Method pmc_read

tinygrad/runtime/ops_amd.py:176–195

(self, buf, sched)

    return self.pmc_reset_counters(en=True)

  def pmc_read(self, buf, sched):
    self.set_grbm()
    self.wreg(self.gc.regCP_PERFMON_CNTL if self.dev.target[0] <= 11 else self.gc.regCP_PERFMON_CNTL_1, perfmon_state=1, perfmon_sample_enable=1)

    for s in sched:
      offset = itertools.count(s.off, step=8)

      for xcc in range(s.xcc):
        with self.pred_exec(xcc_mask=1 << xcc):
          for inst, se_idx, sa_idx, wgp_idx in itertools.product(range(s.inst), range(s.se), range(s.sa), range(s.wgp)):
            loff = next(offset)
            if s.wgp > 1 and not self.dev.iface.is_wgp_active(xcc, se_idx, sa_idx, wgp_idx): continue
            self.set_grbm(**({'instance':inst} if s.inst > 1 else ({'se':se_idx}|({'sh':sa_idx, 'wgp':wgp_idx} if self.dev.target[0] != 9 else {}))))

            # Copy counter to memory (src_sel = perf, dst_sel = tc_l2)
            lo, hi = getattr(self.gc, f'{s.regsample}_LO'), getattr(self.gc, f'{s.regsample}_HI', None)
            self.pkt3(self.pm4.PACKET3_COPY_DATA, (2 << 8) | 4, lo.addr[0], 0, *data64_le(buf.va_addr+loff))
            if hi is not None: self.pkt3(self.pm4.PACKET3_COPY_DATA, (2 << 8) | 4, hi.addr[0], 0, *data64_le(buf.va_addr+loff+4))

    return self.pmc_reset_counters(en=True)

  ### SQTT ###