From 69ebcb55836ebbd9dff51d7fc3eecb28a82971f2 Mon Sep 17 00:00:00 2001
From: foreman <dl.constructicon@amd.com>
Date: Mon, 14 Jul 2014 20:24:58 -0400
Subject: [PATCH] P4 to Git Change 1055054 by gandryey@gera-dev-w7 on
 2014/07/14 20:18:53

	ECR #304775 - Device enqueuing
	- Switch to the single thread scheduler for now(the current version isn't friendly for single thread). Hopefully it's a temporary solution until synchronization issue with multithreaded scheduler will be identified.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#104 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuschedcl.cpp#20 edit


[ROCm/clr commit: 6314b334bade159bbff6427dbb8960762714602b]
---
 .../clr/rocclr/runtime/device/gpu/gpublit.cpp |  2 +-
 .../rocclr/runtime/device/gpu/gpuschedcl.cpp  | 26 ++++++++++++-------
 2 files changed, 18 insertions(+), 10 deletions(-)

diff --git a/projects/clr/rocclr/runtime/device/gpu/gpublit.cpp b/projects/clr/rocclr/runtime/device/gpu/gpublit.cpp
index 42ea2d6753..544dc25bc1 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gpublit.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gpublit.cpp
@@ -2730,7 +2730,7 @@ KernelBlitManager::runScheduler(
 
     size_t  dim = 1;
     size_t  globalWorkOffset[1] = { 0 };
-    size_t  globalWorkSize[1] = { numSlots / 32 };
+    size_t  globalWorkSize[1] = { 1 };
     size_t  localWorkSize[1] = { 1 };
 
     // Program kernels arguments
diff --git a/projects/clr/rocclr/runtime/device/gpu/gpuschedcl.cpp b/projects/clr/rocclr/runtime/device/gpu/gpuschedcl.cpp
index 0708bf975e..db43d6b0be 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gpuschedcl.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gpuschedcl.cpp
@@ -353,16 +353,23 @@ scheduler(
     __global AmdAqlWrap*    wraps = (__global AmdAqlWrap*)&queue[1];
     __global uint*          amask = (__global uint *)queue->aql_slot_mask;
 
-    uint launch = 0;
-    uint loop;  
+    uint launch;
+    do {
+        launch = atomic_load_explicit((__global atomic_uint*)&param->launch,
+            memory_order_acquire, memory_scope_device);
+    } while (launch != 0);
+    uint loop;
 
     do {
-        uint mask = atomic_load_explicit((__global atomic_uint*)(&amask[get_group_id(0)]),
-              memory_order_acquire, memory_scope_device);
+        for (uint m = 0; m < (queue->aql_slot_num >> 5); ++m) {
+            uint mask = atomic_load_explicit((__global atomic_uint*)(&amask[m]),
+                  memory_order_acquire, memory_scope_device);
 
-        if (mask != 0) {
-            int baseIdx = get_group_id(0) * 32;
-            for (int idx = baseIdx + 31 - clz(mask); (idx >= baseIdx) && (launch == 0); --idx) {
+            int baseIdx = m * 32;
+            while (mask != 0) {
+                uint sIdx = ctz(mask);
+                uint idx = baseIdx + sIdx;
+                mask &= ~(1 << sIdx);
                 __global AmdAqlWrap* disp = (__global AmdAqlWrap*)&wraps[idx];
                 uint slotState = atomic_load_explicit((__global atomic_uint*)(&disp->state),
                     memory_order_acquire, memory_scope_device);
@@ -435,7 +442,7 @@ scheduler(
                         // Decrement the child execution counter on the parent
                         atomic_fetch_sub_explicit(
                             (__global atomic_uint*)&parent->child_counter,
-                             1, memory_order_acq_rel, memory_scope_device);
+                            1, memory_order_acq_rel, memory_scope_device);
                         event->state = CL_COMPLETE;
                         disp->state = AQL_WRAP_FREE;
                         release_slot(amask, idx);
@@ -465,7 +472,7 @@ scheduler(
                         // Decrement the child execution counter on the parent
                         atomic_fetch_sub_explicit(
                             (__global atomic_uint*)&parent->child_counter,
-                             1, memory_order_acq_rel, memory_scope_device);
+                            1, memory_order_acq_rel, memory_scope_device);
                         disp->state = AQL_WRAP_FREE;
                         release_slot(amask, idx);
                     }
@@ -474,6 +481,7 @@ scheduler(
                     disp->state = AQL_WRAP_DONE;
                 }
             }
+            if (launch == 1) break;
         }
         barrier(CLK_GLOBAL_MEM_FENCE);