\
\ Some attempts to test and reverse engineer the Xburst media instructions
\
\ Copyright (C) David Kühling <dvdkhlng TA gmx TOD de> 2010
\
\ License: LGPLv3+; NO WARRANTY
\
\ !This depends on the new 2010-08-07 GForth CVS snapshot!
\ Get Gforth package for Nanonote from:
\
\ (compiled) http://mosquito.dyndns.tv/~spock/openwrt
\ (source) git clone git://projects.qi-hardware.com/openwrt-packages.git
\
\ Other information about the media instruction set:
\
\ http://www.gp32x.com/board/index.php?/topic/50216-jzsimd-mxu-instructions-list/
\ http://nmj.sumomo.ne.jp/suz-aa1/Jz47xx/mxu_as
\ http://pastebin.com/f442c9e6e


\ define new assembler opcodes
also assembler definitions
: S32M2I,  ( xr rt -- )  \ move from media register
   $1C ( special2) 26 lshift   $2E ( function) or
   ( rt)  asm-rt 
   ( xr) asm-shamt  l, ;
: S32I2M,  ( xr rt -- )  \ move to media register
   $1C ( special2) 26 lshift   $2F ( function) or
   ( rt)  asm-rt
   ( xr) asm-shamt  l, ;
: S32LDD, ( xr r offs -- )
   \ 12345678901234567890123456789012
   \ SSSSSSRRRRRrOOOOOOOOOOXXXXEEEEEE
   $1C ( special2) 26 lshift   $10 ( function) or
   ( reverse? 1 20 lshift or)
   ( offs)  swap 2/ 2/ 10 lshift  or   \ only 10 bits max!
   ( rt) asm-rs
   ( xrf)  swap 6 lshift or
   l, ;
: S32STD, ( xr r offs -- )
   \ 12345678901234567890123456789012
   \ SSSSSSRRRRRrOOOOOOOOOOXXXXEEEEEE
   $1C ( special2) 26 lshift   $11 ( function) or
   ( reverse? 1 20 lshift or)
   ( offs)  swap 2/ 2/ 10 lshift  or   \ only 10 bits max!
   ( rt) asm-rs
   ( xrf)  swap 6 lshift or
   l, ;

0 constant ww  1 constant lw  2 constant hw  3 constant xw
0 constant aa  1 constant as  2 constant sa  3 constant ss
: Q16ADD,  ( xr1 xr2 xr3 xr4 AA WW -- ) \ see mcu_as
   $1C ( special2) 26 lshift   %001110 ( function) or
   ( ww) swap 22 lshift or
   ( aa) swap 24 lshift or
   ( regs) 6 4 - dup 4 4 * + do  
      swap I lshift or
   4 -loop l, ;
previous definitions

\ read out dsp registers
abi-code dsp>  ( -- dsp31 ... dsp0 )
   $v0  $4  4 -17 *  addiu,
   17 0 [?do]    [i] $t0 S32M2I,   $t0 [i] 4 * $v0  sw,  [loop]
   $ra jr,
   nop,
end-code

: .dsp   dsp>  17 0 do  cr i 3 .r ." = "  hex. loop ;

\ need to write '1' to reg16 (some status reg?) to enable SIMD instructions
abi-code reg16!
   $t0   0 $4  lw,    \ load into normal register first, since
   16 $t0 S32I2M,     \ only S32I2M can access XR16!  
   $ra jr,
   $v0  $4  4  addiu,
end-code

abi-code xrclear  \ clear all 16 QR registers 
   $t0  $zero  move,  
   16 0 [?do]   [i] $t0 S32I2M,  [loop]
   $ra jr,
   $v0  $4  move,
end-code

abi-code testld ( x1 x2 x3 -- )
   1 $4 0 S32LDD,  \ load from stack into XR1..XR3
   2 $4 4 S32LDD,  \ check result using .dsp
   3 $4 8 S32LDD,
   $ra jr,
   $v0  $4  12  addiu,
end-code

abi-code testq16add ( x1 x2 -- x3 x4 )
   1  $4 4 S32LDD,         \ load 2 operands from from stack
   2  $4 0 S32LDD,
   3 1 2 4 AS WW q16add,   \ xr3 = xr1+xr2, xr4 = xr1-xr2,
   3  $4 4 S32STD,         \ load 2 operands from from stack
   4  $4 0 S32STD,   
   $ra jr,
   $v0  $4 move,
end-code

CR .( enabling simd instructions)  1 reg16!

CR .( testing whether it works... )
: test-xburst-media-instructions
   $00040003 $00010002 testq16add
   assert0( swap $00050005 = )
   assert0(      $00030001 = ) ;

test-xburst-media-instructions
.( OK! )