COMPMID-1043: Rework GCGEMMMatrixMultiplyKernel interface and allow auto initialization of the tensors

This patch also:
- removes support for already reshaped weights in GCConvolutionLayer
- makes GCConvolutionLayer similar to CLGEMMConvolutionLayer
- enables usage of the GCGEMM function in GCConvolution instead of calling the
  GEMM kernels directly

Change-Id: I3e4a64335555e86e18585d38d8fda4bfdb44e265
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127696
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 305 insertions, 238 deletions

diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp
index a5f09e8eac..b4bb5470ad 100644
--- a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -31,37 +31,180 @@
 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
 #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
 #include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
 #include <set>
 #include <string>
 
 using namespace arm_compute;
 using namespace arm_compute::gles_compute;
+using namespace arm_compute::misc::shape_calculator;
 
-GCGEMMMatrixMultiplyKernel::GCGEMMMatrixMultiplyKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+namespace
 {
-}
+using ElementsProcessed = Steps;
 
-void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed)
+inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_UNUSED(reshape_info);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
 
     if(!is_interleaved_transposed)
     {
-        ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
+        ARM_COMPUTE_ERROR_ON(input0->dimension(0) != input1->dimension(1));
+
+        if(output->total_size() != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != output->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != output->dimension(1));
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+        }
+    }
+    else
+    {
+        const int m                         = reshape_info.m();
+        const int n                         = reshape_info.n();
+        const int k                         = reshape_info.k();
+        const int mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
+        const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
+
+        TensorShape tensor_shape0{ input0->tensor_shape() };
+        tensor_shape0.set(0, k);
+        tensor_shape0.set(1, m);
+
+        TensorShape tensor_shape1{ input1->tensor_shape() };
+        tensor_shape1.set(0, n);
+        tensor_shape1.set(1, k);
+
+        const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
+        const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
+
+        const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+        const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
+
+        if(output->total_size() != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != static_cast<size_t>(n));
+            ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m));
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output);
+        }
+    }
+
+    return Status{};
+}
+
+inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output,
+                                                               bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
+                                                               GPUTarget gpu_target, ElementsProcessed &num_elements_processed)
+{
+    ARM_COMPUTE_UNUSED(gpu_target);
+
+    // Output tensor auto inizialitation if not yet initialized
+    TensorShape tensor_shape{ input0->tensor_shape() };
+    tensor_shape.set(0, is_interleaved_transposed ? reshape_info.n() : input1->dimension(0));
+    tensor_shape.set(1, is_interleaved_transposed ? reshape_info.m() : input0->dimension(1));
+
+    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(tensor_shape));
+
+    bool   window_changed = false;
+    Window win{};
+
+    const DataType data_type                           = input0->data_type();
+    unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
+    unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
+
+    if(is_interleaved_transposed)
+    {
+        // Configure window kernel
+        num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(data_type);
+        num_elems_processed_per_iteration_y = 4;
+
+        win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+        AccessWindowRectangle input0_access(input0, 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f);
+        AccessWindowTranspose input1_access(input1, 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f);
+        AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+
+        update_window_and_padding(win, input0_access, input1_access, output_access);
+
+        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+    else // The input tensors have not been reshaped
+    {
+        // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor
+
+        switch(data_type)
+        {
+            case DataType::F16:
+                num_elems_processed_per_iteration_x = 4;
+                num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4);
+                break;
+
+            case DataType::F32:
+                num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(data_type);
+                num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4);
+                break;
+
+            default:
+                ARM_COMPUTE_ERROR("Current data type is not supported");
+                break;
+        }
+
+        win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+        AccessWindowStatic    input0_access(input0, 0, 0, ceil_to_multiple(input0->dimension(0), 8), ceil_to_multiple(input0->dimension(1), num_elems_processed_per_iteration_y));
+        AccessWindowStatic    input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1));
+        AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+
+        update_window_and_padding(win, input0_access, input1_access, output_access);
+
+        Coordinates coord;
+        coord.set_num_dimensions(output->num_dimensions());
+        output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape()));
     }
 
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+GCGEMMMatrixMultiplyKernel::GCGEMMMatrixMultiplyKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+{
+}
+
+void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info));
+
     _input0 = input0;
     _input1 = input1;
     _output = output;
 
+    ElementsProcessed num_elements_processed{};
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info, GPUTarget::UNKNOWN, num_elements_processed);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    IGCKernel::configure(win_config.second);
+
+    // Create build options
     std::set<std::string> build_opts;
+    std::string           kernel_name;
     Window                win;
 
     build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
@@ -74,6 +217,12 @@ void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTen
     // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
     if(is_interleaved_transposed)
     {
+        const int mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
+        const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
+
+        build_opts.emplace("#define MULT_TRANSPOSE1XW_WIDTH " + support::cpp11::to_string(mult_transpose1xW_width));
+        build_opts.emplace("#define MULT_INTERLEAVE4X4_HEIGHT " + support::cpp11::to_string(mult_interleave4x4_height));
+
         switch(input0->info()->data_type())
         {
             case DataType::F16:
@@ -91,56 +240,20 @@ void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTen
 
         build_opts.emplace("#define GEMM_MM_INTERLEAVED_TRANSPOSED");
 
-        // Create kernel
-        _kernel = GCKernelLibrary::get().create_kernel(("gemm_mm_interleaved_transposed"), build_opts);
-
-        // Configure window kernel
-        const unsigned int     num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type());
-        constexpr unsigned int num_elems_processed_per_iteration_y = 4;
-
-        win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
-        AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f);
-        AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f);
-        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
-
-        update_window_and_padding(win, input0_access, input1_access, output_access);
-
-        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+        kernel_name = "gemm_mm_interleaved_transposed";
     }
     else
     {
-        ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
-
         // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor
-        unsigned int num_elems_processed_per_iteration_x;
-        unsigned int num_elems_processed_per_iteration_y;
 
         switch(input0->info()->data_type())
         {
             case DataType::F16:
                 build_opts.emplace("#define DATA_TYPE_FP16");
-
-#define MM_PROCESS_4X_OPTIMIZED
-
-#if defined(MM_PROCESS_4X)
-                num_elems_processed_per_iteration_x = 4;
-                num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
-                build_opts.emplace("#define MM_PROCESS_4X");
-#elif defined(MM_PROCESS_4X_OPTIMIZED) /* MM_PROCESS_4X */
-                num_elems_processed_per_iteration_x = 4;
-                num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
                 build_opts.emplace("#define MM_PROCESS_4X_OPTIMIZED");
-#elif defined(MM_PROCESS_8X)           /* MM_PROCESS_4X */
-                num_elems_processed_per_iteration_x = 8;
-                num_elems_processed_per_iteration_y = 1;
-                build_opts.emplace("#define MM_PROCESS_8X");
-#endif                                 /* MM_PROCESS_4X */
                 break;
 
             case DataType::F32:
-                num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type());
-                num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
                 build_opts.emplace("#define DATA_TYPE_FP32");
                 break;
 
@@ -150,31 +263,31 @@ void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTen
         }
 
         build_opts.emplace("#define GEMM_MM_FLOATING_POINT");
-        build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_X " + support::cpp11::to_string(num_elems_processed_per_iteration_x));
-        build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_Y " + support::cpp11::to_string(num_elems_processed_per_iteration_y));
-
-        // Create kernel
-        _kernel = GCKernelLibrary::get().create_kernel("gemm_mm_floating_point", build_opts);
+        build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_X " + support::cpp11::to_string(num_elements_processed.x()));
+        build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_Y " + support::cpp11::to_string(num_elements_processed.y()));
 
-        win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
-#if defined(MM_PROCESS_4X_OPTIMIZED)
-        AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), 8), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y));
-#else  /* MM_PROCESS_4X_OPTIMIZED */
-        AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), num_elems_processed_per_iteration_x), ceil_to_multiple(input0->info()->dimension(1),
-                                         num_elems_processed_per_iteration_y));
-#endif /* MM_PROCESS_4X_OPTIMIZED */
-        AccessWindowStatic    input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1));
-        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
-
-        update_window_and_padding(win, input0_access, input1_access, output_access);
-
-        Coordinates coord;
-        coord.set_num_dimensions(output->info()->num_dimensions());
-        output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
+        kernel_name = "gemm_mm_floating_point";
     }
 
-    IGCKernel::configure(win);
+    // Create kernel
+    _kernel = GCKernelLibrary::get().create_kernel(kernel_name, build_opts);
+}
+
+Status GCGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed,
+                                            const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target)
+{
+    ARM_COMPUTE_UNUSED(alpha);
+    ElementsProcessed num_elements_processed{};
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, is_interleaved_transposed, reshape_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
+                                                              input1->clone().get(),
+                                                              output->clone().get(),
+                                                              is_interleaved_transposed,
+                                                              reshape_info,
+                                                              gpu_target,
+                                                              num_elements_processed)
+                                .first);
+    return Status{};
 }
 
 void GCGEMMMatrixMultiplyKernel::run(const Window &window)
diff --git a/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp
index 4c08873dcf..55bf9b754b 100644
--- a/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp
@@ -31,11 +31,13 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
 
 using namespace arm_compute;
 using namespace arm_compute::gles_compute;
+using namespace arm_compute::misc::shape_calculator;
 
 GCWeightsReshapeKernel::GCWeightsReshapeKernel()
     : _input(nullptr), _biases(nullptr), _output(nullptr)
@@ -47,15 +49,8 @@ void GCWeightsReshapeKernel::configure(const IGCTensor *input, const IGCTensor *
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
-    // Calculate output shape
-    TensorShape output_shape{ input->info()->tensor_shape() };
-    output_shape.collapse(3);
-    const size_t tmp_dim = output_shape[0];
-    output_shape.set(0, output_shape[1]);
-    output_shape.set(1, tmp_dim + (biases != nullptr ? 1 : 0));
-
     // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_weights_reshaped_shape(*input->info(), (biases != nullptr))));
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
diff --git a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
index b1c8665216..dc73eb85e6 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
@@ -37,14 +37,14 @@
 using namespace arm_compute;
 
 GCConvolutionLayerReshapeWeights::GCConvolutionLayerReshapeWeights()
-    : _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false)
+    : _weights_reshape_kernel(), _weights_reshaped()
 {
 }
 
-void GCConvolutionLayerReshapeWeights::configure(const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, bool transpose1xW)
+void GCConvolutionLayerReshapeWeights::configure(const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
     if(biases != nullptr)
@@ -56,75 +56,62 @@ void GCConvolutionLayerReshapeWeights::configure(const IGCTensor *weights, const
     }
 
     const bool       append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
-    const unsigned   bias_element  = (append_biases) ? 1 : 0;
     const IGCTensor *biases_to_use = (append_biases) ? biases : nullptr;
 
-    _transpose1xW = transpose1xW;
-
-    if(transpose1xW)
-    {
-        // Create tensor to store the reshaped weights
-        const unsigned int mat_weights_cols = weights->info()->dimension(3);
-        const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
-        TensorShape        shape_wr(mat_weights_cols, mat_weights_rows);
-        const DataType     dt                   = weights->info()->data_type();
-        const int          fixed_point_position = weights->info()->fixed_point_position();
-        TensorInfo         info_wr(shape_wr, 1, dt, fixed_point_position);
-
-        _weights_reshaped.allocator()->init(info_wr);
-        _weights_reshape_kernel.configure(weights, biases_to_use, &_weights_reshaped);
-        _weights_transposed_kernel.configure(&_weights_reshaped, output);
-        _weights_reshaped.allocator()->allocate();
-    }
-    else
-    {
-        _weights_reshape_kernel.configure(weights, biases_to_use, output);
-    }
+    _weights_reshape_kernel.configure(weights, biases_to_use, output);
 }
 
 void GCConvolutionLayerReshapeWeights::run()
 {
     GCScheduler::get().dispatch(_weights_reshape_kernel);
-    if(_transpose1xW)
-    {
-        GCScheduler::get().dispatch(_weights_transposed_kernel);
-    }
 }
 
 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(), _activationlayer_function(),
-      _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), _is_activationlayer_enabled(false)
+    : _memory_group(std::move(memory_manager)), _reshape_weights(), _input_im2col_kernel(), _mm_gemm(), _output_col2im_kernel(), _fill_border(), _activationlayer_function(), _original_weights(nullptr),
+      _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _is_first_run(true), _is_activationlayer_enabled(false)
 {
 }
 
-void GCConvolutionLayer::configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output, bool is_interleaved_transposed)
+void GCConvolutionLayer::configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output)
 {
-    _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), output->info()));
+
+    _mm_gemm.configure(input, weights, nullptr, output, 1.f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */));
+}
+
+Status GCConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output)
+{
+    // Perform validation step on Matrix multiply function
+    GCGEMM::validate(input, weights, nullptr, output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */));
+    return Status{};
 }
 
 void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
                                    const Size2D &dilation, const ActivationLayerInfo &act_info)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!");
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2));
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
+    _is_first_run     = true;
+    _original_weights = weights;
+
     if(biases != nullptr)
     {
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
-        ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
+        ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
 
     const DataType dt = input->info()->data_type();
 
-    _append_bias          = (biases != nullptr);
-    _are_weights_reshaped = weights_info.are_reshaped();
-
-    const unsigned   bias_element  = (_append_bias) ? 1 : 0;
-    const IGCTensor *biases_to_use = (_append_bias) ? biases : nullptr;
+    const bool       append_bias   = (biases != nullptr);
+    const unsigned   bias_element  = (append_bias) ? 1 : 0;
+    const IGCTensor *biases_to_use = (append_bias) ? biases : nullptr;
 
     // Get parameters from conv_info
     unsigned int stride_x = 0;
@@ -135,57 +122,19 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
 
-    const unsigned int kernel_width  = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0);
-    const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
+    const unsigned int kernel_width  = weights->info()->dimension(0);
+    const unsigned int kernel_height = weights->info()->dimension(1);
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
                                                  conv_info, dilation);
 
-    // Check if its a "fully connected" convolution
-    _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
-    const bool run_interleaved      = (!_is_fully_connected_convolution);
-
     unsigned int mat_weights_cols = weights->info()->dimension(3);
     unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
 
-    // Reshape weights if needed
-    if(_are_weights_reshaped)
-    {
-        if(_is_fully_connected_convolution)
-        {
-            mat_weights_cols = weights->info()->dimension(0);
-            mat_weights_rows = weights->info()->dimension(1);
-        }
-        else
-        {
-            mat_weights_cols                         = weights_info.num_kernels();
-            const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4;
-            mat_weights_rows                         = quarter_reshaped_cols + bias_element;
-        }
-    }
-    else
-    {
-        if(_is_fully_connected_convolution)
-        {
-            // Create tensor to store the reshaped weights
-            int num_elems_read_per_iteration_x = 1;
-            if(dt == DataType::F16)
-            {
-                num_elems_read_per_iteration_x = 2;
-            }
-            TensorShape shape_wr((ceil_to_multiple(mat_weights_cols, num_elems_read_per_iteration_x)), mat_weights_rows);
-            _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wr));
-            _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, false /* 1xW transpose */);
-        }
-        else
-        {
-            // Create tensor to store transposed weights
-            const float transpose_width = 16.0f / input->info()->element_size();
-            TensorShape shape_wt(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
-            _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wt));
-            _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, true /* 1xW transpose */);
-        }
-        weights = &_weights_reshaped;
-    }
+    // _weights_reshaped will be auto configured in the kernel.
+    // Just append biases and do not transpose 1xW as it will be reshaped in GCGEMM
+    _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped);
+
+    weights = &_weights_reshaped;
 
     // Create tensor to store im2col reshaped inputs
     const unsigned int mat_input_cols = mat_weights_rows;
@@ -200,19 +149,6 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     _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)
-    {
-        TensorShape shape_interleaved = shape_im2col;
-        shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-
-        // 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
     TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
     shape_gemm.set(0, mat_weights_cols);
@@ -224,26 +160,18 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     _gemm_output.allocator()->init(info_gemm);
     _memory_group.manage(&_gemm_output);
 
-    // Configure kernels
     if(dt == DataType::F16)
     {
         BorderSize border_size = BorderSize(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
         input->info()->extend_padding(border_size);
         _fill_border.configure(input, border_size, BorderMode::CONSTANT, PixelValue(0)); // for PAD of im2col fp16: consider it as border
     }
-    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias, dilation);
+    // Configure im2col
+    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation);
+
+    // Configure GEMM
+    configure_mm(&_input_im2col_reshaped, weights, &_gemm_output);
 
-    // Configure matrix multiply
-    if(run_interleaved)
-    {
-        _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
-        configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output);
-        _input_interleaved_reshaped.allocator()->allocate();
-    }
-    else
-    {
-        configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, false);
-    }
     _input_im2col_reshaped.allocator()->allocate();
 
     // Configure Col2Im
@@ -253,10 +181,7 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
 
     // Allocate intermediate tensor
-    if(!_are_weights_reshaped)
-    {
-        _weights_reshaped.allocator()->allocate();
-    }
+    _weights_reshaped.allocator()->allocate();
 
     //Configure Activation Layer
     _is_activationlayer_enabled = act_info.enabled();
@@ -265,15 +190,22 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
     {
         _activationlayer_function.configure(output, nullptr, act_info);
     }
+
+    ARM_COMPUTE_UNUSED(weights_info);
 }
 
 void GCConvolutionLayer::run()
 {
     // Run weights reshaping (Runs once for every configure)
-    if(!_are_weights_reshaped)
+    if(_is_first_run)
     {
-        _are_weights_reshaped = true;
+        ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
         _reshape_weights.run();
+        _is_first_run = false;
+
+        // Mark original weights tensor as unused
+        _original_weights->mark_as_unused();
     }
 
     _memory_group.acquire();
@@ -283,16 +215,8 @@ void GCConvolutionLayer::run()
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_input_im2col_kernel);
 
-    if(!_is_fully_connected_convolution)
-    {
-        GCScheduler::get().memory_barrier();
-        // Run interleave4x4
-        GCScheduler::get().dispatch(_input_interleave_kernel);
-    }
-
-    GCScheduler::get().memory_barrier();
-    // Runs matrix multiply on reshaped matrices
-    GCScheduler::get().dispatch(_mm_kernel);
+    // Run gemm on reshaped matrices
+    _mm_gemm.run();
 
     GCScheduler::get().memory_barrier();
     // Reshape output matrix
diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index 9c8568a329..0a75a38c50 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -40,62 +40,82 @@
 using namespace arm_compute;
 using namespace arm_compute::gles_compute;
 
-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)
+namespace
 {
-}
-
-void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
+Status validate_arguments(const ITensorInfo *a, const ITensorInfo *b, const IGCTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info = GEMMInfo())
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
+
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
     ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
     ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
-    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.reshape_b_only_on_first_run(), "Reshape matrix B only on first run is not supported");
-    ARM_COMPUTE_UNUSED(gemm_info);
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
-        ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
-        ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix C");
-        ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(0) != output->info()->dimension(0), "The C matrix must have the same number of rows as the output matrix");
-        ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != output->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info());
+        ARM_COMPUTE_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
+        ARM_COMPUTE_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B");
     }
 
-    ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A");
+    }
 
-    // If the input tensor has less than 16 rows, we run a special version of GEMM without reshaping the input tensors
-    _is_interleaved_transposed = a->info()->dimension(1) > 16;
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_UNUSED(gemm_info);
+    return Status{};
+}
+} // namespace
+
+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),
+      _is_first_run(true), _reshape_b_only_on_first_run(false)
+{
+}
+
+void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(a->info(), b->info(), c, output->info(), alpha, beta, gemm_info));
+
+    // Check if we need to reshape the matrix B only on the first run
+    _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
 
     const IGCTensor *matrix_a = a;
     const IGCTensor *matrix_b = b;
 
+    // Arguments used by GEMMReshapeInfo
+    // If we pass the matrix A and matrix B reshaped to GCGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to GCGEMMReshapeInfo
+    // in order to know how the matrices have been reshaped
+    const int m                         = a->info()->dimension(1);
+    const int n                         = b->info()->dimension(0);
+    const int k                         = a->info()->dimension(0);
+    int       mult_transpose1xW_width   = 1;
+    int       mult_interleave4x4_height = 1;
+
+    // If the input tensor has less than 16 rows, we run a special version of GEMM without reshaping the input tensors
+    _is_interleaved_transposed = a->info()->dimension(1) > 16;
+
     if(_is_interleaved_transposed)
     {
         matrix_a = &_tmp_a;
         matrix_b = &_tmp_b;
 
-        TensorShape shape_tmp_a = a->info()->tensor_shape();
-        TensorShape shape_tmp_b = b->info()->tensor_shape();
-
-        shape_tmp_a.set(0, a->info()->dimension(0) * 4);
-        shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));
-
-        const unsigned int transpose_w = max_gc_vector_width / data_size_from_type(b->info()->data_type());
-        shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
-        shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));
-
-        TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
-        _tmp_a.allocator()->init(info_a);
+        // Manage intermediate buffers
         _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);
-        if(!gemm_info.reshape_b_only_on_first_run())
+        if(!_reshape_b_only_on_first_run)
         {
             _memory_group.manage(&_tmp_b);
         }
+        // _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel
 
         // Configure interleave kernel
         _interleave_kernel.configure(a, &_tmp_a);
@@ -104,7 +124,7 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
         _transpose_kernel.configure(b, &_tmp_b);
     }
 
-    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed);
+    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height));
 
     if(_is_interleaved_transposed)
     {
@@ -121,6 +141,12 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
     }
 }
 
+Status GCGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const IGCTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(a, b, c, output, alpha, beta, gemm_info));
+    return Status{};
+}
+
 void GCGEMM::run()
 {
     _memory_group.acquire();
@@ -129,8 +155,17 @@ void GCGEMM::run()
         // Run interleave kernel
         GCScheduler::get().dispatch(_interleave_kernel, false);
 
-        // Run transpose kernel
-        GCScheduler::get().dispatch(_transpose_kernel, false);
+        if(_is_first_run)
+        {
+            // Run transpose kernel
+            GCScheduler::get().dispatch(_transpose_kernel, false);
+            _is_first_run = false;
+        }
+        else if(!_reshape_b_only_on_first_run)
+        {
+            // Run transpose kernel
+            GCScheduler::get().dispatch(_transpose_kernel, false);
+        }
         GCScheduler::get().memory_barrier();
     }