Duff's device

From Wikipedia, the free encyclopedia
  (Redirected from Duffs device)
Jump to: navigation, search

In the C programming language, Duff's device is a way of (manually) implementing loop unrolling by interleaving two syntactic constructs of C, the do-while loop construct, and a switch statement. Its discovery is credited to Tom Duff in November 1983, who at the time was working for Lucasfilm and used it to speed up a real-time animation program.

Loop unrolling attempts to reduce the overhead of conditional branching needed to check whether a loop is done, by executing a batch of loops bodies per iteration. To handle cases where the number of iterations is not divisible by the unrolled-loop increments, a common technique among assembly programmers is to jump directly into the middle of the unrolled loop body to handle the remainder.[1] Duff was looking to implement this technique in C, and succeeded in doing so by using C's case label fall-through feature to jump into the unrolled body.[2]

When used with modern optimizing compilers, Duff's device may no longer provide any performance improvements.[3]

Original version[edit]

Duff's problem was to copy 16-bit units ("shorts" in C) from an array into a memory-mapped output register, denoted in C by a pointer. His original code, in K&R C, looked as follows:

send(to, from, count)
register short *to, *from;
register count;
{
    do {                          /* count > 0 assumed */
        *to = *from++;
    } while(--count > 0);
}

This code assumes that count > 0. The pointer to is not incremented as it would be required for a memory-to-memory copy. If count were always divisible by eight, unrolling this loop eight-fold would produce the following:

send(to, from, count)
register short *to, *from;
register count;
{
    register n = count / 8;
    do {
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
    } while (--n > 0);
}

Duff realized that to handle cases where count is not divisible by eight, the assembly programmer's technique of jumping into the loop body could be implemented by interlacing the structures of a switch statement and a loop, putting the switch's case labels at the points of the loop body that correspond to the remainder of count / 8:[1]

send(to, from, count)
register short *to, *from;
register count;
{
    register n = (count + 7) / 8;
    switch (count % 8) {
    case 0: do { *to = *from++;
    case 7:      *to = *from++;
    case 6:      *to = *from++;
    case 5:      *to = *from++;
    case 4:      *to = *from++;
    case 3:      *to = *from++;
    case 2:      *to = *from++;
    case 1:      *to = *from++;
            } while (--n > 0);
    }
}

Duff's device can similarly be applied with any other size for the unrolled loop, not just eight as in the example above.

Mechanism[edit]

Based on an algorithm used widely by programmers coding in assembly for minimizing the number of tests and branches during a copy, Duff's device appears out of place when implemented in C. The device is valid C by virtue of two attributes in C:

  1. Relaxed specification of the switch statement in the language's definition. At the time of the device's invention this was the first edition of The C Programming Language which requires only that the body of the switch be a syntactically valid (compound) statement within which case labels can appear prefixing any sub-statement. In conjunction with the fact that, in the absence of a break statement, the flow of control will fall through from a statement controlled by one case label to that controlled by the next, this means that the code specifies a succession of count copies from sequential source addresses to the memory-mapped output port.
  2. The ability to jump into the middle of a loop in C.

This leads to what the Jargon File calls "the most dramatic use yet seen of fall through in C".[4] C's default fall-through in case statements has long been one of its most controversial features; Duff himself said that "This code forms some sort of argument in that debate, but I'm not sure whether it's for or against."[4]

Simplified explanation[edit]

A functionally equivalent version
with switch and do disentangled
send(to, from, count)
register short *to, *from;
register count;
{
    register n = (count + 7) / 8;
    switch (count % 8) {
        case 0: *to = *from++;
        case 7: *to = *from++;
        case 6: *to = *from++;
        case 5: *to = *from++;
        case 4: *to = *from++;
        case 3: *to = *from++;
        case 2: *to = *from++;
        case 1: *to = *from++;
    }
    while (--n > 0) {
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
        *to = *from++;
    }
}

The basic idea of Duff's device is that the CPU time can be saved when running a loop by reducing the number of loop tests. For example, in the case of a loop with only a single instruction in the block code, the loop test will be performed for every iteration of the loop, that is every time the instruction is executed. If, instead, eight copies of the same instruction are placed in the loop, then the test will be performed only every eight iterations, and this will gain time by avoiding seven tests (this method is called loop unwinding). The problem is that for this to work the total number of iterations must be a multiple of eight.[1]

Duff's device provides a solution by first performing the "extra" number of iterations, after which a multiple of eight remains (in fact, the remainder of the integral division by eight), followed by iterating as many times as necessary the groups of eight similar instructions. To obtain the number of "extra" iterations, the code calculates the value of the total number of iterations modulo eight and then applies the following trick: according to that value, the processor will jump to a case statement placed in such a way that it is followed by exactly the number of iterations you need. Once this is done, everything is straightforward – the code continues by doing iterations of groups of eight instructions, this has become possible since the remaining number of iterations is a multiple of eight.[1]

An unusual feature of the code is that, to be able to do that jump, case keywords are required both inside and outside the loop. This is unusual because the contents of a case statement are traditionally thought of as a block of code nested inside the case statement, and a reader would typically expect it to end before the next case statement. According to the specifications of C language, this is not necessary; indeed, case statements can appear anywhere inside the switch code block, and at any depth; the processor will simply jump to the next statement, wherever it may be.

Performance[edit]

Many compilers will optimize the switch into a jump table just as would be done in an assembly implementation.

The primary increase in speed versus a simple, straightforward loop, comes from loop unwinding that reduces the number of performed branches, which are computationally expensive due to the need to flush—​​and hence stall—​​the instruction pipeline. The switch statement is used to handle the remainder of the data not evenly divisible by the number of operations unrolled (in this example, eight byte moves are unrolled, so the switch handles an extra 1–7 bytes automatically).

This automatic handling of the remainder may not be the best solution on all systems and compilers – in some cases two loops may actually be faster (one loop, unrolled, to do the main copy, and a second loop to handle the remainder). The problem appears to come down to the ability of the compiler to correctly optimize the device; it may also interfere with pipelining and branch prediction on some architectures.[5] When numerous instances of Duff's device were removed from the XFree86 Server in version 4.0, there was an improvement in performance and a noticeable reduction in size of the executable.[3] Therefore, when considering using this code, it may be worth running a few benchmarks to verify that it actually is the fastest code on the target architecture, at the target optimization level, with the target compiler.

For the purpose of memory-to-memory copies (which was not the original use of Duff's device, although it can be modified to serve this purpose as described in section below), the standard C library provides function memcpy; it will not perform worse than a memory-to-memory copy version of this code, and may contain architecture-specific optimizations that will make it significantly faster.[6][7]

Stroustrup's version[edit]

The original Device was designed to copy to a memory-mapped I/O port, pointed to by the variable to. To copy one memory location to another, one solution would be to auto-increment to in every assignment statement; that is, change every *to = *from++; to *to++ = *from++;. This modification of the Device appears in a "what does this code do?" exercise in Bjarne Stroustrup's book The C++ Programming Language.[8]

See also[edit]

  • Coroutine – Duff's device can be used to implement coroutines in C/C++

Notes[edit]

References[edit]

  1. ^ a b c d Ralf Holly (2005-08-01). "A Reusable Duff Device". Dr. Dobb's Journal. Retrieved 2015-09-18. 
  2. ^ Tom Duff (1988-08-29). "Subject: Re: Explanation, please!". Lysator. Retrieved 2015-11-03. 
  3. ^ a b Ted Tso (2000-08-22). "Re: [PATCH] Re: Move of input drivers, some word needed from you". lkml.indiana.edu. Linux kernel mailing list. Retrieved 2014-08-22. Jim Gettys has a wonderful explanation of this effect in the X server. It turns out that with branch predictions and the relative speed of CPU vs. memory changing over the past decade, loop unrolling is pretty much pointless. In fact, by eliminating all instances of Duff's Device from the XFree86 4.0 server, the server shrunk in size by _half_ _a_ _megabyte_ (!!!), and was faster to boot, because the elimination of all that excess code meant that the X server wasn't thrashing the cache lines as much. 
  4. ^ a b Duff's device from FOLDOC
  5. ^ James Ralston's USENIX 2003 Journal
  6. ^ Wall, Mike (2002-03-19). "Using Block Prefetch for Optimized Memory Performance" (PDF). mit.edu. Retrieved 2012-09-22. 
  7. ^ Fog, Agner (2012-02-29). "Optimizing subroutines in assembly language" (PDF). Copenhagen University College of Engineering. pp. 100 ff. Retrieved 2012-09-22. 
  8. ^ Bjarne Stroustrup. The C++ Programming Language (3rd ed.). Addison-Wesley. ISBN 0-201-88954-4.  Exercise 15 in Section 6.6 (p.141)

This article is based on material taken from Duff's Device at the Free On-line Dictionary of Computing prior to 1 November 2008 and incorporated under the "relicensing" terms of the GFDL, version 1.3 or later.

Further reading[edit]

External links[edit]