Method configure

src/gpu/cl/operators/ClFullyConnected.cpp:330–462

}

void ClFullyConnected::configure(const CLCompileContext &compile_context,
                                 ITensorInfo            *src,
                                 ITensorInfo            *weights,
                                 ITensorInfo            *biases,
                                 ITensorInfo            *dst,
                                 FullyConnectedLayerInfo fc_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
    const GPUTarget gpu_target = get_arch_from_target(CLScheduler::get().target());

    // Perform validate step
    ARM_COMPUTE_ERROR_THROW_ON(ClFullyConnected::validate(src, weights, biases, dst, fc_info));
    ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, fc_info);

    _transpose_weights  = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
    _is_fc_after_conv   = true;
    _is_quantized       = is_data_type_quantized_asymmetric(src->data_type());
    _is_prepared        = fc_info.retain_internal_weights;
    _weights_to_use     = TensorInfo(*weights);
    _weights_to_use_idx = ACL_SRC_1;

    // When using dynamic weights - use matmul kernels.
    // Note: MatMul is not used in the following cases (Gemm is used as fallback) :
    // 1. When the weights tensor is not dynamic
    // 2. MatMul does not support broadcasting batch dimension, and therefore is disabled if fc is batched.
    // 3. When FC is after convolution and src tensor data layout does not match weights trained data layout (weights conversion kernel is required)
    const bool is_batched_fc_layer = dst->dimension(1) > 1;
    _use_matmul = gpu_target != GPUTarget::MIDGARD && !weights->are_values_constant() && !is_batched_fc_layer &&
                  !(src->num_dimensions() > 1 && (src->data_layout() != fc_info.weights_trained_layout));
    _dynamic_gemm = !weights->are_values_constant() && _transpose_weights && !_use_matmul;

    // With the Fully Connected layer we can have 4 different cases:
    //  1) Convolution layer -> Fully Connected layer without batches
    //  2) Fully Connected layer -> Fully Connected layer without batches
    //  3) Convolution layer -> Fully Connected layer with batches
    //  4) Fully Connected layer -> Fully Connected layer with batches

    // Check if we have a fully connected layer with batches
    if (is_batched_fc_layer)
    {
        _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
                            (std::equal(src->tensor_shape().cbegin() + 3, src->tensor_shape().cend(),
                                        dst->tensor_shape().cbegin() + 1));
    }
    else
    {
        _is_fc_after_conv = src->num_dimensions() > 1;
    }

    ITensorInfo *weights_used = weights;

    // Reshape weights if needed - Not needed when matmul is in use as matmul fuses transpose op.
    if (_transpose_weights && !_use_matmul)
    {
        // Reshape the weights
        _reshape_weights = std::make_unique<ClTranspose>();
        _reshape_weights->configure(compile_context, weights, &_reshaped_weights);
        weights_used        = &_reshaped_weights;