Port ClTranspose to new API

Partially Resolves: COMPMID-4277 (1/2)

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I704c2303135cbe1ba46d2fd5642c84c562204bc7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5194
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 74 insertions, 20 deletions

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 46a90a54b7..176c5bebd5 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,7 +40,7 @@
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
-#include "src/core/CL/kernels/CLTransposeKernel.h"
+#include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 #include "support/Cast.h"
 
 #include <algorithm>
@@ -141,6 +141,20 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I
 }
 } // namespace
 
+struct CLFullyConnectedLayerReshapeWeights::Impl
+{
+    const ITensor                                      *src{ nullptr };
+    ITensor                                            *dst{ nullptr };
+    std::unique_ptr<opencl::kernels::ClTransposeKernel> op{ nullptr };
+};
+
+CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights()
+    : _impl(std::make_unique<Impl>())
+{
+}
+
+CLFullyConnectedLayerReshapeWeights::~CLFullyConnectedLayerReshapeWeights() = default;
+
 void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output);
@@ -148,14 +162,23 @@ void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLT
 
 void CLFullyConnectedLayerReshapeWeights::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
 {
-    auto k = std::make_unique<CLTransposeKernel>();
-    k->configure(compile_context, input, output);
-    _kernel = std::move(k);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<opencl::kernels::ClTransposeKernel>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info());
 }
 
 Status CLFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return CLTransposeKernel::validate(input, output);
+    return opencl::kernels::ClTransposeKernel::validate(input, output);
+}
+
+void CLFullyConnectedLayerReshapeWeights::run()
+{
+    ITensorPack pack{};
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    CLScheduler::get().enqueue_op(*_impl->op.get(), pack, false);
 }
 
 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
@@ -164,6 +187,7 @@ CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> mem
       _are_weights_reshaped(true), _is_fc_after_conv(true), _is_quantized(false), _is_prepared(false), _original_weights(nullptr)
 {
 }
+
 void CLFullyConnectedLayer::configure_mm(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
                                          const FullyConnectedLayerInfo &fc_info)
 {
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 5036126aea..05d459c899 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -41,7 +41,7 @@
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
-#include "src/core/CL/kernels/CLTransposeKernel.h"
+#include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 
 namespace arm_compute
 {
@@ -51,7 +51,7 @@ using namespace arm_compute::utils::info_helpers;
 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
       _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(),
-      _transpose_cell_state(std::make_unique<CLTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
+      _transpose_cell_state(std::make_unique<opencl::kernels::ClTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
       _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(),
       _fully_connected_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(), _ones_fill(), _mean_std_norm_input_gate(), _pixelwise_mul_input_gate_coeff(), _accum_input_gate_bias(), _mean_std_norm_forget_gate(),
@@ -256,7 +256,8 @@ void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTe
     _memory_group.manage(&_cell_state_out1);
     _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
     _memory_group.manage(&_cell_state_out2);
-    _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights, &_cell_state_out2);
+    _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights->info(), _cell_state_out2.info());
+    _recurrent_to_cell_weights = recurrent_to_cell_weights;
     _memory_group.manage(&_cell_state_out3);
     _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
     _cell_state_out2.allocator()->allocate();
@@ -680,7 +681,12 @@ void CLLSTMLayer::run()
     }
 
     _fully_connected_cell_state.run();
-    CLScheduler::get().enqueue(*_transpose_cell_state);
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _recurrent_to_cell_weights);
+    pack.add_tensor(TensorType::ACL_DST, &_cell_state_out2);
+    CLScheduler::get().enqueue_op(*_transpose_cell_state,
+                                  pack,
+                                  false);
     _gemm_cell_state1.run();
     _accum_cell_state1.run();
     if(_is_layer_norm_lstm)
diff --git a/src/runtime/CL/functions/CLPermute.cpp b/src/runtime/CL/functions/CLPermute.cpp
index eeb0169241..c1da2a9eca 100644
--- a/src/runtime/CL/functions/CLPermute.cpp
+++ b/src/runtime/CL/functions/CLPermute.cpp
@@ -43,9 +43,8 @@ CLPermute::CLPermute()
     : _impl(std::make_unique<Impl>())
 {
 }
-CLPermute::CLPermute(CLPermute &&) = default;
-CLPermute &CLPermute::operator=(CLPermute &&) = default;
-CLPermute::~CLPermute()                       = default;
+
+CLPermute::~CLPermute() = default;
 
 void CLPermute::configure(const ICLTensor *input, ICLTensor *output, const PermutationVector &perm)
 {
diff --git a/src/runtime/CL/functions/CLTranspose.cpp b/src/runtime/CL/functions/CLTranspose.cpp
index 67151be9a1..142cf73259 100644
--- a/src/runtime/CL/functions/CLTranspose.cpp
+++ b/src/runtime/CL/functions/CLTranspose.cpp
@@ -23,12 +23,27 @@
  */
 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
 
-#include "src/core/CL/kernels/CLTransposeKernel.h"
-
-#include <utility>
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/runtime/gpu/cl/operators/ClTranspose.h"
 
 namespace arm_compute
 {
+struct CLTranspose::Impl
+{
+    const ICLTensor                     *src{ nullptr };
+    ICLTensor                           *dst{ nullptr };
+    std::unique_ptr<opencl::ClTranspose> op{ nullptr };
+};
+CLTranspose::CLTranspose()
+    : _impl(std::make_unique<Impl>())
+{
+}
+CLTranspose::~CLTranspose() = default;
+
 void CLTranspose::configure(const ICLTensor *input, ICLTensor *output)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output);
@@ -36,13 +51,23 @@ void CLTranspose::configure(const ICLTensor *input, ICLTensor *output)
 
 void CLTranspose::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
 {
-    auto k = std::make_unique<CLTransposeKernel>();
-    k->configure(compile_context, input, output);
-    _kernel = std::move(k);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<opencl::ClTranspose>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info());
 }
 
 Status CLTranspose::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return CLTransposeKernel::validate(input, output);
+    return opencl::ClTranspose::validate(input, output);
+}
+
+void CLTranspose::run()
+{
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    _impl->op->run(pack);
 }
 } // namespace arm_compute
\ No newline at end of file