From 0d526e251eff551530e02535ee2c4282baea790a Mon Sep 17 00:00:00 2001 From: nimlgen <138685161+nimlgen@users.noreply.github.com> Date: Tue, 8 Oct 2024 17:43:58 +0300 Subject: [PATCH] nv sync on gpu before local update (#6954) --- tinygrad/runtime/ops_nv.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tinygrad/runtime/ops_nv.py b/tinygrad/runtime/ops_nv.py index 62d3fa6b8d..f4793f55be 100644 --- a/tinygrad/runtime/ops_nv.py +++ b/tinygrad/runtime/ops_nv.py @@ -527,7 +527,8 @@ class NVDevice(HCQCompiled): bytes_per_tpc = round_up(bytes_per_warp * 48 * 2, 0x8000) self.shader_local_mem = self._gpu_alloc(round_up(bytes_per_tpc * 64, 0x20000), huge_page=True, contig=True) - NVComputeQueue().setup(local_mem=self.shader_local_mem.va_addr, local_mem_tpc_bytes=bytes_per_tpc) \ + NVComputeQueue().wait(self.timeline_signal, self.timeline_value - 1) \ + .setup(local_mem=self.shader_local_mem.va_addr, local_mem_tpc_bytes=bytes_per_tpc) \ .signal(self.timeline_signal, self.timeline_value).submit(self) self.timeline_value += 1