Read HCC_OPT_FLUSH and optimize dispatch accordingly.

If HCC is in this mode, we can use less aggressive flushes in some
cases.

hipamd/src/hip_hcc.cpp

@@ -97,6 +97,8 @@ int HIP_COHERENT_HOST_ALLOC = 0;
 // USE_ HIP_SYNC_HOST_ALLOC
 int HIP_SYNC_HOST_ALLOC = 1;
 
+int HCC_OPT_FLUSH = 0;
+
 
 
 
@@ -1204,6 +1206,9 @@ void HipReadEnv()
 
     READ_ENV_I(release, HIP_COHERENT_HOST_ALLOC, 0, "If set, all host memory will be allocated as fine-grained system memory.  This allows threadfence_system to work but prevents host memory from being cached on GPU which may have performance impact.");
 
+
+    READ_ENV_I(release, HCC_OPT_FLUSH, 0, "Note this flag also impact HCC.  When set, use agent-scope flush rather than system-scope flush when possible.");
+
     // Some flags have both compile-time and runtime flags - generate a warning if user enables the runtime flag but the compile-time flag is disabled.
     if (HIP_DB && !COMPILE_HIP_DB) {
         fprintf (stderr, "warning: env var HIP_DB=0x%x but COMPILE_HIP_DB=0.  (perhaps enable COMPILE_HIP_DB in src code before compiling?)\n", HIP_DB);

hipamd/src/hip_hcc.h

@@ -66,6 +66,9 @@ extern int HIP_COHERENT_HOST_ALLOC;
 // Chicken bits for disabling functionality to work around potential issues:
 extern int HIP_SYNC_HOST_ALLOC;
 
+// TODO - remove when this is standard behavior.
+extern int HCC_OPT_FLUSH;
+
 
 // Class to assign a short TID to each new thread, for HIP debugging purposes.
 class TidInfo {

hipamd/src/hip_module.cpp

@@ -320,11 +320,17 @@ hipError_t hipModuleLaunchKernel(hipFunction_t f,
         aql.kernel_object = f._object;
         aql.setup = 3 << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;
         aql.header =   (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) |
-                                    (1 << HSA_PACKET_HEADER_BARRIER) |
-                                    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
-                                    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
+                       (1 << HSA_PACKET_HEADER_BARRIER);  // TODO - honor queue setting for execute_in_order
 
-        lp.av->dispatch_hsa_kernel(&aql, config[1] /* kernarg*/, kernArgSize);
+        if (HCC_OPT_FLUSH) {
+            aql.header |= (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
+                          (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
+        } else {
+            aql.header |= (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
+                          (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
+        };
+
+        lp.av->dispatch_hsa_kernel(&aql, config[1] /* kernarg*/, kernArgSize, nullptr/*completion_future*/);
 
 
         ihipPostLaunchKernel(f._name, hStream, lp);