SSE-HOWTO

================================================================================

The Pentium III (and later) PC processors have an additional set of
registers and an associated arithmetic unit that can be addressed
through the Streaming SIMD Extension (SSE) instructions. This
instruction set includes cache manipulation functions and instructions
for parallel single precision arithmetic.

An extended instruction set (SSE2) that enables the use of the SSE
registers for double precision arithmetic is supported by the Pentium 4
processor. SSE + SSE2 instructions will also be understood by the
forthcoming generations of AMD processors.

For full details on the architecture and capabilities of the Pentium
processors, the best address is

  http://developer.intel.com/software/idap/processor/

Just click on the processor family you are interested in, and you will be
guided to a host of useful information. In particular, the manuals
that can be downloaded in pdf format are all what is needed to be able
to exploit the advanced features of these processors.

On a Linux PC the SSE unit can be fully used provided

 -  Linux kernel 2.4.0 or later, properly configured for PIII+ machines, 

 -  gcc version 2.95.2 or later, and

 -  binutils snapshot 010122 or later

are installed. The memory prefetch instructions also work with earlier
versions of the kernel, but the SSE registers cannot be accessed
without kernel patch (not recommended). Current stable versions of
binutils support SSE but not SSE2.

If you would like to write your own programs with SSE instructions, I
suggest to first look at the files linalg.c and sse.h in the directory
64bit/. To learn about gcc inline assembly, it suffices to browse
through the gcc info pages (-> C Extensions -> Extended ASM) and the
macros in the file sse.h. A complete list of the available SSE and
SSE2 instructions, with exhaustive explanations on what precisely they
do, is given in the Intel reference manuals that can be downloaded from
the above site (some 800 pages, not to be printed out!).

An important point to note is that the Intel syntax and the gcc syntax
are exactly reversed. For an instruction that requires a source and
destination register, for example, the notations are

Intel:   instruction destination source

gcc:     instruction source, destination

Some instructions have a third argument, a so-called immediate integer,
and the rules in this case are

Intel:   instruction destination source immediate

gcc:     instruction immediate, source, destination

In the gcc format the integer can be specified as $0x1f, where the $
is a syntactic key, 0x switches to hexadecimal base and 1f is the
(hexadecimal) number that is to be passed to the instruction.