Fix for barrier call in global_barrier

PawelJurek · igcbot · commit 55afc375b374 · 2024-08-26T12:34:42.000+02:00
Due to known error, the barrier() call in global_barrier was not
properly inlined. Use internal builtin instead until the issue
is resolved.
diff --git a/IGC/BiFModule/Implementation/barrier.cl b/IGC/BiFModule/Implementation/barrier.cl
@@ -314,7 +314,7 @@ uint __intel_get_local_size( void );
 // Implementation of global_barrier using atomic instructions.
 void __global_barrier_atomic()
 {
-    barrier(CLK_GLOBAL_MEM_FENCE);
+    __intel_workgroup_barrier(Device, AcquireRelease | CrossWorkgroupMemory);
 
     __global volatile int* syncBuffer = (__global volatile int*)__builtin_IB_get_sync_buffer();
 
@@ -335,7 +335,7 @@ void __global_barrier_atomic()
         while (atomic_or(syncBuffer, 0) != 0) {}
     }
 
-    barrier(CLK_GLOBAL_MEM_FENCE);
+    __intel_workgroup_barrier(Device, AcquireRelease | CrossWorkgroupMemory);
 }
 
 // Implementation of global_barrier without using atomic instructions except for fences.

Original file line number	Diff line number	Diff line change
`@@ -314,7 +314,7 @@ uint __intel_get_local_size( void );`
`314`	`314`	`// Implementation of global_barrier using atomic instructions.`
`315`	`315`	`void __global_barrier_atomic()`
`316`	`316`	`{`
`317`		`- barrier(CLK_GLOBAL_MEM_FENCE);`
	`317`	`+ __intel_workgroup_barrier(Device, AcquireRelease \| CrossWorkgroupMemory);`
`318`	`318`
`319`	`319`	`__global volatile int* syncBuffer = (__global volatile int*)__builtin_IB_get_sync_buffer();`
`320`	`320`
`@@ -335,7 +335,7 @@ void __global_barrier_atomic()`
`335`	`335`	`while (atomic_or(syncBuffer, 0) != 0) {}`
`336`	`336`	`}`
`337`	`337`
`338`		`- barrier(CLK_GLOBAL_MEM_FENCE);`
	`338`	`+ __intel_workgroup_barrier(Device, AcquireRelease \| CrossWorkgroupMemory);`
`339`	`339`	`}`
`340`	`340`
`341`	`341`	`// Implementation of global_barrier without using atomic instructions except for fences.`