add PING to RemoteCmd (#15371)

* add PING to RemoteCmd

* cleanup

extra/remote/bench.py

@@ -16,12 +16,14 @@ if __name__ == "__main__":
   pci = RemotePCIDevice("BN", devs[0])
   print(f"connected to {os.environ['REMOTE']}, device: {devs[0]}\n")
 
-  # latency
-  for _ in range(10): pci.read_config(0, 4)
+  # ping (minimal server round-trip, no device I/O)
+  from tinygrad.runtime.support.system import RemoteCmd
+  sock = pci.sock
+  for _ in range(10): RemotePCIDevice._rpc(sock, 0, RemoteCmd.PING)
   st = time.perf_counter()
-  for _ in range(LAT_N_RUNS): pci.read_config(0, 4)
-  lat = (time.perf_counter() - st) / LAT_N_RUNS
-  print(f"RPC latency: {lat*1e6:.1f} us ({1/lat:,.0f} ops/sec)\n")
+  for _ in range(LAT_N_RUNS): RemotePCIDevice._rpc(sock, 0, RemoteCmd.PING)
+  ping_lat = (time.perf_counter() - st) / LAT_N_RUNS
+  print(f"PING latency: {ping_lat*1e6:.1f} us ({1/ping_lat:,.0f} ops/sec)\n")
 
   # throughput
   sysmem, _ = pci.alloc_sysmem(max(SIZES))

extra/remote/serve.py

@@ -12,6 +12,9 @@ mapped_bars: dict[tuple[int, int], object] = {}
 sysmem_allocs: list[tuple] = []
 
 def handle(conn, cmd, dev_id, bar, arg0, arg1, arg2):
+  if cmd == RemoteCmd.PING:
+    return conn.sendall(resp())
+
   if cmd == RemoteCmd.PROBE:
     payload = conn.recv(arg1, socket.MSG_WAITALL) if arg1 > 0 else b""
     filter_devices: dict[int, list[int]] = {}

tinygrad/runtime/support/system.py

@@ -284,7 +284,7 @@ class PCIIfaceBase:
 # *** Remote PCI Devices
 
 class RemoteCmd(enum.IntEnum):
-  PROBE, MAP_BAR, MAP_SYSMEM_FD, CFG_READ, CFG_WRITE, RESET, MMIO_READ, MMIO_WRITE, MAP_SYSMEM, SYSMEM_READ, SYSMEM_WRITE, RESIZE_BAR = range(12)
+  PROBE,MAP_BAR,MAP_SYSMEM_FD,CFG_READ,CFG_WRITE,RESET,MMIO_READ,MMIO_WRITE,MAP_SYSMEM,SYSMEM_READ,SYSMEM_WRITE,RESIZE_BAR,PING = range(13)
 
 class RemoteMMIOInterface(MMIOInterface):
   def __init__(self, dev:RemotePCIDevice, residx:int, nbytes:int, fmt='B', off=0, rd_cmd=RemoteCmd.MMIO_READ, wr_cmd=RemoteCmd.MMIO_WRITE):