Method __call__

tinygrad/runtime/ops_amd.py:611–641

(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int|None, ...]=(),
               wait=False, timeout:int|None=None)

    weakref.finalize(self, self._fini, self.dev, self.lib_gpu, buf_spec)

  def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int|None, ...]=(),
               wait=False, timeout:int|None=None):
    if self.dev.sqtt_enabled: cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).sqtt_start(self.dev.sqtt_buffers).submit(self.dev)
    res = super().__call__(*bufs, global_size=global_size, local_size=local_size, vals=vals, wait=wait, timeout=timeout)
    if self.dev.pmc_enabled:
      cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).pmc_read(self.dev.pmc_buffer, self.dev.pmc_sched) \
                                                          .signal(self.dev.timeline_signal, self.dev.next_timeline()).submit(self.dev)
      self.dev.allocator._copyout(pmc_buf:=memoryview(bytearray(self.dev.pmc_buffer.size)), self.dev.pmc_buffer)
      Compiled.profile_events += [ProfilePMCEvent(self.dev.device, self.prof_prg_counter, self.dev.pmc_sched, bytes(pmc_buf),
                                                  self.dev.prof_exec_counter)]
    if self.dev.sqtt_enabled:
      cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).sqtt_stop(self.dev.sqtt_wptrs) \
                                                          .signal(self.dev.timeline_signal, self.dev.next_timeline()).submit(self.dev)
      self.dev.synchronize()

      for se, buf in enumerate(self.dev.sqtt_buffers):
        wptr = (self.dev.sqtt_wptrs.cpu_view().view(fmt='I')[se] & 0x1FFFFFFF) * 32
        if self.dev.target[:2] == (11, 0): wptr -= ((buf.va_addr // 32) & 0x1FFFFFFF) * 32

        if DEBUG >= 5: print(f'\t{self.dev.device}: SE {se} blob size {wptr:#x}')
        assert wptr >= 0 and wptr <= buf.size, f"{wptr} > {buf.size}, should never happen"

        # When sqtt buffer overflows, wptr stops at the last dword
        if wptr >= buf.size - 32:
          print(colored(f"{self.dev.device}: Warning: SQTT buffer is full (SE {se})! Increase SQTT buffer with SQTT_BUFFER_SIZE=X (in MB)", "yellow"))

        self.dev.allocator._copyout(sqtt_mv:=memoryview(bytearray(wptr)), buf)
        resbuf = (struct.pack('<Q', 0x11 | (4 << 13) | (0xf << 16) | (se << 24)) + bytes(sqtt_mv)) if self.dev.target[0] == 9 else bytes(sqtt_mv)
        Compiled.profile_events += [ProfileSQTTEvent(self.dev.device, self.prof_prg_counter, se, resbuf,
                                                     bool((SQTT_ITRACE_SE_MASK.value >> se) & 1), self.dev.prof_exec_counter)]
    return res

class AMDAllocator(HCQAllocator['AMDDevice']):
  def __init__(self, dev:AMDDevice):