Result of Deconvolution not correct

[daoxian]

Output of 'strings libarm_compute.so | grep arm_compute_version': arm_compute_version=v22.08 Build options: {'compiler_prefix': '/usr/bin/', 'toolchain_prefix': '/usr/bin/', 'Werror': '0', 'debug': '1', 'neon': '1', 'opencl': '0', 'os': 'linux', 'arch': 'armv8.2-a', 'benchmark_tests': '0', 'validation_tests': '0', 'examples': '1', 'extra_cxx_flags': '-fPIC'} Git hash=unknown

Platform: Arm v8.2

Operating System: CentOS 7

Problem description: I cannot get the correct result when using NEDevolutionLayer with stride=2 and pad=1 :

        int stride=2, pad=1;
        npy0.init_tensor(src0, DataType::F32);
        src0.allocator()->allocate();
        npy0.fill_tensor(src0);

        npy1.init_tensor(src1, DataType::F32);
        src1.allocator()->allocate();
        npy1.fill_tensor(src1);

        const arm_compute::PadStrideInfo padstride_info(stride, stride, pad, pad, pad, pad, arm_compute::DimensionRoundingType::FLOOR);
        auto out_dim = arm_compute::deconvolution_output_dimensions(src0.info()->tensor_shape().x(), src0.info()->tensor_shape().y(), src1.info()->tensor_shape().x(), src1.info()->tensor_shape().y(), padstride_info);
        TensorShape output_shape = arm_compute::misc::shape_calculator::compute_deconvolution_output_shape(out_dim, *src0.info(), *src1.info());
        dst.allocator()->init(TensorInfo(output_shape, 1, DataType::F32));
        dst.allocator()->allocate();
        deconv.configure(&src0, &src1, nullptr, &dst, padstride_info);

The input and output data are as follows:

data.zip All data layouts are NCHW. Any clues will be appreciated! Thanks!

[daoxian]

I've seen the reason: weight (src1) should be in the [width, height, IFM, OFM] layout.