[libclc] Add v3 variants of async_work_group_copy/async_work_group_strided_copy/prefetch (#137932)

3-component vector type is supported for them per OpenCL spec.

Author: wenju-he

libclc/generic/include/clc/async/async_work_group_copy.h

@@ -9,15 +9,21 @@
 #define __CLC_DST_ADDR_SPACE local
 #define __CLC_SRC_ADDR_SPACE global
 #define __CLC_BODY <clc/async/async_work_group_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+#define __CLC_BODY <clc/async/async_work_group_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY
 #undef __CLC_DST_ADDR_SPACE
 #undef __CLC_SRC_ADDR_SPACE
-#undef __CLC_BODY
 
 #define __CLC_DST_ADDR_SPACE global
 #define __CLC_SRC_ADDR_SPACE local
 #define __CLC_BODY <clc/async/async_work_group_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+#define __CLC_BODY <clc/async/async_work_group_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY
 #undef __CLC_DST_ADDR_SPACE
 #undef __CLC_SRC_ADDR_SPACE
-#undef __CLC_BODY

libclc/generic/include/clc/async/async_work_group_strided_copy.h

@@ -9,15 +9,21 @@
 #define __CLC_DST_ADDR_SPACE local
 #define __CLC_SRC_ADDR_SPACE global
 #define __CLC_BODY <clc/async/async_work_group_strided_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+#define __CLC_BODY <clc/async/async_work_group_strided_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY
 #undef __CLC_DST_ADDR_SPACE
 #undef __CLC_SRC_ADDR_SPACE
-#undef __CLC_BODY
 
 #define __CLC_DST_ADDR_SPACE global
 #define __CLC_SRC_ADDR_SPACE local
 #define __CLC_BODY <clc/async/async_work_group_strided_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+#define __CLC_BODY <clc/async/async_work_group_strided_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY
 #undef __CLC_DST_ADDR_SPACE
 #undef __CLC_SRC_ADDR_SPACE
-#undef __CLC_BODY

libclc/generic/include/clc/async/gentype.inc

@@ -7,5 +7,9 @@
 //===----------------------------------------------------------------------===//
 
 #define __CLC_BODY <clc/async/prefetch.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+
+#define __CLC_BODY <clc/async/prefetch.inc>
+#include <clc/math/gentype.inc>
 #undef __CLC_BODY

libclc/generic/include/clc/async/prefetch.h

@@ -9,4 +9,9 @@
 #include <clc/clc.h>
 
 #define __CLC_BODY <async_work_group_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+
+#define __CLC_BODY <async_work_group_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY

libclc/generic/lib/async/async_work_group_copy.cl

@@ -9,4 +9,9 @@
 #include <clc/clc.h>
 
 #define __CLC_BODY <async_work_group_strided_copy.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+
+#define __CLC_BODY <async_work_group_strided_copy.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY

libclc/generic/lib/async/async_work_group_strided_copy.cl

@@ -9,4 +9,9 @@
 #include <clc/clc.h>
 
 #define __CLC_BODY <prefetch.inc>
-#include <clc/async/gentype.inc>
+#include <clc/integer/gentype.inc>
+#undef __CLC_BODY
+
+#define __CLC_BODY <prefetch.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY