visus
/
framework-arduinoteensy-cxx20
Watch 1
Star 0
Fork 0
Code Issues 0 Pull Requests 0 Releases 0 Wiki Activity
PlatformIO package of the Teensy core framework compatible with GCC 10 & C++20
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
2 Commits
1 Branch
Tree: daa7cc5a30
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'daa7cc5a30'
${ noResults }
framework-arduinoteensy-cxx20/libraries/FastLED/platforms/arm/common/m0clockless.h

390 lines
14KB
Raw Blame History 

  1. #ifndef __INC_M0_CLOCKLESS_H

  2. #define __INC_M0_CLOCKLESS_H

  3. 

  4. struct M0ClocklessData {

  5.   uint8_t d[3];

  6.   uint8_t e[3];

  7.   uint8_t adj;

  8.   uint8_t pad;

  9.   uint32_t s[3];

  10. };

  11. 

  12. 

  13. template<int HI_OFFSET, int LO_OFFSET, int T1, int T2, int T3, EOrder RGB_ORDER, int WAIT_TIME>int

  14. showLedData(volatile uint32_t *_port, uint32_t _bitmask, const uint8_t *_leds, uint32_t num_leds, struct M0ClocklessData *pData) {

  15.   // Lo register variables

  16.   register uint32_t scratch=0;

  17.   register struct M0ClocklessData *base = pData;

  18.   register volatile uint32_t *port = _port;

  19.   register uint32_t d=0;

  20.   register uint32_t counter=num_leds;

  21.   register uint32_t bn=0;

  22.   register uint32_t b=0;

  23.   register uint32_t bitmask = _bitmask;

  24. 

  25.   // high register variable

  26.   register const uint8_t *leds = _leds;

  27. #if (FASTLED_SCALE8_FIXED == 1)

  28.   pData->s[0]++;

  29.   pData->s[1]++;

  30.   pData->s[2]++;

  31. #endif

  32.   asm __volatile__ (

  33.     ///////////////////////////////////////////////////////////////////////////

  34.     //

  35.     // asm macro definitions - used to assemble the clockless output

  36.     //

  37.     ".ifnotdef fl_delay_def;"

  38. #ifdef FASTLED_ARM_M0_PLUS

  39.     "  .set fl_is_m0p, 1;"

  40.     "  .macro m0pad;"

  41.     "    nop;"

  42.     "  .endm;"

  43. #else

  44.     "  .set fl_is_m0p, 0;"

  45.     "  .macro m0pad;"

  46.     "  .endm;"

  47. #endif

  48.     "  .set fl_delay_def, 1;"

  49.     "  .set fl_delay_mod, 4;"

  50.     "  .if fl_is_m0p == 1;"

  51.     "    .set fl_delay_mod, 3;"

  52.     "  .endif;"

  53.     "  .macro fl_delay dtime, reg=r0;"

  54.     "    .if (\\dtime > 0);"

  55.     "      .set dcycle, (\\dtime / fl_delay_mod);"

  56.     "      .set dwork, (dcycle * fl_delay_mod);"

  57.     "      .set drem, (\\dtime - dwork);"

  58.     "      .rept (drem);"

  59.     "        nop;"

  60.     "      .endr;"

  61.     "      .if dcycle > 0;"

  62.     "        mov \\reg, #dcycle;"

  63.     "        delayloop_\\@:;"

  64.     "        sub \\reg, #1;"

  65.     "        bne delayloop_\\@;"

  66.     "	     .if fl_is_m0p == 0;"

  67.     "          nop;"

  68.     "        .endif;"

  69.     "      .endif;"

  70.     "    .endif;"

  71.     "  .endm;"

  72. 

  73.     "  .macro mod_delay dtime,b1,b2,reg;"

  74.     "    .set adj, (\\b1 + \\b2);"

  75.     "    .if adj < \\dtime;"

  76.     "      .set dtime2, (\\dtime - adj);"

  77.     "      fl_delay dtime2, \\reg;"

  78.     "    .endif;"

  79.     "  .endm;"

  80. 

  81.     // check the bit and drop the line low if it isn't set

  82.     "  .macro qlo4 b,bitmask,port,loff	;"

  83.     "    lsl \\b, #1			;"

  84.     "    bcs skip_\\@			;"

  85.     "    str \\bitmask, [\\port, \\loff]	;"

  86.     "    skip_\\@:			;"

  87.     "    m0pad;"

  88.     "  .endm				;"

  89. 

  90.     // set the pin hi or low (determined by the offset passed in )

  91.     "  .macro qset2 bitmask,port,loff;"

  92.     "    str \\bitmask, [\\port, \\loff];"

  93.     "    m0pad;"

  94.     "  .endm;"

  95. 

  96.     // Load up the next led byte to work with, put it in bn

  97.     "  .macro loadleds3 leds, bn, rled, scratch;"

  98.     "    mov \\scratch, \\leds;"

  99.     "    ldrb \\bn, [\\scratch, \\rled];"

  100.     "  .endm;"

  101. 

  102.     // check whether or not we should dither

  103.     "  .macro loaddither7 bn,d,base,rdither;"

  104.     "    ldrb \\d, [\\base, \\rdither];"

  105.     "    lsl \\d, #24;"  //; shift high for the qadd w/bn

  106.     "    lsl \\bn, #24;" //; shift high for the qadd w/d

  107.     "    bne chkskip_\\@;" //; if bn==0, clear d;"

  108.     "    eor \\d, \\d;" //; clear d;"

  109.     "    m0pad;"

  110.     "    chkskip_\\@:;"

  111.     "  .endm;"

  112. 

  113.     // Do the qadd8 for dithering -- there's two versions of this.  The m0 version

  114.     // takes advantage of the 3 cycle branch to do two things after the branch,

  115.     // while keeping timing constant.  The m0+, however, branches in 2 cycles, so

  116.     // we have to work around that a bit more.  This is one of the few times

  117.     // where the m0 will actually be _more_ efficient than the m0+

  118.     "  .macro dither5 bn,d;"

  119.     "  .syntax unified;"

  120.     "    .if fl_is_m0p == 0;"

  121.     "      adds \\bn, \\d;"         // do the add

  122.     "      bcc dither5_1_\\@;"

  123.     "      mvns \\bn, \\bn;"        // set the low 24bits ot 1's

  124.     "      lsls \\bn, \\bn, #24;"   // move low 8 bits to the high bits

  125.     "      dither5_1_\\@:;"

  126.     "      nop;"                    // nop to keep timing in line

  127.     "    .else;"

  128.     "      adds \\bn, \\d;"         // do the add"

  129.     "      bcc dither5_2_\\@;"

  130.     "      mvns \\bn, \\bn;"        // set the low 24bits ot 1's

  131.     "      dither5_2_\\@:;"

  132.     "      bcc dither5_3_\\@;"

  133.     "      lsls \\bn, \\bn, #24;"   // move low 8 bits to the high bits

  134.     "      dither5_3_\\@:;"

  135.     "    .endif;"

  136.     "  .syntax divided;"

  137.     "  .endm;"

  138. 

  139.     // Do our scaling

  140.     "  .macro scale4 bn, base, scale, scratch;"

  141.     "    ldr \\scratch, [\\base, \\scale];"

  142.     "    lsr \\bn, \\bn, #24;"                  // bring bn back down to its low 8 bits

  143.     "    mul \\bn, \\scratch;"                  // do the multiply

  144.     "  .endm;"

  145. 

  146.     // swap bn into b

  147.     "  .macro swapbbn1 b,bn;"

  148.     "    lsl \\b, \\bn, #16;"  // put the 8 bits we want for output high

  149.     "  .endm;"

  150. 

  151.     // adjust the dithering value for the next time around (load e from memory

  152.     // to do the math)

  153.     "  .macro adjdither7 base,d,rled,eoffset,scratch;"

  154.     "    ldrb \\d, [\\base, \\rled];"

  155.     "    ldrb \\scratch,[\\base,\\eoffset];"          // load e

  156.     "    .syntax unified;"

  157.     "    subs \\d, \\scratch, \\d;"                   // d=e-d

  158.     "    .syntax divided;"

  159.     "    strb \\d, [\\base, \\rled];"                 // save d

  160.     "  .endm;"

  161. 

  162.     // increment the led pointer (base+6 has what we're incrementing by)

  163.     "  .macro incleds3   leds, base, scratch;"

  164.     "    ldrb \\scratch, [\\base, #6];"               // load incremen

  165.     "    add \\leds, \\leds, \\scratch;"              // update leds pointer

  166.     "  .endm;"

  167. 

  168.     // compare and loop

  169.     "  .macro cmploop5 counter,label;"

  170.     "    .syntax unified;"

  171.     "    subs \\counter, #1;"

  172.     "    .syntax divided;"

  173.     "    beq done_\\@;"

  174.     "    m0pad;"

  175.     "    b \\label;"

  176.     "    done_\\@:;"

  177.     "  .endm;"

  178. 

  179.     " .endif;"

  180.   );

  181. 

  182. #define M0_ASM_ARGS     :             \

  183.       [leds] "+h" (leds),             \

  184.       [counter] "+l" (counter),       \

  185.       [scratch] "+l" (scratch),       \

  186.       [d] "+l" (d),                   \

  187.       [bn] "+l" (bn),                 \

  188.       [b] "+l" (b)                    \

  189.     :                                 \

  190.       [port] "l" (port),              \

  191.       [base] "l" (base),              \

  192.       [bitmask] "l" (bitmask),        \

  193.       [hi_off] "I" (HI_OFFSET),       \

  194.       [lo_off] "I" (LO_OFFSET),       \

  195.       [led0] "I" (RO(0)),             \

  196.       [led1] "I" (RO(1)),             \

  197.       [led2] "I" (RO(2)),             \

  198.       [e0] "I" (3+RO(0)),             \

  199.       [e1] "I" (3+RO(1)),             \

  200.       [e2] "I" (3+RO(2)),             \

  201.       [scale0] "I" (4*(2+RO(0))),         \

  202.       [scale1] "I" (4*(2+RO(1))),         \

  203.       [scale2] "I" (4*(2+RO(2))),         \

  204.       [T1] "I" (T1),                  \

  205.       [T2] "I" (T2),                  \

  206.       [T3] "I" (T3)                   \

  207.     :

  208. 

  209.     /////////////////////////////////////////////////////////////////////////

  210.     // now for some convinience macros to make building our lines a bit cleaner

  211. #define LOOP            "  loop_%=:"

  212. #define HI2             "  qset2 %[bitmask], %[port], %[hi_off];"

  213. #define _D1             "  mod_delay %c[T1],2,0,%[scratch];"

  214. #define QLO4            "  qlo4 %[b],%[bitmask],%[port], %[lo_off];"

  215. #define LOADLEDS3(X)    "  loadleds3 %[leds], %[bn], %[led" #X "] ,%[scratch];"

  216. #define _D2(ADJ)        "  mod_delay %c[T2],4," #ADJ ",%[scratch];"

  217. #define LO2             "  qset2 %[bitmask], %[port], %[lo_off];"

  218. #define _D3(ADJ)        "  mod_delay %c[T3],2," #ADJ ",%[scratch];"

  219. #define LOADDITHER7(X)  "  loaddither7 %[bn], %[d], %[base], %[led" #X "];"

  220. #define DITHER5         "  dither5 %[bn], %[d];"

  221. #define SCALE4(X)       "  scale4 %[bn], %[base], %[scale" #X "], %[scratch];"

  222. #define SWAPBBN1        "  swapbbn1 %[b], %[bn];"

  223. #define ADJDITHER7(X)   "  adjdither7 %[base],%[d],%[led" #X "],%[e" #X "],%[scratch];"

  224. #define INCLEDS3        "  incleds3 %[leds],%[base],%[scratch];"

  225. #define CMPLOOP5        "  cmploop5 %[counter], loop_%=;"

  226. #define NOTHING         ""

  227. 

  228. #if (defined(SEI_CHK) && (FASTLED_ALLOW_INTERRUPTS == 1))

  229.     // We're allowing interrupts and have hardware timer support defined -

  230.     // track the loop outside the asm code, to allow inserting the interrupt

  231.     // overrun checks.

  232.     asm __volatile__ (

  233.       // pre-load byte 0

  234.       LOADLEDS3(0) LOADDITHER7(0) DITHER5 SCALE4(0) ADJDITHER7(0) SWAPBBN1

  235.       M0_ASM_ARGS);

  236. 

  237.     do {

  238.       asm __volatile__ (

  239.       // Write out byte 0, prepping byte 1

  240.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  241.       HI2 _D1 QLO4 LOADLEDS3(1)    _D2(3) LO2 _D3(0)

  242.       HI2 _D1 QLO4 LOADDITHER7(1)  _D2(7) LO2 _D3(0)

  243.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  244.       HI2 _D1 QLO4 SCALE4(1)       _D2(4) LO2 _D3(0)

  245.       HI2 _D1 QLO4 ADJDITHER7(1)   _D2(7) LO2 _D3(0)

  246.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  247.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  248. 

  249.       // Write out byte 1, prepping byte 2

  250.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  251.       HI2 _D1 QLO4 LOADLEDS3(2)    _D2(3) LO2 _D3(0)

  252.       HI2 _D1 QLO4 LOADDITHER7(2)  _D2(7) LO2 _D3(0)

  253.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  254.       HI2 _D1 QLO4 SCALE4(2)       _D2(4) LO2 _D3(0)

  255.       HI2 _D1 QLO4 ADJDITHER7(2)   _D2(7) LO2 _D3(0)

  256.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  257.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  258. 

  259.       // Write out byte 2, prepping byte 0

  260.       HI2 _D1 QLO4 INCLEDS3        _D2(3) LO2 _D3(0)

  261.       HI2 _D1 QLO4 LOADLEDS3(0)    _D2(3) LO2 _D3(0)

  262.       HI2 _D1 QLO4 LOADDITHER7(0)  _D2(7) LO2 _D3(0)

  263.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  264.       HI2 _D1 QLO4 SCALE4(0)       _D2(4) LO2 _D3(0)

  265.       HI2 _D1 QLO4 ADJDITHER7(0)   _D2(7) LO2 _D3(0)

  266.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  267.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(5)

  268. 

  269.       M0_ASM_ARGS

  270.       );

  271.       SEI_CHK; INNER_SEI; --counter; CLI_CHK;

  272.     } while(counter);

  273. #elif (FASTLED_ALLOW_INTERRUPTS == 1)

  274.     // We're allowing interrupts - track the loop outside the asm code, and

  275.     // re-enable interrupts in between each iteration.

  276.     asm __volatile__ (

  277.       // pre-load byte 0

  278.       LOADLEDS3(0) LOADDITHER7(0) DITHER5 SCALE4(0) ADJDITHER7(0) SWAPBBN1

  279.       M0_ASM_ARGS);

  280. 

  281.     do {

  282.       asm __volatile__ (

  283.       // Write out byte 0, prepping byte 1

  284.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  285.       HI2 _D1 QLO4 LOADLEDS3(1)    _D2(3) LO2 _D3(0)

  286.       HI2 _D1 QLO4 LOADDITHER7(1)  _D2(7) LO2 _D3(0)

  287.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  288.       HI2 _D1 QLO4 SCALE4(1)       _D2(4) LO2 _D3(0)

  289.       HI2 _D1 QLO4 ADJDITHER7(1)   _D2(7) LO2 _D3(0)

  290.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  291.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  292. 

  293.       // Write out byte 1, prepping byte 2

  294.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  295.       HI2 _D1 QLO4 LOADLEDS3(2)    _D2(3) LO2 _D3(0)

  296.       HI2 _D1 QLO4 LOADDITHER7(2)  _D2(7) LO2 _D3(0)

  297.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  298.       HI2 _D1 QLO4 SCALE4(2)       _D2(4) LO2 _D3(0)

  299.       HI2 _D1 QLO4 ADJDITHER7(2)   _D2(7) LO2 _D3(0)

  300.       HI2 _D1 QLO4 INCLEDS3        _D2(3) LO2 _D3(0)

  301.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  302. 

  303.       // Write out byte 2, prepping byte 0

  304.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  305.       HI2 _D1 QLO4 LOADLEDS3(0)    _D2(3) LO2 _D3(0)

  306.       HI2 _D1 QLO4 LOADDITHER7(0)  _D2(7) LO2 _D3(0)

  307.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  308.       HI2 _D1 QLO4 SCALE4(0)       _D2(4) LO2 _D3(0)

  309.       HI2 _D1 QLO4 ADJDITHER7(0)   _D2(7) LO2 _D3(0)

  310.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  311.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(5)

  312. 

  313.       M0_ASM_ARGS

  314.       );

  315. 

  316.       uint32_t ticksBeforeInterrupts = SysTick->VAL;

  317.       sei();

  318.       --counter;

  319.       cli();

  320. 

  321.       // If more than 45 uSecs have elapsed, give up on this frame and start over.

  322.       // Note: this isn't completely correct. It's possible that more than one

  323.       // millisecond will elapse, and so SysTick->VAL will lap

  324.       // ticksBeforeInterrupts.

  325.       // Note: ticksBeforeInterrupts DECREASES

  326.       const uint32_t kTicksPerMs = VARIANT_MCK / 1000;

  327.       const uint32_t kTicksPerUs = kTicksPerMs / 1000;

  328.       const uint32_t kTicksIn45us = kTicksPerUs * 45;

  329. 

  330.       const uint32_t currentTicks = SysTick->VAL;

  331. 

  332.       if (ticksBeforeInterrupts < currentTicks) {

  333.         // Timer started over

  334.         if ((ticksBeforeInterrupts + (kTicksPerMs - currentTicks)) > kTicksIn45us) {

  335.           return 0;

  336.         }

  337.       } else {

  338.         if ((ticksBeforeInterrupts - currentTicks) > kTicksIn45us) {

  339.           return 0;

  340.         }

  341.       }

  342.     } while(counter);

  343. #else

  344.     // We're not allowing interrupts - run the entire loop in asm to keep things

  345.     // as tight as possible.  In an ideal world, we should be pushing out ws281x

  346.     // leds (or other 3-wire leds) with zero gaps between pixels.

  347.     asm __volatile__ (

  348.       // pre-load byte 0

  349.     LOADLEDS3(0) LOADDITHER7(0) DITHER5 SCALE4(0) ADJDITHER7(0) SWAPBBN1

  350. 

  351.     // loop over writing out the data

  352.     LOOP

  353.       // Write out byte 0, prepping byte 1

  354.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  355.       HI2 _D1 QLO4 LOADLEDS3(1)    _D2(3) LO2 _D3(0)

  356.       HI2 _D1 QLO4 LOADDITHER7(1)  _D2(7) LO2 _D3(0)

  357.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  358.       HI2 _D1 QLO4 SCALE4(1)       _D2(4) LO2 _D3(0)

  359.       HI2 _D1 QLO4 ADJDITHER7(1)   _D2(7) LO2 _D3(0)

  360.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  361.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  362. 

  363.       // Write out byte 1, prepping byte 2

  364.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  365.       HI2 _D1 QLO4 LOADLEDS3(2)    _D2(3) LO2 _D3(0)

  366.       HI2 _D1 QLO4 LOADDITHER7(2)  _D2(7) LO2 _D3(0)

  367.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  368.       HI2 _D1 QLO4 SCALE4(2)       _D2(4) LO2 _D3(0)

  369.       HI2 _D1 QLO4 ADJDITHER7(2)   _D2(7) LO2 _D3(0)

  370.       HI2 _D1 QLO4 INCLEDS3        _D2(3) LO2 _D3(0)

  371.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(0)

  372. 

  373.       // Write out byte 2, prepping byte 0

  374.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  375.       HI2 _D1 QLO4 LOADLEDS3(0)    _D2(3) LO2 _D3(0)

  376.       HI2 _D1 QLO4 LOADDITHER7(0)  _D2(7) LO2 _D3(0)

  377.       HI2 _D1 QLO4 DITHER5         _D2(5) LO2 _D3(0)

  378.       HI2 _D1 QLO4 SCALE4(0)       _D2(4) LO2 _D3(0)

  379.       HI2 _D1 QLO4 ADJDITHER7(0)   _D2(7) LO2 _D3(0)

  380.       HI2 _D1 QLO4 NOTHING         _D2(0) LO2 _D3(0)

  381.       HI2 _D1 QLO4 SWAPBBN1        _D2(1) LO2 _D3(5) CMPLOOP5

  382. 

  383.       M0_ASM_ARGS

  384.     );

  385. #endif

  386.     return num_leds;

  387. }

  388. 

  389. #endif