[Offload][AMDGPU] Fix timing issue in hostexec packet processing

saiislam · David Salinas · commit 3985999df1a8 · 2024-10-17T18:29:38.000-04:00
The host processing of packets on gfx12 was removing the packets
too early from the stack. This sleep instruction fixes the timing
and thus pagefaulting issue.

Change-Id: I75af63c09c5b00b4c5cc538ea952cb6accb6043b
diff --git a/offload/hostexec/services/amdgcn_hostexec.cpp b/offload/hostexec/services/amdgcn_hostexec.cpp
@@ -346,6 +346,7 @@ struct consumer_t {
           if (flag == 0)
             continue;
           execute_service(service_id, buffer->device_id, payload->slots[wi]);
+          std::this_thread::sleep_for(std::chrono::nanoseconds(1));
         }
       }
       __atomic_store_n(&header->control, reset_ready_flag(header->control),

Original file line number	Diff line number	Diff line change
`@@ -346,6 +346,7 @@ struct consumer_t {`
`346`	`346`	`if (flag == 0)`
`347`	`347`	`continue;`
`348`	`348`	`execute_service(service_id, buffer->device_id, payload->slots[wi]);`
	`349`	`+ std::this_thread::sleep_for(std::chrono::nanoseconds(1));`
`349`	`350`	`}`
`350`	`351`	`}`
`351`	`352`	`__atomic_store_n(&header->control, reset_ready_flag(header->control),`