COMPMID-810 Add NHWC data format support for NEON convolution

Change-Id: I2a7b49a12da7f3bc3f04749243b1dc111160de6e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129348
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

2 files changed, 17 insertions, 8 deletions

diff --git a/arm_compute/runtime/NEON/AssemblyHelper.h b/arm_compute/runtime/NEON/AssemblyHelper.h
index 3db419e148..ecaf35ac3e 100644
--- a/arm_compute/runtime/NEON/AssemblyHelper.h
+++ b/arm_compute/runtime/NEON/AssemblyHelper.h
@@ -84,7 +84,12 @@ public:
         const int ldb = _b->info()->strides_in_bytes().y() / sizeof(TypeInput);
         const int ldd = _d->info()->strides_in_bytes().y() / sizeof(TypeOutput);
 
-        const int batch_stride_a = _a->info()->strides_in_bytes().z() / sizeof(TypeInput);
+        // In the case of NHWC we want to interpret the output shape as 3D. Thus, the batch stride for A is
+        // the relevant multiple of the row stride.
+        const bool is_nhwc           = _a->info()->data_layout() == DataLayout::NHWC;
+        const int  stride_in_bytes_a = is_nhwc ? _a->info()->strides_in_bytes().y() * _d->info()->dimension(1) : _a->info()->strides_in_bytes().z();
+
+        const int batch_stride_a = stride_in_bytes_a / sizeof(TypeInput);
         const int batch_stride_d = _d->info()->strides_in_bytes().z() / sizeof(TypeOutput);
 
         const int multi_stride_a = _a->info()->strides_in_bytes()[3] / sizeof(TypeInput);
@@ -158,7 +163,7 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d
     const int      M           = d->info()->tensor_shape().y();
     const int      N           = d->info()->tensor_shape().x();
     const int      K           = a->info()->tensor_shape().x();
-    const int      batches     = a->info()->tensor_shape().total_size_upper(2);
+    const int      batches     = d->info()->tensor_shape().total_size_upper(2);
     const int      multis      = b->info()->tensor_shape().z();
     unsigned int   num_threads = NEScheduler::get().num_threads();
 
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 752693188c..d64fd9e771 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -26,6 +26,7 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
+#include "arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h"
 #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h"
@@ -176,6 +177,7 @@ private:
     NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
     NECol2ImKernel                                      _output_col2im_kernel;
     NEActivationLayer                                   _activationlayer_function;
+    NEArithmeticAdditionKernel                          _add_bias_kernel;
 
     const ITensor *_original_weights;
 
@@ -187,12 +189,14 @@ private:
     Tensor _workspace;
     Tensor _B_pretransposed;
 
-    bool _append_bias;
-    bool _is_fully_connected_convolution;
-    bool _are_weights_reshaped;
-    bool _is_quantized;
-    bool _is_interleaved;
-    bool _is_activationlayer_enabled;
+    DataLayout _data_layout;
+    bool       _append_bias;
+    bool       _is_fully_connected_convolution;
+    bool       _are_weights_reshaped;
+    bool       _is_quantized;
+    bool       _is_interleaved;
+    bool       _is_activationlayer_enabled;
+    bool       _skip_im2col;
 };
 }
 #endif /* __ARM_COMPUTE_NECONVOLUTIONGEMMLAYER_H__ */