[Openmp-commits] [openmp] [amdgpu][openmp] Avoiding writing to packet header twice (PR #70695)

[Jon Chesterfield via Openmp-commits]

================
@@ -222,34 +222,36 @@ hsa_status_t launch_kernel(hsa_agent_t dev_agent, hsa_executable_t executable,
   // Set up the packet for exeuction on the device. We currently only launch
   // with one thread on the device, forcing the rest of the wavefront to be
   // masked off.
-  std::memset(packet, 0, sizeof(hsa_kernel_dispatch_packet_t));
-  packet->setup = (1 + (params.num_blocks_y * params.num_threads_y != 1) +
-                   (params.num_blocks_z * params.num_threads_z != 1))
-                  << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;
+  uint16_t setup = (1 + (params.num_blocks_y * params.num_threads_y != 1) +
+                    (params.num_blocks_z * params.num_threads_z != 1))
+                   << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;
   packet->workgroup_size_x = params.num_threads_x;
----------------
JonChesterfield wrote:

To what end? The packet assignments following are complete, and if they weren't, that would be a problem. It's probably only the first byte that is critical, but might be the first four bytes in practice.

https://github.com/llvm/llvm-project/pull/70695