diff mbox

[hsa] Redesign busy loop waiting so that a kernel dispatch signal can be reused

Message ID 56547A51.5040209@suse.cz
State New
Headers show

Commit Message

Martin Liška Nov. 24, 2015, 2:55 p.m. UTC 
Hello.

Following patch is a workaround for Carrizo devices that tend to have problems
with propagation of signal values due to an issue with L2.

Commited to the branch.

Martin
diff mbox

Patch

From ca4475aedb47e49b4bdc0a8980f200ec93b31d61 Mon Sep 17 00:00:00 2001
From: marxin <mliska@suse.cz>
Date: Tue, 24 Nov 2015 10:41:54 +0100
Subject: [PATCH 1/2] Redesign busy loop waiting so that a kernel dispatch
 signal can be reused

libgomp/ChangeLog:

2015-11-24  Martin Liska  <mliska@suse.cz>

	* plugin/plugin-hsa.c (GOMP_OFFLOAD_run): Rewrite busy loop
	that does a workaround for Carrizo machines.
---
 libgomp/plugin/plugin-hsa.c | 34 ++++++++++++++++++----------------
 1 file changed, 18 insertions(+), 16 deletions(-)

diff --git a/libgomp/plugin/plugin-hsa.c b/libgomp/plugin/plugin-hsa.c
index b866a78..99ec8e1 100644
--- a/libgomp/plugin/plugin-hsa.c
+++ b/libgomp/plugin/plugin-hsa.c
@@ -1219,24 +1219,26 @@  GOMP_OFFLOAD_run (int n, void *fn_ptr, void *vars, void** args)
   __atomic_store_n ((uint16_t*)(&packet->header), header, __ATOMIC_RELEASE);
   hsa_signal_store_release (agent->command_q->doorbell_signal, index);
 
-  /* TODO: fixup, following workaround is necessary to run kernel from
-     kernel dispatch mechanism on a Carrizo machine.  */
-
-  for (unsigned i = 0; i < shadow->kernel_dispatch_count; i++)
-    {
-      hsa_signal_t child_s;
-      child_s.handle = shadow->children_dispatches[i]->signal;
-
-      HSA_DEBUG ("Waiting for children completion signal: %lu\n",
-		 shadow->children_dispatches[i]->signal);
-      while (hsa_signal_wait_acquire
-	     (child_s, HSA_SIGNAL_CONDITION_LT, 1, UINT64_MAX,
-	      HSA_WAIT_STATE_BLOCKED) != 0);
-    }
+  /* TODO: GPU agents in Carrizo APUs cannot properly update L2 cache for
+     signal wait and signal load operations on their own and we need to
+     periodically call the hsa_signal_load_acquire on completion signals of
+     children kernels in the CPU to make that happen.  As soon the
+     limitation will be resolved, this workaround can be removed.  */
 
   HSA_DEBUG ("Kernel dispatched, waiting for completion\n");
-  while (hsa_signal_wait_acquire (s, HSA_SIGNAL_CONDITION_LT, 1,
-				  UINT64_MAX, HSA_WAIT_STATE_BLOCKED) != 0);
+
+  /* Root signal waits with 1ms timeout.  */
+  while (hsa_signal_wait_acquire (s, HSA_SIGNAL_CONDITION_LT, 1, 1000 * 1000,
+				  HSA_WAIT_STATE_BLOCKED) != 0)
+    for (unsigned i = 0; i < shadow->kernel_dispatch_count; i++)
+      {
+	hsa_signal_t child_s;
+	child_s.handle = shadow->children_dispatches[i]->signal;
+
+	HSA_DEBUG ("Waiting for children completion signal: %lu\n",
+		   shadow->children_dispatches[i]->signal);
+	hsa_signal_load_acquire (child_s);
+      }
 
   release_kernel_dispatch (shadow);
 
-- 
2.6.3