{
  "registry": "export_demo",
  "kernel_count": 3,
  "kernels": [
    {
      "tile": "128x128x32",
      "dtypes": {
        "A": "fp16",
        "B": "fp16",
        "C": "fp16"
      },
      "layout": "rcr",
      "pipeline": "compv4",
      "target": "gfx942"
    },
    {
      "tile": "256x256x64",
      "dtypes": {
        "A": "fp16",
        "B": "fp16",
        "C": "fp16"
      },
      "layout": "rcr",
      "pipeline": "compv4",
      "target": "gfx942"
    },
    {
      "tile": "64x64x32",
      "dtypes": {
        "A": "fp16",
        "B": "fp16",
        "C": "fp16"
      },
      "layout": "rcr",
      "pipeline": "compv4",
      "target": "gfx942"
    }
  ],
  "cpp_registry": {
    "metadata": {
      "timestamp": "Dec  4 2025 06:23:15",
      "total_kernels": 1,
      "export_version": "1.0",
      "dispatcher_version": "1.0.0"
    },
    "statistics": {
      "by_datatype": {},
      "by_pipeline": {},
      "by_scheduler": {}
    },
    "kernels": [
      {
        "identifier": "128x128x32_2x2x1_32x32x16_nopers",
        "name": "gemm_fp16_rcrr_compv4_cshuffle_intrawave_False_False_False_False_128x128x32_2x2x1_32x32x16",
        "algorithm": {
          "tile_shape": {
            "m": 128,
            "n": 128,
            "k": 32
          },
          "wave_shape": {
            "m": 2,
            "n": 2,
            "k": 1
          },
          "warp_tile_shape": {
            "m": 32,
            "n": 32,
            "k": 16
          },
          "block_size": 256,
          "persistent": false,
          "double_buffer": true,
          "preshuffle": false,
          "transpose_c": false
        }
      }
    ]
  }
}