Fix CLDepthwiseConvolutionLayer 3x3 QASYMM8

Fix errors when computing tensors with one element only

- Replace Tensor3D with raw pointers so to get rid of offset to first element for NCHW layout
- Add stronger out of bound constraints for NHWC layout
- Set the border size to the input's padding for NHWC
- Fill the strides == 0 with the largest stride, so to avoid accessing empty strides and multiplying by 0

Resolve COMPMID-4088

Change-Id: I751a4e6d7094b3c42306ff7f53af848fd35f19ac
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5024
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 3 insertions, 3 deletions

diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
index 1f296f8e26..de986de9f6 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -352,8 +352,8 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::Co
 
     if(_input1->info()->num_dimensions() < 3)
     {
-        // The stride_z for matrix B must be zero if we do not slice
-        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
+        // The stride_w for matrix B must be the same as stride_z if we do not slice
+        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != _input1->info()->strides_in_bytes()[2]);
     }
 
     const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u;