better oor message in profiles (#15516)

* better oor message

* x

tinygrad/viz/serve.py

@@ -180,9 +180,9 @@ def enum_str(s, cache:dict[str, int]) -> int:
 
 def option(s:int|None) -> int: return 0 if s is None else s+1
 
-def rel_ts(ts:int|Decimal, start_ts:int) -> int:
+def rel_ts(ts:int|Decimal, start_ts:int, ctx:str="") -> int:
   val = int(ts) - start_ts
-  if val < 0 or val > 0xFFFFFFFF: raise ValueError(f"timestamp out of range: ts={ts} start={start_ts} {val}")
+  if val < 0 or val > 0xFFFFFFFF: raise ValueError(f"timestamp out of range: {ctx} diff={val} (ts={ts} start={start_ts})")
   return val
 
 # Profiler API
@@ -228,7 +228,8 @@ def timeline_layout(dev_events:list[tuple[int, int, float, DevEvent]], start_ts:
         membw = e.name.ret / (dur * 1e-6)
         fmt.append(f"{membw*1e-9:.0f} GB/s" if membw < 1e13 else f"{membw*1e-12:.0f} TB/s")
       elif e.name.tb: fmt.append("TB:"+json.dumps(e.name.tb))
-    events.append(struct.pack("<IIIIfI", enum_str(name, scache), option(ref), option(key), rel_ts(st,start_ts), dur, enum_str("\n".join(fmt),scache)))
+    events.append(struct.pack("<IIIIfI", enum_str(name, scache), option(ref), option(key), rel_ts(st,start_ts, f"'{name}' on {e.device}"),
+                              dur, enum_str("\n".join(fmt),scache)))
   return struct.pack("<BI", 0, len(events))+b"".join(events) if events else None
 
 def encode_mem_free(key:int, ts:int, execs:list[ProfilePointEvent], scache:dict) -> bytes:
@@ -242,7 +243,7 @@ def encode_mem_free(key:int, ts:int, execs:list[ProfilePointEvent], scache:dict)
 def graph_layout(k:str, dev_events:list[tuple[int, int, float, DevEvent]], start_ts:int, end_ts:int, peaks:list[int], dtype_size:dict[str, int],
                  scache:dict[str, int]) -> tuple[str, bytes|None]:
   if k.startswith("LINE:"):
-    xy = [(rel_ts(e.ts, start_ts), e.key) for st,_,_,e in dev_events if isinstance(e, ProfilePointEvent)]
+    xy = [(rel_ts(e.ts, start_ts, f"line '{k}' on {e.device}"), e.key) for st,_,_,e in dev_events if isinstance(e, ProfilePointEvent)]
     peaks.append(peak:=max([y for _,y in xy]))
     return k.replace("LINE:", ""), struct.pack("<BIBQ", 1, len(xy), 1, peak)+b"".join(struct.pack("<IQ", x, y) for x,y in xy)
   peak, mem = 0, 0
@@ -253,7 +254,7 @@ def graph_layout(k:str, dev_events:list[tuple[int, int, float, DevEvent]], start
     if not isinstance(e, ProfilePointEvent): continue
     if e.name == "alloc":
       safe_sz = min(1_000_000_000_000, e.arg["sz"])
-      events.append(struct.pack("<BIIIQ", 1, rel_ts(e.ts, start_ts), e.key, enum_str(e.arg["dtype"].name, scache), safe_sz))
+      events.append(struct.pack("<BIIIQ", 1, rel_ts(e.ts, start_ts, f"alloc on {e.device}"), e.key, enum_str(e.arg["dtype"].name, scache), safe_sz))
       dtype_size.setdefault(e.arg["dtype"].name, e.arg["dtype"].itemsize)
       temp[e.key] = nbytes = safe_sz*e.arg["dtype"].itemsize
       mem += nbytes
@@ -261,9 +262,9 @@ def graph_layout(k:str, dev_events:list[tuple[int, int, float, DevEvent]], start
     if e.name == "exec" and e.arg["bufs"]:
       for b in e.arg["bufs"]: buf_ei.setdefault(b, []).append(e)
     if e.name == "free":
-      events.append(encode_mem_free(e.key, rel_ts(e.ts, start_ts), buf_ei.pop(e.key, []), scache))
+      events.append(encode_mem_free(e.key, rel_ts(e.ts, start_ts, f"free on {e.device}"), buf_ei.pop(e.key, []), scache))
       mem -= temp.pop(e.key)
-  for t in temp: events.append(encode_mem_free(t, rel_ts(end_ts, start_ts), buf_ei.pop(t, []), scache))
+  for t in temp: events.append(encode_mem_free(t, rel_ts(end_ts, start_ts, f"end_ts for {k}"), buf_ei.pop(t, []), scache))
   peaks.append(peak)
   return f"{k} Memory", struct.pack("<BIBQ", 1, len(events), 0, peak)+b"".join(events) if events else None
 
@@ -482,9 +483,10 @@ def get_profile(profile:list[ProfileEvent], sort_fn:Callable[[str], Any]=device_
     layout.update([graph_layout(k, v, start_ts, unwrap(end_ts), peaks, dtype_size, scache)])
   sorted_layout = sorted([k for k,v in layout.items() if v is not None], key=sort_fn)
   ret = [b"".join([struct.pack("<B", len(k)), k.encode(), unwrap(layout[k])]) for k in sorted_layout]
-  index = json.dumps({"strings":list(scache), "dtypeSize":dtype_size, "markers":[{"ts":rel_ts(e.ts, start_ts), **e.arg} for e in markers],
+  index = json.dumps({"strings":list(scache), "dtypeSize":dtype_size,
+                      "markers":[{"ts":rel_ts(e.ts, start_ts, f"marker '{e.arg.get('name','?')}'"), **e.arg} for e in markers],
                       **ext_data}).encode()
-  return struct.pack("<IQII", rel_ts(unwrap(end_ts), start_ts), max(peaks,default=0), len(index), len(ret))+index+b"".join(ret)
+  return struct.pack("<IQII", rel_ts(unwrap(end_ts), start_ts, "end_ts"), max(peaks,default=0), len(index), len(ret))+index+b"".join(ret)
 
 # ** PMA counters