[clang] 0d7f8d4 - [OpenCL] Remove argument names from async copy builtins

[Sven van Haastregt via cfe-commits]

Author: Sven van Haastregt
Date: 2022-05-25T10:05:25+01:00
New Revision: 0d7f8d42fd170a434006a928a7106b894bdbdd16

URL: https://github.com/llvm/llvm-project/commit/0d7f8d42fd170a434006a928a7106b894bdbdd16
DIFF: https://github.com/llvm/llvm-project/commit/0d7f8d42fd170a434006a928a7106b894bdbdd16.diff

LOG: [OpenCL] Remove argument names from async copy builtins

This simplifies completeness comparisons against OpenCLBuiltins.td and
also makes the header no longer "claim" the argument name identifiers.

Continues the direction set out in D119560.

Added: 
    

Modified: 
    clang/lib/Headers/opencl-c.h

Removed: 
    


################################################################################
diff  --git a/clang/lib/Headers/opencl-c.h b/clang/lib/Headers/opencl-c.h
index 69e2c85610bae..1942da77f5e50 100644
--- a/clang/lib/Headers/opencl-c.h
+++ b/clang/lib/Headers/opencl-c.h
@@ -12515,141 +12515,141 @@ cl_mem_fence_flags __ovld get_fence(void *ptr);
  * synchronization of source data such as using a
  * barrier before performing the copy.
  */
-event_t __ovld async_work_group_copy(__local char *dst, const __global char *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar *dst, const __global uchar *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short *dst, const __global short *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort *dst, const __global ushort *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int *dst, const __global int *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint *dst, const __global uint *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long *dst, const __global long *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong *dst, const __global ulong *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float *dst, const __global float *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local char2 *dst, const __global char2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar2 *dst, const __global uchar2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short2 *dst, const __global short2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort2 *dst, const __global ushort2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int2 *dst, const __global int2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint2 *dst, const __global uint2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long2 *dst, const __global long2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong2 *dst, const __global ulong2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float2 *dst, const __global float2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local char3 *dst, const __global char3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar3 *dst, const __global uchar3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short3 *dst, const __global short3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort3 *dst, const __global ushort3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int3 *dst, const __global int3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint3 *dst, const __global uint3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long3 *dst, const __global long3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong3 *dst, const __global ulong3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float3 *dst, const __global float3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local char4 *dst, const __global char4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar4 *dst, const __global uchar4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short4 *dst, const __global short4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort4 *dst, const __global ushort4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int4 *dst, const __global int4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint4 *dst, const __global uint4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long4 *dst, const __global long4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong4 *dst, const __global ulong4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float4 *dst, const __global float4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local char8 *dst, const __global char8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar8 *dst, const __global uchar8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short8 *dst, const __global short8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort8 *dst, const __global ushort8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int8 *dst, const __global int8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint8 *dst, const __global uint8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long8 *dst, const __global long8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong8 *dst, const __global ulong8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float8 *dst, const __global float8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local char16 *dst, const __global char16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uchar16 *dst, const __global uchar16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local short16 *dst, const __global short16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ushort16 *dst, const __global ushort16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local int16 *dst, const __global int16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local uint16 *dst, const __global uint16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local long16 *dst, const __global long16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local ulong16 *dst, const __global ulong16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local float16 *dst, const __global float16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char *dst, const __local char *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar *dst, const __local uchar *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short *dst, const __local short *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort *dst, const __local ushort *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int *dst, const __local int *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint *dst, const __local uint *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long *dst, const __local long *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong *dst, const __local ulong *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float *dst, const __local float *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char2 *dst, const __local char2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar2 *dst, const __local uchar2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short2 *dst, const __local short2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort2 *dst, const __local ushort2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int2 *dst, const __local int2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint2 *dst, const __local uint2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long2 *dst, const __local long2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong2 *dst, const __local ulong2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float2 *dst, const __local float2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char3 *dst, const __local char3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar3 *dst, const __local uchar3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short3 *dst, const __local short3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort3 *dst, const __local ushort3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int3 *dst, const __local int3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint3 *dst, const __local uint3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long3 *dst, const __local long3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong3 *dst, const __local ulong3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float3 *dst, const __local float3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char4 *dst, const __local char4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar4 *dst, const __local uchar4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short4 *dst, const __local short4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort4 *dst, const __local ushort4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int4 *dst, const __local int4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint4 *dst, const __local uint4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long4 *dst, const __local long4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong4 *dst, const __local ulong4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float4 *dst, const __local float4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char8 *dst, const __local char8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar8 *dst, const __local uchar8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short8 *dst, const __local short8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort8 *dst, const __local ushort8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int8 *dst, const __local int8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint8 *dst, const __local uint8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long8 *dst, const __local long8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong8 *dst, const __local ulong8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float8 *dst, const __local float8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global char16 *dst, const __local char16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uchar16 *dst, const __local uchar16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global short16 *dst, const __local short16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ushort16 *dst, const __local ushort16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global int16 *dst, const __local int16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global uint16 *dst, const __local uint16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global long16 *dst, const __local long16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global ulong16 *dst, const __local ulong16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global float16 *dst, const __local float16 *src, size_t num_elements, event_t event);
+event_t __ovld async_work_group_copy(__local char *, const __global char *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar *, const __global uchar *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short *, const __global short *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort *, const __global ushort *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int *, const __global int *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint *, const __global uint *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long *, const __global long *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong *, const __global ulong *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float *, const __global float *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local char2 *, const __global char2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar2 *, const __global uchar2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short2 *, const __global short2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort2 *, const __global ushort2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int2 *, const __global int2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint2 *, const __global uint2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long2 *, const __global long2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong2 *, const __global ulong2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float2 *, const __global float2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local char3 *, const __global char3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar3 *, const __global uchar3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short3 *, const __global short3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort3 *, const __global ushort3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int3 *, const __global int3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint3 *, const __global uint3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long3 *, const __global long3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong3 *, const __global ulong3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float3 *, const __global float3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local char4 *, const __global char4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar4 *, const __global uchar4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short4 *, const __global short4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort4 *, const __global ushort4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int4 *, const __global int4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint4 *, const __global uint4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long4 *, const __global long4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong4 *, const __global ulong4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float4 *, const __global float4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local char8 *, const __global char8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar8 *, const __global uchar8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short8 *, const __global short8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort8 *, const __global ushort8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int8 *, const __global int8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint8 *, const __global uint8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long8 *, const __global long8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong8 *, const __global ulong8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float8 *, const __global float8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local char16 *, const __global char16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uchar16 *, const __global uchar16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local short16 *, const __global short16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ushort16 *, const __global ushort16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local int16 *, const __global int16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local uint16 *, const __global uint16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local long16 *, const __global long16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local ulong16 *, const __global ulong16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local float16 *, const __global float16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char *, const __local char *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar *, const __local uchar *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short *, const __local short *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort *, const __local ushort *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int *, const __local int *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint *, const __local uint *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long *, const __local long *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong *, const __local ulong *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float *, const __local float *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char2 *, const __local char2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar2 *, const __local uchar2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short2 *, const __local short2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort2 *, const __local ushort2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int2 *, const __local int2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint2 *, const __local uint2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long2 *, const __local long2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong2 *, const __local ulong2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float2 *, const __local float2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char3 *, const __local char3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar3 *, const __local uchar3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short3 *, const __local short3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort3 *, const __local ushort3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int3 *, const __local int3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint3 *, const __local uint3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long3 *, const __local long3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong3 *, const __local ulong3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float3 *, const __local float3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char4 *, const __local char4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar4 *, const __local uchar4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short4 *, const __local short4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort4 *, const __local ushort4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int4 *, const __local int4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint4 *, const __local uint4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long4 *, const __local long4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong4 *, const __local ulong4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float4 *, const __local float4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char8 *, const __local char8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar8 *, const __local uchar8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short8 *, const __local short8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort8 *, const __local ushort8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int8 *, const __local int8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint8 *, const __local uint8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long8 *, const __local long8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong8 *, const __local ulong8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float8 *, const __local float8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global char16 *, const __local char16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uchar16 *, const __local uchar16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global short16 *, const __local short16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ushort16 *, const __local ushort16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global int16 *, const __local int16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global uint16 *, const __local uint16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global long16 *, const __local long16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global ulong16 *, const __local ulong16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global float16 *, const __local float16 *, size_t, event_t);
 #ifdef cl_khr_fp64
-event_t __ovld async_work_group_copy(__local double *dst, const __global double *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local double2 *dst, const __global double2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local double3 *dst, const __global double3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local double4 *dst, const __global double4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local double8 *dst, const __global double8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local double16 *dst, const __global double16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double *dst, const __local double *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double2 *dst, const __local double2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double3 *dst, const __local double3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double4 *dst, const __local double4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double8 *dst, const __local double8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global double16 *dst, const __local double16 *src, size_t num_elements, event_t event);
+event_t __ovld async_work_group_copy(__local double *, const __global double *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local double2 *, const __global double2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local double3 *, const __global double3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local double4 *, const __global double4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local double8 *, const __global double8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local double16 *, const __global double16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double *, const __local double *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double2 *, const __local double2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double3 *, const __local double3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double4 *, const __local double4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double8 *, const __local double8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global double16 *, const __local double16 *, size_t, event_t);
 #endif //cl_khr_fp64
 #ifdef cl_khr_fp16
-event_t __ovld async_work_group_copy(__local half *dst, const __global half *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local half2 *dst, const __global half2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local half3 *dst, const __global half3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local half4 *dst, const __global half4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local half8 *dst, const __global half8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__local half16 *dst, const __global half16 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half *dst, const __local half *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half2 *dst, const __local half2 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half3 *dst, const __local half3 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half4 *dst, const __local half4 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half8 *dst, const __local half8 *src, size_t num_elements, event_t event);
-event_t __ovld async_work_group_copy(__global half16 *dst, const __local half16 *src, size_t num_elements, event_t event);
+event_t __ovld async_work_group_copy(__local half *, const __global half *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local half2 *, const __global half2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local half3 *, const __global half3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local half4 *, const __global half4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local half8 *, const __global half8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__local half16 *, const __global half16 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half *, const __local half *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half2 *, const __local half2 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half3 *, const __local half3 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half4 *, const __local half4 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half8 *, const __local half8 *, size_t, event_t);
+event_t __ovld async_work_group_copy(__global half16 *, const __local half16 *, size_t, event_t);
 #endif //cl_khr_fp16
 
 /**
@@ -12678,141 +12678,141 @@ event_t __ovld async_work_group_copy(__global half16 *dst, const __local half16
  * synchronization of source data such as using a
  * barrier before performing the copy.
  */
-event_t __ovld async_work_group_strided_copy(__local char *dst, const __global char *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar *dst, const __global uchar *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short *dst, const __global short *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort *dst, const __global ushort *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int *dst, const __global int *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint *dst, const __global uint *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long *dst, const __global long *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong *dst, const __global ulong *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float *dst, const __global float *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local char2 *dst, const __global char2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar2 *dst, const __global uchar2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short2 *dst, const __global short2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort2 *dst, const __global ushort2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int2 *dst, const __global int2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint2 *dst, const __global uint2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long2 *dst, const __global long2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong2 *dst, const __global ulong2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float2 *dst, const __global float2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local char3 *dst, const __global char3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar3 *dst, const __global uchar3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short3 *dst, const __global short3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort3 *dst, const __global ushort3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int3 *dst, const __global int3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint3 *dst, const __global uint3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long3 *dst, const __global long3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong3 *dst, const __global ulong3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float3 *dst, const __global float3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local char4 *dst, const __global char4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar4 *dst, const __global uchar4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short4 *dst, const __global short4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort4 *dst, const __global ushort4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int4 *dst, const __global int4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint4 *dst, const __global uint4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long4 *dst, const __global long4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong4 *dst, const __global ulong4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float4 *dst, const __global float4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local char8 *dst, const __global char8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar8 *dst, const __global uchar8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short8 *dst, const __global short8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort8 *dst, const __global ushort8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int8 *dst, const __global int8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint8 *dst, const __global uint8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long8 *dst, const __global long8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong8 *dst, const __global ulong8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float8 *dst, const __global float8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local char16 *dst, const __global char16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uchar16 *dst, const __global uchar16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local short16 *dst, const __global short16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ushort16 *dst, const __global ushort16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local int16 *dst, const __global int16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local uint16 *dst, const __global uint16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local long16 *dst, const __global long16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local ulong16 *dst, const __global ulong16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local float16 *dst, const __global float16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char *dst, const __local char *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar *dst, const __local uchar *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short *dst, const __local short *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort *dst, const __local ushort *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int *dst, const __local int *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint *dst, const __local uint *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long *dst, const __local long *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong *dst, const __local ulong *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float *dst, const __local float *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char2 *dst, const __local char2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar2 *dst, const __local uchar2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short2 *dst, const __local short2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort2 *dst, const __local ushort2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int2 *dst, const __local int2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint2 *dst, const __local uint2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long2 *dst, const __local long2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong2 *dst, const __local ulong2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float2 *dst, const __local float2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char3 *dst, const __local char3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar3 *dst, const __local uchar3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short3 *dst, const __local short3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort3 *dst, const __local ushort3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int3 *dst, const __local int3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint3 *dst, const __local uint3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long3 *dst, const __local long3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong3 *dst, const __local ulong3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float3 *dst, const __local float3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char4 *dst, const __local char4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar4 *dst, const __local uchar4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short4 *dst, const __local short4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort4 *dst, const __local ushort4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int4 *dst, const __local int4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint4 *dst, const __local uint4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long4 *dst, const __local long4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong4 *dst, const __local ulong4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float4 *dst, const __local float4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char8 *dst, const __local char8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar8 *dst, const __local uchar8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short8 *dst, const __local short8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort8 *dst, const __local ushort8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int8 *dst, const __local int8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint8 *dst, const __local uint8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long8 *dst, const __local long8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong8 *dst, const __local ulong8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float8 *dst, const __local float8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global char16 *dst, const __local char16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uchar16 *dst, const __local uchar16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global short16 *dst, const __local short16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ushort16 *dst, const __local ushort16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global int16 *dst, const __local int16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global uint16 *dst, const __local uint16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global long16 *dst, const __local long16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global ulong16 *dst, const __local ulong16 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global float16 *dst, const __local float16 *src, size_t num_elements, size_t dst_stride, event_t event);
+event_t __ovld async_work_group_strided_copy(__local char *, const __global char *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar *, const __global uchar *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short *, const __global short *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort *, const __global ushort *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int *, const __global int *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint *, const __global uint *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long *, const __global long *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong *, const __global ulong *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float *, const __global float *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local char2 *, const __global char2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar2 *, const __global uchar2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short2 *, const __global short2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort2 *, const __global ushort2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int2 *, const __global int2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint2 *, const __global uint2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long2 *, const __global long2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong2 *, const __global ulong2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float2 *, const __global float2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local char3 *, const __global char3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar3 *, const __global uchar3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short3 *, const __global short3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort3 *, const __global ushort3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int3 *, const __global int3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint3 *, const __global uint3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long3 *, const __global long3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong3 *, const __global ulong3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float3 *, const __global float3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local char4 *, const __global char4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar4 *, const __global uchar4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short4 *, const __global short4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort4 *, const __global ushort4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int4 *, const __global int4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint4 *, const __global uint4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long4 *, const __global long4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong4 *, const __global ulong4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float4 *, const __global float4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local char8 *, const __global char8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar8 *, const __global uchar8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short8 *, const __global short8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort8 *, const __global ushort8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int8 *, const __global int8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint8 *, const __global uint8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long8 *, const __global long8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong8 *, const __global ulong8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float8 *, const __global float8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local char16 *, const __global char16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uchar16 *, const __global uchar16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local short16 *, const __global short16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ushort16 *, const __global ushort16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local int16 *, const __global int16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local uint16 *, const __global uint16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local long16 *, const __global long16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local ulong16 *, const __global ulong16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local float16 *, const __global float16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char *, const __local char *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar *, const __local uchar *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short *, const __local short *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort *, const __local ushort *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int *, const __local int *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint *, const __local uint *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long *, const __local long *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong *, const __local ulong *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float *, const __local float *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char2 *, const __local char2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar2 *, const __local uchar2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short2 *, const __local short2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort2 *, const __local ushort2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int2 *, const __local int2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint2 *, const __local uint2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long2 *, const __local long2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong2 *, const __local ulong2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float2 *, const __local float2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char3 *, const __local char3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar3 *, const __local uchar3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short3 *, const __local short3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort3 *, const __local ushort3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int3 *, const __local int3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint3 *, const __local uint3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long3 *, const __local long3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong3 *, const __local ulong3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float3 *, const __local float3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char4 *, const __local char4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar4 *, const __local uchar4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short4 *, const __local short4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort4 *, const __local ushort4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int4 *, const __local int4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint4 *, const __local uint4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long4 *, const __local long4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong4 *, const __local ulong4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float4 *, const __local float4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char8 *, const __local char8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar8 *, const __local uchar8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short8 *, const __local short8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort8 *, const __local ushort8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int8 *, const __local int8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint8 *, const __local uint8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long8 *, const __local long8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong8 *, const __local ulong8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float8 *, const __local float8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global char16 *, const __local char16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uchar16 *, const __local uchar16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global short16 *, const __local short16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ushort16 *, const __local ushort16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global int16 *, const __local int16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global uint16 *, const __local uint16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global long16 *, const __local long16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global ulong16 *, const __local ulong16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global float16 *, const __local float16 *, size_t, size_t, event_t);
 #ifdef cl_khr_fp64
-event_t __ovld async_work_group_strided_copy(__local double *dst, const __global double *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local double2 *dst, const __global double2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local double3 *dst, const __global double3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local double4 *dst, const __global double4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local double8 *dst, const __global double8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local double16 *dst, const __global double16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double *dst, const __local double *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double2 *dst, const __local double2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double3 *dst, const __local double3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double4 *dst, const __local double4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double8 *dst, const __local double8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global double16 *dst, const __local double16 *src, size_t num_elements, size_t dst_stride, event_t event);
+event_t __ovld async_work_group_strided_copy(__local double *, const __global double *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local double2 *, const __global double2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local double3 *, const __global double3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local double4 *, const __global double4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local double8 *, const __global double8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local double16 *, const __global double16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double *, const __local double *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double2 *, const __local double2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double3 *, const __local double3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double4 *, const __local double4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double8 *, const __local double8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global double16 *, const __local double16 *, size_t, size_t, event_t);
 #endif //cl_khr_fp64
 #ifdef cl_khr_fp16
-event_t __ovld async_work_group_strided_copy(__local half *dst, const __global half *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local half2 *dst, const __global half2 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local half3 *dst, const __global half3 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local half4 *dst, const __global half4 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local half8 *dst, const __global half8 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__local half16 *dst, const __global half16 *src, size_t num_elements, size_t src_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half *dst, const __local half *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half2 *dst, const __local half2 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half3 *dst, const __local half3 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half4 *dst, const __local half4 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half8 *dst, const __local half8 *src, size_t num_elements, size_t dst_stride, event_t event);
-event_t __ovld async_work_group_strided_copy(__global half16 *dst, const __local half16 *src, size_t num_elements, size_t dst_stride, event_t event);
+event_t __ovld async_work_group_strided_copy(__local half *, const __global half *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local half2 *, const __global half2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local half3 *, const __global half3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local half4 *, const __global half4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local half8 *, const __global half8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__local half16 *, const __global half16 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half *, const __local half *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half2 *, const __local half2 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half3 *, const __local half3 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half4 *, const __local half4 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half8 *, const __local half8 *, size_t, size_t, event_t);
+event_t __ovld async_work_group_strided_copy(__global half16 *, const __local half16 *, size_t, size_t, event_t);
 #endif //cl_khr_fp16
 
 /**
@@ -12826,7 +12826,7 @@ event_t __ovld async_work_group_strided_copy(__global half16 *dst, const __local
  * the same num_events and event objects specified
  * in event_list; otherwise the results are undefined.
  */
-void __ovld wait_group_events(int num_events, event_t *event_list);
+void __ovld wait_group_events(int, event_t *);
 
 /**
  * Prefetch num_elements * sizeof(gentype)
@@ -12835,75 +12835,75 @@ void __ovld wait_group_events(int num_events, event_t *event_list);
  * and does not affect the functional
  * behavior of the kernel.
  */
-void __ovld prefetch(const __global char *p, size_t num_elements);
-void __ovld prefetch(const __global uchar *p, size_t num_elements);
-void __ovld prefetch(const __global short *p, size_t num_elements);
-void __ovld prefetch(const __global ushort *p, size_t num_elements);
-void __ovld prefetch(const __global int *p, size_t num_elements);
-void __ovld prefetch(const __global uint *p, size_t num_elements);
-void __ovld prefetch(const __global long *p, size_t num_elements);
-void __ovld prefetch(const __global ulong *p, size_t num_elements);
-void __ovld prefetch(const __global float *p, size_t num_elements);
-void __ovld prefetch(const __global char2 *p, size_t num_elements);
-void __ovld prefetch(const __global uchar2 *p, size_t num_elements);
-void __ovld prefetch(const __global short2 *p, size_t num_elements);
-void __ovld prefetch(const __global ushort2 *p, size_t num_elements);
-void __ovld prefetch(const __global int2 *p, size_t num_elements);
-void __ovld prefetch(const __global uint2 *p, size_t num_elements);
-void __ovld prefetch(const __global long2 *p, size_t num_elements);
-void __ovld prefetch(const __global ulong2 *p, size_t num_elements);
-void __ovld prefetch(const __global float2 *p, size_t num_elements);
-void __ovld prefetch(const __global char3 *p, size_t num_elements);
-void __ovld prefetch(const __global uchar3 *p, size_t num_elements);
-void __ovld prefetch(const __global short3 *p, size_t num_elements);
-void __ovld prefetch(const __global ushort3 *p, size_t num_elements);
-void __ovld prefetch(const __global int3 *p, size_t num_elements);
-void __ovld prefetch(const __global uint3 *p, size_t num_elements);
-void __ovld prefetch(const __global long3 *p, size_t num_elements);
-void __ovld prefetch(const __global ulong3 *p, size_t num_elements);
-void __ovld prefetch(const __global float3 *p, size_t num_elements);
-void __ovld prefetch(const __global char4 *p, size_t num_elements);
-void __ovld prefetch(const __global uchar4 *p, size_t num_elements);
-void __ovld prefetch(const __global short4 *p, size_t num_elements);
-void __ovld prefetch(const __global ushort4 *p, size_t num_elements);
-void __ovld prefetch(const __global int4 *p, size_t num_elements);
-void __ovld prefetch(const __global uint4 *p, size_t num_elements);
-void __ovld prefetch(const __global long4 *p, size_t num_elements);
-void __ovld prefetch(const __global ulong4 *p, size_t num_elements);
-void __ovld prefetch(const __global float4 *p, size_t num_elements);
-void __ovld prefetch(const __global char8 *p, size_t num_elements);
-void __ovld prefetch(const __global uchar8 *p, size_t num_elements);
-void __ovld prefetch(const __global short8 *p, size_t num_elements);
-void __ovld prefetch(const __global ushort8 *p, size_t num_elements);
-void __ovld prefetch(const __global int8 *p, size_t num_elements);
-void __ovld prefetch(const __global uint8 *p, size_t num_elements);
-void __ovld prefetch(const __global long8 *p, size_t num_elements);
-void __ovld prefetch(const __global ulong8 *p, size_t num_elements);
-void __ovld prefetch(const __global float8 *p, size_t num_elements);
-void __ovld prefetch(const __global char16 *p, size_t num_elements);
-void __ovld prefetch(const __global uchar16 *p, size_t num_elements);
-void __ovld prefetch(const __global short16 *p, size_t num_elements);
-void __ovld prefetch(const __global ushort16 *p, size_t num_elements);
-void __ovld prefetch(const __global int16 *p, size_t num_elements);
-void __ovld prefetch(const __global uint16 *p, size_t num_elements);
-void __ovld prefetch(const __global long16 *p, size_t num_elements);
-void __ovld prefetch(const __global ulong16 *p, size_t num_elements);
-void __ovld prefetch(const __global float16 *p, size_t num_elements);
+void __ovld prefetch(const __global char *, size_t);
+void __ovld prefetch(const __global uchar *, size_t);
+void __ovld prefetch(const __global short *, size_t);
+void __ovld prefetch(const __global ushort *, size_t);
+void __ovld prefetch(const __global int *, size_t);
+void __ovld prefetch(const __global uint *, size_t);
+void __ovld prefetch(const __global long *, size_t);
+void __ovld prefetch(const __global ulong *, size_t);
+void __ovld prefetch(const __global float *, size_t);
+void __ovld prefetch(const __global char2 *, size_t);
+void __ovld prefetch(const __global uchar2 *, size_t);
+void __ovld prefetch(const __global short2 *, size_t);
+void __ovld prefetch(const __global ushort2 *, size_t);
+void __ovld prefetch(const __global int2 *, size_t);
+void __ovld prefetch(const __global uint2 *, size_t);
+void __ovld prefetch(const __global long2 *, size_t);
+void __ovld prefetch(const __global ulong2 *, size_t);
+void __ovld prefetch(const __global float2 *, size_t);
+void __ovld prefetch(const __global char3 *, size_t);
+void __ovld prefetch(const __global uchar3 *, size_t);
+void __ovld prefetch(const __global short3 *, size_t);
+void __ovld prefetch(const __global ushort3 *, size_t);
+void __ovld prefetch(const __global int3 *, size_t);
+void __ovld prefetch(const __global uint3 *, size_t);
+void __ovld prefetch(const __global long3 *, size_t);
+void __ovld prefetch(const __global ulong3 *, size_t);
+void __ovld prefetch(const __global float3 *, size_t);
+void __ovld prefetch(const __global char4 *, size_t);
+void __ovld prefetch(const __global uchar4 *, size_t);
+void __ovld prefetch(const __global short4 *, size_t);
+void __ovld prefetch(const __global ushort4 *, size_t);
+void __ovld prefetch(const __global int4 *, size_t);
+void __ovld prefetch(const __global uint4 *, size_t);
+void __ovld prefetch(const __global long4 *, size_t);
+void __ovld prefetch(const __global ulong4 *, size_t);
+void __ovld prefetch(const __global float4 *, size_t);
+void __ovld prefetch(const __global char8 *, size_t);
+void __ovld prefetch(const __global uchar8 *, size_t);
+void __ovld prefetch(const __global short8 *, size_t);
+void __ovld prefetch(const __global ushort8 *, size_t);
+void __ovld prefetch(const __global int8 *, size_t);
+void __ovld prefetch(const __global uint8 *, size_t);
+void __ovld prefetch(const __global long8 *, size_t);
+void __ovld prefetch(const __global ulong8 *, size_t);
+void __ovld prefetch(const __global float8 *, size_t);
+void __ovld prefetch(const __global char16 *, size_t);
+void __ovld prefetch(const __global uchar16 *, size_t);
+void __ovld prefetch(const __global short16 *, size_t);
+void __ovld prefetch(const __global ushort16 *, size_t);
+void __ovld prefetch(const __global int16 *, size_t);
+void __ovld prefetch(const __global uint16 *, size_t);
+void __ovld prefetch(const __global long16 *, size_t);
+void __ovld prefetch(const __global ulong16 *, size_t);
+void __ovld prefetch(const __global float16 *, size_t);
 #ifdef cl_khr_fp64
-void __ovld prefetch(const __global double *p, size_t num_elements);
-void __ovld prefetch(const __global double2 *p, size_t num_elements);
-void __ovld prefetch(const __global double3 *p, size_t num_elements);
-void __ovld prefetch(const __global double4 *p, size_t num_elements);
-void __ovld prefetch(const __global double8 *p, size_t num_elements);
-void __ovld prefetch(const __global double16 *p, size_t num_elements);
+void __ovld prefetch(const __global double *, size_t);
+void __ovld prefetch(const __global double2 *, size_t);
+void __ovld prefetch(const __global double3 *, size_t);
+void __ovld prefetch(const __global double4 *, size_t);
+void __ovld prefetch(const __global double8 *, size_t);
+void __ovld prefetch(const __global double16 *, size_t);
 #endif //cl_khr_fp64
 #ifdef cl_khr_fp16
-void __ovld prefetch(const __global half *p, size_t num_elements);
-void __ovld prefetch(const __global half2 *p, size_t num_elements);
-void __ovld prefetch(const __global half3 *p, size_t num_elements);
-void __ovld prefetch(const __global half4 *p, size_t num_elements);
-void __ovld prefetch(const __global half8 *p, size_t num_elements);
-void __ovld prefetch(const __global half16 *p, size_t num_elements);
+void __ovld prefetch(const __global half *, size_t);
+void __ovld prefetch(const __global half2 *, size_t);
+void __ovld prefetch(const __global half3 *, size_t);
+void __ovld prefetch(const __global half4 *, size_t);
+void __ovld prefetch(const __global half8 *, size_t);
+void __ovld prefetch(const __global half16 *, size_t);
 #endif // cl_khr_fp16
 
 // OpenCL v1.1 s6.11.1, v1.2 s6.12.11 - Atomic Functions