2ORC - The Oil Runtime Compiler
5(and OIL stands for Optimized Inner Loops)
7Entropy Wave Inc (http://entropywave.com/) presents Orc, the sucessor
8to Liboil - The Library of Optimized Inner Loops.
10Orc is a library and set of tools for compiling and executing
11very simple programs that operate on arrays of data. The "language"
12is a generic assembly language that represents many of the features
13available in SIMD architectures, including saturated addition and
14subtraction, and many arithmetic operations.
16At this point, developers interested in using Orc should look at the
17examples and try out a few Orc programs in an experimental branch
18of their own projects. And provide feedback on how it works. There
19will likely be some major changes in ease of use from a developer's
20perspective over the next few releases.
22The 0.4 series of Orc releases will be API and ABI compatible, and
23will be incompatible with the 0.5 series when it comes out. The first
24release of the 0.5 series is anticipated to coincide with the release
25of GStreamer 1.0.
30 - Users can create, compile, and run simple programs that use the
31 vector extensions of the CPU, all directly from an application.
33 - Users can compile Orc programs to assembly source code to be
34 compiled and used without linking against the Orc library.
36 - The generic assembly language can be extended by an application
37 by adding new opcodes.
39 - An application can add rules for converting existing or new opcodes
40 to binary code for a specific target.
42 - Current targets: SSE, MMX, MIPS, Altivec, NEON, and TI C64x+.
43 (The c64x target only produces source code.)
45 - Programs can optionally be emulated, which is useful for testing, or
46 if no rules are available to convert Orc opcodes to executable code.
51 Web: http://code.entropywave.com/projects/orc/
52 Download: http://code.entropywave.com/download/orc/
55Questions and Answers:
57 - Q: Why not let gcc vectorize my code?
59 A: Two reasons: first, since Orc's assembly language is much more
60 restrictive than C, Orc can generate better code than gcc, and
61 second, Orc can generate code for functions you define at runtime.
62 Many algorithms require gluing together several stages of operations,
63 and if each stage has several options, the total amount of code to
64 cover all combinations could be inconveniently large.
66 - Q: Why not use compiler intrinsics for SIMD code?
68 A: Compiler intrinsics only work for one target, and need to be
69 hand written. Plus, some compilers are very picky about source
70 code that uses intrinsics, and will silently produce slow code.
71 And, of course, you can't compile intrinsics at runtime.
73 - Q: How big is the Orc library?
75 A: For embedded users, the --enable-backend configure option can
76 be used to disable irrelvant targets. Compiled with only one target
77 (SSE), the library size is about 150 kB uncompressed, or 48 kB
78 compressed. The goal was to keep the uncompressed size under
79 about 100 kB (but that failed!). A typical build with all targets
80 and the full ABI is around 350 kB.
83Caveats (Known Bugs):
85 - ?
90 - Addition of more complex loop control and array structures.
92 - Addition of an option to compile the Orc library with only the
93 runtime features for a single target, e.g., for embedded systems.
95 - Addition of rewrite rules, which convert an instruction that cannot
96 be converted to binary code into a series of instructions that can.
97 This is necessary since assembly instructions on most targets do
98 not cover all the features of the Orc assembly language.
101About Entropy Wave:
103Entropy Wave creates tools that allow content producers and distributors
104use open video technology. Through use of open source software like
105GStreamer and Dirac, Entropy Wave's customers save money on licensing
106costs for encoding and streaming video on the web. Entropy Wave was
107founded in 2008 by long-time open source developer David Schleef.