Method alloc

tinygrad/runtime/ops_amd.py:748–776

(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, cpu_addr=None)

    self.queue_event_arr_ptr = ctypes.addressof(self.queue_event_arr)

  def alloc(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, cpu_addr=None) -> HCQBuffer:
    flags = kfd.KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE | kfd.KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE | kfd.KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE

    if uncached: flags |= kfd.KFD_IOC_ALLOC_MEM_FLAGS_COHERENT | kfd.KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED | kfd.KFD_IOC_ALLOC_MEM_FLAGS_GTT
    else: flags |= (kfd.KFD_IOC_ALLOC_MEM_FLAGS_USERPTR if host else kfd.KFD_IOC_ALLOC_MEM_FLAGS_VRAM)

    # Make mapped cpu address to be uncachable
    if cpu_addr is not None: flags |= kfd.KFD_IOC_ALLOC_MEM_FLAGS_COHERENT | kfd.KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED

    if cpu_access or host: flags |= kfd.KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC

    if flags & kfd.KFD_IOC_ALLOC_MEM_FLAGS_USERPTR:
      buf = addr = cpu_addr or FileIOInterface.anon_mmap(0, size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | mmap.MAP_ANONYMOUS, 0)
    else: buf, addr = 0, FileIOInterface.anon_mmap(0, size, 0, mmap.MAP_PRIVATE | mmap.MAP_ANONYMOUS | MAP_NORESERVE, 0)

    try: mem = kfd.AMDKFD_IOC_ALLOC_MEMORY_OF_GPU(self.kfd, va_addr=addr, size=size, gpu_id=self.gpu_id, flags=flags, mmap_offset=buf)
    except OSError as e:
      if e.errno == errno.EINVAL and (flags & kfd.KFD_IOC_ALLOC_MEM_FLAGS_VRAM) and cpu_access:
        raise MemoryError("Cannot allocate host-visible VRAM. Ensure the resizable BAR option is enabled on your system.") from e
      if e.errno == errno.ENOMEM: raise MemoryError(f"Cannot allocate {size} bytes: no memory is available.") from e
      raise

    if not (flags & kfd.KFD_IOC_ALLOC_MEM_FLAGS_USERPTR):
      buf = self.drm_fd.mmap(mem.va_addr, mem.size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | MAP_FIXED, mem.mmap_offset)
      assert addr == buf == mem.va_addr

    view = MMIOInterface(mem.va_addr, mem.size, fmt='B') if cpu_access or host else None
    self.map(hcqbuf:=HCQBuffer(mem.va_addr, mem.size, meta=mem, view=view, owner=self.dev))
    return hcqbuf

  def free(self, mem):
    gpus = (ctypes.c_int32 * 1)(self.gpu_id)