Shared Graph-Scoped Tables - 2025.2 English - UG1079

Sometimes, multiple kernels use the same table definition. The AI Engine architecture is a distributed address-space architecture. This means each processor binary image that executes such a kernel needs to have that table defined in its own local memory. To ensure correct graph linkage spread across multiple processors, declare the tables as extern in both the kernel source file and the graph class definition file. Then, specify the table definition in a separate header file that is attached as a property to the kernel as shown below.

#include <adf.h>

extern int16 lutarray[8]; // extern in graph context for parameter::array binding

class shared_table_graph : public adf::graph {
public:
  adf::kernel k1;
  adf::kernel k2;

  // One parameter per kernel consumer so we can place each in the correct tile
  adf::parameter p1;
  adf::parameter p2;

  adf::input_plio in1, in2;
  adf::output_plio out1, out2;

  shared_table_graph() {
    // Create kernels
    k1 = adf::kernel::create(simple1);
    k2 = adf::kernel::create(simple2);

    // Create parameters bound to the same symbol
    p1 = adf::parameter::array(lutarray);
    p2 = adf::parameter::array(lutarray);

    // Connect parameters to kernels (graph-level linkage)
    adf::connect(p1, k1);
    adf::connect(p2, k2);

This ensures the header file that defines the table is included in the final binary link wherever this kernel is used, without causing re-definition errors.