Method _encode_decode

tinygrad/runtime/ops_nv.py:347–358

(self, bufout:HCQBuffer, bufin:HCQBuffer, desc_buf:HCQBuffer, hist:list[HCQBuffer], shape:tuple[int,...], frame_pos:int)

  def _map(self, buf:HCQBuffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)

  def _encode_decode(self, bufout:HCQBuffer, bufin:HCQBuffer, desc_buf:HCQBuffer, hist:list[HCQBuffer], shape:tuple[int,...], frame_pos:int):
    assert all(h.va_addr % 0x100 == 0 for h in hist + [bufin, bufout, desc_buf]), "all buffers must be 0x100 aligned"

    h, w = ((2 * shape[0]) // 3 if shape[0] % 3 == 0 else (2 * shape[0] - 1) // 3), shape[1]
    self.dev._ensure_has_vid_hw(w, h)

    q = NVVideoQueue().wait(self.dev.timeline_signal, self.dev.timeline_value - 1)
    with hcq_profile(self.dev, queue=q, desc="HEVC Decode", enabled=PROFILE, dev_suff="NVDEC"):
      q.decode_hevc_chunk(desc_buf, bufin, bufout, frame_pos, hist, [(frame_pos-x) % (len(hist) + 1) for x in range(len(hist), 0, -1)],
                          round_up(w, 64)*round_up(h, 64), self.dev.vid_coloc_buf, self.dev.vid_filter_buf, self.dev.intra_top_off,
                          self.dev.intra_unk_off, self.dev.vid_stat_buf)
    q.signal(self.dev.timeline_signal, self.dev.next_timeline()).submit(self.dev)

@dataclass
class GPFifo: