Skip to content

Commit

Permalink
Optimize LPF12_2 for ARM64
Browse files Browse the repository at this point in the history
  • Loading branch information
@starg2
starg2 committed Oct 23, 2024
1 parent 0b1c682 commit 7924e37
Show file tree
Hide file tree
Showing 3 changed files with 158 additions and 4 deletions.
154 changes: 152 additions & 2 deletions timidity/filter.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5238,7 +5238,69 @@ static void recalc_filter_LPF_BW_batch(int batch_size, FilterCoefficients **fcs)

#endif

#if (USE_X86_EXT_INTRIN >= 10) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)
#if (USE_ARM64_EXT_INTRIN >= 1) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)

static void sample_filter_LPF12_2_batch(int batch_size, FILTER_T** dcs, FILTER_T** dbs, DATA_T** sps, int32* counts)
{
for (int i = 0; i < MIX_VOICE_BATCH_SIZE; i += 2) {
if (i >= batch_size)
break;

int32 acounts[2] = {counts[i], i + 1 < batch_size ? counts[i + 1] : 0};
int32x2_t vcounts = vld1_s32(acounts);

float64x2_t vdb01[2];
vdb01[0] = vld1q_f64(&dbs[i][0]);
vdb01[1] = i + 1 < batch_size ? vld1q_f64(&dbs[i + 1][0]) : vdupq_n_f64(0.0);

float64x2_t vdb[2];
vdb[0] = vtrn1q_f64(vdb01[0], vdb01[1]);
vdb[1] = vtrn2q_f64(vdb01[0], vdb01[1]);

float64x2_t vdc01[2];
vdc01[0] = vld1q_f64(&dcs[i][0]);
vdc01[1] = i + 1 < batch_size ? vld1q_f64(&dcs[i + 1][0]) : vdupq_n_f64(0.0);

float64x2_t vdc[2];
vdc[0] = vtrn1q_f64(vdc01[0], vdc01[1]);
vdc[1] = vtrn2q_f64(vdc01[0], vdc01[1]);

int32 count_max = acounts[0] < acounts[1] ? acounts[1] : acounts[0];

for (int32 j = 0; j < count_max; j += 2) {
float64x2_t vsp01[2];
vsp01[0] = j < counts[i] ? vld1q_f64(&sps[i][j]) : vdupq_n_f64(0.0);
vsp01[1] = i + 1 < batch_size && j < counts[i + 1] ? vld1q_f64(&sps[i + 1][j]) : vdupq_n_f64(0.0);

float64x2_t vsps[2];
vsps[0] = vtrn1q_f64(vsp01[0], vsp01[1]);
vsps[1] = vtrn2q_f64(vsp01[0], vsp01[1]);

for (int k = 0; k < 2; k++) {
int32x2_t vmask32 = vreinterpret_s32_u32(vclt_s32(vdup_n_s32(j + k), vcounts));
uint64x2_t vmask = vreinterpretq_u64_s64(vmovl_s32(vmask32));

vdb[1] = vbslq_f64(vmask, vfmaq_f64(vdb[1], vsubq_f64(vsps[k], vdb[0]), vdc[1]), vdb[1]);
vdb[0] = vbslq_f64(vmask, vaddq_f64(vdb[0], vdb[1]), vdb[0]);
vdb[1] = vbslq_f64(vmask, vmulq_f64(vdb[1], vdc[0]), vdb[1]);
vsps[k] = vdb[0];
}

if (j < counts[i])
vst1q_f64(&sps[i][j], vtrn1q_f64(vsps[0], vsps[1]));

if (i + 1 < batch_size && j < counts[i + 1])
vst1q_f64(&sps[i + 1][j], vtrn2q_f64(vsps[0], vsps[1]));
}

vst1q_f64(dbs[i], vtrn1q_f64(vdb[0], vdb[1]));

if (i + 1 < batch_size)
vst1q_f64(dbs[i + 1], vtrn2q_f64(vdb[0], vdb[1]));
}
}

#elif (USE_X86_EXT_INTRIN >= 10) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)

static void sample_filter_LPF12_2_batch(int batch_size, FILTER_T **dcs, FILTER_T **dbs, DATA_T **sps, int32 *counts)
{
Expand Down Expand Up @@ -5458,7 +5520,91 @@ static void sample_filter_LPF12_2_batch(int batch_size, FILTER_T **dcs, FILTER_T

#endif

#if (USE_X86_EXT_INTRIN >= 10) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)
#if (USE_ARM64_EXT_INTRIN >= 1) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)

static void recalc_filter_LPF12_2_batch(int batch_size, FilterCoefficients** fcs)
{
for (int i = 0; i < MIX_VOICE_BATCH_SIZE; i += 2) {
if (i >= batch_size)
break;

float64x2_t vfcrange01[2];
vfcrange01[0] = vld1q_f64(&fcs[i]->range[0]); // [r00, r01]
vfcrange01[1] = i + 1 < batch_size ? vld1q_f64(&fcs[i + 1]->range[0]) : vdupq_n_f64(0); // [r10, r11]

float64x2_t vfcrange23[2];
vfcrange23[0] = vld1q_f64(&fcs[i]->range[2]); // [r02, r03]
vfcrange23[1] = i + 1 < batch_size ? vld1q_f64(&fcs[i + 1]->range[2]) : vdupq_n_f64(0); // [r12, r13]

float64x2_t vfcrange[4];
vfcrange[0] = vtrn1q_f64(vfcrange01[0], vfcrange01[1]); // [r00, r10]
vfcrange[1] = vtrn2q_f64(vfcrange01[0], vfcrange01[1]); // [r01, r11]
vfcrange[2] = vtrn1q_f64(vfcrange23[0], vfcrange23[1]); // [r02, r12]
vfcrange[3] = vtrn2q_f64(vfcrange23[0], vfcrange23[1]); // [r03, r13]

double afcfreq[2] = {fcs[i]->freq, i + 1 < batch_size ? fcs[i + 1]->freq : 0.0};
float64x2_t vfcfreq = vld1q_f64(afcfreq);

double afcreso_DB[2] = {fcs[i]->reso_dB, i + 1 < batch_size ? fcs[i + 1]->reso_dB : 0.0};
float64x2_t vfcreso_dB = vld1q_f64(afcreso_DB);

uint64x2_t vmask = vorrq_u64(
vorrq_u64(vcltq_f64(vfcfreq, vfcrange[0]), vcgtq_f64(vfcfreq, vfcrange[1])),
vorrq_u64(vcltq_f64(vfcreso_dB, vfcrange[2]), vcgtq_f64(vfcreso_dB, vfcrange[3]))
);

int amask[2] = {!!vmask.n128_u64[0], i + 1 < batch_size ? !!vmask.n128_u64[1] : 0};

if (amask[0] || amask[1]) {
// vfcfreq * (1.0 - ext_filter_margin) = vfcfreq - vfcfreq * ext_filter_margin
vfcrange[0] = vfmsq_n_f64(vfcfreq, vfcfreq, ext_filter_margin);
vfcrange[1] = vfmaq_n_f64(vfcfreq, vfcfreq, ext_filter_margin);
vfcrange[2] = vfmsq_n_f64(vfcreso_dB, vfcreso_dB, ext_filter_margin);
vfcrange[3] = vfmaq_n_f64(vfcreso_dB, vfcreso_dB, ext_filter_margin);

vfcrange01[0] = vtrn1q_f64(vfcrange[0], vfcrange[1]);
vfcrange01[1] = vtrn2q_f64(vfcrange[0], vfcrange[1]);
vfcrange23[0] = vtrn1q_f64(vfcrange[2], vfcrange[3]);
vfcrange23[1] = vtrn2q_f64(vfcrange[2], vfcrange[3]);

if (amask[0]) {
vst1q_f64(fcs[i]->range, vfcrange01[0]);
vst1q_f64(&fcs[i]->range[2], vfcrange23[0]);
}

if (amask[1]) {
vst1q_f64(fcs[i + 1]->range, vfcrange01[1]);
vst1q_f64(&fcs[i + 1]->range[2], vfcrange23[1]);
}

double afcdiv_flt_rate[2] = {fcs[i]->div_flt_rate, i + 1 < batch_size ? fcs[i + 1]->div_flt_rate : fcs[i]->div_flt_rate};
float64x2_t vf = vmulq_f64(vmulq_n_f64(vfcfreq, M_PI2), vld1q_f64(afcdiv_flt_rate));

FLOAT_T reso_db_cf_p = RESO_DB_CF_P(fcs[i]->reso_dB);
double areso_db_cf_p[2] = {reso_db_cf_p, i + 1 < batch_size ? RESO_DB_CF_P(fcs[i + 1]->reso_dB) : reso_db_cf_p};

float64x2_t v1 = vdupq_n_f64(1.0);
float64x2_t v2 = vdupq_n_f64(2.0);
float64x2_t v0_5 = vdupq_n_f64(0.5);

float64x2_t vq = vsubq_f64(v1, vdivq_f64(vf, vfmaq_f64(vsubq_f64(vf, v2), vaddq_f64(vld1q_f64(areso_db_cf_p), vdivq_f64(v0_5, vaddq_f64(v1, vf))), v2)));

float64x2_t vdc[2];
vdc[0] = vmulq_f64(vq, vq);

double acosf[2] = {cos(vf.n128_f64[0]), cos(vf.n128_f64[1])};
vdc[1] = vfmsq_f64(vaddq_f64(vdc[0], v1), vmulq_f64(v2, vld1q_f64(acosf)), vq);

if (amask[0])
vst1q_f64(fcs[i]->dc, vtrn1q_f64(vdc[0], vdc[1]));

if (amask[1])
vst1q_f64(fcs[i + 1]->dc, vtrn2q_f64(vdc[0], vdc[1]));
}
}
}

#elif (USE_X86_EXT_INTRIN >= 10) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)

static void recalc_filter_LPF12_2_batch(int batch_size, FilterCoefficients **fcs)
{
Expand Down Expand Up @@ -6028,6 +6174,7 @@ void buffer_filter_batch(int batch_size, FilterCoefficients **fcs, DATA_T **sps,
}

switch (fcs[0]->type) {
#if USE_X86_EXT_INTRIN >= 3
case FILTER_LPF24:
recalc_filter_LPF24_batch(batch_size, fcs);
sample_filter_LPF24_batch(batch_size, dcs, dbs, sps, counts);
Expand All @@ -6037,16 +6184,19 @@ void buffer_filter_batch(int batch_size, FilterCoefficients **fcs, DATA_T **sps,
recalc_filter_LPF_BW_batch(batch_size, fcs);
sample_filter_LPF_BW_batch(batch_size, dcs, dbs, sps, counts);
break;
#endif

case FILTER_LPF12_2:
recalc_filter_LPF12_2_batch(batch_size, fcs);
sample_filter_LPF12_2_batch(batch_size, dcs, dbs, sps, counts);
break;

#if USE_X86_EXT_INTRIN >= 3
case FILTER_HPF12_2:
recalc_filter_LPF12_2_batch(batch_size, fcs);
sample_filter_HPF12_2_batch(batch_size, dcs, dbs, sps, counts);
break;
#endif

default:
ctl->cmsg(CMSG_ERROR, VERB_NORMAL, "buffer_filter_batch(): error: unsupported filter type");
Expand Down
6 changes: 5 additions & 1 deletion timidity/mix.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -247,11 +247,15 @@ static int mix_voice_batch_is_filter_type_supported(int8 type)
{
switch (type) {
case FILTER_NONE:
#if USE_X86_EXT_INTRIN >= 3
case FILTER_LPF24:
case FILTER_LPF_BW:
#endif
case FILTER_LPF12_2:
#if USE_X86_EXT_INTRIN >= 3
case FILTER_HPF12_2:
return 1;
#endif
return 1;

default:
return 0;
Expand Down
2 changes: 1 addition & 1 deletion timidity/mix.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@
#ifndef ___MIX_H_
#define ___MIX_H_

#if (USE_X86_EXT_INTRIN >= 3) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)
#if ((USE_ARM64_EXT_INTRIN >= 1) || (USE_X86_EXT_INTRIN >= 3)) && defined(DATA_T_DOUBLE) && defined(FLOAT_T_DOUBLE)
#define MIX_VOICE_BATCH
#define MIX_VOICE_BATCH_SIZE 32
#endif
Expand Down

0 comments on commit 7924e37

Please sign in to comment.