Selecting one of the following will take you directly to that section:
Portability flag.
Use C99 language features.
Chooses generally optimal flags for the target platform. As of the PGI 7.0 release, the flags "-fast" and "-fastsse" are equivlent for 64-bit compilations. For 32-bit compilations "-fast" does not include "-Mscalarsse", "-Mcache_align", or "-Mvect=sse".
Disable C++ exception handling support.
Disable C++ run time type information support.
Generate zero-overhead C++ exception handlers.
Specify the type of the target processor as 64-bit mode.
Specify the type of the target processor as 32-bit mode.
Disable loop autoparallelization within a acc parallel region.
Enable OpenACC directives.
Inline functions declared with the inline keyword.
Disable inlining of functions declared with the inline keyword. <Default>
Align "unconstrained" data objects of size greater than or equal to 16 bytes on cache-line boundaries. An "unconstrained" object is a variable or array that is not a member of an aggregate structure or common block, is not allocatable, and is not an automatic array. On by default on 64-bit Linux systems.
Align doubles on double alignment boundaries
Do not align doubles on double alignment boundaries. <Default>
Set SSE to flush-to-zero mode; if a floating-point underflow occurs, the value is set to zero.
Treat denormalized numbers as zero. Included with "-fast" on Intel based systems. For AMD based systems, "-Mdaz" is not included by default with "-fast".
Generate code to set up a stack frame. <Default>
Eliminates operations that set up a true stack frame pointer for every function. With this option enabled, you cannot perform a traceback on the generated code and you cannot access local variables.
Instructs the compiler to use relaxed precision in the calculation of floating-point reciprocal square root (1/sqrt). Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to use relaxed precision in the calculation of floating-point square root. Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to use relaxed precision in the calculation of floating-point division. Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to allow floating-point expression reordering, including factoring. Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to use relaxed precision in the calculation of some intrinsic functions. Can result in improved performance at the expense of numerical accuracy. The default on an AMD system is "-Mfprelaxed=sqrt,rsqrt,order". The default on an Intel system is "-Mfprelaxed=rsqrt,sqrt,div,order"
Instructs the compiler to use low-precision approximation in the calculation of reciprocal square root (1/sqrt). Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to use low-precision approximation in the calculation of square root. Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to use low-precision approximation in the calculation of divides. Can result in improved performance at the expense of numerical accuracy.
Instructs the compiler to perform low-precision approximation in the calculation of floating-point division, square-root, and reciprocal square root. Can result in improved performance at the expense of numerical accuracy.
Enable aggressive partial redundancy elimination.
Enable partial redundancy elimination.
Set the fetch-ahead distance for prefetch instructions to N cache lines
Set maximum number of prefetch instructions to generate for a given loop to N.
Use the prefetchnta instruction.
Use the prefetch instruction.
Use the prefetcht0 instruction.
Use the AMD-specific prefetchw instruction.
Enable generation of prefetch instructions on processors where they are supported.
Disable generation of prefetch instructions. <Default>
Use SSE/SSE2 instructions to perform scalar floating-point arithmetic on targets where these instructions are supported. This option is default on SSE2 enabled target architectures.
Do not use SSE/SSE2 instructions to perform scalar floating-point arithmetic; use x87 operations instead.
Instructs the compiler to extend the sign bit that is set as a result of an object's conversion from one data type to an object of a larger signed data type.
Aligns or does not align innermost loops on 32 byte boundaries with -tp barcelona. Small loops on barcelona systems may run fast if aligned on 32-byte boundaries; however, in practice, most assemblers do not yet implement efficient padding causing some programs to run more slowly with this as default. Use -Mloop32 on systems with an assembler tuned for barcleona. The default is -Mnoloop32.
Treat individual array element references as candidates for possible loop-carried redundancy elimination. The default is to eliminate only redundant expressions involving two or more operands.
Allow expression re-association; specifying this sub-option can increase opportunities for loop-carried redundancy elimination.
Disable expression re-association.
Enables loop-carried redundancy elimination, an optimization that can reduce the number of arithmetic operations and memory references in loops.
Disable loop-carried redundancy elimination. <Default>
Instructs the compiler not to perform idiom recognition or introduce calls to hand-optimized vector functions.
Generate profile-feedback instrumentation (PFI); this includes extra code to collect run-time statistics and dump them to a trace file for use in a subsequent compilation. PFI gathers information about a program's execution and data values but does not gather information from hardware performance counters. PFI does gather data for optimizations which are unique to profile-feedback optimization.
The indirect sub-option enables collection of indirect function call targets, which can be used for indirect function call inlining.
Enable profile-feedback optimizations including indirect function call inlining. This option requires a pgfi.out file generated from a binary built with -Mpfi=indirect.
Generate profile-feedback instrumentation (PFI); this includes extra code to collect run-time statistics and dump them to a trace file for use in a subsequent compilation. PFI gathers information about a program's execution and data values but does not gather information from hardware performance counters. PFI does gather data for optimizations which are unique to profile-feedback optimization.
Enable profile-feedback optimizations.
Print compiler feedback messages for accelerator directives.
Print compiler feedback messages.
Interprocedural Analysis option: Recognize when targets of pointer dummy are aligned.
Interprocedural Analysis option: Disable recognizition when targets of pointer dummy are aligned.
Interprocedural Analysis option: Remove arguments replaced by -Mipa=ptr,const
Interprocedural Analysis option: Do not remove arguments replaced by -Mipa=ptr,const
Interprocedural Analysis option: Generate call graph information for pgicg tool.
Interprocedural Analysis option: Do not generate call graph information for pgicg tool.
Interprocedural Analysis option: Enable interprocedural constant propagation.
Interprocedural Analysis option: Disable interprocedural constant propagation.
Interprocedural Analysis option: Used with -Mipa=inline to specify functions which should not be inlined.
Instructs the compiler to perform interprocedural analysis. Equivalant to -Mipa=align,arg,const,f90ptr,shape,globals,libc,localarg,ptr,pure.
Interprocedural Analysis option: Force all objects to recompile regardless whether IPA information has changed.
Interprocedural Analysis option: Optimize references to global values.
Interprocedural Analysis option: Do not optimize references to global values.
Interprocedural Analysis option: Automatically determine which functions to inline, limit to N levels where N is a supplied constant value. If no value is suppiled, then the default value of 2 is used. IPA-based function inlining is performed from leaf routines upward.
Interprocedural Analysis option: Automatically determine which functions to inline, limit to 2 levels (default). IPA-based function inlining is performed from leaf routines upward.
Interprocedural Analysis option: Automatically determine which functions to inline, independent of information gathered from profile guided feedback (-Mpfi), limit to N levels where N is a supplied constant value. If no value is suppiled, then the default value of 2 is used. IPA-based function inlining is performed from leaf routines upward.
Interprocedural Analysis option: Automatically determine which functions to inline, independent of information gathered from profile guided feedback (-Mpfi), limit to 2 levels (default). IPA-based function inlining is performed from leaf routines upward.
Interprocedural Analysis option: Inline static functions which are outside the scope of the current file.
Interprocedural Analysis option: Allow inlining of routines from libraries.
Interprocedural Analysis option: Do not inline routines from libraries.
Interprocedural Analysis option: Used to optimize calls to certain functions from the system's standard C library, libc.
Interprocedural Analysis option: Allow recompiling and optimization of routines from libraries using IPA information.
Interprocedural Analysis option: Don't optimize routines in libraries.
Interprocedural Analysis option: -Mipa=arg plus externalizes local pointer targets.
Interprocedural Analysis option: -Mipa=arg plus externalizes local pointer targets.
Interprocedural Analysis option: Do not externalize local pointer targets.
Interprocedural Analysis option: Enable pointer disambiguation across procedure calls.
Interprocedural Analysis option: Disable pointer disambiguation.
Interprocedural Analysis option: Fortran 90/95 Pointer disambiguation across calls.
Interprocedural Analysis option: Disable Fortran 90/95 pointer disambiguation
Interprocedural Analysis option: Pure function detection.
Interprocedural Analysis option: Disable pure function detection.
Interprocedural Analysis option: Allows inlining in Fortran even when array shapes do not match.
Interprocedural Analysis option: Perform Fortran 90 array shape propagation.
Interprocedural Analysis option: Disable Fortran 90 array shape propagation.
Interprocedural Analysis option: Remove functions that are never called.
Interprocedural Analysis option: Do not remove functions that are never called.
Enable Interprocedural Analysis.
Instructs the PGI runtime libraries to create a number of threads equal to the number of cores on the system during parallel execution. Note that the settings of the environment variables OMP_NUM_THREADS or NCPUS will take precedence.
Instructs the PGI runtime libraries to automatically bind threads to the executing system's cores.
Instructs the parallelizer to generate alternate serial code for parallelized loops. Without arguments, the parallelizer determines an appropriate cutoff length and generates serial code to be executed whenever the loop count is less than or equal to that length.
Instructs the parallelizer to generate alternate serial code for parallelized loops. With arguments, the serial altcode is executed whenever the loop count is less than or equal to N.
Always execute the parallelized version of a loop regardless of the loop count.
Disables parallelization of loops with reductions.
Assume loops containing calls are safe to parallelize and allows loops containing calls to be candidates for parallelization. Also, no minimum loop count threshold must be satisfied before parallelization will occur, and last values of scalars are assumed to be safe.
Do not assume loops containing calls are safe to parallelize.
Parallelize with block distribution. Contiguous blocks of iterations of a parallelizable loop are assigned to the available processors.
Parallelize with cyclic distribution. The outermost parallelizable loop in any loop nest is parallelized. If a parallelized loop is innermost, its iterations are allocated to processors cyclically. For example, if there are 3 processors executing a loop, processor 0 performs iterations 0, 3, 6, etc.; processor 1 performs iterations 1, 4, 7, etc.; and processor 2 performs iterations 2, 5, 8, etc.
Enable parallelization of innermost loops.
Disable parallelization of innermost loops.
Parallelize loops nested at most $1 levels deep
Instructs the compiler to enable auto-concurrentization of loops. If -Mconcur is specified, multiple processors will be used to execute loops that the compiler determines to be parallelizable.
Instructs the inliner to inline the functions within the library filename.ext.
Instructs the inliner to inline all eligible functions except foo, a function in the source text. Multiple functions can be listed, comma-separated.
Instructs the inliner to inline function func.
Allows inlining in Fortran even when array shapes do not match.
Instructs the inliner to inline functions with N or fewer statements where N is a supplied constant value.
Instructs the inliner to perform N levels of inlining where N is a supplied constant value. If no value is suppiled, then the default value of 2 is used.
Instructs the inliner to perform 1 level of inlining.
Disable constant propagation from assertions derived from equality conditionals.
Link with the huge page runtime library and allocate a maximum of N huge pages where N is a supplied constant value. If no constant value is supplied, then the maximum number of huge pages the application can use is limited by the number of huge pages the operating system has available or the value of the environment variable PGI_HUGE_PAGES. Note that setting PGI_HUGE_PAGES will override the value of N.
Link with the huge page runtime library. The maximum number of huge pages the application can use is limited by the number of huge pages the operating system has available or the value of the environment variable PGI_HUGE_PAGES.
Link with the huge page runtime library. Use huge pages for an executable's .BSS section.
Adds a call to the routine "mallopt" in the main routine. This option can have a dramatic impact on the performance of programs that dynamically allocate memory, especially for those which have a few large mallocs. To be effective, this switch must be specified when compiling the file containing the Fortran, C, or C++ main routine.
Link with PGI's Alloc library which replaces the system's Malloc, Calloc, Realloc, and Free functions with PGI versions. Programs using -Msmartalloc must be compiled and linked with "-Bdynamic".
Disable support for large (> 2GB) addresses on 64-bit Windows.
Assume all pointers and arrays are independent and safe for aggressive optimizations, and in particular that no pointers or arrays overlap of conflict with each other.
Instructs the compiler that arrays and pointers are treated with the same copyin and copyout semantics as Fortran dummy arguments.
Instructs the compiler that local pointers and arrays do not overlap or conflict with each other and are independent.
Instructs the compiler that local pointers and arrays do not overlap or conflict with each other and are independent.
Instructs the compiler that static pointers and arrays do not overlap or conflict with each other and are independent.
Instructs the compiler that global or external pointers and arrays do not overlap or conflict with each other and are independent.
Instructs the C/C++ compiler to override data dependencies between pointers of a given storage class.
Instructs the compiler to completely unroll loops with a constant loop count of less than or equal to N where N is a supplied constant value. If no constant value is given, then a default of 4 is used. A value of 1 inhibits the complete unrolling of loops with constant loop counts.
"-Munroll=n:n" instructs the compiler to unroll loops N times where N is a supplied constant value. If no constant value is given, then a default of 4 is used.
"-Munroll=m:n" instructs the compiler to unroll loops with multiple blocks N times where N is a supplied constant value. If no constant value is given, then a default of 4 is used.
Instructs the compiler to unroll loops with multiple blocks using the default value of 4 times
Invokes the loop unroller.
Disable loop unrolling. <Default>
Don't check dependence relations for vector or parallel code.
Allow parallelization of loops with conditional scalar assignments.
Generate code to check for zero loop increments.
Enable an optional post-pass instruction scheduling.
Disable an optional post-pass instruction scheduling. <Default>
Disable automatic vector pipelining. <Default>
Instructs the vectorizer to generate alternate code for vectorized loops when appropriate. For each vectorized loop the compiler decides whether to generate altcode and what type or types to generate, which may be any or all of:
The compiler also determines suitable loop count and array alignment conditions for executing the altcode.
Disables alternate code generation for vectorized loops.
Instructs the vectorizer to enable certain associativity conversions that can change the results of a computations due to roundoff error. A typical optimization is to change an arithmetic operation to an arithmetic opteration that is mathmatically correct, but can be computationally different, due to round-off error.
Instructs the vectorizer to disable associativity conversions.
Instructs the vectorizer, when performing cache tiling optimizations, to assume a cache size of N. The default size is processor dependent.
Instructs the vectorizer to enable loop fusion.
Instructs the vectorizer to disable vectorization of indirect array references.
Instructs the vectorizer to enable idiom recognition.
Instructs the vectorizer to disable idiom recognition.
Generate vector loops for all loops where possible regardless of the number of statements in the loop. This overrides a heuristic in the vectorizer that ordinarily prevents vectorization of loops with a number of statements that exceed a certain threshold.
Instructs the vectorizer to generate partial vectorization.
Instructs the vectorizer to generate prefetch instructions.
Enables generation of packed SSE instructions for short vector operations that arise from scalar code outside of loops or within the body of a loop iteration.
Instructs the vectorizer to search for vectorizable loops and, where possible, make use of SSE, SSE2, and prefetch instructions.
Instructs the driver to disable the -Mvect=sse option which is part of the "-fast" option.
Enable automatic vector pipelining.
Disables -Ktrap=fp.
-Ktrap is only processed by the compilers when compiling main functions' programs. The options inv, denorm, divz, ovf, unf, and inexact correspond to the processor's exception mask bits invalid operation, denormalized operand, divide-by-zero, overflow, underflow, and precision, respectively. Normally, the processor's exception mask bits are on (floating-point exceptions are masked the processor recovers from the exceptions and continues). If a floating-point exception occurs and its corresponding mask bit is off (or unmasked ), execution terminates with an arithmetic exception (C's SIGFPE signal). -Ktrap=fp is equivalent to -Ktrap=inv,divz,ovf.
Enable long branches.
Link with the AMD Core Math Library. Available from www.amd.com
Use the -mp option to instruct the compiler to interpret user-inserted OpenMP shared-memory parallel programming directives and generate an executable file which will utilize multiple processors in a shared-memory parallel system. When used strictly as a linker flag, the PGI OpenMP runtime will be linked and users can use the environment variables MP_BIND and MP_BLIST to bind a serial program to a CPU.
The align sub-option to -mp forces loop iterations to be allocated to OpenMP processes using an algorithm that maximizes alignment of vector sub-sections in loops that are both parallelized and vectorized for SSE. This can improve performance in program units that include many such loops. It can result in load-balancing problems that significantly decrease performance in program units with relatively short loops that contain a large amount of work in each iteration.
The nonuma suboption to -mp tells the driver to not link with libnuma.
(For use only on 64-bit Linux targets) Generate code for the medium memory model in the linux86-64 execution environment. The default small memory model of the linux86-64 environment limits the combined area for a user's object or executable to 1GB, with the Linux kernel managing usage of the second 1GB of address for system routines, shared libraries, stacks, etc. Programs are started at a fixed address, and the program can use a single instruction to make most memory references. The medium memory model allows for larger than 2GB data areas, or .bss sections. Program units compiled using either -mcmodel=medium or -fpic require additional instructions to reference memory. The effect on performance is a function of the data-use of the application. The -mcmodel=medium switch must be used at both compile time and link time to create 64-bit executables. Program units compiled for the default small memory model can be linked into medium memory model executables as long as they are compiled -fpic, or position-independent.
Enable support for 64-bit indexing and single static data objects larger than 2GB in size. This option is default in the presence of -mcmodel=medium. Can be used separately together with the default small memory model for certain 64-bit applications that manage their own memory space.
Enable dead store elimination.
Enable optimizations using ANSI C type-based pointer disambiguation.
A basic block is generated for each C statement. No scheduling is done between statements. No global optimizations are performed.
Level-one optimization specifies local optimization (-O1). The compiler performs scheduling of basic blocks as well as register allocation. This optimization level is a good choice when the code is very irregular; that is it contains many short statements containing IF statements and the program does not contain loops (DO or DO WHILE statements). For certain types of code, this optimization level may perform better than level-two (-O2) although this case rarely occurs.
The PGI compilers perform many different types of local optimizations, including but not limited to:
Level-two optimization (-O2 or -O) specifies global optimization. The -fast option generally will specify global optimization; however, the -fast switch will vary from release to release depending on a reasonable selection of switches for any one particular release. The -O or -O2 level performs all level-one local optimizations as well as global optimizations. Control flow analysis is applied and global registers are allocated for all functions and subroutines. Loop regions are given special consideration. This optimization level is a good choice when the program contains loops, the loops are short, and the structure of the code is regular.
The PGI compilers perform many different types of global optimizations, including but not limited to:
All level 1 and 2 optimizations are performed. In addition, this level enables more aggressive code hoisting and scalar replacement optimizations that may or may not be profitable.
Target NVIDIA GPUs.
Target NVIDIA GPUs.
Target NVIDIA GPUs.
Optimize for Nvidia..
Use NVIDIA CUDA 4.2 to compile device code.
Use NVIDIA CUDA 5.0 to compile device code.
Use NVIDIA CUDA 5.5 to compile device code.
Use NVIDIA CUDA 7.0 to compile device code.
Target ATI GPUs with the Tahiti architecture.
Target NVIDA GPUs with compute capability 2.0
Enable beta code generation features
Target NVIDA GPUs with compute capability 3.0
Target NVIDA GPUs with compute capability 3.5
Use the fast math library on the device.
Use LLVM to compile device code.
Set maximum number of registers to use on the GPU.
Use pinned host data for memory transfers.
Create a Unified Binary using multiple targets.
Specify the type of the target processor as AMD64 Processor 32-bit mode.
Specify the type of the target processor as AMD64 Processor 64-bit mode.
Specify the type of the target processor as AMD64 Barcelona Processor 64-bit mode.
Specify the type of the target processor as AMD64 Barcelona Processor 32-bit mode.
Specify the type of the target processor as AMD64 Shangahi Processor 64-bit mode.
Specify the type of the target processor as AMD64 Shanghai Processor 32-bit mode.
Specify the type of the target processor as AMD64 Barcelona Processor 32-bit mode.
Specify the type of the target processor as Intel Penryn Processor 64-bit mode.
Specify the type of the target processor as Intel Penryn Processor 32-bit mode.
Specify the type of the target processor as Intel Nehalem Processor 64-bit mode.
Specify the type of the target processor as Intel Nehalem Processor 32-bit mode.
Specify the type of the target processor as Intel haswell Processor 64-bit mode.
Specify the type of the target processor as Intel haswell Processor 32-bit mode.
Specify the type of the target processor as AMD Istanbul Processor 64-bit mode.
Specify the type of the target processor as AMD Istanbul Processor 32-bit mode.
Specify the type of the target processor as Intel Penryn Processor 32-bit mode.
Specify the type of the target processor as Intel P7 Architecture with EM64t, 64-bit mode.
Specify the type of the target processor as Intel P7 Architecture (Pentium 4, Xeon, Centrino).
Specify the type of the target processor as Intel Core 2 EM64T or compatible architecture using 64-bit mode.
Specify the type of the target processor as Intel Core 2 or compatible architecture using 32-bit mode.
Use the unified AMD/Intel 64-bit mode.
Experimental flags.
Link with static libraries.
Staticily link with the PGI runtime libraries. System libraries may still be dynamically linked.
Link with dynamic libraries.
Link with Dynamic Link Libraries (DLL). Note that -Bdynamic must also be used during compilation as well as linking. Implies -D_DLL.
Disable runtime stack checking and set the stack's reservere size to N bytes and commit size to bytes at link time.
Pass the flag "-force:multiple" to the linker to create an output file whether or not the linker finds more than one definition for a symbol. This flag is required when linking with either PGI's Alloc library (-Msmartalloc) or Microquill's Smartheap library. Both libraries replace the system's Malloc, Calloc, Realloc, and Free functions.
Print the compiler version.
Use explicitly compiler Version $1.$2.
Link using FFTW 3.3.3 library for Linux. Description from FFTW:
FFTW is a C subroutine library for computing the discrete Fourier transform (DFT) in one or more dimensions, of arbitrary input size, and of both real and complex data (as well as of even/odd data, i.e. the discrete cosine/sine transforms or DCT/DST).
Link using MicroQuill's SmartHeap 8 (32-bit) library for Linux. Description from Microquill:
SmartHeap is a fast (3X-100X faster than compiler-supplied libraries), portable (Windows, Linux, Solaris, HP-UX, IBM-AIX, Dec OSF Tru64, SGI Irix), reliable, ANSI-compliant malloc/operator new library. SmartHeap supports multiple memory pools, includes a fixed-size allocator, and is thread-safe. SmartHeap also includes comprehensive memory debugging APIs to detect leakage, overwrites, double-frees, wild pointers, out of memory, references to previously freed memory, and other memory errors.
Link using MicroQuill's SmartHeap 8.1 (32-bit) library for Windows. Description from Microquill:
SmartHeap is a fast (3X-100X faster than compiler-supplied libraries), portable (Windows, Linux, Solaris, HP-UX, IBM-AIX, Dec OSF Tru64, SGI Irix), reliable, ANSI-compliant malloc/operator new library. SmartHeap supports multiple memory pools, includes a fixed-size allocator, and is thread-safe. SmartHeap also includes comprehensive memory debugging APIs to detect leakage, overwrites, double-frees, wild pointers, out of memory, references to previously freed memory, and other memory errors.
Link using MicroQuill's SmartHeap 8.1 (64-bit) library for Windows. Description from Microquill:
SmartHeap is a fast (3X-100X faster than compiler-supplied libraries), portable (Windows, Linux, Solaris, HP-UX, IBM-AIX, Dec OSF Tru64, SGI Irix), reliable, ANSI-compliant malloc/operator new library. SmartHeap supports multiple memory pools, includes a fixed-size allocator, and is thread-safe. SmartHeap also includes comprehensive memory debugging APIs to detect leakage, overwrites, double-frees, wild pointers, out of memory, references to previously freed memory, and other memory errors.
Add GNUC specific options from a header file.
Use the Microsoft C++ configuration for Boost. Used with the "boost_msvc_config" src.alt.
400.perlbench Portability flag to get correct definition of "tm" struct.
The PGI C++ GNU compatible compiler for Linux.
The PGI C compiler for Windows.
The PGI C++ compiler for Windows.
The PGI Fortran 95 compiler for Windows.
The PGI Fortran compiler for Windows.
The PGI C compiler for Linux.
The PGI C++ compiler for Linux.
The PGI Fortran 95 compiler for Linux.
The PGI Fortran compiler for Linux.
Disable warning messages.
Interprocedural Analysis option: Specifies the number of concurent IPA second pass compliation proccess that may be performed. This option speeds-up the compilation time on multi-core systems but does not perform any optimizations.
Specifies a directory to search for include files. Use -I to add directories to the search path for include files.
Specifies a directory to search for include files. Use -I to add directories to the search path for include files.
Specifies a directory to search for libraries. Use -L to add directories to the search path for library files. Multiple -L options are valid. However, the position of multiple -L options is important relative to -l options supplied.
Link with the OpenCL library.