l_copy_and_adjust_arguments_for_device: reduce scope of device-side buffer increment

This ensures the buffer increment is initialized to zero for every
kernel argument, not only the first one. This has no functional impact
since all kernel argument types use a non-default buffer increment.

Signed-off-by: Peter Colberg <[email protected]>

Author: pcolberg

src/acl_kernel.cpp

@@ -2758,8 +2758,6 @@ l_copy_and_adjust_arguments_for_device(cl_kernel kernel, cl_device_id device,
   cl_uint dev_local_ptr_size = 4; // Always.
   cl_uint dev_global_ptr_size = device->address_bits >> 3;
 
-  cl_uint buf_incr = 0;
-
   // Bump allocator pointer for each local aspace
   // Maps the aspace ID to the next available local memory address.
   std::unordered_map<unsigned, size_t> next_local;
@@ -2784,6 +2782,9 @@ l_copy_and_adjust_arguments_for_device(cl_kernel kernel, cl_device_id device,
     const acl_kernel_arg_info_t *arg_info =
         &(kernel->accel_def->iface.args[iarg]);
 
+    // Exclude kernel argument value from device-side buffer by default.
+    cl_uint buf_incr = 0;
+
     if (arg_info->addr_space == ACL_ARG_ADDR_LOCAL) {
 #ifdef MEM_DEBUG_MSG
       printf("local");