amdgcn: Implement {read_,write_,}mem_fence builtin
authorJan Vesely <jan.vesely@rutgers.edu>
Wed, 16 Aug 2017 17:08:56 +0000 (17:08 +0000)
committerJan Vesely <jan.vesely@rutgers.edu>
Wed, 16 Aug 2017 17:08:56 +0000 (17:08 +0000)
v2: add more detailed comment about waitcnt instruction

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
Tested-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 311021

libclc/amdgcn/lib/SOURCES
libclc/amdgcn/lib/mem_fence/fence.cl [new file with mode: 0644]
libclc/amdgcn/lib/mem_fence/waitcnt.ll [new file with mode: 0644]
libclc/generic/include/clc/clc.h
libclc/generic/include/clc/explicit_fence/explicit_memory_fence.h [new file with mode: 0644]

index 1ff5fd1..24f5949 100644 (file)
@@ -1,4 +1,6 @@
 math/ldexp.cl
+mem_fence/fence.cl
+mem_fence/waitcnt.ll
 synchronization/barrier_impl.ll
 workitem/get_global_offset.cl
 workitem/get_group_id.cl
diff --git a/libclc/amdgcn/lib/mem_fence/fence.cl b/libclc/amdgcn/lib/mem_fence/fence.cl
new file mode 100644 (file)
index 0000000..408ffc3
--- /dev/null
@@ -0,0 +1,39 @@
+#include <clc/clc.h>
+
+void __clc_amdgcn_s_waitcnt(unsigned flags);
+
+// s_waitcnt takes 16bit argument with a combined number of maximum allowed
+// pending operations:
+// [12:8] LGKM -- LDS, GDS, Konstant (SMRD), Messages
+// [7] -- undefined
+// [6:4] -- exports, GDS, and mem write
+// [3:0] -- vector memory operations
+
+// Newer clang supports __builtin_amdgcn_s_waitcnt
+#if __clang_major__ >= 5
+#  define __waitcnt(x) __builtin_amdgcn_s_waitcnt(x)
+#else
+#  define __waitcnt(x) __clc_amdgcn_s_waitcnt(x)
+#endif
+
+_CLC_DEF void mem_fence(cl_mem_fence_flags flags)
+{
+       if (flags & CLK_GLOBAL_MEM_FENCE) {
+               // scalar loads are counted with LGKM but we don't know whether
+               // the compiler turned any loads to scalar
+               __waitcnt(0);
+       } else if (flags & CLK_LOCAL_MEM_FENCE)
+               __waitcnt(0xff); // LGKM is [12:8]
+}
+#undef __waitcnt
+
+// We don't have separate mechanism for read and write fences
+_CLC_DEF void read_mem_fence(cl_mem_fence_flags flags)
+{
+       mem_fence(flags);
+}
+
+_CLC_DEF void write_mem_fence(cl_mem_fence_flags flags)
+{
+       mem_fence(flags);
+}
diff --git a/libclc/amdgcn/lib/mem_fence/waitcnt.ll b/libclc/amdgcn/lib/mem_fence/waitcnt.ll
new file mode 100644 (file)
index 0000000..8be7f18
--- /dev/null
@@ -0,0 +1,11 @@
+declare void @llvm.amdgcn.s.waitcnt(i32) #0
+
+; Export waitcnt intrinsic for clang < 5
+define void @__clc_amdgcn_s_waitcnt(i32 %flags) #1 {
+entry:
+  tail call void @llvm.amdgcn.s.waitcnt(i32 %flags)
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind alwaysinline }
index 5130632..deb9d70 100644 (file)
 #include <clc/synchronization/cl_mem_fence_flags.h>
 #include <clc/synchronization/barrier.h>
 
+/* 6.11.9 Explicit Memory Fence Functions */
+#include <clc/explicit_fence/explicit_memory_fence.h>
+
 /* 6.11.10 Async Copy and Prefetch Functions */
 #include <clc/async/async_work_group_copy.h>
 #include <clc/async/async_work_group_strided_copy.h>
diff --git a/libclc/generic/include/clc/explicit_fence/explicit_memory_fence.h b/libclc/generic/include/clc/explicit_fence/explicit_memory_fence.h
new file mode 100644 (file)
index 0000000..8e046b1
--- /dev/null
@@ -0,0 +1,3 @@
+_CLC_DECL void mem_fence(cl_mem_fence_flags flags);
+_CLC_DECL void read_mem_fence(cl_mem_fence_flags flags);
+_CLC_DECL void write_mem_fence(cl_mem_fence_flags flags);