wirtz
/
music-synthesizer-for-android
mirror of https://github.com/dcoredump/music-synthesizer-for-android.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Add NEON intrinsics for FM

Browse Source 
Ported the old assembly language file to intrinsics, so it builds on
64 bit as well as 32 bit.
master
Raph Levien 8 years ago
parent 076ab6eab8
commit cd8d8fbbaa
4 changed files with 142 additions and 9 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 5
      app/build.gradle
  2. 126
      app/src/main/jni/fm_op_kernel.cc
  3. 2
      app/src/main/jni/resofilter.cc
  4. 18
      app/src/main/jni/synth.h

5
app/build.gradle
Unescape View File

@ -22,6 +22,11 @@ model {
moduleName "synth"
ldLibs.addAll(['log', 'OpenSLES'])
}
abis {
create("armeabi-v7a") {
cppFlags.add('-mfpu=neon')
}
}
sources {
main {
jni {

126
app/src/main/jni/fm_op_kernel.cc
Unescape View File

@ -21,12 +21,132 @@
#include "sin.h"
#include "fm_op_kernel.h"
#ifdef HAVE_NEON
#ifdef HAVE_NEON_INTRINSICS
extern "C"
void neon_fm_kernel(const int *in, const int *busin, int *out, int count,
int32_t phase0, int32_t freq, int32_t gain1, int32_t dgain);
const int32_t __attribute__ ((aligned(16))) const_0_1_2_3_arg[4] = {0, 1, 2, 3};
const int32_t __attribute__ ((aligned(16))) mask23_arg = 0x7fffff;
const float32_t __attribute__ ((aligned(16))) coeffs_arg[4] = {
-0.01880853017455781, 0.25215252666796095, -1.2333439964934032, 1.0
};
void neon_fm_kernel(const int *in, const int *busin, int *out, int count,
int32_t phase0, int32_t freq_arg, int32_t gain1_arg, int32_t dgain_arg) {
int32x4_t phase = vld1q_dup_s32(&phase0);
int32x4_t freq = vld1q_dup_s32(&freq_arg);
int32x4_t const_0_1_2_3 = vld1q_s32(const_0_1_2_3_arg);
phase = vmlaq_s32(phase, freq, const_0_1_2_3);
int32x4_t gain1 = vld1q_dup_s32(&gain1_arg);
int32x4_t dgain = vld1q_dup_s32(&dgain_arg);
gain1 = vmlaq_s32(gain1, dgain, const_0_1_2_3);
int32x4_t mask23 = vld1q_dup_s32(&mask23_arg);
float32x4_t coeffs = vld1q_f32(coeffs_arg);
float32x4_t gainf = vcvtq_n_f32_s32(gain1, 24);
int32x4_t freq4 = vshlq_n_s32(freq, 2);
float32x4_t dgainf = vcvtq_n_f32_s32(dgain, 22);
count -= 4;
int32x4_t q15 = vmovq_n_s32(0x800000);
int32x4_t q7 = vmovq_n_s32(0x400000);
while (true) {
int32x4_t phase4 = vaddq_s32(phase, freq4);
int32x4_t phase8 = vaddq_s32(phase4, freq4);
int32x4_t data1a = vld1q_s32(in);
data1a = vaddq_s32(data1a, phase);
int32x4_t data1b = vld1q_s32(in + 4);
data1b = vaddq_s32(data1b, phase4);
int32x4_t data1c = vld1q_s32(in + 8);
data1c = vaddq_s32(data1c, phase8);
phase = vaddq_s32(phase8, freq4);
in += 12;
int32x4_t data4a = (int32x4_t)vtstq_s32(data1a, q15);
int32x4_t data4b = (int32x4_t)vtstq_s32(data1b, q15);
int32x4_t data4c = (int32x4_t)vtstq_s32(data1c, q15);
data1a = vandq_s32(data1a, mask23);
data1b = vandq_s32(data1b, mask23);
data1c = vandq_s32(data1c, mask23);
data1a = vsubq_s32(data1a, q7);
data1b = vsubq_s32(data1b, q7);
data1c = vsubq_s32(data1c, q7);
float32x4_t fdata1a = vcvtq_n_f32_s32(data1a, 22);
float32x4_t fdata1b = vcvtq_n_f32_s32(data1b, 22);
float32x4_t fdata1c = vcvtq_n_f32_s32(data1c, 22);
fdata1a = vmulq_f32(fdata1a, fdata1a);
fdata1b = vmulq_f32(fdata1b, fdata1b);
fdata1c = vmulq_f32(fdata1c, fdata1c);
float32x4_t fdata2a = vdupq_lane_f32(vget_low_f32(coeffs), 1);
float32x4_t fdata2b = vdupq_lane_f32(vget_low_f32(coeffs), 1);
float32x4_t fdata2c = vdupq_lane_f32(vget_low_f32(coeffs), 1);
fdata2a = vmlaq_lane_f32(fdata2a, fdata1a, vget_low_f32(coeffs), 0);
fdata2b = vmlaq_lane_f32(fdata2b, fdata1b, vget_low_f32(coeffs), 0);
fdata2c = vmlaq_lane_f32(fdata2c, fdata1c, vget_low_f32(coeffs), 0);
float32x4_t fdata3a = vdupq_lane_f32(vget_high_f32(coeffs), 0);
float32x4_t fdata3b = vdupq_lane_f32(vget_high_f32(coeffs), 0);
float32x4_t fdata3c = vdupq_lane_f32(vget_high_f32(coeffs), 0);
fdata3a = vmlaq_f32(fdata3a, fdata1a, fdata2a);
fdata3b = vmlaq_f32(fdata3b, fdata1b, fdata2b);
fdata3c = vmlaq_f32(fdata3c, fdata1c, fdata2c);
fdata2a = vdupq_lane_f32(vget_high_f32(coeffs), 1);
fdata2b = vdupq_lane_f32(vget_high_f32(coeffs), 1);
fdata2c = vdupq_lane_f32(vget_high_f32(coeffs), 1);
fdata2a = vmlaq_f32(fdata2a, fdata1a, fdata3a);
fdata2b = vmlaq_f32(fdata2b, fdata1b, fdata3b);
fdata2c = vmlaq_f32(fdata2c, fdata1c, fdata3c);
fdata3a = vaddq_f32(gainf, dgainf);
fdata3b = vaddq_f32(fdata3a, dgainf);
fdata2a = vmulq_f32(fdata2a, gainf);
fdata2b = vmulq_f32(fdata2b, fdata3a);
fdata2c = vmulq_f32(fdata2c, fdata3b);
gainf = vaddq_f32(fdata3b, dgainf);
int32x4_t data3a = vcvtq_n_s32_f32(fdata2a, 24);
int32x4_t data3b = vcvtq_n_s32_f32(fdata2b, 24);
int32x4_t data3c = vcvtq_n_s32_f32(fdata2c, 24);
data1a = vld1q_s32(busin);
data1b = vld1q_s32(busin + 4);
data1c = vld1q_s32(busin + 8);
busin += 12;
data3a = veorq_s32(data3a, data4a);
data3b = veorq_s32(data3b, data4b);
data3c = veorq_s32(data3c, data4c);
data3a = vaddq_s32(data3a, data1a);
data3b = vaddq_s32(data3b, data1b);
data3c = vaddq_s32(data3c, data1c);
vst1q_s32(out, data3a);
vst1q_s32(out + 4, data3b);
vst1q_s32(out + 8, data3c);
out += 12;
count -= 12;
if (count <= 0) {
if (count == 0) {
// finish last chunk of 4
data1a = vld1q_s32(in);
data1a = vaddq_s32(data1a, phase);
data4a = (int32x4_t)vtstq_s32(data1a, q15);
data1a = vandq_s32(data1a, mask23);
data1a = vsubq_s32(data1a, q7);
fdata1a = vcvtq_n_f32_s32(data1a, 22);
fdata1a = vmulq_f32(fdata1a, fdata1a);
fdata2a = vdupq_lane_f32(vget_low_f32(coeffs), 1);
fdata2a = vmlaq_lane_f32(fdata2a, fdata1a, vget_low_f32(coeffs), 0);
fdata3a = vdupq_lane_f32(vget_high_f32(coeffs), 0);
fdata3a = vmlaq_f32(fdata3a, fdata1a, fdata2a);
fdata2a = vdupq_lane_f32(vget_high_f32(coeffs), 1);
fdata2a = vmlaq_f32(fdata2a, fdata1a, fdata3a);
fdata2a = vmulq_f32(fdata2a, gainf);
data3a = vcvtq_n_s32_f32(fdata2a, 24);
data1a = vld1q_s32(busin);
data3a = veorq_s32(data3a, data4a);
data3a = vaddq_s32(data3a, data1a);
vst1q_s32(out, data3a);
}
break;
}
}
}
const int32_t __attribute__ ((aligned(16))) zeros[N] = {0};
#endif
@ -38,7 +158,7 @@ void FmOpKernel::compute(int32_t *output, const int32_t *input,
int32_t gain = gain1;
int32_t phase = phase0;
if (hasNeon()) {
#ifdef HAVE_NEON
#ifdef HAVE_NEON_INTRINSICS
neon_fm_kernel(input, add ? output : zeros, output, N,
phase0, freq, gain, dgain);
#endif
@ -68,7 +188,7 @@ void FmOpKernel::compute_pure(int32_t *output, int32_t phase0, int32_t freq,
int32_t gain = gain1;
int32_t phase = phase0;
if (hasNeon()) {
#ifdef HAVE_NEON
#ifdef HAVE_NEON_INTRINSICS
neon_fm_kernel(zeros, add ? output : zeros, output, N,
phase0, freq, gain, dgain);
#endif

2
app/src/main/jni/resofilter.cc
Unescape View File

@ -202,7 +202,7 @@ void ResoFilter::process(const int32_t **inbufs, const int32_t *control_in,
float overdrive = control_in[2] * (1.0 / (1 << 24));
const int32_t *ibuf = inbufs[0];
int32_t *obuf = outbufs[0];
bool useneon = hasNeon();
bool useneon = false; // TODO: hasNeon();
if (useneon) {
#ifdef HAVE_NEON
float params[2];

18
app/src/main/jni/synth.h
Unescape View File

@ -51,11 +51,21 @@ inline static T max(const T& a, const T& b) {
return a > b ? a : b;
}
#ifdef HAVE_NEON
#include <cpu-features.h>
#ifdef __aarch64__
#define HAVE_NEON_INTRINSICS
#include <arm_neon.h>
static inline bool hasNeon() {
return true;
}
#elif __ARM_NEON__
#define HAVE_NEON_INTRINSICS
#include <arm_neon.h>
// TODO: get this working (although maybe we don't even care about non-NEON ARM any more)
//#include <cpu-features.h>
static inline bool hasNeon() {
return (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
return true;
//return (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
}
#else
static inline bool hasNeon() {
@ -63,6 +73,4 @@ static inline bool hasNeon() {
}
#endif
#endif // __SYNTH_H

Write Preview
Loading…
Cancel
Save