Mixed kernels set_ffts_base_addr and pto.section versus cube/vector func.func

[fiskrt]

set_ffts_base_addr

When should set_ffts_base_addr be generated? For example it's used in the examples in pto-isa with the TPUSH/TPOP funcionality (https://gitcode.com/cann/pto-isa/blob/master/tests/npu/a2a3/src/st/testcase/tpushpop_cv/tpushpop_cv_kernel.cpp). While in

module {
  func.func @cube_push_gm(%gm_slot_buffer: memref<256xf32, #pto.address_space<gm>>,
                          %c2v_consumer_buf: i32) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
    %pipe = pto.initialize_l2g2l_pipe {
      dir_mask = 1,
      slot_size = 1024,
      slot_num = 8,
      local_slot_num = 8,
      flag_base = 0
    }(%gm_slot_buffer : memref<256xf32, #pto.address_space<gm>>, %c2v_consumer_buf : i32) -> !pto.pipe
    %acc_tile = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=16, v_row=16, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
    pto.tpush(%acc_tile, %pipe : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=16, v_row=16, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.pipe) {split = 0}
    return
  }

  func.func @vector_pop_gm(%gm_slot_buffer: memref<256xf32, #pto.address_space<gm>>,
                           %c2v_consumer_buf: i32) attributes {pto.kernel_kind = #pto.kernel_kind<vector>} {
    %pipe = pto.initialize_l2g2l_pipe {dir_mask = 1, slot_size = 1024, slot_num = 8, local_slot_num = 8, flag_base = 0}
(%gm_slot_buffer : memref<256xf32, #pto.address_space<gm>>, %c2v_consumer_buf : i32) -> !pto.pipe
    %fifo_tile = pto.declare_tile -> !pto.tile_buf<loc=vec, dtype=f32, rows=8, cols=16, v_row=8, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
    pto.tpop(%fifo_tile, %pipe : !pto.tile_buf<loc=vec, dtype=f32, rows=8, cols=16, v_row=8, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.pipe) {split = 1}
    pto.tfree(%pipe : !pto.pipe) {split = 2}
    return
  }
}

ptoas does not generate any set_ffts_base_addr for this mlir code.

Separate functions vs pto.section

If we have a single func.func with pto.section.vector and pto.section.cube can we use the buffers with TPUSH/TPOP. Or do the mixed kernels only support separate functions with attributes {pto.kernel_kind = #pto.kernel_kind<vector/cube>}?

error: 'pto.reserve_buffer' op expects at most one reserve_buffer in the function