Function mul_q8_sme_possible

src/cpu/kernels/CpuMulKernel.cpp:339–375

}

bool mul_q8_sme_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, float scale)
{
    ARM_COMPUTE_UNUSED(src0, src1, dst, scale);
#ifdef ARM_COMPUTE_ENABLE_SME2
    const auto        &in0_shape = src0->tensor_shape();
    const auto        &in1_shape = src1->tensor_shape();
    const unsigned int dst_dims  = dst->num_dimensions();

    // Calculate Scale
    const auto iq0        = src0->quantization_info().uniform();
    const auto iq1        = src1->quantization_info().uniform();
    const auto oq         = dst->quantization_info().uniform();
    const auto multiplier = ((iq0.scale * iq1.scale) / oq.scale) * scale;
    const auto max_result = multiplier * (127) * (127) + static_cast<float>(oq.offset);
    const auto min_result = multiplier * (-128) * (-128) + static_cast<float>(oq.offset);

    // Does not support broadcasting on x
    // Does not support dims > 4D output, unless input shapes are identical (therefore collapsible)
    // Checks whether CPU has SME2 Available
    if (in0_shape.x() == in1_shape.x() && CPUInfo::get().has_sme2() && (in0_shape == in1_shape || dst_dims <= 4))
    {
        // Check if multiplier cannot be stored as a 14.18 signed fixed-point number
        if (multiplier < -8191.f || multiplier > 8191.f)
        {
            return false;
        }
        // It might not be possible to store the result as a 14.18 signed fixed-point number.
        if (max_result > 8191.f || min_result < -8191.f)
        {
            return false;
        }
        // Passed all checks
        return true;
    }
#endif // ARM_COMPUTE_ENABLE_SME2
    return false;
}

bool mul_q8_neon_fixedpoint_possible(const ITensorInfo *src0,
                                     const ITensorInfo *src1,