ROCm 6.1.2 updates

Author: dayatsin-amd

src/core/runtime/amd_blit_sdma.cpp

@@ -164,10 +164,11 @@ hsa_status_t BlitSdma<RingIndexTy, HwIndexMonotonic, SizeToCountOffset, useGCR>:
   }
 
   // HDP flush supported on gfx900 and forward.
-  // FIXME: Not working on gfx10, raises SRBM write protection interrupt.
   // gfx90a can support xGMI host to device connections so bypass HDP flush
   // in this case.
-  if (agent_->isa()->GetMajorVersion() == 9) {
+  // gfx101x seems to have issues with HDP flushes
+  if (agent_->isa()->GetMajorVersion() >= 9 &&
+      !(agent_->isa()->GetMajorVersion() == 10 && agent_->isa()->GetMinorVersion() == 1)) {
     hdp_flush_support_ = link.info.link_type != HSA_AMD_LINK_INFO_TYPE_XGMI;
   }
 

src/core/util/lnx/os_linux.cpp

@@ -111,24 +111,6 @@ class os_thread {
         return;
       }
     }
-\
-    err = pthread_create(&thread, &attrib, ThreadTrampoline, args.get());
-
-    // Probably a stack size error since system limits can be different from PTHREAD_STACK_MIN
-    // Attempt to grow the stack within reason.
-    if ((err == EINVAL) && stackSize != 0) {
-      while (stackSize < 20 * 1024 * 1024) {
-        stackSize *= 2;
-        err = pthread_attr_setstacksize(&attrib, stackSize);
-        if (err != 0) {
-          fprintf(stderr, "pthread_attr_setstacksize failed: %s\n", strerror(err));
-          return;
-        }
-        err = pthread_create(&thread, &attrib, ThreadTrampoline, args.get());
-        if (err != EINVAL) break;
-        debug_print("pthread_create returned EINVAL, doubling stack size\n");
-      }
-    }
 
     if (core::Runtime::runtime_singleton_->flag().override_cpu_affinity()) {
       int cores = get_nprocs_conf();
@@ -141,14 +123,32 @@ class os_thread {
       for (int i = 0; i < cores; i++) {
         CPU_SET_S(i, CPU_ALLOC_SIZE(cores), cpuset);
       }
-      err = pthread_setaffinity_np(thread, CPU_ALLOC_SIZE(cores), cpuset);
+      err = pthread_attr_setaffinity_np(&attrib, CPU_ALLOC_SIZE(cores), cpuset);
       CPU_FREE(cpuset);
       if (err != 0) {
         fprintf(stderr, "pthread_attr_setaffinity_np failed: %s\n", strerror(err));
         return;
       }
     }
 
+    err = pthread_create(&thread, &attrib, ThreadTrampoline, args.get());
+
+    // Probably a stack size error since system limits can be different from PTHREAD_STACK_MIN
+    // Attempt to grow the stack within reason.
+    if ((err == EINVAL) && stackSize != 0) {
+      while (stackSize < 20 * 1024 * 1024) {
+        stackSize *= 2;
+        err = pthread_attr_setstacksize(&attrib, stackSize);
+        if (err != 0) {
+          fprintf(stderr, "pthread_attr_setstacksize failed: %s\n", strerror(err));
+          return;
+        }
+        err = pthread_create(&thread, &attrib, ThreadTrampoline, args.get());
+        if (err != EINVAL) break;
+        debug_print("pthread_create returned EINVAL, doubling stack size\n");
+      }
+    }
+
     if (err == 0)
       args.release();
     else