Update reference file after merge.

reference/opt/shaders-msl/comp/shared-array-of-arrays.comp

@@ -10,7 +10,7 @@ struct SSBO
     float out_data[1];
 };
 
-kernel void main0(uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]], uint gl_LocalInvocationIndex [[thread_index_in_threadgroup]], device SSBO& _67 [[buffer(0)]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
+kernel void main0(device SSBO& _67 [[buffer(0)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]], uint gl_LocalInvocationIndex [[thread_index_in_threadgroup]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
 {
     threadgroup float foo[4][4];
     foo[gl_LocalInvocationID.x][gl_LocalInvocationID.y] = float(gl_LocalInvocationIndex);

reference/shaders-msl/comp/shared-array-of-arrays.comp

@@ -24,7 +24,7 @@ void work(threadgroup float (&foo)[4][4], thread uint3& gl_LocalInvocationID, th
     v_67.out_data[gl_GlobalInvocationID.x] = x;
 }
 
-kernel void main0(uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]], uint gl_LocalInvocationIndex [[thread_index_in_threadgroup]], device SSBO& v_67 [[buffer(0)]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
+kernel void main0(device SSBO& v_67 [[buffer(0)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]], uint gl_LocalInvocationIndex [[thread_index_in_threadgroup]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
 {
     threadgroup float foo[4][4];
     work(foo, gl_LocalInvocationID, gl_LocalInvocationIndex, v_67, gl_GlobalInvocationID);