TheRock/rocm-systems/emulation/rocjitsu/configs/gfx950_cdna4.json

{
  "max_ticks": 100000,
  "num_threads": 1,
  "exec_mode": "functional",
  "vm": {
    "arch": "cdna4",
    "gpu": {
      "device": {
        "gpu_id": 38144,
        "gfx_target_version": 90500,
        "vendor_id": 4098,
        "device_id": 5892,
        "family_id": 160,
        "unique_id": 5929628898254127105,
        "marketing_name": "AMD Instinct MI350X",
        "simd_count": 1024,
        "max_waves_per_simd": 8,
        "num_shader_engines": 4,
        "num_shader_arrays_per_engine": 2,
        "num_cu_per_sh": 4,
        "simd_per_cu": 4,
        "wave_front_size": 64,
        "local_mem_size": 309237645312,
        "lds_size_kb": 160,
        "mem_width": 8192,
        "mem_clk_max": 1600,
        "l2_size_kb": 4096,
        "num_sdma_engines": 5,
        "num_sdma_xgmi_engines": 12,
        "num_cp_queues": 128,
        "max_engine_clk_fcompute": 2700,
        "capability": 2889327232,
        "capability2": 1,
        "debug_prop": 1494
      }
    }
  },
  "topology": {
    "root": {
      "name": "soc", "type": "soc",
      "children": [
        { "name": "vram", "type": "gpu_memory" },
        {
          "name": "iod[0:2]", "type": "iod",
          "config": [{ "key": "num_hbm_stacks", "value": "4" }]
        },
        {
          "name": "xcd[0:8]", "type": "xcd",
          "children": [
            { "name": "l2", "type": "l2_cache" },
            { "name": "cp", "type": "command_processor" },
            {
              "name": "se[0:4]", "type": "shader_engine",
              "children": [{
                "name": "cu[0:8]", "type": "compute_unit",
                "config": [
                  { "key": "num_wf_slots", "value": "32" },
                  { "key": "sgprs_per_wf", "value": "104" },
                  { "key": "vgprs_per_wf", "value": "512" },
                  { "key": "lds_size_kb", "value": "160" }
                ]
              }]
            }
          ]
        }
      ]
    },
    "links": [
      {
        "pattern": "xcd[i].cp.req_[j*8+k] -> xcd[i].se[j].cu[k].cpl",
        "for_ranges": [
          { "var_name": "i", "start": 0, "end": 8 },
          { "var_name": "j", "start": 0, "end": 4 },
          { "var_name": "k", "start": 0, "end": 8 }
        ],
        "latency": 1, "weight": 2
      },
      {
        "pattern": "xcd[i].se[j].cu[k].req -> xcd[i].l2.cpl_[j*8+k]",
        "for_ranges": [
          { "var_name": "i", "start": 0, "end": 8 },
          { "var_name": "j", "start": 0, "end": 4 },
          { "var_name": "k", "start": 0, "end": 8 }
        ],
        "latency": 1, "weight": 10
      },
      {
        "pattern": "xcd[i].l2.req -> iod[i/4].msc.cpl_[i%4]",
        "for_ranges": [{ "var_name": "i", "start": 0, "end": 8 }],
        "latency": 1, "weight": 3
      },
      {
        "pattern": "iod[i].peer_req -> iod[j].peer_cpl",
        "for_ranges": [
          { "var_name": "i", "start": 0, "end": 2 },
          { "var_name": "j", "start": 0, "end": 2 }
        ],
        "where_expr": "i != j",
        "latency": 1, "weight": 1
      },
      {
        "pattern": "xcd[i].se[j].cu[k].adj_req -> xcd[i].se[j].cu[k+1].adj_cpl",
        "for_ranges": [
          { "var_name": "i", "start": 0, "end": 8 },
          { "var_name": "j", "start": 0, "end": 4 },
          { "var_name": "k", "start": 0, "end": 7 }
        ],
        "latency": 1, "weight": 2
      },
      {
        "pattern": "xcd[i].se[j].cu[k+1].adj_req_r -> xcd[i].se[j].cu[k].adj_cpl_r",
        "for_ranges": [
          { "var_name": "i", "start": 0, "end": 8 },
          { "var_name": "j", "start": 0, "end": 4 },
          { "var_name": "k", "start": 0, "end": 7 }
        ],
        "latency": 1, "weight": 2
      }
    ]
  }
}