CLR TVF performs slower when used via cross apply than when used in direct function call

The code I am using

   create table largecsv (val nchar(1000))    insert into largecsv(val) values ('aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee')               select * from dbo.SqlFunction1('aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee') -- 23%    select * from csvtotable('aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee,aaa,bbb,ccc,ddd,eee',',') -- 77%        select F.Val  from largecsv v CROSS APPLY dbo.SqlFunction1(val) F --72%    select F.RESULT  from largecsv CROSS APPLY csvtotable(val,',') F -- 28%

`csvtotable` comes from http://www.codeproject.com/Tips/625872/Convert-a-CSV-delimited-string-to-table-column-in where I changed the `BIGINT` to nchar(1000)

   CREATE FUNCTION [dbo].[CSVtoTable]    (        @LIST nchar(1000),        @Delimeter varchar(10)    )    RETURNS @RET1 TABLE (RESULT nchar(1000))    AS    BEGIN        DECLARE @RET TABLE(RESULT nchar(1000))                IF LTRIM(RTRIM(@LIST))='' RETURN             DECLARE @START BIGINT        DECLARE @LASTSTART BIGINT        SET @LASTSTART=0        SET @START=CHARINDEX(@Delimeter,@LIST,0)            IF @START=0        INSERT INTO @RET VALUES(SUBSTRING(@LIST,0,LEN(@LIST)+1))            WHILE(@START >0)        BEGIN            INSERT INTO @RET VALUES(SUBSTRING(@LIST,@LASTSTART,@START-@LASTSTART))            SET @LASTSTART=@START+1            SET @START=CHARINDEX(@Delimeter,@LIST,@START+1)            IF(@START=0)            INSERT INTO @RET VALUES(SUBSTRING(@LIST,@LASTSTART,LEN(@LIST)+1))        END                INSERT INTO @RET1 SELECT * FROM @RET        RETURN     END

and `SqlFunction1` is a CLR function with the definition

   [SqlFunction(Name="SqlFunction1",FillRowMethodName = "FillRow",TableDefinition="val nvarchar(1000)")]        public static IEnumerable SqlFunction1(SqlString val)        {            string[] splitStr = val.Value.Split(',');            return splitStr;        }        private static void FillRow(Object obj, out SqlString str)        {            str = new SqlString((string)obj);        }

Why is it that when I pass a static string, I get significantly better performance with the CLR function, but if I do a cross apply, even on a table with just a single row, I get much better performance with the TSQL counterpart?

The relative execution cost is:

Static string:  CLR:23% (Query 1),TSQL:77% (Query 2)

Cross apply from table with single row: CLR:72% (Query 3),TSQL:28%(Query 4)

This is running on SQL Server 2012

Looking at the Execution plan, with the cross apply there is a Nested loop join which has a relative cost of 49% in the CLRUDF, but 0% in the SQL UDF.