gpu-compute: enable flexible control of kernel boundary syncs

Kernel end release was turned on for VIPER protocol, which
is in fact write-through based and thus no need to have
release operation. This changeset splits the option
'impl_kern_boundary_sync' into 'impl_kern_launch_acq'
and 'impl_kern_end_rel', and turns off release on VIPER.

Change-Id: I5490019b6765a25bd801cc78fb7445b90eb02a3d
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29917
Reviewed-by: Anthony Gutierrez <anthony.gutierrez@amd.com>
Reviewed-by: Xianwei Zhang <xianwei.zhang@amd.com>
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>

src/arch/gcn3/insts/instructions.cc

@@ -3759,9 +3759,13 @@ namespace Gcn3ISA
             // the last workgroup in the kernel).
             bool kernelEnd =
                 wf->computeUnit->shader->dispatcher().isReachingKernelEnd(wf);
+            // further check whether 'release @ kernel end' is needed
+            bool relNeeded =
+                wf->computeUnit->shader->impl_kern_end_rel;
 
-            // if it is not a kernel end, then retire the workgroup directly
-            if (!kernelEnd) {
+            // if not a kernel end or no release needed, retire the workgroup
+            // directly
+            if (!kernelEnd || !relNeeded) {
                 wf->computeUnit->shader->dispatcher().notifyWgCompl(wf);
                 wf->setStatus(Wavefront::S_STOPPED);
                 wf->computeUnit->completedWGs++;
@@ -3770,8 +3774,8 @@ namespace Gcn3ISA
             }
 
             /**
-             * If it is a kernel end, inject a memory sync and retire the
-             * workgroup after receving response.
+             * If a kernel end and release needed, inject a memory sync and
+             * retire the workgroup after receving all acks.
              */
             setFlag(MemSync);
             setFlag(GlobalSegment);

src/gpu-compute/GPU.py

@@ -213,8 +213,10 @@ class Shader(ClockedObject):
     gpu_cmd_proc = Param.GPUCommandProcessor('Command processor for GPU')
     dispatcher = Param.GPUDispatcher('GPU workgroup dispatcher')
     n_wf = Param.Int(10, 'Number of wavefront slots per SIMD')
-    impl_kern_boundary_sync = Param.Bool(True, """Insert acq/rel packets into
-                                         ruby at kernel boundaries""")
+    impl_kern_launch_acq = Param.Bool(True, """Insert acq packet into
+                                         ruby at kernel launch""")
+    impl_kern_end_rel = Param.Bool(False, """Insert rel packet into
+                                         ruby at kernel end""")
     globalmem = Param.MemorySize('64kB', 'Memory size')
     timing = Param.Bool(False, 'timing memory accesses')
 

src/gpu-compute/dispatcher.cc

@@ -166,12 +166,12 @@ GPUDispatcher::exec()
         auto task = hsaQueueEntries[exec_id];
         bool launched(false);
 
-        // invalidate is needed before starting dispatch
-        if (shader->impl_kern_boundary_sync) {
+        // acq is needed before starting dispatch
+        if (shader->impl_kern_launch_acq) {
             // try to invalidate cache
             shader->prepareInvalidate(task);
         } else {
-            // kern boundary sync is not set, skip invalidate
+            // kern launch acquire is not set, skip invalidate
             task->markInvDone();
         }
 

src/gpu-compute/shader.cc

@@ -56,7 +56,8 @@ Shader::Shader(const Params *p) : ClockedObject(p),
     tickEvent([this]{ execScheduledAdds(); }, "Shader scheduled adds event",
           false, Event::CPU_Tick_Pri),
     timingSim(p->timing), hsail_mode(SIMT),
-    impl_kern_boundary_sync(p->impl_kern_boundary_sync),
+    impl_kern_launch_acq(p->impl_kern_launch_acq),
+    impl_kern_end_rel(p->impl_kern_end_rel),
     coissue_return(1),
     trace_vgpr_all(1), n_cu((p->CUs).size()), n_wf(p->n_wf),
     globalMemSize(p->globalmem),

src/gpu-compute/shader.hh

@@ -216,7 +216,9 @@ class Shader : public ClockedObject
     hsail_mode_e hsail_mode;
 
     // If set, issue acq packet @ kernel launch
-    int impl_kern_boundary_sync;
+    int impl_kern_launch_acq;
+    // If set, issue rel packet @ kernel end
+    int impl_kern_end_rel;
     // If set, fetch returns may be coissued with instructions
     int coissue_return;
     // If set, always dump all 64 gprs to trace