Chris@42: @node Other Important Topics, FFTW Reference, Tutorial, Top Chris@42: @chapter Other Important Topics Chris@42: @menu Chris@42: * SIMD alignment and fftw_malloc:: Chris@42: * Multi-dimensional Array Format:: Chris@42: * Words of Wisdom-Saving Plans:: Chris@42: * Caveats in Using Wisdom:: Chris@42: @end menu Chris@42: Chris@42: @c ------------------------------------------------------------ Chris@42: @node SIMD alignment and fftw_malloc, Multi-dimensional Array Format, Other Important Topics, Other Important Topics Chris@42: @section SIMD alignment and fftw_malloc Chris@42: Chris@42: SIMD, which stands for ``Single Instruction Multiple Data,'' is a set of Chris@42: special operations supported by some processors to perform a single Chris@42: operation on several numbers (usually 2 or 4) simultaneously. SIMD Chris@42: floating-point instructions are available on several popular CPUs: Chris@42: SSE/SSE2/AVX/AVX2/AVX512/KCVI on some x86/x86-64 processors, AltiVec and Chris@42: VSX on some POWER/PowerPCs, NEON on some ARM models. FFTW can be Chris@42: compiled to support the SIMD instructions on any of these systems. Chris@42: @cindex SIMD Chris@42: @cindex SSE Chris@42: @cindex SSE2 Chris@42: @cindex AVX Chris@42: @cindex AVX2 Chris@42: @cindex AVX512 Chris@42: @cindex AltiVec Chris@42: @cindex VSX Chris@42: @cindex precision Chris@42: Chris@42: Chris@42: A program linking to an FFTW library compiled with SIMD support can Chris@42: obtain a nonnegligible speedup for most complex and r2c/c2r Chris@42: transforms. In order to obtain this speedup, however, the arrays of Chris@42: complex (or real) data passed to FFTW must be specially aligned in Chris@42: memory (typically 16-byte aligned), and often this alignment is more Chris@42: stringent than that provided by the usual @code{malloc} (etc.) Chris@42: allocation routines. Chris@42: Chris@42: @cindex portability Chris@42: In order to guarantee proper alignment for SIMD, therefore, in case Chris@42: your program is ever linked against a SIMD-using FFTW, we recommend Chris@42: allocating your transform data with @code{fftw_malloc} and Chris@42: de-allocating it with @code{fftw_free}. Chris@42: @findex fftw_malloc Chris@42: @findex fftw_free Chris@42: These have exactly the same interface and behavior as Chris@42: @code{malloc}/@code{free}, except that for a SIMD FFTW they ensure Chris@42: that the returned pointer has the necessary alignment (by calling Chris@42: @code{memalign} or its equivalent on your OS). Chris@42: Chris@42: You are not @emph{required} to use @code{fftw_malloc}. You can Chris@42: allocate your data in any way that you like, from @code{malloc} to Chris@42: @code{new} (in C++) to a fixed-size array declaration. If the array Chris@42: happens not to be properly aligned, FFTW will not use the SIMD Chris@42: extensions. Chris@42: @cindex C++ Chris@42: Chris@42: @findex fftw_alloc_real Chris@42: @findex fftw_alloc_complex Chris@42: Since @code{fftw_malloc} only ever needs to be used for real and Chris@42: complex arrays, we provide two convenient wrapper routines Chris@42: @code{fftw_alloc_real(N)} and @code{fftw_alloc_complex(N)} that are Chris@42: equivalent to @code{(double*)fftw_malloc(sizeof(double) * N)} and Chris@42: @code{(fftw_complex*)fftw_malloc(sizeof(fftw_complex) * N)}, Chris@42: respectively (or their equivalents in other precisions). Chris@42: Chris@42: @c ------------------------------------------------------------ Chris@42: @node Multi-dimensional Array Format, Words of Wisdom-Saving Plans, SIMD alignment and fftw_malloc, Other Important Topics Chris@42: @section Multi-dimensional Array Format Chris@42: Chris@42: This section describes the format in which multi-dimensional arrays Chris@42: are stored in FFTW. We felt that a detailed discussion of this topic Chris@42: was necessary. Since several different formats are common, this topic Chris@42: is often a source of confusion. Chris@42: Chris@42: @menu Chris@42: * Row-major Format:: Chris@42: * Column-major Format:: Chris@42: * Fixed-size Arrays in C:: Chris@42: * Dynamic Arrays in C:: Chris@42: * Dynamic Arrays in C-The Wrong Way:: Chris@42: @end menu Chris@42: Chris@42: @c =========> Chris@42: @node Row-major Format, Column-major Format, Multi-dimensional Array Format, Multi-dimensional Array Format Chris@42: @subsection Row-major Format Chris@42: @cindex row-major Chris@42: Chris@42: The multi-dimensional arrays passed to @code{fftw_plan_dft} etcetera Chris@42: are expected to be stored as a single contiguous block in Chris@42: @dfn{row-major} order (sometimes called ``C order''). Basically, this Chris@42: means that as you step through adjacent memory locations, the first Chris@42: dimension's index varies most slowly and the last dimension's index Chris@42: varies most quickly. Chris@42: Chris@42: To be more explicit, let us consider an array of rank @math{d} whose Chris@42: dimensions are @ndims{}. Now, we specify a location in the array by a Chris@42: sequence of @math{d} (zero-based) indices, one for each dimension: Chris@42: @tex Chris@42: $(i_0, i_1, i_2, \ldots, i_{d-1})$. Chris@42: @end tex Chris@42: @ifinfo Chris@42: (i[0], i[1], ..., i[d-1]). Chris@42: @end ifinfo Chris@42: @html Chris@42: (i0, i1, i2,..., id-1). Chris@42: @end html Chris@42: If the array is stored in row-major Chris@42: order, then this element is located at the position Chris@42: @tex Chris@42: $i_{d-1} + n_{d-1} (i_{d-2} + n_{d-2} (\ldots + n_1 i_0))$. Chris@42: @end tex Chris@42: @ifinfo Chris@42: i[d-1] + n[d-1] * (i[d-2] + n[d-2] * (... + n[1] * i[0])). Chris@42: @end ifinfo Chris@42: @html Chris@42: id-1 + nd-1 * (id-2 + nd-2 * (... + n1 * i0)). Chris@42: @end html Chris@42: Chris@42: Note that, for the ordinary complex DFT, each element of the array Chris@42: must be of type @code{fftw_complex}; i.e. a (real, imaginary) pair of Chris@42: (double-precision) numbers. Chris@42: Chris@42: In the advanced FFTW interface, the physical dimensions @math{n} from Chris@42: which the indices are computed can be different from (larger than) Chris@42: the logical dimensions of the transform to be computed, in order to Chris@42: transform a subset of a larger array. Chris@42: @cindex advanced interface Chris@42: Note also that, in the advanced interface, the expression above is Chris@42: multiplied by a @dfn{stride} to get the actual array index---this is Chris@42: useful in situations where each element of the multi-dimensional array Chris@42: is actually a data structure (or another array), and you just want to Chris@42: transform a single field. In the basic interface, however, the stride Chris@42: is 1. Chris@42: @cindex stride Chris@42: Chris@42: @c =========> Chris@42: @node Column-major Format, Fixed-size Arrays in C, Row-major Format, Multi-dimensional Array Format Chris@42: @subsection Column-major Format Chris@42: @cindex column-major Chris@42: Chris@42: Readers from the Fortran world are used to arrays stored in Chris@42: @dfn{column-major} order (sometimes called ``Fortran order''). This is Chris@42: essentially the exact opposite of row-major order in that, here, the Chris@42: @emph{first} dimension's index varies most quickly. Chris@42: Chris@42: If you have an array stored in column-major order and wish to Chris@42: transform it using FFTW, it is quite easy to do. When creating the Chris@42: plan, simply pass the dimensions of the array to the planner in Chris@42: @emph{reverse order}. For example, if your array is a rank three Chris@42: @code{N x M x L} matrix in column-major order, you should pass the Chris@42: dimensions of the array as if it were an @code{L x M x N} matrix Chris@42: (which it is, from the perspective of FFTW). This is done for you Chris@42: @emph{automatically} by the FFTW legacy-Fortran interface Chris@42: (@pxref{Calling FFTW from Legacy Fortran}), but you must do it Chris@42: manually with the modern Fortran interface (@pxref{Reversing array Chris@42: dimensions}). Chris@42: @cindex Fortran interface Chris@42: Chris@42: @c =========> Chris@42: @node Fixed-size Arrays in C, Dynamic Arrays in C, Column-major Format, Multi-dimensional Array Format Chris@42: @subsection Fixed-size Arrays in C Chris@42: @cindex C multi-dimensional arrays Chris@42: Chris@42: A multi-dimensional array whose size is declared at compile time in C Chris@42: is @emph{already} in row-major order. You don't have to do anything Chris@42: special to transform it. For example: Chris@42: Chris@42: @example Chris@42: @{ Chris@42: fftw_complex data[N0][N1][N2]; Chris@42: fftw_plan plan; Chris@42: ... Chris@42: plan = fftw_plan_dft_3d(N0, N1, N2, &data[0][0][0], &data[0][0][0], Chris@42: FFTW_FORWARD, FFTW_ESTIMATE); Chris@42: ... Chris@42: @} Chris@42: @end example Chris@42: Chris@42: This will plan a 3d in-place transform of size @code{N0 x N1 x N2}. Chris@42: Notice how we took the address of the zero-th element to pass to the Chris@42: planner (we could also have used a typecast). Chris@42: Chris@42: However, we tend to @emph{discourage} users from declaring their Chris@42: arrays in this way, for two reasons. First, this allocates the array Chris@42: on the stack (``automatic'' storage), which has a very limited size on Chris@42: most operating systems (declaring an array with more than a few Chris@42: thousand elements will often cause a crash). (You can get around this Chris@42: limitation on many systems by declaring the array as Chris@42: @code{static} and/or global, but that has its own drawbacks.) Chris@42: Second, it may not optimally align the array for use with a SIMD Chris@42: FFTW (@pxref{SIMD alignment and fftw_malloc}). Instead, we recommend Chris@42: using @code{fftw_malloc}, as described below. Chris@42: Chris@42: @c =========> Chris@42: @node Dynamic Arrays in C, Dynamic Arrays in C-The Wrong Way, Fixed-size Arrays in C, Multi-dimensional Array Format Chris@42: @subsection Dynamic Arrays in C Chris@42: Chris@42: We recommend allocating most arrays dynamically, with Chris@42: @code{fftw_malloc}. This isn't too hard to do, although it is not as Chris@42: straightforward for multi-dimensional arrays as it is for Chris@42: one-dimensional arrays. Chris@42: Chris@42: Creating the array is simple: using a dynamic-allocation routine like Chris@42: @code{fftw_malloc}, allocate an array big enough to store N Chris@42: @code{fftw_complex} values (for a complex DFT), where N is the product Chris@42: of the sizes of the array dimensions (i.e. the total number of complex Chris@42: values in the array). For example, here is code to allocate a Chris@42: @threedims{5,12,27} rank-3 array: Chris@42: @findex fftw_malloc Chris@42: Chris@42: @example Chris@42: fftw_complex *an_array; Chris@42: an_array = (fftw_complex*) fftw_malloc(5*12*27 * sizeof(fftw_complex)); Chris@42: @end example Chris@42: Chris@42: Accessing the array elements, however, is more tricky---you can't Chris@42: simply use multiple applications of the @samp{[]} operator like you Chris@42: could for fixed-size arrays. Instead, you have to explicitly compute Chris@42: the offset into the array using the formula given earlier for Chris@42: row-major arrays. For example, to reference the @math{(i,j,k)}-th Chris@42: element of the array allocated above, you would use the expression Chris@42: @code{an_array[k + 27 * (j + 12 * i)]}. Chris@42: Chris@42: This pain can be alleviated somewhat by defining appropriate macros, Chris@42: or, in C++, creating a class and overloading the @samp{()} operator. Chris@42: The recent C99 standard provides a way to reinterpret the dynamic Chris@42: array as a ``variable-length'' multi-dimensional array amenable to Chris@42: @samp{[]}, but this feature is not yet widely supported by compilers. Chris@42: @cindex C99 Chris@42: @cindex C++ Chris@42: Chris@42: @c =========> Chris@42: @node Dynamic Arrays in C-The Wrong Way, , Dynamic Arrays in C, Multi-dimensional Array Format Chris@42: @subsection Dynamic Arrays in C---The Wrong Way Chris@42: Chris@42: A different method for allocating multi-dimensional arrays in C is Chris@42: often suggested that is incompatible with FFTW: @emph{using it will Chris@42: cause FFTW to die a painful death}. We discuss the technique here, Chris@42: however, because it is so commonly known and used. This method is to Chris@42: create arrays of pointers of arrays of pointers of @dots{}etcetera. Chris@42: For example, the analogue in this method to the example above is: Chris@42: Chris@42: @example Chris@42: int i,j; Chris@42: fftw_complex ***a_bad_array; /* @r{another way to make a 5x12x27 array} */ Chris@42: Chris@42: a_bad_array = (fftw_complex ***) malloc(5 * sizeof(fftw_complex **)); Chris@42: for (i = 0; i < 5; ++i) @{ Chris@42: a_bad_array[i] = Chris@42: (fftw_complex **) malloc(12 * sizeof(fftw_complex *)); Chris@42: for (j = 0; j < 12; ++j) Chris@42: a_bad_array[i][j] = Chris@42: (fftw_complex *) malloc(27 * sizeof(fftw_complex)); Chris@42: @} Chris@42: @end example Chris@42: Chris@42: As you can see, this sort of array is inconvenient to allocate (and Chris@42: deallocate). On the other hand, it has the advantage that the Chris@42: @math{(i,j,k)}-th element can be referenced simply by Chris@42: @code{a_bad_array[i][j][k]}. Chris@42: Chris@42: If you like this technique and want to maximize convenience in accessing Chris@42: the array, but still want to pass the array to FFTW, you can use a Chris@42: hybrid method. Allocate the array as one contiguous block, but also Chris@42: declare an array of arrays of pointers that point to appropriate places Chris@42: in the block. That sort of trick is beyond the scope of this Chris@42: documentation; for more information on multi-dimensional arrays in C, Chris@42: see the @code{comp.lang.c} Chris@42: @uref{http://c-faq.com/aryptr/dynmuldimary.html, FAQ}. Chris@42: Chris@42: @c ------------------------------------------------------------ Chris@42: @node Words of Wisdom-Saving Plans, Caveats in Using Wisdom, Multi-dimensional Array Format, Other Important Topics Chris@42: @section Words of Wisdom---Saving Plans Chris@42: @cindex wisdom Chris@42: @cindex saving plans to disk Chris@42: Chris@42: FFTW implements a method for saving plans to disk and restoring them. Chris@42: In fact, what FFTW does is more general than just saving and loading Chris@42: plans. The mechanism is called @dfn{wisdom}. Here, we describe Chris@42: this feature at a high level. @xref{FFTW Reference}, for a less casual Chris@42: but more complete discussion of how to use wisdom in FFTW. Chris@42: Chris@42: Plans created with the @code{FFTW_MEASURE}, @code{FFTW_PATIENT}, or Chris@42: @code{FFTW_EXHAUSTIVE} options produce near-optimal FFT performance, Chris@42: but may require a long time to compute because FFTW must measure the Chris@42: runtime of many possible plans and select the best one. This setup is Chris@42: designed for the situations where so many transforms of the same size Chris@42: must be computed that the start-up time is irrelevant. For short Chris@42: initialization times, but slower transforms, we have provided Chris@42: @code{FFTW_ESTIMATE}. The @code{wisdom} mechanism is a way to get the Chris@42: best of both worlds: you compute a good plan once, save it to Chris@42: disk, and later reload it as many times as necessary. The wisdom Chris@42: mechanism can actually save and reload many plans at once, not just Chris@42: one. Chris@42: @ctindex FFTW_MEASURE Chris@42: @ctindex FFTW_PATIENT Chris@42: @ctindex FFTW_EXHAUSTIVE Chris@42: @ctindex FFTW_ESTIMATE Chris@42: Chris@42: Chris@42: Whenever you create a plan, the FFTW planner accumulates wisdom, which Chris@42: is information sufficient to reconstruct the plan. After planning, Chris@42: you can save this information to disk by means of the function: Chris@42: @example Chris@42: int fftw_export_wisdom_to_filename(const char *filename); Chris@42: @end example Chris@42: @findex fftw_export_wisdom_to_filename Chris@42: (This function returns non-zero on success.) Chris@42: Chris@42: The next time you run the program, you can restore the wisdom with Chris@42: @code{fftw_import_wisdom_from_filename} (which also returns non-zero on success), Chris@42: and then recreate the plan using the same flags as before. Chris@42: @example Chris@42: int fftw_import_wisdom_from_filename(const char *filename); Chris@42: @end example Chris@42: @findex fftw_import_wisdom_from_filename Chris@42: Chris@42: Wisdom is automatically used for any size to which it is applicable, as Chris@42: long as the planner flags are not more ``patient'' than those with which Chris@42: the wisdom was created. For example, wisdom created with Chris@42: @code{FFTW_MEASURE} can be used if you later plan with Chris@42: @code{FFTW_ESTIMATE} or @code{FFTW_MEASURE}, but not with Chris@42: @code{FFTW_PATIENT}. Chris@42: Chris@42: The @code{wisdom} is cumulative, and is stored in a global, private Chris@42: data structure managed internally by FFTW. The storage space required Chris@42: is minimal, proportional to the logarithm of the sizes the wisdom was Chris@42: generated from. If memory usage is a concern, however, the wisdom can Chris@42: be forgotten and its associated memory freed by calling: Chris@42: @example Chris@42: void fftw_forget_wisdom(void); Chris@42: @end example Chris@42: @findex fftw_forget_wisdom Chris@42: Chris@42: Wisdom can be exported to a file, a string, or any other medium. Chris@42: For details, see @ref{Wisdom}. Chris@42: Chris@42: @node Caveats in Using Wisdom, , Words of Wisdom-Saving Plans, Other Important Topics Chris@42: @section Caveats in Using Wisdom Chris@42: @cindex wisdom, problems with Chris@42: Chris@42: @quotation Chris@42: @html Chris@42: Chris@42: @end html Chris@42: For in much wisdom is much grief, and he that increaseth knowledge Chris@42: increaseth sorrow. Chris@42: @html Chris@42: Chris@42: @end html Chris@42: [Ecclesiastes 1:18] Chris@42: @cindex Ecclesiastes Chris@42: @end quotation Chris@42: @iftex Chris@42: @medskip Chris@42: @end iftex Chris@42: Chris@42: @cindex portability Chris@42: There are pitfalls to using wisdom, in that it can negate FFTW's Chris@42: ability to adapt to changing hardware and other conditions. For Chris@42: example, it would be perfectly possible to export wisdom from a Chris@42: program running on one processor and import it into a program running Chris@42: on another processor. Doing so, however, would mean that the second Chris@42: program would use plans optimized for the first processor, instead of Chris@42: the one it is running on. Chris@42: Chris@42: It should be safe to reuse wisdom as long as the hardware and program Chris@42: binaries remain unchanged. (Actually, the optimal plan may change even Chris@42: between runs of the same binary on identical hardware, due to Chris@42: differences in the virtual memory environment, etcetera. Users Chris@42: seriously interested in performance should worry about this problem, Chris@42: too.) It is likely that, if the same wisdom is used for two Chris@42: different program binaries, even running on the same machine, the Chris@42: plans may be sub-optimal because of differing code alignments. It is Chris@42: therefore wise to recreate wisdom every time an application is Chris@42: recompiled. The more the underlying hardware and software changes Chris@42: between the creation of wisdom and its use, the greater grows Chris@42: the risk of sub-optimal plans. Chris@42: Chris@42: Nevertheless, if the choice is between using @code{FFTW_ESTIMATE} or Chris@42: using possibly-suboptimal wisdom (created on the same machine, but for a Chris@42: different binary), the wisdom is likely to be better. For this reason, Chris@42: we provide a function to import wisdom from a standard system-wide Chris@42: location (@code{/etc/fftw/wisdom} on Unix): Chris@42: @cindex wisdom, system-wide Chris@42: Chris@42: @example Chris@42: int fftw_import_system_wisdom(void); Chris@42: @end example Chris@42: @findex fftw_import_system_wisdom Chris@42: Chris@42: FFTW also provides a standalone program, @code{fftw-wisdom} (described Chris@42: by its own @code{man} page on Unix) with which users can create wisdom, Chris@42: e.g. for a canonical set of sizes to store in the system wisdom file. Chris@42: @xref{Wisdom Utilities}. Chris@42: @cindex fftw-wisdom utility Chris@42: