COMPMID-1480 Add support for NHWC QASYMM8/FP32(non-optimized) to NEON DepthwiseConvolution

Change-Id: I751f5d3fb74085d2e67f610ecf52da4736d0cfb5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/143870
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 199 insertions, 76 deletions

diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 3b54ed62c7..d1727fc878 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -36,8 +36,8 @@ using namespace arm_compute::misc;
 using namespace arm_compute::misc::shape_calculator;
 
 NEDepthwiseConvolutionLayer3x3::NEDepthwiseConvolutionLayer3x3()
-    : _dwc_kernel(), _output_stage_kernel(), _border_handler(), _permute_input(), _permute_weights(), _permute_output(), _accumulator(), _input_nhwc(), _weights_hwio(), _output_nhwc(), _has_bias(false),
-      _is_quantized(false), _is_optimized(false), _are_weights_reshaped(false), _is_nchw(true), _is_first_run(true)
+    : _dwc_kernel(), _output_stage_kernel(), _border_handler(), _permute_input(), _permute_weights(), _permute_output(), _accumulator(), _permuted_input(), _permuted_weights(), _permuted_output(),
+      _has_bias(false), _is_quantized(false), _is_optimized(false), _are_weights_reshaped(false), _is_nchw(true), _is_first_run(true), _permute(false)
 {
 }
 
@@ -57,29 +57,31 @@ void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *we
                                                                                           input->info()->data_layout());
     _are_weights_reshaped = false;
     _is_nchw              = input->info()->data_layout() == DataLayout::NCHW;
-
-    ARM_COMPUTE_ERROR_ON(!_is_optimized && !_is_nchw);
+    _permute              = _is_optimized == _is_nchw;
 
     if(_is_optimized)
     {
         if(_is_nchw)
         {
             // Configure the function to transform the input tensor from NCHW -> NHWC
-            _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U));
+            _permute_input.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
+            _permuted_input.info()->set_data_layout(DataLayout::NHWC);
 
             // Configure the function to transform the weights tensor from IHW -> HWI
-            _permute_weights.configure(weights, &_weights_hwio, PermutationVector(2U, 0U, 1U));
+            _permute_weights.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
+            _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
 
             // Configure optimized depthwise
-            _dwc_kernel.configure(&_input_nhwc, &_weights_hwio, &_output_nhwc, conv_info, depth_multiplier, DataLayout::NHWC);
+            _dwc_kernel.configure(&_permuted_input, &_permuted_weights, &_permuted_output, conv_info, depth_multiplier, DataLayout::NHWC);
 
             // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
-            _permute_output.configure(&_output_nhwc, output, PermutationVector(1U, 2U, 0U));
+            _permute_output.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
+            _permuted_output.info()->set_data_layout(DataLayout::NCHW);
 
             // Allocate tensors
-            _input_nhwc.allocator()->allocate();
-            _weights_hwio.allocator()->allocate();
-            _output_nhwc.allocator()->allocate();
+            _permuted_input.allocator()->allocate();
+            _permuted_weights.allocator()->allocate();
+            _permuted_output.allocator()->allocate();
         }
         else
         {
@@ -91,36 +93,70 @@ void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *we
         // Allocate the intermediate accumulator tensor in case of quantized input
         if(_is_quantized)
         {
-            _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, DataType::S32));
+            TensorShape accum_shape = output->info()->tensor_shape();
+
+            if(!_is_nchw)
+            {
+                permute(accum_shape, PermutationVector(1U, 2U, 0U));
+            }
+
+            _accumulator.allocator()->init(TensorInfo(accum_shape, 1, DataType::S32));
             _accumulator.info()->set_quantization_info(input->info()->quantization_info());
             zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().offset));
         }
 
-        // Configure depthwise convolution kernel
-        _dwc_kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info, depth_multiplier);
+        if(!_is_nchw)
+        {
+            // Configure the function to transform the input tensor from NHWC -> NCHW
+            _permute_input.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
+            _permuted_input.info()->set_data_layout(DataLayout::NCHW);
 
-        // Configure border handler
-        _border_handler.configure(input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value);
-    }
+            // Configure the function to transform the weights tensor from HWI -> IHW
+            _permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
+            _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
 
-    // Configure biases accumulation
-    if(_has_bias || _is_quantized)
-    {
-        if(_is_quantized)
-        {
-            const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+            // Configure optimized depthwise
+            _dwc_kernel.configure(&_permuted_input, &_permuted_weights, (_is_quantized) ? &_accumulator : &_permuted_output, conv_info, depth_multiplier);
+
+            // Configure border handler
+            _border_handler.configure(&_permuted_input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value);
 
-            float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
-            int   output_multiplier, output_shift;
-            quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-            _output_stage_kernel.configure(&_accumulator, biases, output, output_multiplier, output_shift, output_quant_info.offset);
-            _accumulator.allocator()->allocate();
+            // Allocate tensors
+            _permuted_input.allocator()->allocate();
+            _permuted_weights.allocator()->allocate();
         }
         else
         {
-            _output_stage_kernel.configure(output, biases);
+            // Configure depthwise convolution kernel
+            _dwc_kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info, depth_multiplier);
+
+            // Configure border handler
+            _border_handler.configure(input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value);
         }
     }
+
+    // Configure biases accumulation
+    if(_is_quantized)
+    {
+        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        int   output_multiplier, output_shift;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+        _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, output_quant_info.offset);
+        _accumulator.allocator()->allocate();
+    }
+    else if(_has_bias)
+    {
+        _output_stage_kernel.configure((_is_nchw || _is_optimized) ? output : &_permuted_output, biases);
+    }
+
+    if(!_is_optimized && !_is_nchw)
+    {
+        // Configure the function to transform the convoluted output to NHWC
+        _permute_output.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
+        _permuted_output.allocator()->allocate();
+    }
 }
 
 Status NEDepthwiseConvolutionLayer3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
@@ -142,32 +178,29 @@ void NEDepthwiseConvolutionLayer3x3::run()
         _dwc_kernel.generate_convolver();
     }
 
-    // Permute weights in HWIO format if the optimized kernel will be executedd
-    if(!_are_weights_reshaped && _is_optimized && _is_nchw)
-    {
-        _are_weights_reshaped = true;
-        _permute_weights.run();
-    }
-
-    // Handle input
-    if(_is_optimized)
+    // Permute weights
+    if(_permute)
     {
-        if(_is_nchw)
+        if(!_are_weights_reshaped)
         {
-            // Permute input to NHWC format execution
-            _permute_input.run();
+            _are_weights_reshaped = true;
+            _permute_weights.run();
         }
+
+        _permute_input.run();
     }
-    else
+
+    // Handle input
+    if(!_is_optimized)
     {
-        // Fill border in NCHW format execution
+        // Fill border
         NEScheduler::get().schedule(&_border_handler, Window::DimX);
     }
 
     // Execute depthwise convolution
     NEScheduler::get().schedule(&_dwc_kernel, Window::DimX);
 
-    // Permute output to ACL's native NCHW format in case of NHWC execution
+    // Permute output
     if(_is_optimized && _is_nchw)
     {
         _permute_output.run();
@@ -178,27 +211,54 @@ void NEDepthwiseConvolutionLayer3x3::run()
     {
         NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX);
     }
+
+    // Permute output
+    if(!_is_optimized && !_is_nchw)
+    {
+        _permute_output.run();
+    }
 }
 
 NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayer()
-    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
-      _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_prepared(false), _is_quantized(false), _original_weights(nullptr)
+    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _permute_input(),
+      _permute_weights(), _permute_output(), _input_reshaped(), _weights_reshaped(), _v2mm_output(), _output_reshaped(), _permuted_input(), _permuted_weights(), _permuted_output(), _is_prepared(false),
+      _is_quantized(false), _is_nhwc(false), _original_weights(nullptr)
 {
 }
 
 void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier)
 {
+    const unsigned int channel_idx = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+    ARM_COMPUTE_UNUSED(channel_idx);
+
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_ERROR_ON((input->info()->dimension(2) * depth_multiplier) != weights->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON((input->info()->dimension(channel_idx) * depth_multiplier) != weights->info()->dimension(channel_idx));
+
+    _is_nhwc = input->info()->data_layout() == DataLayout::NHWC;
+
+    ITensor       *input_to_use   = input;
+    const ITensor *weights_to_use = weights;
+    ITensor       *output_to_use  = output;
+
+    if(_is_nhwc)
+    {
+        _permute_input.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
+        _permuted_input.info()->set_data_layout(DataLayout::NCHW);
+        input_to_use = &_permuted_input;
+
+        _permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
+        _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
+        weights_to_use = &_permuted_weights;
+    }
 
-    const size_t weights_w = weights->info()->dimension(0);
-    const size_t weights_h = weights->info()->dimension(1);
-    const size_t weights_z = weights->info()->dimension(2);
+    const size_t weights_w = weights_to_use->info()->dimension(0);
+    const size_t weights_h = weights_to_use->info()->dimension(1);
+    const size_t weights_z = weights_to_use->info()->dimension(2);
 
     _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
     _is_prepared      = false;
-    _original_weights = weights;
+    _original_weights = weights_to_use;
 
     // Should bias be appended ?
     bool append_bias = (biases != nullptr) && !_is_quantized;
@@ -210,6 +270,14 @@ void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weigh
     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
 
+    if(_is_nhwc)
+    {
+        permute(output_shape, PermutationVector(1U, 2U, 0U));
+        _permuted_output.allocator()->init(output->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
+        _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+        output_to_use = &_permuted_output;
+    }
+
     // Output width and height
     const unsigned int conv_w = output_shape.x();
     const unsigned int conv_h = output_shape.y();
@@ -219,41 +287,50 @@ void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weigh
     const size_t conv_size  = conv_w * conv_h;
 
     // Im2Col configuration
-    TensorShape shape_im2col = input->info()->tensor_shape();
+    TensorShape shape_im2col = input_to_use->info()->tensor_shape();
     shape_im2col.set(0, patch_size);
     shape_im2col.set(1, conv_size);
     shape_im2col.set(2, weights_z);
-    _input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
-    _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier);
+    _input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col).set_data_layout(DataLayout::NCHW));
+    _im2col_kernel.configure(input_to_use, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier);
 
     // Weights reshape configuration
     const TensorShape shape_weights_reshape(patch_size, weights_z);
-    _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
-    _weights_reshape_kernel.configure(weights, &_weights_reshaped, append_bias ? biases : nullptr);
+    _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape).set_data_layout(DataLayout::NCHW));
+    _weights_reshape_kernel.configure(weights_to_use, &_weights_reshaped, append_bias ? biases : nullptr);
 
     // GEMV configuration
     DataType    v2mm_dt        = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
-    TensorShape shape_v2mm_out = input->info()->tensor_shape();
+    TensorShape shape_v2mm_out = input_to_use->info()->tensor_shape();
     shape_v2mm_out.set(0, conv_size * weights_z);
     shape_v2mm_out.set(1, 1);
     shape_v2mm_out.set(2, 1);
-    _v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
+    _v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out).set_data_layout(DataLayout::NCHW));
     _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
     _output_reshaped.allocator()->init(_v2mm_output.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
-    _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
+    _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output_to_use, conv_w, conv_h);
 
     // Output staged configuration
     if(_is_quantized)
     {
-        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+        const QuantizationInfo output_quant_info = output->info()->quantization_info();
 
         float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
         int   output_multiplier, output_shift;
         quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-        _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output_quant_info.offset);
+        _output_stage_kernel.configure(&_output_reshaped, biases, output_to_use, output_multiplier, output_shift, output_quant_info.offset);
         _output_reshaped.allocator()->allocate();
     }
 
+    if(_is_nhwc)
+    {
+        _permute_output.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
+
+        _permuted_input.allocator()->allocate();
+        _permuted_weights.allocator()->allocate();
+        _permuted_output.allocator()->allocate();
+    }
+
     // Fill borders on inputs
     PixelValue zero_in(static_cast<int32_t>(0));
     PixelValue zero_w(static_cast<int32_t>(0));
@@ -273,53 +350,84 @@ void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weigh
     _v2mm_output.allocator()->allocate();
 }
 
-Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &conv_info,
                                              unsigned int depth_multiplier)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW && input->data_layout() != DataLayout::NHWC);
 
+    const ITensorInfo *input_to_use   = input;
+    const ITensorInfo *weights_to_use = weights;
+    const ITensorInfo *output_to_use  = output;
+
+    TensorShape permuted_input_shape   = input->tensor_shape();
+    TensorShape permuted_weights_shape = weights->tensor_shape();
+    TensorInfo  permuted_input;
+    TensorInfo  permuted_weights;
+
+    if(input->data_layout() == DataLayout::NHWC)
+    {
+        permute(permuted_input_shape, PermutationVector(1U, 2U, 0U));
+        permute(permuted_weights_shape, PermutationVector(1U, 2U, 0U));
+
+        permuted_input   = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NCHW));
+        permuted_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NCHW));
+
+        input_to_use   = &permuted_input;
+        weights_to_use = &permuted_weights;
+    }
+
     const bool         is_quantized = is_data_type_quantized_asymmetric(input->data_type());
     const bool         append_bias  = (biases != nullptr) && !is_quantized;
-    const TensorShape  output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
-    const size_t       weights_w    = weights->dimension(0);
-    const size_t       weights_h    = weights->dimension(1);
-    const size_t       weights_z    = weights->dimension(2);
+    TensorShape        output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+    const size_t       weights_w    = weights_to_use->dimension(0);
+    const size_t       weights_h    = weights_to_use->dimension(1);
+    const size_t       weights_z    = weights_to_use->dimension(2);
     const unsigned int conv_w       = output_shape.x();
     const unsigned int conv_h       = output_shape.y();
     const size_t       patch_size   = weights_w * weights_h + (append_bias ? 1 : 0);
     const size_t       conv_size    = conv_w * conv_h;
 
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
 
+    TensorInfo permuted_output;
+    if(input->data_layout() == DataLayout::NHWC)
+    {
+        permute(output_shape, PermutationVector(1U, 2U, 0U));
+        permuted_output = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape).set_data_layout(DataLayout::NCHW));
+        output_to_use   = &permuted_output;
+    }
+
     // Im2Col configuration
-    TensorShape shape_im2col = input->tensor_shape();
+    TensorShape shape_im2col = input_to_use->tensor_shape();
     shape_im2col.set(0, patch_size);
     shape_im2col.set(1, conv_size);
     shape_im2col.set(2, weights_z);
-    TensorInfo input_reshaped(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
-    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseIm2ColKernel::validate(input, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier));
+    TensorInfo input_reshaped(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col).set_data_layout(DataLayout::NCHW));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseIm2ColKernel::validate(input_to_use, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier));
 
     // Weights reshape configuration
     const TensorShape shape_weights_reshape(patch_size, weights_z);
-    TensorInfo        weights_reshaped(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
-    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseWeightsReshapeKernel::validate(weights, &weights_reshaped, append_bias ? biases : nullptr));
+    TensorInfo        weights_reshaped(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape).set_data_layout(DataLayout::NCHW));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseWeightsReshapeKernel::validate(weights_to_use, &weights_reshaped, append_bias ? biases : nullptr));
 
     // GEMV configuration
     DataType    v2mm_dt        = (input->data_type() == DataType::QASYMM8) ? DataType::S32 : input->data_type();
-    TensorShape shape_v2mm_out = input->tensor_shape();
+    TensorShape shape_v2mm_out = input_to_use->tensor_shape();
     shape_v2mm_out.set(0, conv_size * weights_z);
     shape_v2mm_out.set(1, 1);
     shape_v2mm_out.set(2, 1);
-    TensorInfo v2mm_output(input->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
+    TensorInfo v2mm_output(input->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out).set_data_layout(DataLayout::NCHW));
     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixVectorMultiplyKernel::validate(&input_reshaped, &weights_reshaped, &v2mm_output));
 
-    TensorInfo output_reshaped(v2mm_output.clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
-    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseVectorToTensorKernel::validate(&v2mm_output, (is_quantized) ? &output_reshaped : output, conv_w, conv_h));
+    TensorInfo output_reshaped(v2mm_output.clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_to_use->tensor_shape()));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseVectorToTensorKernel::validate(&v2mm_output, (is_quantized) ? &output_reshaped : output_to_use, conv_w, conv_h));
 
     if(is_quantized)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&output_reshaped, biases, output));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&output_reshaped, biases, output_to_use));
     }
 
     return Status{};
@@ -329,6 +437,11 @@ void NEDepthwiseConvolutionLayer::run()
 {
     prepare();
 
+    if(_is_nhwc)
+    {
+        _permute_input.run();
+    }
+
     NEScheduler::get().schedule(&_im2col_kernel, Window::DimX);
     NEScheduler::get().schedule(&_v2mm_input_fill_border, Window::DimX);
     NEScheduler::get().schedule(&_v2mm_kernel, Window::DimX);
@@ -337,6 +450,11 @@ void NEDepthwiseConvolutionLayer::run()
     {
         NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX);
     }
+
+    if(_is_nhwc)
+    {
+        _permute_output.run();
+    }
 }
 
 void NEDepthwiseConvolutionLayer::prepare()
@@ -345,6 +463,11 @@ void NEDepthwiseConvolutionLayer::prepare()
     {
         ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 
+        if(_is_nhwc)
+        {
+            _permute_weights.run();
+        }
+
         // Run reshape and mark original weights as unused
         _weights_reshaped.allocator()->allocate();
         NEScheduler::get().schedule(&_weights_reshape_kernel, Window::DimX);