[SYCL][CUDA] Add missing device scope to atomic fence (#9824)

npmiller · web-flow · commit 93eb9ffe5779 · 2023-06-13T15:29:29.000+01:00
This patch maps `Device` scope fence to the right NVVM built-in. It
would previously incorrectly use the CTA (threadblock) variant.
diff --git a/libclc/ptx-nvidiacl/libspirv/synchronization/barrier.cl b/libclc/ptx-nvidiacl/libspirv/synchronization/barrier.cl
@@ -13,6 +13,8 @@ _CLC_OVERLOAD _CLC_DEF void __spirv_MemoryBarrier(unsigned int memory,
                                                   unsigned int semantics) {
   if (memory == CrossDevice) {
     __nvvm_membar_sys();
+  } else if (memory == Device) {
+    __nvvm_membar_gl();
   } else {
     __nvvm_membar_cta();
   }

Original file line number	Diff line number	Diff line change
`@@ -13,6 +13,8 @@ _CLC_OVERLOAD _CLC_DEF void __spirv_MemoryBarrier(unsigned int memory,`
`13`	`13`	`unsigned int semantics) {`
`14`	`14`	`if (memory == CrossDevice) {`
`15`	`15`	`__nvvm_membar_sys();`
	`16`	`+ } else if (memory == Device) {`
	`17`	`+ __nvvm_membar_gl();`
`16`	`18`	`} else {`
`17`	`19`	`__nvvm_membar_cta();`
`18`	`20`	`}`