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>

2 files changed, 122 insertions, 163 deletions

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();
     }