Method execute

src/jit/impl/mlir/executable_cpu.cpp:76–119

}

void MLIRCPUExecutable::execute(JITExecutor* fusion_opr) {
    auto&& args = fusion_opr->args();
    std::vector<void*> args_array(args.inputs.size() + args.outputs.size());
    std::vector<void*> args_array_pointer(args.inputs.size() + args.outputs.size());
    size_t idx = 0;
    for (size_t i = 0; i < args.inputs.size(); i++) {
        args_array[idx] = tensor2memref(
                {args.inputs[i].from->dev_tensor().raw_ptr(), args.inputs[i].layout});
        args_array_pointer[idx] = &args_array[idx];
        idx++;
    }
    int64_t nr_elements = 0;
    for (size_t i = 0; i < args.outputs.size(); i++) {
        if (nr_elements == 0) {
            nr_elements = args.outputs[i].layout.total_nr_elems();
        } else {
            mgb_assert(
                    static_cast<size_t>(nr_elements) ==
                            args.outputs[i].layout.total_nr_elems(),
                    "The number of elements of outputs mismatch, expected: "
                    "%zu got: %zu(%s)",
                    static_cast<size_t>(nr_elements),
                    args.outputs[i].layout.total_nr_elems(),
                    args.outputs[i].layout.to_string().c_str());
        }
        args_array[idx] = tensor2memref(
                {args.outputs[i].from->dev_tensor().raw_ptr(), args.outputs[i].layout});
        args_array_pointer[idx] = &args_array[idx];
        idx++;
    }

    args_array_pointer.push_back(&nr_elements);
    std::string adapter_name = std::string("_mlir_ciface_") + m_kernel_name;
    auto err = m_engine->invoke(
            adapter_name, llvm::MutableArrayRef<void*>(args_array_pointer));
    if (err) {
        mgb_throw(
                InternalError, "failed to run MLIR kernel %s\n", m_kernel_name.c_str());
    }

    for (size_t i = 0; i < args_array.size(); i++) {
        free(args_array[i]);
    }
}

MLIRCPUExecutable::~MLIRCPUExecutable() {}