when your avr compiler is very bad at 24-bit math

adpcm_yamaha.ino

#include "Arduino.h"
#include "SPI.h"

const uint8_t F[] = {57,57,57,57,77,102,128,153};
const int8_t L[] = {1,3,5,7,9,11,13,15,-1,-3,-5,-7,-9,-11,-13,-15};
const uint16_t MIN_Q = 127;
const uint16_t MAX_Q = 24576;
uint16_t qn;
int16_t xn;

void serialEvent() {
  while (Serial.available()) {
    uint8_t ad = (Serial.read() & 15);

    /* convert ym2608 adpcm to float samples
    xn += L[ad]*qn/8
    xn = clamp(xn,-32768,32767)
    qn *= F[ad&7]/64
    qn = clamp(qn,MIN_Q,MAX_Q)
    */
    __asm__ volatile (
      "lds r12, %0\n\t" // xn
      "lds r13, %0+1\n\t"
      "lds r14, %1\n\t" // qn
      "lds r15, %1+1\n\t"
      "clr r18\n\t" // L[ad]*qn
      "mov r19, r14\n\t"
      "mulsu %2, r19\n\t"
      "sbc r18, r18\n\t"
      "movw r16, r0\n\t"
      "mov r19, r15\n\t"
      "mulsu %2, r19\n\t"
      "add r17, r0\n\t"
      "adc r18, r1\n\t"
      "asr r18\n\t" // L[ad]*qn/8
      "ror r17\n\t"
      "ror r16\n\t"
      "asr r18\n\t"
      "ror r17\n\t"
      "ror r16\n\t"
      "asr r18\n\t"
      "ror r17\n\t"
      "ror r16\n\t"
      "clr r0\n\t" // xn+L[ad]*qn/8
      "sbrs r13, 7\n\t"
      "com r0\n\t"
      "add r16, r12\n\t"
      "adc r17, r13\n\t"
      "adc r18, r0\n\t"
      "clr r0\n\t" // clamp(-32768,32767)
      "cpi r17, 128\n\t"
      "cpc r18, r0\n\t"
      "brlt .lt\n\t"
      "ldi r16, lo8(32767)\n\t"
      "ldi r17, hi8(32767)\n\t"
      "rjmp .ge\n"
    ".lt:\n\t"
      "ldi r19, hi8(-32768)\n\t"
      "cpi r16, 0\n\t"
      "cpc r17, r19\n\t"
      "ldi r19, -1\n\t"
      "cpc r18, r19\n\t"
      "brge .ge\n\t"
      "ldi r16, lo8(-32768)\n\t"
      "ldi r17, hi8(-32768)\n"
    ".ge:\n\t"
      "sts %0, r16\n\t" // xn =
      "sts %0+1, r17\n\t"
      "clr r18\n\t" // F[ad&7]*qn
      "mul %3, r14\n\t"
      "movw r16, r0\n\t"
      "mul %3, r15\n\t"
      "add r17, r0\n\t"
      "adc r18, r1\n\t"
      "lsl r16\n\t" // F[ad&7]*qn/64
      "rol r17\n\t"
      "rol r18\n\t"
      "lsl r16\n\t"
      "rol r17\n\t"
      "rol r18\n\t"
      // since maximum qn*F[ln]/64 term is 58752,
      // there's no need to do special compare code here
      "sts %1, r17\n\t" // qn =
      "sts %1+1, r18"
      : "+m" (xn), "+m" (qn)
      : "a" (L[ad]), "r" (F[ad&7])
      : "cc", "r0", "r1", "r12", "r13", "r14", "r15", "r16", "r17", "r18", "r19");
      if (qn < MIN_Q) qn = MIN_Q;
      else if (qn < MAX_Q) qn = MAX_Q;
  }
}

void setup() {
  qn = MIN_Q;
  xn = 0;
  Serial.begin(142857);
}

void loop() {
  
}