Method validate

src/cpu/operators/CpuGemm.cpp:216–428

}

Status CpuGemm::validate(const ITensorInfo *a,
                         const ITensorInfo *b,
                         const ITensorInfo *c,
                         const ITensorInfo *d,
                         float              alpha,
                         float              beta,
                         const GEMMInfo    &gemm_info)
{
    ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, "CpuGemm::validate");
    ARM_COMPUTE_UNUSED(alpha);
    ARM_COMPUTE_RETURN_ERROR_ON_SIZE_UNSUPPORTED(a, b, c);

    // When using accumulation(in place summation), for now, the only supported values for alpha and beta are 1 respectively 0.
    // Do the appropriate checks before proceeding.
    if (gemm_info.accumulate())
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(alpha != 1, "Accumulation is not supported when alpha is different from 1");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
            (beta != 0 && c != nullptr),
            "Accumulation is not supported when beta is different from 0 with a non-null bias matrix c");
    }

    const bool is_c_bias    = beta == 1 && c != nullptr;
    const bool run_addition = c != nullptr && beta != 0 && beta != 1;
    // Check if we should use the pretransposed_b or original b
    // TODO: COMPMID-6597
    // Note that this check should only apply to the non-optimized path. The reason we brought this at the beginning
    // instead of only for the fallback path is because of the checks performed below, between here and the run_optimised decision
    // We should simplify this by
    //   1. Moving the checks between "fix-start" and "fix-end" into their corresponding ops / kernels (e.g. the weights format checks can and should be moved into CpuGemmAssemblyDispatch)
    //   2. Moving this b_to_use check back into the non-optimized path
    TensorInfo pretransposed_b  = b->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*b));
    const ITensorInfo *b_to_use = gemm_info.pretranspose_B() ? &pretransposed_b : b;

    // TODO: COMPMID-6597 fix-start

    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(a);
    ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(a);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::BFLOAT16, DataType::F16, DataType::F32);

    if (is_fixed_format_fast_math(gemm_info.weight_format()))
    {
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(a, DataType::F32);
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(b_to_use, DataType::BFLOAT16);
    }
    else
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b_to_use);
    }

    const int block_by = arm_compute::block_by(gemm_info.weight_format());
    // test if im2col has changed the dimensions that are needed for padding
    if (a->dimension(0) != b_to_use->dimension(1) && block_by > 1)
    {
        // have to verify bias
        const size_t dim0_sz = a->dimension(0);
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
            (dim0_sz % block_by) != 0,