amd autogened kfd ioctls (#5757)

* amd autogened kio

* unused import

* linter

autogen_stubs.sh

@@ -126,6 +126,11 @@ generate_comgr() {
 
 generate_kfd() {
   clang2py /usr/include/linux/kfd_ioctl.h -o $BASE/kfd.py -k cdefstum
+  awk '/^#define AMDKFD_IOC_/ { if ($0 ~ /\\$/) { getline nextline; $0 = $0 nextline }
+    if (match($0, /AMDKFD_IOC_([A-Z_]+).*AMDKFD_(IOW?R?)\(0x([0-9A-F]+),.*struct ([a-z_]+)/, arr)) {
+      print "AMDKFD_IOC_" arr[1] " = (\"" arr[2] "\", 0x" arr[3] ", struct_" arr[4] ")"
+    }
+  }' /usr/include/linux/kfd_ioctl.h >> $BASE/kfd.py
   fixup $BASE/kfd.py
   sed -i "s\import ctypes\import ctypes, os\g" $BASE/kfd.py
   python3 -c "import tinygrad.runtime.autogen.kfd"

tinygrad/runtime/autogen/kfd.py

@@ -810,3 +810,35 @@ __all__ = \
     'struct_kfd_ioctl_wait_events_args',
     'struct_kfd_memory_exception_failure',
     'struct_kfd_process_device_apertures', 'union_kfd_event_data_0']
+AMDKFD_IOC_GET_VERSION = ("IOR", 0x01, struct_kfd_ioctl_get_version_args)
+AMDKFD_IOC_CREATE_QUEUE = ("IOWR", 0x02, struct_kfd_ioctl_create_queue_args)
+AMDKFD_IOC_DESTROY_QUEUE = ("IOWR", 0x03, struct_kfd_ioctl_destroy_queue_args)
+AMDKFD_IOC_SET_MEMORY_POLICY = ("IOW", 0x04, struct_kfd_ioctl_set_memory_policy_args)
+AMDKFD_IOC_GET_CLOCK_COUNTERS = ("IOWR", 0x05, struct_kfd_ioctl_get_clock_counters_args)
+AMDKFD_IOC_GET_PROCESS_APERTURES = ("IOR", 0x06, struct_kfd_ioctl_get_process_apertures_args)
+AMDKFD_IOC_UPDATE_QUEUE = ("IOW", 0x07, struct_kfd_ioctl_update_queue_args)
+AMDKFD_IOC_CREATE_EVENT = ("IOWR", 0x08, struct_kfd_ioctl_create_event_args)
+AMDKFD_IOC_DESTROY_EVENT = ("IOW", 0x09, struct_kfd_ioctl_destroy_event_args)
+AMDKFD_IOC_SET_EVENT = ("IOW", 0x0A, struct_kfd_ioctl_set_event_args)
+AMDKFD_IOC_RESET_EVENT = ("IOW", 0x0B, struct_kfd_ioctl_reset_event_args)
+AMDKFD_IOC_WAIT_EVENTS = ("IOWR", 0x0C, struct_kfd_ioctl_wait_events_args)
+AMDKFD_IOC_DBG_REGISTER = ("IOW", 0x0D, struct_kfd_ioctl_dbg_register_args)
+AMDKFD_IOC_DBG_UNREGISTER = ("IOW", 0x0E, struct_kfd_ioctl_dbg_unregister_args)
+AMDKFD_IOC_DBG_ADDRESS_WATCH = ("IOW", 0x0F, struct_kfd_ioctl_dbg_address_watch_args)
+AMDKFD_IOC_DBG_WAVE_CONTROL = ("IOW", 0x10, struct_kfd_ioctl_dbg_wave_control_args)
+AMDKFD_IOC_SET_SCRATCH_BACKING_VA = ("IOWR", 0x11, struct_kfd_ioctl_set_scratch_backing_va_args)
+AMDKFD_IOC_GET_TILE_CONFIG = ("IOWR", 0x12, struct_kfd_ioctl_get_tile_config_args)
+AMDKFD_IOC_SET_TRAP_HANDLER = ("IOW", 0x13, struct_kfd_ioctl_set_trap_handler_args)
+AMDKFD_IOC_ACQUIRE_VM = ("IOW", 0x15, struct_kfd_ioctl_acquire_vm_args)
+AMDKFD_IOC_ALLOC_MEMORY_OF_GPU = ("IOWR", 0x16, struct_kfd_ioctl_alloc_memory_of_gpu_args)
+AMDKFD_IOC_FREE_MEMORY_OF_GPU = ("IOW", 0x17, struct_kfd_ioctl_free_memory_of_gpu_args)
+AMDKFD_IOC_MAP_MEMORY_TO_GPU = ("IOWR", 0x18, struct_kfd_ioctl_map_memory_to_gpu_args)
+AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU = ("IOWR", 0x19, struct_kfd_ioctl_unmap_memory_from_gpu_args)
+AMDKFD_IOC_SET_CU_MASK = ("IOW", 0x1A, struct_kfd_ioctl_set_cu_mask_args)
+AMDKFD_IOC_GET_QUEUE_WAVE_STATE = ("IOWR", 0x1B, struct_kfd_ioctl_get_queue_wave_state_args)
+AMDKFD_IOC_GET_DMABUF_INFO = ("IOWR", 0x1C, struct_kfd_ioctl_get_dmabuf_info_args)
+AMDKFD_IOC_IMPORT_DMABUF = ("IOWR", 0x1D, struct_kfd_ioctl_import_dmabuf_args)
+AMDKFD_IOC_ALLOC_QUEUE_GWS = ("IOWR", 0x1E, struct_kfd_ioctl_alloc_queue_gws_args)
+AMDKFD_IOC_SMI_EVENTS = ("IOWR", 0x1F, struct_kfd_ioctl_smi_events_args)
+AMDKFD_IOC_SVM = ("IOWR", 0x20, struct_kfd_ioctl_svm_args)
+AMDKFD_IOC_SET_XNACK_MODE = ("IOWR", 0x21, struct_kfd_ioctl_set_xnack_mode_args)

tinygrad/runtime/ops_amd.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 from typing import Tuple, List, Any, cast
-import os, fcntl, ctypes, ctypes.util, functools, re, pathlib, mmap, errno, time, array
+import os, fcntl, ctypes, ctypes.util, functools, pathlib, mmap, errno, time, array, contextlib
 from dataclasses import dataclass
 from tinygrad.device import HCQCompiled, HCQAllocator, HCQBuffer, HWComputeQueue, HWCopyQueue, \
                             HCQSignal, HCQProgram, BufferOptions
@@ -13,31 +13,16 @@ if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl  # noqa: F401 # pylint
 if getenv("MOCKGPU"): import extra.mockgpu.mockgpu # noqa: F401 # pylint: disable=unused-import
 
 def is_usable_gpu(gpu_id):
-  try:
-    with gpu_id.open() as f:
-      return int(f.read()) != 0
-  except OSError:
-    return False
+  with contextlib.suppress(OSError): return int(pathlib.Path(gpu_id).read_text()) != 0
+  return False
 
 def kfd_ioctl(idir, nr, user_struct, fd, **kwargs):
   ret = fcntl.ioctl(fd, (idir<<30) | (ctypes.sizeof(made := user_struct(**kwargs))<<16) | (ord('K')<<8) | nr, made)
   if ret != 0: raise RuntimeError(f"ioctl returned {ret}")
   return made
 
-def ioctls_from_header():
-  #hdr = pathlib.Path("/usr/include/linux/kfd_ioctl.h").read_text().replace("\\\n", "")
-  #pattern = r'#define\s+(AMDKFD_IOC_[A-Z0-9_]+)\s+AMDKFD_(IOW?R?)\((0x[0-9a-fA-F]+),\s+struct\s([A-Za-z0-9_]+)\)'
-  #matches = re.findall(pattern, hdr, re.MULTILINE)
-  # get this from python instead
-  hdrpy = (pathlib.Path(__file__).parent / "autogen" / "kfd.py").read_text()
-  pattern = r'# (AMDKFD_IOC_[A-Z0-9_]+)\s=\s_(IOW?R?).*\(( 0x[0-9a-fA-F]+) ,\s+struct\s([A-Za-z0-9_]+)\s+\)'
-  matches = re.findall(pattern, hdrpy, re.MULTILINE)
-  idirs = {"IOW": 1, "IOR": 2, "IOWR": 3}
-  fxns = {name.replace("AMDKFD_IOC_", "").lower():
-          functools.partial(kfd_ioctl, idirs[idir], int(nr, 0x10), getattr(kfd, "struct_"+sname))
-          for name, idir, nr, sname in matches}
-  return type("KIO", (object, ), fxns)
-kio = ioctls_from_header()
+kio:Any = type("KIO", (object,), {name[11:].lower(): functools.partial(kfd_ioctl, {"IOW": 1, "IOR": 2, "IOWR": 3}[p[0]], p[1], p[2])
+                              for name,p in kfd.__dict__.items() if name.startswith("AMDKFD_IOC_")})
 
 regBIF_BX_PF1_GPU_HDP_FLUSH_REQ, regBIF_BX_PF1_GPU_HDP_FLUSH_DONE = 0x0106, 0x0107