  */
 class CLDirectDeconvolutionLayer : public IFunction
@@ -119,11 +118,12 @@ private:
     CLMemoryGroup                _memory_group;
     CLDeconvolutionLayerUpsample _scale_f;
     CLConvolutionLayer           _conv_f;
-    CPPFlipWeightsKernel         _flip_weights;
+    CLReverse                    _flip_weights;
 
     CLTensor   _scaled_output;
     ICLTensor *_original_weights;
     CLTensor   _weights_flipped;
+    CLTensor   _flip_axis;
 
     bool _is_prepared;
 };
diff --git a/arm_compute/runtime/CL/functions/CLLSTMLayer.h b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
index a804a4af5b..8bd47cbf8e 100644
--- a/arm_compute/runtime/CL/functions/CLLSTMLayer.h
+++ b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
@@ -29,6 +29,7 @@
 #include "arm_compute/core/CL/kernels/CLActivationLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLCopyKernel.h"
 #include "arm_compute/core/CL/kernels/CLElementwiseOperationKernel.h"
+#include "arm_compute/core/CL/kernels/CLMemsetKernel.h"
 #include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h"
 #include "arm_compute/core/CL/kernels/CLWidthConcatenate2TensorsKernel.h"
 #include "arm_compute/core/Types.h"
@@ -188,6 +189,7 @@ private:
     CLWidthConcatenate2TensorsKernel     _concat_weights_forget_gate;
     CLWidthConcatenate2TensorsKernel     _concat_weights_input_gate;
     CLWidthConcatenate2TensorsKernel     _concat_weights_output;
+    CLMemsetKernel                       _ones_memset_kernel;
     CLTensor                             _input_gate_out1;
     CLTensor                             _input_gate_out2;
     CLTensor                             _input_gate_out3;
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp b/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
index ce8667d656..c66dff01c5 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -27,14 +27,11 @@
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 
-#include <cmath>
-#include <memory>
-#include <tuple>
-
-using namespace arm_compute;
-
+namespace arm_compute
+{
 CLDeconvolutionLayerUpsample::CLDeconvolutionLayerUpsample() // NOLINT
     : _upsample(),
+      _memset(),
       _output(nullptr)
 {
 }
@@ -51,22 +48,13 @@ void CLDeconvolutionLayerUpsample::configure(ICLTensor *input, ICLTensor *output
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     _output = output;
+    _memset.configure(_output, PixelValue(0, _output->info()->data_type(), _output->info()->quantization_info()));
     _upsample.configure(input, _output, inner_border, info);
 }
 
 void CLDeconvolutionLayerUpsample::run()
 {
-    _output->map(CLScheduler::get().queue(), true);
-    if(is_data_type_quantized_asymmetric(_output->info()->data_type()))
-    {
-        const uint8_t quantized_zero = _output->info()->quantization_info().offset;
-        std::fill_n(_output->buffer(), _output->info()->total_size(), quantized_zero);
-    }
-    else
-    {
-        memset(_output->buffer(), 0, _output->info()->total_size());
-    }
-    _output->unmap(CLScheduler::get().queue());
-
-    CLScheduler::get().enqueue(_upsample, false);
+    CLScheduler::get().enqueue(_memset, false);
+    CLScheduler::get().enqueue(_upsample, true);
 }
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
index c01588a164..ee76248e35 100644
--- a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
@@ -28,7 +28,6 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
-#include "arm_compute/runtime/CPP/CPPScheduler.h"
 #include "utils/TypePrinter.h"
 
 #include <memory>
@@ -46,6 +45,7 @@ CLDirectDeconvolutionLayer::CLDirectDeconvolutionLayer(std::shared_ptr<IMemoryMa
       _scaled_output(),
       _original_weights(nullptr),
       _weights_flipped(),
+      _flip_axis(),
       _is_prepared(false)
 {
 }
@@ -120,8 +120,9 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights,
     const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
 
     _original_weights = weights;
+    _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
     _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
-    _flip_weights.configure(weights, &_weights_flipped);
+    _flip_weights.configure(weights, &_weights_flipped, &_flip_axis);
 
     auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h),
                                                     info.pad().first, info.pad().second, stride_x, stride_y);
@@ -151,10 +152,18 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights,
     const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
     _scale_f.configure(input, &_scaled_output, BorderSize(), upsample_info);
 
-    // setup the function to convolve the upscaled output
+    // Setup the function to convolve the upscaled output
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
     _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info, weights_info);
     _scaled_output.allocator()->allocate();
+
+    // Setup flip axis data
+    _flip_axis.allocator()->allocate();
+    _flip_axis.map(true);
+    auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
+    axis_data[0]   = 0;
+    axis_data[1]   = 1;
+    _flip_axis.unmap();
 }
 
 void CLDirectDeconvolutionLayer::run()
@@ -177,16 +186,13 @@ void CLDirectDeconvolutionLayer::prepare()
 
         // Run weights flipping and mark original weights tensor as unused
         _weights_flipped.allocator()->allocate();
-        _weights_flipped.map(true);
-        _original_weights->map(CLScheduler::get().queue(), true);
-        CPPScheduler::get().schedule(&_flip_weights, Window::DimZ);
-        _weights_flipped.unmap();
-        _original_weights->unmap(CLScheduler::get().queue());
+        _flip_weights.run();
         _original_weights->mark_as_unused();
 
         // Prepare convolution
         _conv_f.prepare();
 
+        // Free flipped weights
         if(!_weights_flipped.is_used())
         {
             _weights_flipped.allocator()->free();
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index a004762a4e..13c4871148 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -44,9 +44,10 @@ CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
       _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(), _gemm_output(), _pixelwise_mul_output_state1(), _transpose_output(),
       _accum_output1(), _accum_output2(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _gemm_output_state(), _accum_output_state(),
       _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(), _concat_weights_input_gate(), _concat_weights_output(),
-      _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(),
-      _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _cell_state_activation(),
-      _output_state1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false), _is_prepared(false)
+      _ones_memset_kernel(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(),
+      _forget_gate_out5(), _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(), _output2(), _output3(), _output4(),
+      _cell_state_activation(), _output_state1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false),
+      _is_prepared(false)
 {
 }
 
@@ -104,7 +105,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     std::vector<const ICLTensor *> inputs_vector;
     inputs_vector.emplace_back(input);
     inputs_vector.emplace_back(output_state_in);
-    const TensorShape concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    const TensorShape concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0);
     _forget_gate_out2.allocator()->init(TensorInfo(concat_shape, 1, input->info()->data_type()));
 
     _memory_group.manage(&_forget_gate_out2);
@@ -114,7 +115,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
 
     weights_vector.emplace_back(input_to_forget_weights);
     weights_vector.emplace_back(recurrent_to_forget_weights);
-    const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(weights_vector);
+    const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(weights_vector, 0);
     _forget_gate_out6.allocator()->init(TensorInfo(weights_concat_shape, 1, input->info()->data_type()));
 
     _concat_weights_forget_gate.configure(input_to_forget_weights, recurrent_to_forget_weights, &_forget_gate_out6);
@@ -155,6 +156,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     {
         _memory_group.manage(&_input_gate_out1);
         _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _ones_memset_kernel.configure(&_ones, PixelValue(1, _ones.info()->data_type()));
         _subtract_input_gate.configure(ArithmeticOperation::SUB, &_ones, forget_gate_out, &_input_gate_out1, ConvertPolicy::SATURATE);
         _ones.allocator()->allocate();
         _run_cifg_opt = true;
@@ -167,7 +169,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
         std::vector<const ICLTensor *> lstm_weights;
         lstm_weights.emplace_back(lstm_params.input_to_input_weights());
         lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
-        TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(lstm_weights);
+        TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
         _input_gate_out2.allocator()->init(TensorInfo(lstm_weights_concat_shape, 1, input->info()->data_type()));
 
         _concat_weights_input_gate.configure(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), &_input_gate_out2);
@@ -237,7 +239,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     std::vector<const ICLTensor *> in_out_weights;
     in_out_weights.emplace_back(input_to_output_weights);
     in_out_weights.emplace_back(recurrent_to_output_weights);
-    TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(in_out_weights);
+    TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
     _output2.allocator()->init(TensorInfo(in_out_weights_concat_shape, 1, input->info()->data_type()));
 
     _concat_weights_output.configure(input_to_output_weights, recurrent_to_output_weights, &_output2);
@@ -392,7 +394,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     std::vector<const ITensorInfo *> inputs_vector;
     inputs_vector.emplace_back(input);
     inputs_vector.emplace_back(output_state_in);
-    const TensorShape concat_shape       = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    const TensorShape concat_shape       = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0);
     TensorInfo        forget_gate_concat = TensorInfo(concat_shape, 1, input->data_type());
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(input, output_state_in, &forget_gate_concat));
@@ -417,7 +419,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
         std::vector<const ITensorInfo *> lstm_weights;
         lstm_weights.emplace_back(lstm_params.input_to_input_weights());
         lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
-        TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(lstm_weights);
+        TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
         TensorInfo  lstm_gate_concat          = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
         ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), &lstm_gate_concat));
 
@@ -454,7 +456,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     std::vector<const ITensorInfo *> in_out_weights;
     in_out_weights.emplace_back(input_to_output_weights);
     in_out_weights.emplace_back(recurrent_to_output_weights);
-    TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(in_out_weights);
+    TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
     TensorInfo  in_out_gate_concat          = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
     ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(input_to_output_weights, recurrent_to_output_weights, &in_out_gate_concat));
     // Validate output gate tmp
@@ -518,16 +520,7 @@ void CLLSTMLayer::run()
 
     if(_run_cifg_opt)
     {
-        _ones.map(true);
-        if(_ones.info()->data_type() == DataType::F16)
-        {
-            std::fill_n(reinterpret_cast<half *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1);
-        }
-        else
-        {
-            std::fill_n(reinterpret_cast<float *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1);
-        }
-        _ones.unmap();
+        CLScheduler::get().enqueue(_ones_memset_kernel);
         CLScheduler::get().enqueue(_subtract_input_gate);
     }
     else