From a6c810eb03d19255da8f85f6fcb8fc1b48c306cd Mon Sep 17 00:00:00 2001
From: qazal <77887910+Qazalin@users.noreply.github.com>
Date: Mon, 31 Aug 2026 11:25:48 +0800
Subject: [PATCH] mi350p: fix pmc (#17858)

---
 tinygrad/runtime/ops_amd.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/tinygrad/runtime/ops_amd.py b/tinygrad/runtime/ops_amd.py
index 8d8c48f8de..f73b3cf0a9 100644
--- a/tinygrad/runtime/ops_amd.py
+++ b/tinygrad/runtime/ops_amd.py
@@ -1065,7 +1065,8 @@ class AMDDevice(HCQCompiled):
       for k in (PMC_COUNTERS:=getenv("PMC_COUNTERS", pmc_default).split(",")):
         if k not in self.pmc_counters: raise RuntimeError(f"PMC counter {k} is not supported. Available: {','.join(self.pmc_counters.keys())}")
 
-      cast(AMDComputeQueue, unwrap(self.hw_compute_queue_t)()).pmc_start([(k, *self.pmc_counters[k]) for k in PMC_COUNTERS]).submit(self)
+      with (q:=cast(AMDComputeQueue, unwrap(self.hw_compute_queue_t)())).pred_exec((1 << self.xccs) - 1):
+        q.pmc_start([(k, *self.pmc_counters[k]) for k in PMC_COUNTERS]).submit(self)
       self.pmc_buffer = self.allocator.alloc(self.pmc_sched[-1].off + self.pmc_sched[-1].size, BufferSpec(nolru=True, uncached=True))
       self.allocator._copyin(self.pmc_buffer, memoryview(bytearray(self.pmc_buffer.size))) # zero pmc buffers, some counters have only lo part.