#!/usr/bin/env python3 """Does any PTE attribute change what a GPU engine reads from the TMR? TMZ (bit 3) did not. This varies the cacheability/allocation attributes -- MTYPE (NV10 mask, bits 48-50) and NOALLOC (bit 58) -- on the same mapping. The idea: if the fence were expressed at a cache/coherence level rather than as a hard physical-address block, an uncached or no-allocate access might behave differently. Also maps the fence BOUNDARY from the GPU path, which is independently useful: the TMR range is documented as 0x300000..0x1600000 from the SMU experiment, and confirming those edges with a different initiator either corroborates it or finds the real edges. Same validated harness as the TMZ run: every configuration is preceded by a control read of a page we own. """ import time import numpy as np from tinygrad import Device, Tensor from tinygrad.runtime.support.memory import AddrSpace, PageTableTraverseContext import tinygrad.runtime.autogen.am.am as am OUT = "/tmp/pteattr.txt" def say(s): with open(OUT, "a") as f: f.write(s + "\n") print(s, flush=True) d = Device["AMD"] adev = d.iface.dev_impl mm = adev.mm hb = adev.reg("regMP0_SMN_C2PMSG_81") SZ = 0x1000 def alive(): a = hb.read(); time.sleep(0.2); return hb.read() != a src = (Tensor.zeros(SZ, dtype="uint8").contiguous() + 1).contiguous().realize() va = src.uop.buffer._buf scratch = mm.palloc(SZ, zero=True, boot=False) adev.vram[scratch:scratch + SZ] = bytes([0x3C]) * SZ say("\n######## PTE attribute / boundary probe ########") say("VA %#x scratch PA %#x (0x3C fill)" % (va, scratch)) def remap(pa, orbits=0, clearbits=0): mm.unmap_range(va, SZ) mm.map_range(va, SZ, [(pa, SZ)], aspace=AddrSpace.PHYS) if orbits or clearbits: ctx = PageTableTraverseContext(adev, mm.root_page_table, va, inspect=True) for _o, pt, i, cnt, _c in ctx.next(SZ): for k in range(cnt): pt.entries[i + k] = (pt.entries[i + k] & ~clearbits) | orbits adev.gmc.flush_tlb(ip='GC', vmid=0) def read(tag): try: b = (src + 0).realize().numpy().astype(np.uint8) u = np.unique(b) say(" %-30s %s unique=%d" % (tag, bytes(b[:12]).hex(), len(u))) return b except Exception as e: say(" %-30s FAILED %s" % (tag, type(e).__name__)) return None MT = am.AMDGPU_PTE_MTYPE_NV10_MASK NOALLOC = am.AMDGPU_PTE_NOALLOC TMR_PA = 0x900000 remap(scratch) c = read("control scratch") if c is None or np.unique(c)[0] != 0x3C or len(np.unique(c)) != 1: say("control invalid -- stop"); raise SystemExit(0) say(" control VALID") say("\nA) attribute sweep on the TMR page") for name, orb, clr in (("default", 0, 0), ("MTYPE cleared", 0, MT), ("MTYPE=1<<48", 1 << 48, MT), ("MTYPE=2<<48", 2 << 48, MT), ("MTYPE=3<<48", 3 << 48, MT), ("NOALLOC", NOALLOC, 0), ("NOALLOC+MTYPE cleared", NOALLOC, MT)): remap(TMR_PA, orb, clr) read(name) if not alive(): say(" heartbeat stopped"); break say(" psp alive %s" % alive()) say("\nB) fence boundary from the GPU path (documented TMR 0x300000..0x1600000)") for pa in (0x200000, 0x2FF000, 0x300000, 0x400000, 0x900000, 0x15FF000, 0x1600000, 0x1700000, 0x2000000): remap(pa) b = read("PA %#09x" % pa) say(" psp alive %s" % alive())