|
 Math Products
GoFast® for SPARC C Compilers
Fast Floating Point Library
Features
- Conforms to IEEE 754 Format
- Includes single and double precision
- "Link and Go" compiler support
- Supports SPARClite designs
- Includes test programs and make files
Description
GoFast® for SPARC is designed to allow programs developed for the SPARC
architecture to access high performance floating point numerics. GoFast for SPARC
provides "link and go" compatibility with the Microtec C, SUN and GNU C SPARC
compilers. GoFast provides full IEEE and ANSI compatible floating point support.
Floating Point Technology
GoFast for SPARC is based on U S Software's "Architecture Independent
Technology" (AIT) and proven floating point algorithms the company has been
developing for over decade. The algorithms have been thoroughly tested using
automated methods. GoFast was specifically optimized and integrated for high
performance on SPARC processors.
GoFast libraries are designed for "link and go" operation with each supported
compiler.
Processor Support
- V8 - SPARC version 8: floating-point unit, 32-bit divide
- V7 - SPARC version 7: no floating-point unit, 32-bit divide
- splite - Fujitsu MB86930 (SPARClite): no floating point unit, 1-bit
divide, scan instructions
- fuj934 - Fujitsu MB86934: floating point unit, 1-bit divide, scan
instruction.
Each version also emulates floating-point instructions and 32-bit divide instructions. Two register interfaces are also supported: o0 and f0. The difference between these interfaces is in which registers are used for return values
Functionality
GoFast supports C runtime floating point including the following single and double
precision routines:
- + - * /
- conversion operations
- sqrt, sin, cos, tan
- asin, acos, atan, atan2
- sinh, cosh, tanh
- log, log10
- exp, pow
- floor, ceil, fabs
- modf, fmod, frexp, ldexp
Additional support includes test programs and make files.
Timings
The following table gives approximate timings, in microseconds, for some emulated
floating-point instructions. The times were measured on the 49 MHz Fujitsu SPARClite
evaluation board, using static RAM with cache both disabled and enabled.
Instruction
|
SRAM single cache |
SRAM double cache |
| branch, not taken |
1.60 |
0.85 |
2.55 |
0.85 |
| branch, taken |
1.75 |
0.90 |
2.70 |
0.90 |
| load |
2.70 |
1.30 |
2.65 |
1.35 |
| store |
2.55 |
1.30 |
2.95 |
1.55 |
| add |
6.80 |
3.30 |
7.90 |
4.25 |
| subtract |
6.80 |
3.30 |
9.05 |
4.45 |
| multiply |
6.60 |
3.25 |
8.20 |
4.30 |
| divide |
8.35 |
4.75 |
12.10 |
6.10 |
| compare |
4.75 |
2.35 |
4.65 |
2.25 |
| integer to float |
4.40 |
2.10 |
4.50 |
2.25 |
| float to integer |
4.75 |
2.35 |
5.40 |
4.15 |
| double to single |
5.45 |
2.95 |
|
|
| single to double |
|
|
4.15 |
2.30 |
| square root |
12.95 |
7.30 |
22.60 |
13.05 |
Compiler Support
GoFast for SPARC includes libraries for "link and go", seamless operation with theSUN, and GNU C compilers.
The GoFast routines directly replace the compiler's floating point runtime library
routines. In addition, routines normally not supported by the compiler are available with
GoFast. An example might be the single precision version of the functions.
Capabilities and Accuracy
GoFast is reentrant. These is no need to perform any special state save/restore.
GoFast is ROMable. GoFast is IEEE 754 and ANSI C compatible.
The basic functions + - * / and the conversions perform the IEEE 754 "nearest or even"
rounding exactly. The other functions are accurate within one mantissa unit.
GoFast routines make no distinction between quiet and signaling Not-a-Numbers
(NaNs). In the case of an invalid operation, the answer is always a standard quiet NaN.
GoFast routines support IEEE 754 masked exception handling for overflows and
invalid operations. No unmasked exceptions are supported. Underflow and loss of
precision are not reported. Division by zero is treated as an invalid operation.
|
