hcq2: fix ib reuse (#17330)

* hcq2: initialize IB reuse counters at link

* x

extra/hcq2/ops_amd2.py

@@ -158,7 +158,8 @@ def pm4_submit(ctx, lin):
 
   ib = UOp.placeholder((size_dw + 2,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
   done_idx, submit_idx = UOp.const(size_dw + 0, dtypes.int), UOp.const(size_dw + 1, dtypes.int)
-  submitted = (counter:=ib.after(make_patches(ib, [((size_dw + i) * 4, UOp.const(0, dtypes.uint32)) for i in range(2)])).index(submit_idx)).load()
+  init_counters = make_patches(ib, [((size_dw + i) * 4, UOp.const(0, dtypes.uint32)) for i in range(2)]).rtag("link")
+  submitted = (counter:=ib.after(init_counters).index(submit_idx)).load()
   completed = ib.after(loop:=UOp.loop(0)).index(done_idx).load()
   ib_free = completed.end(loop, completed != submitted)
 

tinygrad/runtime/support/hcq2.py

@@ -252,6 +252,7 @@ def is_value_known_at_link(val:UOp) -> bool:
   return not val.variables() and not runtime_reads and all(b.op is not Ops.PARAM or b.tag is not None for b in addressed_bufs)
 
 def is_link_patch(p:UOp, jit:bool) -> bool:
+  if p.tag == "link": return True
   store = p.src[0] if (is_binary_patch:=(p.op is Ops.END and p.src[0].op is Ops.STORE)) else p
   if not jit: return store.buf_uop.tag == "program"
   return is_binary_patch or (store.op is Ops.STORE and is_value_known_at_link(store.src[1]))