Skip to content

Commit

Permalink
Improved YM7128 performance
Browse files Browse the repository at this point in the history
  • Loading branch information
@TexZK
TexZK committed Apr 13, 2024
1 parent 7f9df53 commit 6fcd10b
Show file tree
Hide file tree
Showing 7 changed files with 209 additions and 179 deletions.
4 changes: 2 additions & 2 deletions include/aymo_ym7128_arm_neon.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,11 +42,11 @@ struct aymo_(chip) {
uint8_t align_[sizeof(vi16x8_t) - sizeof(struct aymo_ym7128_chip)];

// 128-bit data
int16_t xxv[8];
vi16x8_t xxv;
vi16x8_t kk1;
vi16x8_t kk2;
vi16x8_t kkm;
int16_t tiv[8];
vi16x8_t ti;
vi16x8_t kgl;
vi16x8_t kgr;
vi16x8_t kv;
Expand Down
4 changes: 2 additions & 2 deletions include/aymo_ym7128_x86_avx.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,11 +42,11 @@ struct aymo_(chip) {
uint8_t align_[sizeof(vi16x8_t) - sizeof(struct aymo_ym7128_chip)];

// Vector data
int16_t xxv[8];
vi16x8_t xxv;
vi16x8_t kk1;
vi16x8_t kk2;
vi16x8_t kkm;
int16_t tiv[8];
vi16x8_t ti;
vi16x8_t kgl;
vi16x8_t kgr;
vi16x8_t kv;
Expand Down
4 changes: 2 additions & 2 deletions include/aymo_ym7128_x86_sse41.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,11 +42,11 @@ struct aymo_(chip) {
uint8_t align_[sizeof(vi16x8_t) - sizeof(struct aymo_ym7128_chip)];

// Vector data
int16_t xxv[8];
vi16x8_t xxv;
vi16x8_t kk1;
vi16x8_t kk2;
vi16x8_t kkm;
int16_t tiv[8];
vi16x8_t ti;
vi16x8_t kgl;
vi16x8_t kgr;
vi16x8_t kv;
Expand Down
124 changes: 67 additions & 57 deletions src/aymo_ym7128_arm_neon.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -63,9 +63,7 @@ void aymo_(ctor)(struct aymo_(chip)* chip)
aymo_memset((&chip->parent.vt + 1u), 0, (sizeof(*chip) - sizeof(chip->parent.vt)));

// Initialize input stage coefficients (-1 as a placeholder for computed values)
chip->xxv[2] = 1;
chip->xxv[3] = 1;

chip->xxv = vseta(0, 0, 0, 0, 1, 1, 0, 0);
chip->kk1 = vseta(0, -1, -1, -1, -0x8000, -0x8000, -0x8000, -0x8000);
chip->kk2 = vseta(0, -1, -0x8000, 0, 0, 0, 0x8000, -0x8000);
chip->kkm = vseta(0, 0x7FFF, 0x7FFF, 0, 0, 0, AYMO_YM7128_DELAY_LENGTH, AYMO_YM7128_DELAY_LENGTH);
Expand Down Expand Up @@ -151,17 +149,17 @@ void aymo_(write)(struct aymo_(chip)* chip, uint16_t address, uint8_t value)
else if (address <= (uint16_t)aymo_ym7128_reg_t8) {
value &= 0x1Fu;
int16_t t = aymo_ym7128_tap[value];
int16_t hi = chip->xxv[1]; // hi
int16_t hi = vextract(chip->xxv, 1); // hi
t = (hi - t);
if (t < 0) {
t += AYMO_YM7128_DELAY_LENGTH;
}
if (address == (uint16_t)aymo_ym7128_reg_t0) {
chip->xxv[0] = t; // ti0
chip->xxv = vinsert(chip->xxv, t, 0); // ti0
}
else {
uint16_t i = (address - (uint16_t)aymo_ym7128_reg_t1);
chip->tiv[i] = t;
chip->ti = vinsertn(chip->ti, t, i);
}
}

Expand All @@ -178,46 +176,61 @@ void aymo_(process_i16)(struct aymo_(chip)* chip, uint32_t count, const int16_t
assert(y);
if AYMO_UNLIKELY(!count) return;

int16_t AYMO_ALIGN_V128 vv[8] = {0};

int16_t ti0 = chip->xxv[0];
int16_t t0 = chip->uh[ti0];
chip->xxv[4] = t0;

const int16_t* xe = &x[count];

while AYMO_LIKELY(x != xe) {
chip->xxv[5] = (*x++ & AYMO_YM7128_SIGNAL_MASK);

vsfence();
vi16x8_t xx = vload(chip->xxv);
chip->xxv[6] = t0; // t0d = t0
xx = vmulhrs(xx, chip->kk1);
vi16x8_t xxv = chip->xxv;
vi16x8_t kk1 = chip->kk1;
vi16x8_t kk2 = chip->kk2;
vi16x8_t kkm = chip->kkm;
vi16x8_t ti = chip->ti;
vi16x8_t kgl = chip->kgl;
vi16x8_t kgr = chip->kgr;
vi16x8_t kv = chip->kv;

vi16x8_t zc = chip->zc;
vi16x8_t zb = chip->zb;
vi16x8_t kf = chip->kf;
vi16x8_t ke = chip->ke;
vi16x8_t za = chip->za;
vi16x8_t kd = chip->kd;
vi16x8_t kc = chip->kc;
vi16x8_t kb = chip->kb;
vi16x8_t ka = chip->ka;

do {
int16_t ti0 = vextract(xxv, 0);
int16_t t0 = chip->uh[ti0];
xxv = vinsert(xxv, t0, 4);

xxv = vinsert(xxv, (*x++ & AYMO_YM7128_SIGNAL_MASK), 5);

vi16x8_t xx = xxv;
xxv = vinsert(xxv, t0, 6); // t0d = t0
xx = vmulhrs(xx, kk1);
xx = vaddsi(xx, vrevv(xx));
xx = vmulhrs(xx, chip->kk2);
xx = vand(xx, vcmpgt(chip->kkm, xx));
xx = vmulhrs(xx, kk2);
xx = vand(xx, vcmpgt(kkm, xx));
xx = vaddsi(xx, vrev64q_s16(xx));
vstore(vv, xx);
vi16x8_t ti = vload(chip->tiv);
vi16x8_t vv = xx;
vi16x8_t tj = vsub(ti, vset1(-1));
vi16x8_t tm = vcmpgt(vset1(AYMO_YM7128_DELAY_LENGTH - 1), ti); // tj < DL
vstore(chip->tiv, vand(tj, tm));
vsfence();
ti = vand(tj, tm);

chip->xxv[0] = vv[0]; // ti0'
int16_t hj = vv[1];
chip->xxv[1] = hj; // hi'
int16_t u = vv[5];
xxv = vinsert(xxv, vextract(vv, 0), 0); // ti0'
int16_t hj = vextract(vv, 1);
xxv = vinsert(xxv, hj, 1); // hi'
int16_t u = vextract(vv, 5);
chip->uh[hj] = u;
int16_t AYMO_ALIGN_V128 tuv[8];
for (unsigned i = 0u; i < 8u; ++i) {
tuv[i] = chip->uh[chip->tiv[i]];
}
vsfence();
vi16x8_t tu = vload(tuv);

vi16x8_t gl = vmulhrs(tu, chip->kgl);
vi16x8_t gr = vmulhrs(tu, chip->kgr);
vi16x8_t tu = vsetx();
tu = vinsert(tu, chip->uh[vextract(ti, 0)], 0);
tu = vinsert(tu, chip->uh[vextract(ti, 1)], 1);
tu = vinsert(tu, chip->uh[vextract(ti, 2)], 2);
tu = vinsert(tu, chip->uh[vextract(ti, 3)], 3);
tu = vinsert(tu, chip->uh[vextract(ti, 4)], 4);
tu = vinsert(tu, chip->uh[vextract(ti, 5)], 5);
tu = vinsert(tu, chip->uh[vextract(ti, 6)], 6);
tu = vinsert(tu, chip->uh[vextract(ti, 7)], 7);

vi16x8_t gl = vmulhrs(tu, kgl);
vi16x8_t gr = vmulhrs(tu, kgr);
vi32x4_t ggl = vpaddlq_s16(gl);
vi32x4_t ggr = vpaddlq_s16(gr);
ggl = vvadd(ggl, vvext(ggl, ggl, 2));
Expand All @@ -226,28 +239,22 @@ void aymo_(process_i16)(struct aymo_(chip)* chip, uint32_t count, const int16_t
ggr = vvadd(ggr, vvrev(ggr));
vi16x8_t ggrl = vvpacks(ggr, ggl);
vi16x8_t gglr = vext(ggrl, ggrl, 1);
vi16x8_t vlr = vmulhrs(gglr, chip->kv);
vi16x8_t vlr = vmulhrs(gglr, kv);

vi16x8_t zc = chip->zc;
vi16x8_t zb = chip->zb;
zc = vext(zb, zc, 6); // '543210..'
chip->zc = zc;

vi16x8_t y1 = vmulhrs(zc, chip->kf);
vi16x8_t y0 = vmulhrs(zc, chip->ke);
vi16x8_t y1 = vmulhrs(zc, kf);
vi16x8_t y0 = vmulhrs(zc, ke);

vi16x8_t za = chip->za;
zb = vext(za, zb, 6); // '543210..'
chip->zb = zb;

y1 = vaddsi(y1, vmulhrs(zb, chip->kd));
y0 = vaddsi(y0, vmulhrs(zb, chip->kc));
y1 = vaddsi(y1, vmulhrs(zb, kd));
y0 = vaddsi(y0, vmulhrs(zb, kc));

za = vext(vlr, za, 6); // '543210..'
chip->za = za;

y1 = vaddsi(y1, vmulhrs(za, chip->kb));
y0 = vaddsi(y0, vmulhrs(za, chip->ka));
y1 = vaddsi(y1, vmulhrs(za, kb));
y0 = vaddsi(y0, vmulhrs(za, ka));

vi16x4_t yy0 = vqadd_s16(vgetlo(y0), vgethi(y0));
vi16x4_t yy1 = vqadd_s16(vgetlo(y1), vgethi(y1));
Expand All @@ -257,11 +264,14 @@ void aymo_(process_i16)(struct aymo_(chip)* chip, uint32_t count, const int16_t
vi16x4_t yy = vext_s16(yy0, yy1, 2);
yy = vand_s16(yy, vdup_n_s16((int16_t)AYMO_YM7128_SIGNAL_MASK));
vst1_s16(y, yy); y += 4u;
} while (--count);

ti0 = chip->xxv[0];
t0 = chip->uh[ti0];
chip->xxv[4] = t0;
}
chip->xxv = xxv;
chip->ti = ti;

chip->zc = zc;
chip->zb = zb;
chip->za = za;
}


Expand Down
Loading

0 comments on commit 6fcd10b

Please sign in to comment.