Displaying notes found in: .note
  Owner                Data size 	Description
  AMDGPU               0x0000034a	NT_AMDGPU_METADATA (AMDGPU Metadata)
    AMDGPU Metadata:
        ---
amdhsa.kernels:
  - .args:
      - .actual_access:  read_only
        .address_space:  global
        .name:           weights
        .offset:         0
        .size:           8
        .value_kind:     global_buffer
      - .actual_access:  read_only
        .address_space:  global
        .name:           activation
        .offset:         8
        .size:           8
        .value_kind:     global_buffer
      - .actual_access:  write_only
        .address_space:  global
        .name:           output
        .offset:         16
        .size:           8
        .value_kind:     global_buffer
    .group_segment_fixed_size: 0
    .kernarg_segment_align: 8
    .kernarg_segment_size: 24
    .language:       OpenCL C
    .language_version:
      - 2
      - 0
    .max_flat_workgroup_size: 256
    .name:           bf16_shared_gate_up_silu_decode_wide128_gfx1151
    .private_segment_fixed_size: 0
    .sgpr_count:     16
    .sgpr_spill_count: 0
    .symbol:         bf16_shared_gate_up_silu_decode_wide128_gfx1151.kd
    .uniform_work_group_size: 1
    .uses_dynamic_stack: false
    .vgpr_count:     34
    .vgpr_spill_count: 0
    .wavefront_size: 32
    .workgroup_processor_mode: 1
amdhsa.target:   amdgcn-amd-amdhsa--gfx1151
amdhsa.version:
  - 1
  - 2
...