void AudioBlockPanStereoToStereo_NEON(const float aInputL[WEBAUDIO_BLOCK_SIZE],
const float aInputR[WEBAUDIO_BLOCK_SIZE],
float aGainL, float aGainR,
bool aIsOnTheLeft,
float aOutputL[WEBAUDIO_BLOCK_SIZE],
float aOutputR[WEBAUDIO_BLOCK_SIZE]) {
ASSERT_ALIGNED(aInputL);
ASSERT_ALIGNED(aInputR);
ASSERT_ALIGNED(aOutputL);
ASSERT_ALIGNED(aOutputR);
float32x4_t vinL0, vinL1;
float32x4_t vinR0, vinR1;
float32x4_t voutL0, voutL1;
float32x4_t voutR0, voutR1;
float32x4_t vscaleL = vmovq_n_f32(aGainL);
float32x4_t vscaleR = vmovq_n_f32(aGainR);
if (aIsOnTheLeft) {
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
voutL0 = vmlaq_f32(vinL0, vinR0, vscaleL);
voutL1 = vmlaq_f32(vinL1, vinR1, vscaleL);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
voutR0 = vmulq_f32(vinR0, vscaleR);
voutR1 = vmulq_f32(vinR1, vscaleR);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
} else {
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
voutL0 = vmulq_f32(vinL0, vscaleL);
voutL1 = vmulq_f32(vinL1, vscaleL);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
voutR0 = vmlaq_f32(vinR0, vinL0, vscaleR);
voutR1 = vmlaq_f32(vinR1, vinL1, vscaleR);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
}
}
static void test_fma() {
for(int i=0; i<1020 * 4; i++) {
data_f[i] = i;
}
float32x4_t c0_02 = vdupq_n_f32(0.02f);
float32x4_t c0_04 = vdupq_n_f32(0.04f);
float32x4_t c0_05 = vdupq_n_f32(0.05f);
float32x4_t c0_10 = vdupq_n_f32(0.1f);
float32x4_t c0_20 = vdupq_n_f32(0.2f);
float32x4_t c1_00 = vdupq_n_f32(1.0f);
startTime();
// Do ~1 billion ops
for (int ct=0; ct < (1000 * (1000 / 80)); ct++) {
for (int i=0; i < 1000; i++) {
float32x4_t t;
t = vmulq_f32(vld1q_f32((float32_t *)&data_f[i]), c0_02);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+4]), c0_04);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+8]), c0_05);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+12]), c0_10);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+16]), c0_20);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+20]), c0_20);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+24]), c0_10);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+28]), c0_05);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+32]), c0_04);
t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+36]), c0_02);
t = vaddq_f32(t, c1_00);
vst1q_f32((float32_t *)&data_f[i], t);
}
}
endTime("neon fma", 1e9);
}
static void cft1st_128_neon(float* a) {
const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign);
int j, k2;
for (k2 = 0, j = 0; j < 128; j += 16, k2 += 4) {
float32x4_t a00v = vld1q_f32(&a[j + 0]);
float32x4_t a04v = vld1q_f32(&a[j + 4]);
float32x4_t a08v = vld1q_f32(&a[j + 8]);
float32x4_t a12v = vld1q_f32(&a[j + 12]);
float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v));
float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v));
float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v));
float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v));
const float32x4_t wk1rv = vld1q_f32(&rdft_wk1r[k2]);
const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2]);
const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2]);
const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2]);
const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2]);
const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2]);
float32x4_t x0v = vaddq_f32(a01v, a23v);
const float32x4_t x1v = vsubq_f32(a01v, a23v);
const float32x4_t x2v = vaddq_f32(a45v, a67v);
const float32x4_t x3v = vsubq_f32(a45v, a67v);
const float32x4_t x3w = vrev64q_f32(x3v);
float32x4_t x0w;
a01v = vaddq_f32(x0v, x2v);
x0v = vsubq_f32(x0v, x2v);
x0w = vrev64q_f32(x0v);
a45v = vmulq_f32(wk2rv, x0v);
a45v = vmlaq_f32(a45v, wk2iv, x0w);
x0v = vmlaq_f32(x1v, x3w, vec_swap_sign);
x0w = vrev64q_f32(x0v);
a23v = vmulq_f32(wk1rv, x0v);
a23v = vmlaq_f32(a23v, wk1iv, x0w);
x0v = vmlsq_f32(x1v, x3w, vec_swap_sign);
x0w = vrev64q_f32(x0v);
a67v = vmulq_f32(wk3rv, x0v);
a67v = vmlaq_f32(a67v, wk3iv, x0w);
a00v = vcombine_f32(vget_low_f32(a01v), vget_low_f32(a23v));
a04v = vcombine_f32(vget_low_f32(a45v), vget_low_f32(a67v));
a08v = vcombine_f32(vget_high_f32(a01v), vget_high_f32(a23v));
a12v = vcombine_f32(vget_high_f32(a45v), vget_high_f32(a67v));
vst1q_f32(&a[j + 0], a00v);
vst1q_f32(&a[j + 4], a04v);
vst1q_f32(&a[j + 8], a08v);
vst1q_f32(&a[j + 12], a12v);
}
}
void * scaled_sumf_thread_NEON(void * argument)
{
jsize i = 0;
struct scaled_sumfneon_thread_data * data = (struct scaled_sumfneon_thread_data *) argument;
float32_t * r = (float32_t *)data->r;
const float32_t * x = (const float32_t *)data->x;
const float32_t * y = (const float32_t *)data->y;
const float32_t a = (const float32_t)data->a;
const jsize size = data->size;
float32x4_t rx4, xx4, yx4, ax4;
ax4 = vdupq_n_f32(a);
for(i; i < size ; i += 4)
{
xx4 = vld1q_f32(&(x[i]));
yx4 = vld1q_f32(&(y[i]));
rx4 = vmlaq_f32(xx4, ax4, yx4);
vst1q_f32(&(r[i]), rx4);
}
}
template <bool align> SIMD_INLINE void SquaredDifferenceSum16f(const uint16_t * a, const uint16_t * b, size_t offset, float32x4_t & sum)
{
float32x4_t _a = vcvt_f32_f16((float16x4_t)LoadHalf<align>(a + offset));
float32x4_t _b = vcvt_f32_f16((float16x4_t)LoadHalf<align>(b + offset));
float32x4_t _d = vsubq_f32(_a, _b);
sum = vmlaq_f32(sum, _d, _d);
}
template <bool align> SIMD_INLINE void SquaredDifferenceSum32f(const float * a, const float * b, size_t offset, float32x4_t & sum)
{
float32x4_t _a = Load<align>(a + offset);
float32x4_t _b = Load<align>(b + offset);
float32x4_t _d = vsubq_f32(_a, _b);
sum = vmlaq_f32(sum, _d, _d);
}
static void ScaleErrorSignalNEON(int extended_filter_enabled,
float normal_mu,
float normal_error_threshold,
float x_pow[PART_LEN1],
float ef[2][PART_LEN1]) {
const float mu = extended_filter_enabled ? kExtendedMu : normal_mu;
const float error_threshold = extended_filter_enabled ?
kExtendedErrorThreshold : normal_error_threshold;
const float32x4_t k1e_10f = vdupq_n_f32(1e-10f);
const float32x4_t kMu = vmovq_n_f32(mu);
const float32x4_t kThresh = vmovq_n_f32(error_threshold);
int i;
// vectorized code (four at once)
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t x_pow_local = vld1q_f32(&x_pow[i]);
const float32x4_t ef_re_base = vld1q_f32(&ef[0][i]);
const float32x4_t ef_im_base = vld1q_f32(&ef[1][i]);
const float32x4_t xPowPlus = vaddq_f32(x_pow_local, k1e_10f);
float32x4_t ef_re = vdivq_f32(ef_re_base, xPowPlus);
float32x4_t ef_im = vdivq_f32(ef_im_base, xPowPlus);
const float32x4_t ef_re2 = vmulq_f32(ef_re, ef_re);
const float32x4_t ef_sum2 = vmlaq_f32(ef_re2, ef_im, ef_im);
const float32x4_t absEf = vsqrtq_f32(ef_sum2);
const uint32x4_t bigger = vcgtq_f32(absEf, kThresh);
const float32x4_t absEfPlus = vaddq_f32(absEf, k1e_10f);
const float32x4_t absEfInv = vdivq_f32(kThresh, absEfPlus);
uint32x4_t ef_re_if = vreinterpretq_u32_f32(vmulq_f32(ef_re, absEfInv));
uint32x4_t ef_im_if = vreinterpretq_u32_f32(vmulq_f32(ef_im, absEfInv));
uint32x4_t ef_re_u32 = vandq_u32(vmvnq_u32(bigger),
vreinterpretq_u32_f32(ef_re));
uint32x4_t ef_im_u32 = vandq_u32(vmvnq_u32(bigger),
vreinterpretq_u32_f32(ef_im));
ef_re_if = vandq_u32(bigger, ef_re_if);
ef_im_if = vandq_u32(bigger, ef_im_if);
ef_re_u32 = vorrq_u32(ef_re_u32, ef_re_if);
ef_im_u32 = vorrq_u32(ef_im_u32, ef_im_if);
ef_re = vmulq_f32(vreinterpretq_f32_u32(ef_re_u32), kMu);
ef_im = vmulq_f32(vreinterpretq_f32_u32(ef_im_u32), kMu);
vst1q_f32(&ef[0][i], ef_re);
vst1q_f32(&ef[1][i], ef_im);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
float abs_ef;
ef[0][i] /= (x_pow[i] + 1e-10f);
ef[1][i] /= (x_pow[i] + 1e-10f);
abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
if (abs_ef > error_threshold) {
abs_ef = error_threshold / (abs_ef + 1e-10f);
ef[0][i] *= abs_ef;
ef[1][i] *= abs_ef;
}
// Stepsize factor
ef[0][i] *= mu;
ef[1][i] *= mu;
}
}
f64 dotProduct(const Size2D &_size,
const f32 * src0Base, ptrdiff_t src0Stride,
const f32 * src1Base, ptrdiff_t src1Stride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
Size2D size(_size);
if (src0Stride == src1Stride &&
src0Stride == (ptrdiff_t)(size.width * sizeof(f32)))
{
size.width *= size.height;
size.height = 1;
}
#define DOT_FLOAT_BLOCKSIZE (1 << 13)
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const f32 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const f32 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
while(i + 4 <= size.width)
{
size_t lim = std::min(i + DOT_FLOAT_BLOCKSIZE, size.width) - 4;
float32x4_t v_sum = vdupq_n_f32(0.0f);
for( ; i <= lim; i += 4 )
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
v_sum = vmlaq_f32(v_sum, vld1q_f32(src0 + i), vld1q_f32(src1 + i));
}
float32x2_t vres = vpadd_f32(vget_low_f32(v_sum),vget_high_f32(v_sum));
result += vget_lane_f32(vres, 0) + vget_lane_f32(vres, 1);
}
if(i + 2 <= size.width)
{
float32x2_t vres = vmul_f32(vld1_f32(src0 + i), vld1_f32(src1 + i));
result += vget_lane_f32(vres, 0) + vget_lane_f32(vres, 1);
i += 2;
}
for (; i < size.width; ++i)
result += src0[i] * src1[i];
}
return result;
#else
(void)_size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
return 0;
#endif
}
static void SubbandCoherenceNEON(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1],
float* fft,
float* cohde,
float* cohxd,
int* extreme_filter_divergence) {
int i;
SmoothedPSD(aec, efw, dfw, xfw, extreme_filter_divergence);
{
const float32x4_t vec_1eminus10 = vdupq_n_f32(1e-10f);
// Subband coherence
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_sd = vld1q_f32(&aec->sd[i]);
const float32x4_t vec_se = vld1q_f32(&aec->se[i]);
const float32x4_t vec_sx = vld1q_f32(&aec->sx[i]);
const float32x4_t vec_sdse = vmlaq_f32(vec_1eminus10, vec_sd, vec_se);
const float32x4_t vec_sdsx = vmlaq_f32(vec_1eminus10, vec_sd, vec_sx);
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_cohde = vmulq_f32(vec_sde.val[0], vec_sde.val[0]);
float32x4_t vec_cohxd = vmulq_f32(vec_sxd.val[0], vec_sxd.val[0]);
vec_cohde = vmlaq_f32(vec_cohde, vec_sde.val[1], vec_sde.val[1]);
vec_cohde = vdivq_f32(vec_cohde, vec_sdse);
vec_cohxd = vmlaq_f32(vec_cohxd, vec_sxd.val[1], vec_sxd.val[1]);
vec_cohxd = vdivq_f32(vec_cohxd, vec_sdsx);
vst1q_f32(&cohde[i], vec_cohde);
vst1q_f32(&cohxd[i], vec_cohxd);
}
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
cohde[i] =
(aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) /
(aec->sd[i] * aec->se[i] + 1e-10f);
cohxd[i] =
(aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) /
(aec->sx[i] * aec->sd[i] + 1e-10f);
}
}
void test_vmlaQf32 (void)
{
float32x4_t out_float32x4_t;
float32x4_t arg0_float32x4_t;
float32x4_t arg1_float32x4_t;
float32x4_t arg2_float32x4_t;
out_float32x4_t = vmlaq_f32 (arg0_float32x4_t, arg1_float32x4_t, arg2_float32x4_t);
}
static int PartitionDelayNEON(const AecCore* aec) {
// Measures the energy in each filter partition and returns the partition with
// highest energy.
// TODO(bjornv): Spread computational cost by computing one partition per
// block?
float wfEnMax = 0;
int i;
int delay = 0;
for (i = 0; i < aec->num_partitions; i++) {
int j;
int pos = i * PART_LEN1;
float wfEn = 0;
float32x4_t vec_wfEn = vdupq_n_f32(0.0f);
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t vec_wfBuf0 = vld1q_f32(&aec->wfBuf[0][pos + j]);
const float32x4_t vec_wfBuf1 = vld1q_f32(&aec->wfBuf[1][pos + j]);
vec_wfEn = vmlaq_f32(vec_wfEn, vec_wfBuf0, vec_wfBuf0);
vec_wfEn = vmlaq_f32(vec_wfEn, vec_wfBuf1, vec_wfBuf1);
}
{
float32x2_t vec_total;
// A B C D
vec_total = vpadd_f32(vget_low_f32(vec_wfEn), vget_high_f32(vec_wfEn));
// A+B C+D
vec_total = vpadd_f32(vec_total, vec_total);
// A+B+C+D A+B+C+D
wfEn = vget_lane_f32(vec_total, 0);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
wfEn += aec->wfBuf[0][pos + j] * aec->wfBuf[0][pos + j] +
aec->wfBuf[1][pos + j] * aec->wfBuf[1][pos + j];
}
if (wfEn > wfEnMax) {
wfEnMax = wfEn;
delay = i;
}
}
return delay;
}
int Scale_arm::forward(const Mat& bottom_blob, Mat& top_blob) const
{
int w = bottom_blob.w;
int h = bottom_blob.h;
int channels = bottom_blob.c;
int size = w * h;
top_blob.create(w, h, channels);
if (top_blob.empty())
return -100;
if (bias_term)
{
const float* scale_ptr = scale_data;
const float* bias_ptr = bias_data;
#pragma omp parallel for
for (int q=0; q<channels; q++)
{
const float* ptr = bottom_blob.channel(q);
float* outptr = top_blob.channel(q);
float s = scale_ptr[q];
float bias = bias_ptr[q];
#if __ARM_NEON
int nn = size >> 2;
int remain = size - (nn << 2);
#else
int remain = size;
#endif // __ARM_NEON
#if __ARM_NEON
float32x4_t _s = vdupq_n_f32(s);
float32x4_t _bias = vdupq_n_f32(bias);
for (; nn>0; nn--)
{
float32x4_t _p = vld1q_f32(ptr);
_p = vmlaq_f32(_bias, _p, _s);
vst1q_f32(outptr, _p);
ptr += 4;
outptr += 4;
}
#endif // __ARM_NEON
for (; remain>0; remain--)
{
*outptr = *ptr * s + bias;
ptr++;
outptr++;
}
}
}
else
{
template <bool align> SIMD_INLINE void SquaredDifferenceKahanSum32f(const float * a, const float * b, size_t offset, float32x4_t & sum, float32x4_t & correction)
{
float32x4_t _a = Load<align>(a + offset);
float32x4_t _b = Load<align>(b + offset);
float32x4_t _d = vsubq_f32(_a, _b);
float32x4_t term = vmlaq_f32(correction, _d, _d);
float32x4_t temp = vaddq_f32(sum, term);
correction = vsubq_f32(vmulq_f32(temp, sum), term);
sum = temp;
}
void AudioBufferAddWithScale_NEON(const float* aInput,
float aScale,
float* aOutput,
uint32_t aSize)
{
ASSERT_ALIGNED(aInput);
ASSERT_ALIGNED(aOutput);
float32x4_t vin0, vin1, vin2, vin3;
float32x4_t vout0, vout1, vout2, vout3;
float32x4_t vscale = vmovq_n_f32(aScale);
uint32_t dif = aSize % 16;
aSize -= dif;
unsigned i = 0;
for (; i < aSize; i+=16) {
vin0 = vld1q_f32(ADDRESS_OF(aInput, i));
vin1 = vld1q_f32(ADDRESS_OF(aInput, i+4));
vin2 = vld1q_f32(ADDRESS_OF(aInput, i+8));
vin3 = vld1q_f32(ADDRESS_OF(aInput, i+12));
vout0 = vld1q_f32(ADDRESS_OF(aOutput, i));
vout1 = vld1q_f32(ADDRESS_OF(aOutput, i+4));
vout2 = vld1q_f32(ADDRESS_OF(aOutput, i+8));
vout3 = vld1q_f32(ADDRESS_OF(aOutput, i+12));
vout0 = vmlaq_f32(vout0, vin0, vscale);
vout1 = vmlaq_f32(vout1, vin1, vscale);
vout2 = vmlaq_f32(vout2, vin2, vscale);
vout3 = vmlaq_f32(vout3, vin3, vscale);
vst1q_f32(ADDRESS_OF(aOutput, i), vout0);
vst1q_f32(ADDRESS_OF(aOutput, i+4), vout1);
vst1q_f32(ADDRESS_OF(aOutput, i+8), vout2);
vst1q_f32(ADDRESS_OF(aOutput, i+12), vout3);
}
for (unsigned j = 0; j < dif; ++i, ++j) {
aOutput[i] += aInput[i]*aScale;
}
}
static void FilterFarNEON(
int num_partitions,
int x_fft_buf_block_pos,
float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
float y_fft[2][PART_LEN1]) {
int i;
for (i = 0; i < num_partitions; i++) {
int j;
int xPos = (i + x_fft_buf_block_pos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + x_fft_buf_block_pos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t x_fft_buf_re = vld1q_f32(&x_fft_buf[0][xPos + j]);
const float32x4_t x_fft_buf_im = vld1q_f32(&x_fft_buf[1][xPos + j]);
const float32x4_t h_fft_buf_re = vld1q_f32(&h_fft_buf[0][pos + j]);
const float32x4_t h_fft_buf_im = vld1q_f32(&h_fft_buf[1][pos + j]);
const float32x4_t y_fft_re = vld1q_f32(&y_fft[0][j]);
const float32x4_t y_fft_im = vld1q_f32(&y_fft[1][j]);
const float32x4_t a = vmulq_f32(x_fft_buf_re, h_fft_buf_re);
const float32x4_t e = vmlsq_f32(a, x_fft_buf_im, h_fft_buf_im);
const float32x4_t c = vmulq_f32(x_fft_buf_re, h_fft_buf_im);
const float32x4_t f = vmlaq_f32(c, x_fft_buf_im, h_fft_buf_re);
const float32x4_t g = vaddq_f32(y_fft_re, e);
const float32x4_t h = vaddq_f32(y_fft_im, f);
vst1q_f32(&y_fft[0][j], g);
vst1q_f32(&y_fft[1][j], h);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
y_fft[0][j] += MulRe(x_fft_buf[0][xPos + j],
x_fft_buf[1][xPos + j],
h_fft_buf[0][pos + j],
h_fft_buf[1][pos + j]);
y_fft[1][j] += MulIm(x_fft_buf[0][xPos + j],
x_fft_buf[1][xPos + j],
h_fft_buf[0][pos + j],
h_fft_buf[1][pos + j]);
}
}
}
static void FilterFarNEON(AecCore* aec, float yf[2][PART_LEN1]) {
int i;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t xfBuf_re = vld1q_f32(&aec->xfBuf[0][xPos + j]);
const float32x4_t xfBuf_im = vld1q_f32(&aec->xfBuf[1][xPos + j]);
const float32x4_t wfBuf_re = vld1q_f32(&aec->wfBuf[0][pos + j]);
const float32x4_t wfBuf_im = vld1q_f32(&aec->wfBuf[1][pos + j]);
const float32x4_t yf_re = vld1q_f32(&yf[0][j]);
const float32x4_t yf_im = vld1q_f32(&yf[1][j]);
const float32x4_t a = vmulq_f32(xfBuf_re, wfBuf_re);
const float32x4_t e = vmlsq_f32(a, xfBuf_im, wfBuf_im);
const float32x4_t c = vmulq_f32(xfBuf_re, wfBuf_im);
const float32x4_t f = vmlaq_f32(c, xfBuf_im, wfBuf_re);
const float32x4_t g = vaddq_f32(yf_re, e);
const float32x4_t h = vaddq_f32(yf_im, f);
vst1q_f32(&yf[0][j], g);
vst1q_f32(&yf[1][j], h);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
yf[0][j] += MulRe(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
yf[1][j] += MulIm(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
}
}
}
/**
* @brief vector_dot_vector.
*
* @param dst[out] the output element(1*1)
* @param src1[in] the input vector(1*n)
* src2[in] the input vector(1*n)
* dimN[in] size of vector
*
* @return void
*/
void neon_VecdotVec(float *dst,
const float *src1,
const float *src2,
const int dimN)
{
float *mat0 = (float *)src1;
float *mat1 = (float *)src2;
float32x4_t q0 = vld1q_f32(mat0);
float32x4_t q1 = vld1q_f32(mat1);
q0 = vmulq_f32(q0, q1);
int j = 4;
for (; j <= dimN - 4; j += 4)
{
float32x4_t q2 = vld1q_f32(mat0 + j);
float32x4_t q3 = vld1q_f32(mat1 + j);
q0 = vmlaq_f32(q0, q2, q3);
}
float32x2_t d0 = vpadd_f32(vget_low_f32(q0), vget_high_f32(q0));
d0 = vpadd_f32(d0, d0);
*dst = *((float *)&d0);
for (; j < dimN; j++) {
*dst += src1[j] * src2[j];
}
}
static void ne10_fft16_backward_float32_neon (ne10_fft_cpx_float32_t * Fout,
ne10_fft_cpx_float32_t * Fin,
ne10_fft_cpx_float32_t * twiddles)
{
ne10_fft_cpx_float32_t *tw1, *tw2, *tw3;
// the first stage
float32_t *p_src0, *p_src4, *p_src8, *p_src12;
float32x4x2_t q2_in_0123, q2_in_4567, q2_in_89ab, q2_in_cdef;
float32x4_t q_t0_r, q_t0_i, q_t1_r, q_t1_i, q_t2_r, q_t2_i, q_t3_r, q_t3_i;
float32x4_t q_out_r048c, q_out_i048c, q_out_r159d, q_out_i159d;
float32x4_t q_out_r26ae, q_out_i26ae, q_out_r37bf, q_out_i37bf;
p_src0 = (float32_t*) (& (Fin[0]));
p_src4 = (float32_t*) (& (Fin[4]));
p_src8 = (float32_t*) (& (Fin[8]));
p_src12 = (float32_t*) (& (Fin[12]));
q2_in_0123 = vld2q_f32 (p_src0);
q2_in_4567 = vld2q_f32 (p_src4);
q2_in_89ab = vld2q_f32 (p_src8);
q2_in_cdef = vld2q_f32 (p_src12);
q_t2_r = vsubq_f32 (q2_in_0123.val[0], q2_in_89ab.val[0]);
q_t2_i = vsubq_f32 (q2_in_0123.val[1], q2_in_89ab.val[1]);
q_t3_r = vaddq_f32 (q2_in_0123.val[0], q2_in_89ab.val[0]);
q_t3_i = vaddq_f32 (q2_in_0123.val[1], q2_in_89ab.val[1]);
q_t0_r = vaddq_f32 (q2_in_4567.val[0], q2_in_cdef.val[0]);
q_t0_i = vaddq_f32 (q2_in_4567.val[1], q2_in_cdef.val[1]);
q_t1_r = vsubq_f32 (q2_in_4567.val[0], q2_in_cdef.val[0]);
q_t1_i = vsubq_f32 (q2_in_4567.val[1], q2_in_cdef.val[1]);
q_out_r26ae = vsubq_f32 (q_t3_r, q_t0_r);
q_out_i26ae = vsubq_f32 (q_t3_i, q_t0_i);
q_out_r048c = vaddq_f32 (q_t3_r, q_t0_r);
q_out_i048c = vaddq_f32 (q_t3_i, q_t0_i);
q_out_r159d = vsubq_f32 (q_t2_r, q_t1_i);
q_out_i159d = vaddq_f32 (q_t2_i, q_t1_r);
q_out_r37bf = vaddq_f32 (q_t2_r, q_t1_i);
q_out_i37bf = vsubq_f32 (q_t2_i, q_t1_r);
// second stages
float32_t *p_dst0, *p_dst1, *p_dst2, *p_dst3;
float32_t *p_tw1, *p_tw2, *p_tw3;
float32x4_t q_s0_r, q_s0_i, q_s1_r, q_s1_i, q_s2_r, q_s2_i;
float32x4_t q_s3_r, q_s3_i, q_s4_r, q_s4_i, q_s5_r, q_s5_i;
float32x4x2_t q2_tmp_0, q2_tmp_1, q2_tmp_2, q2_tmp_3;
float32x4_t q_in_r0123, q_in_r4567, q_in_r89ab, q_in_rcdef;
float32x4_t q_in_i0123, q_in_i4567, q_in_i89ab, q_in_icdef;
float32x4x2_t q2_tw1, q2_tw2, q2_tw3;
float32x4x2_t q2_out_0123, q2_out_4567, q2_out_89ab, q2_out_cdef;
float32x4_t q_one_by_nfft;
tw1 = twiddles;
tw2 = twiddles + 4;
tw3 = twiddles + 8;
p_dst0 = (float32_t*) (&Fout[0]);
p_dst1 = (float32_t*) (&Fout[4]);
p_dst2 = (float32_t*) (&Fout[8]);
p_dst3 = (float32_t*) (&Fout[12]);
p_tw1 = (float32_t*) tw1;
p_tw2 = (float32_t*) tw2;
p_tw3 = (float32_t*) tw3;
q2_tmp_0 = vzipq_f32 (q_out_r048c, q_out_r159d);
q2_tmp_1 = vzipq_f32 (q_out_i048c, q_out_i159d);
q2_tmp_2 = vzipq_f32 (q_out_r26ae, q_out_r37bf);
q2_tmp_3 = vzipq_f32 (q_out_i26ae, q_out_i37bf);
q_in_r0123 = vcombine_f32 (vget_low_f32 (q2_tmp_0.val[0]), vget_low_f32 (q2_tmp_2.val[0]));
q_in_i0123 = vcombine_f32 (vget_low_f32 (q2_tmp_1.val[0]), vget_low_f32 (q2_tmp_3.val[0]));
q_in_r4567 = vcombine_f32 (vget_high_f32 (q2_tmp_0.val[0]), vget_high_f32 (q2_tmp_2.val[0]));
q_in_i4567 = vcombine_f32 (vget_high_f32 (q2_tmp_1.val[0]), vget_high_f32 (q2_tmp_3.val[0]));
q_in_r89ab = vcombine_f32 (vget_low_f32 (q2_tmp_0.val[1]), vget_low_f32 (q2_tmp_2.val[1]));
q_in_i89ab = vcombine_f32 (vget_low_f32 (q2_tmp_1.val[1]), vget_low_f32 (q2_tmp_3.val[1]));
q_in_rcdef = vcombine_f32 (vget_high_f32 (q2_tmp_0.val[1]), vget_high_f32 (q2_tmp_2.val[1]));
q_in_icdef = vcombine_f32 (vget_high_f32 (q2_tmp_1.val[1]), vget_high_f32 (q2_tmp_3.val[1]));
q2_tw1 = vld2q_f32 (p_tw1);
q2_tw2 = vld2q_f32 (p_tw2);
q2_tw3 = vld2q_f32 (p_tw3);
q_s0_r = vmulq_f32 (q_in_r4567, q2_tw1.val[0]);
q_s0_i = vmulq_f32 (q_in_i4567, q2_tw1.val[0]);
q_s1_r = vmulq_f32 (q_in_r89ab, q2_tw2.val[0]);
q_s1_i = vmulq_f32 (q_in_i89ab, q2_tw2.val[0]);
q_s2_r = vmulq_f32 (q_in_rcdef, q2_tw3.val[0]);
q_s2_i = vmulq_f32 (q_in_icdef, q2_tw3.val[0]);
q_s0_r = vmlaq_f32 (q_s0_r, q_in_i4567, q2_tw1.val[1]);
q_s0_i = vmlsq_f32 (q_s0_i, q_in_r4567, q2_tw1.val[1]);
q_s1_r = vmlaq_f32 (q_s1_r, q_in_i89ab, q2_tw2.val[1]);
q_s1_i = vmlsq_f32 (q_s1_i, q_in_r89ab, q2_tw2.val[1]);
q_s2_r = vmlaq_f32 (q_s2_r, q_in_icdef, q2_tw3.val[1]);
q_s2_i = vmlsq_f32 (q_s2_i, q_in_rcdef, q2_tw3.val[1]);
q_s5_r = vsubq_f32 (q_in_r0123, q_s1_r);
q_s5_i = vsubq_f32 (q_in_i0123, q_s1_i);
q2_out_0123.val[0] = vaddq_f32 (q_in_r0123, q_s1_r);
q2_out_0123.val[1] = vaddq_f32 (q_in_i0123, q_s1_i);
q_s3_r = vaddq_f32 (q_s0_r, q_s2_r);
q_s3_i = vaddq_f32 (q_s0_i, q_s2_i);
q_s4_r = vsubq_f32 (q_s0_r, q_s2_r);
q_s4_i = vsubq_f32 (q_s0_i, q_s2_i);
q_one_by_nfft = vdupq_n_f32 (0.0625f);
q2_out_89ab.val[0] = vsubq_f32 (q2_out_0123.val[0], q_s3_r);
q2_out_89ab.val[1] = vsubq_f32 (q2_out_0123.val[1], q_s3_i);
q2_out_0123.val[0] = vaddq_f32 (q2_out_0123.val[0], q_s3_r);
q2_out_0123.val[1] = vaddq_f32 (q2_out_0123.val[1], q_s3_i);
q2_out_4567.val[0] = vsubq_f32 (q_s5_r, q_s4_i);
q2_out_4567.val[1] = vaddq_f32 (q_s5_i, q_s4_r);
q2_out_cdef.val[0] = vaddq_f32 (q_s5_r, q_s4_i);
q2_out_cdef.val[1] = vsubq_f32 (q_s5_i, q_s4_r);
q2_out_89ab.val[0] = vmulq_f32 (q2_out_89ab.val[0], q_one_by_nfft);
q2_out_89ab.val[1] = vmulq_f32 (q2_out_89ab.val[1], q_one_by_nfft);
q2_out_0123.val[0] = vmulq_f32 (q2_out_0123.val[0], q_one_by_nfft);
q2_out_0123.val[1] = vmulq_f32 (q2_out_0123.val[1], q_one_by_nfft);
q2_out_4567.val[0] = vmulq_f32 (q2_out_4567.val[0], q_one_by_nfft);
q2_out_4567.val[1] = vmulq_f32 (q2_out_4567.val[1], q_one_by_nfft);
q2_out_cdef.val[0] = vmulq_f32 (q2_out_cdef.val[0], q_one_by_nfft);
q2_out_cdef.val[1] = vmulq_f32 (q2_out_cdef.val[1], q_one_by_nfft);
vst2q_f32 (p_dst0, q2_out_0123);
vst2q_f32 (p_dst1, q2_out_4567);
vst2q_f32 (p_dst2, q2_out_89ab);
vst2q_f32 (p_dst3, q2_out_cdef);
}
void AudioBlockPanStereoToStereo_NEON(
const float aInputL[WEBAUDIO_BLOCK_SIZE],
const float aInputR[WEBAUDIO_BLOCK_SIZE], float aGainL[WEBAUDIO_BLOCK_SIZE],
float aGainR[WEBAUDIO_BLOCK_SIZE],
const bool aIsOnTheLeft[WEBAUDIO_BLOCK_SIZE],
float aOutputL[WEBAUDIO_BLOCK_SIZE], float aOutputR[WEBAUDIO_BLOCK_SIZE]) {
ASSERT_ALIGNED(aInputL);
ASSERT_ALIGNED(aInputR);
ASSERT_ALIGNED(aGainL);
ASSERT_ALIGNED(aGainR);
ASSERT_ALIGNED(aIsOnTheLeft);
ASSERT_ALIGNED(aOutputL);
ASSERT_ALIGNED(aOutputR);
float32x4_t vinL0, vinL1;
float32x4_t vinR0, vinR1;
float32x4_t voutL0, voutL1;
float32x4_t voutR0, voutR1;
float32x4_t vscaleL0, vscaleL1;
float32x4_t vscaleR0, vscaleR1;
float32x4_t onleft0, onleft1, notonleft0, notonleft1;
float32x4_t zero = vmovq_n_f32(0);
uint8x8_t isOnTheLeft;
// Although MSVC throws uninitialized value warning for voutL0 and voutL1,
// since we fill all lanes by vsetq_lane_f32, we can ignore it. But to avoid
// compiler warning, set zero.
voutL0 = zero;
voutL1 = zero;
for (uint32_t i = 0; i < WEBAUDIO_BLOCK_SIZE; i += 8) {
vinL0 = vld1q_f32(ADDRESS_OF(aInputL, i));
vinL1 = vld1q_f32(ADDRESS_OF(aInputL, i + 4));
vinR0 = vld1q_f32(ADDRESS_OF(aInputR, i));
vinR1 = vld1q_f32(ADDRESS_OF(aInputR, i + 4));
vscaleL0 = vld1q_f32(ADDRESS_OF(aGainL, i));
vscaleL1 = vld1q_f32(ADDRESS_OF(aGainL, i + 4));
vscaleR0 = vld1q_f32(ADDRESS_OF(aGainR, i));
vscaleR1 = vld1q_f32(ADDRESS_OF(aGainR, i + 4));
// Load output with boolean "on the left" values. This assumes that
// bools are stored as a single byte.
isOnTheLeft = vld1_u8((uint8_t*)&aIsOnTheLeft[i]);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 0), voutL0, 0);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 1), voutL0, 1);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 2), voutL0, 2);
voutL0 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 3), voutL0, 3);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 4), voutL1, 0);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 5), voutL1, 1);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 6), voutL1, 2);
voutL1 = vsetq_lane_f32(vget_lane_u8(isOnTheLeft, 7), voutL1, 3);
// Convert the boolean values into masks by setting all bits to 1
// if true.
voutL0 = (float32x4_t)vcgtq_f32(voutL0, zero);
voutL1 = (float32x4_t)vcgtq_f32(voutL1, zero);
// The right output masks are the same as the left masks
voutR0 = voutL0;
voutR1 = voutL1;
// Calculate left channel assuming isOnTheLeft
onleft0 = vmlaq_f32(vinL0, vinR0, vscaleL0);
onleft1 = vmlaq_f32(vinL1, vinR1, vscaleL0);
// Calculate left channel assuming not isOnTheLeft
notonleft0 = vmulq_f32(vinL0, vscaleL0);
notonleft1 = vmulq_f32(vinL1, vscaleL1);
// Write results using previously stored masks
voutL0 = vbslq_f32((uint32x4_t)voutL0, onleft0, notonleft0);
voutL1 = vbslq_f32((uint32x4_t)voutL1, onleft1, notonleft1);
// Calculate right channel assuming isOnTheLeft
onleft0 = vmulq_f32(vinR0, vscaleR0);
onleft1 = vmulq_f32(vinR1, vscaleR1);
// Calculate right channel assuming not isOnTheLeft
notonleft0 = vmlaq_f32(vinR0, vinL0, vscaleR0);
notonleft1 = vmlaq_f32(vinR1, vinL1, vscaleR1);
// Write results using previously stored masks
voutR0 = vbslq_f32((uint32x4_t)voutR0, onleft0, notonleft0);
voutR1 = vbslq_f32((uint32x4_t)voutR1, onleft1, notonleft1);
vst1q_f32(ADDRESS_OF(aOutputL, i), voutL0);
vst1q_f32(ADDRESS_OF(aOutputL, i + 4), voutL1);
vst1q_f32(ADDRESS_OF(aOutputR, i), voutR0);
vst1q_f32(ADDRESS_OF(aOutputR, i + 4), voutR1);
}
}
static void cftmdl_128_neon(float* a) {
int j;
const int l = 8;
const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign);
float32x4_t wk1rv = vld1q_f32(cftmdl_wk1r);
for (j = 0; j < l; j += 2) {
const float32x2_t a_00 = vld1_f32(&a[j + 0]);
const float32x2_t a_08 = vld1_f32(&a[j + 8]);
const float32x2_t a_32 = vld1_f32(&a[j + 32]);
const float32x2_t a_40 = vld1_f32(&a[j + 40]);
const float32x4_t a_00_32 = vcombine_f32(a_00, a_32);
const float32x4_t a_08_40 = vcombine_f32(a_08, a_40);
const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40);
const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40);
const float32x2_t a_16 = vld1_f32(&a[j + 16]);
const float32x2_t a_24 = vld1_f32(&a[j + 24]);
const float32x2_t a_48 = vld1_f32(&a[j + 48]);
const float32x2_t a_56 = vld1_f32(&a[j + 56]);
const float32x4_t a_16_48 = vcombine_f32(a_16, a_48);
const float32x4_t a_24_56 = vcombine_f32(a_24, a_56);
const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56);
const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56);
const float32x4_t xx0 = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1);
const float32x4_t x1_x3_add =
vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x4_t x1_x3_sub =
vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x2_t yy0_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 0);
const float32x2_t yy0_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 0);
const float32x4_t yy0_as = vcombine_f32(yy0_a, yy0_s);
const float32x2_t yy1_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 1);
const float32x2_t yy1_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 1);
const float32x4_t yy1_as = vcombine_f32(yy1_a, yy1_s);
const float32x4_t yy0 = vmlaq_f32(yy0_as, vec_swap_sign, yy1_as);
const float32x4_t yy4 = vmulq_f32(wk1rv, yy0);
const float32x4_t xx1_rev = vrev64q_f32(xx1);
const float32x4_t yy4_rev = vrev64q_f32(yy4);
vst1_f32(&a[j + 0], vget_low_f32(xx0));
vst1_f32(&a[j + 32], vget_high_f32(xx0));
vst1_f32(&a[j + 16], vget_low_f32(xx1));
vst1_f32(&a[j + 48], vget_high_f32(xx1_rev));
a[j + 48] = -a[j + 48];
vst1_f32(&a[j + 8], vget_low_f32(x1_x3_add));
vst1_f32(&a[j + 24], vget_low_f32(x1_x3_sub));
vst1_f32(&a[j + 40], vget_low_f32(yy4));
vst1_f32(&a[j + 56], vget_high_f32(yy4_rev));
}
{
const int k = 64;
const int k1 = 2;
const int k2 = 2 * k1;
const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2 + 0]);
const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2 + 0]);
const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2 + 0]);
const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2 + 0]);
const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2 + 0]);
wk1rv = vld1q_f32(&rdft_wk1r[k2 + 0]);
for (j = k; j < l + k; j += 2) {
const float32x2_t a_00 = vld1_f32(&a[j + 0]);
const float32x2_t a_08 = vld1_f32(&a[j + 8]);
const float32x2_t a_32 = vld1_f32(&a[j + 32]);
const float32x2_t a_40 = vld1_f32(&a[j + 40]);
const float32x4_t a_00_32 = vcombine_f32(a_00, a_32);
const float32x4_t a_08_40 = vcombine_f32(a_08, a_40);
const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40);
const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40);
const float32x2_t a_16 = vld1_f32(&a[j + 16]);
const float32x2_t a_24 = vld1_f32(&a[j + 24]);
const float32x2_t a_48 = vld1_f32(&a[j + 48]);
const float32x2_t a_56 = vld1_f32(&a[j + 56]);
const float32x4_t a_16_48 = vcombine_f32(a_16, a_48);
const float32x4_t a_24_56 = vcombine_f32(a_24, a_56);
const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56);
const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56);
const float32x4_t xx = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1);
const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1);
const float32x4_t x1_x3_add =
vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
const float32x4_t x1_x3_sub =
vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1);
float32x4_t xx4 = vmulq_f32(wk2rv, xx1);
float32x4_t xx12 = vmulq_f32(wk1rv, x1_x3_add);
float32x4_t xx22 = vmulq_f32(wk3rv, x1_x3_sub);
xx4 = vmlaq_f32(xx4, wk2iv, vrev64q_f32(xx1));
xx12 = vmlaq_f32(xx12, wk1iv, vrev64q_f32(x1_x3_add));
xx22 = vmlaq_f32(xx22, wk3iv, vrev64q_f32(x1_x3_sub));
vst1_f32(&a[j + 0], vget_low_f32(xx));
vst1_f32(&a[j + 32], vget_high_f32(xx));
vst1_f32(&a[j + 16], vget_low_f32(xx4));
vst1_f32(&a[j + 48], vget_high_f32(xx4));
vst1_f32(&a[j + 8], vget_low_f32(xx12));
vst1_f32(&a[j + 40], vget_high_f32(xx12));
vst1_f32(&a[j + 24], vget_low_f32(xx22));
vst1_f32(&a[j + 56], vget_high_f32(xx22));
}
}
}
/*******************************************************************************
* PROCEDURE: gaussian_smooth
* PURPOSE: Blur an image with a gaussian filter.
* NAME: Mike Heath
* DATE: 2/15/96
*******************************************************************************/
short int* gaussian_smooth(unsigned char *image, int rows, int cols, float sigma)
{
int r, c, rr, cc, /* Counter variables. */
windowsize, /* Dimension of the gaussian kernel. */
center; /* Half of the windowsize. */
float *tempim,*tempim1, /* Buffer for separable filter gaussian smoothing. */
*kernel, /* A one dimensional gaussian kernel. */
dot, /* Dot product summing variable. */
sum; /* Sum of the kernel weights variable. */
/****************************************************************************
* Create a 1-dimensional gaussian smoothing kernel.
****************************************************************************/
if(VERBOSE) printf(" Computing the gaussian smoothing kernel.\n");
make_gaussian_kernel(sigma, &kernel, &windowsize);
center = windowsize / 2;
/****************************************************************************
* Allocate a temporary buffer image and the smoothed image.
****************************************************************************/
if((tempim = (float *) malloc(rows*cols* sizeof(float))) == NULL)
{
fprintf(stderr, "Error allocating the buffer image.\n");
exit(1);
}
short int* smoothedim;
if(((smoothedim) = (short int *) malloc(rows*cols*sizeof(short int))) == NULL)
{
fprintf(stderr, "Error allocating the smoothed image.\n");
exit(1);
}
startTimer(&totalTime);
//Neon impelementation of gaussian smooth starts here
/****************************************************************************
* Blur in the x - direction.
****************************************************************************/
int loop;
int floop;
//Modification of input image for neon implementation
//For Filter 1
float * new_image;
//For Filter 2
float *new_image_col;
//kernel is changed to 17 from 15 for neon (two 0s at the beginning and the end)
float new_kernel[17];
//Generating now kernel filter
for (floop = 0 ; floop < 17 ; floop++)
{
if(floop == 0 || floop == 16 )
new_kernel[floop] = 0 ;
else
new_kernel [floop] = kernel[floop -1];
}
//For filter 1, new cols number for neon
unsigned int new_cols;
new_cols=cols+16;
unsigned int i, k;
unsigned int a;
unsigned int m;
unsigned int n, j;
//Malloc of new image used by neon
new_image = (float*)malloc(new_cols*rows*sizeof(float));
for( i =0; i<rows; i++){
memset(&new_image[i*new_cols],0,8*sizeof(float));
for( k=0; k<cols;k++){
new_image[i*new_cols+8+k] = (float)image[i*cols+k];
}
memset(&new_image[i*new_cols+8+cols],0,8*sizeof(float));
}
// Neon handles four piexel at a time
float32x4_t neon_input;
float32x4_t neon_filter;
float32x4_t temp_sum;
float32x2_t tempUpper;
float32x2_t tempLower;
float32_t zero = 0;
float32_t temp_output;
float Basekernel = 0.0f;
float kernelSum;
//When using the new filter, we always assume the image has more than 9 pixels in a row
//Base sum for the filter
for( a=8; a<=16; a++){
Basekernel += new_kernel[a];
}
//Filter 1, filtering row by row
for(m=0; m<rows; m++){
for( n=0; n<cols; n++){
temp_sum = vdupq_n_f32(0);
if(n==0){
kernelSum = Basekernel;
}
else if(n <=8){
kernelSum += new_kernel[8-n];
}
else if(n>=cols-8){
kernelSum -=new_kernel[cols-n+8];
}
//For each pixel, filtering is performed four times
for( j=0; j<4; j++)
{
int kk=0;
if(j>=2)
{
kk=1;
}
neon_input = vld1q_f32(&new_image[m*new_cols+n+j*4+kk]);
neon_filter = vld1q_f32(&new_kernel[j*4+kk]);
temp_sum = vmlaq_f32(temp_sum,neon_input,neon_filter);
}
unsigned int t;
for( t=0; t<=3; t++){
temp_output += vgetq_lane_f32(temp_sum,t );
}
temp_output += new_image[m*new_cols+n+8] * new_kernel[8];
temp_output /= kernelSum;
tempim[m*cols+n] = temp_output;
temp_output=0;
}
}
for(r=0; r<rows; r++)
{
for(c=0; c<cols; c++)
{
dot = 0.0;
sum = 0.0;
for(cc=(-center); cc<=center; cc++)
{
if(((c+cc) >= 0) && ((c+cc) < cols))
{
dot += (float)image[r*cols+(c+cc)] * kernel[center+cc];
sum += kernel[center+cc];
}
}
tempim1[r*cols+c] = dot/sum;
}
}
/****************************************************************************
* Blur in the y - direction.
****************************************************************************/
unsigned int new_rows;
new_rows=rows+16;
new_image_col = (float*)malloc(new_rows*cols*sizeof(float));
if(VERBOSE) printf(" Bluring the image in the Y-direction.\n");
for( i =0; i<cols; i++){//actually nember of new rows are the number of columns here
memset(&new_image_col[i*new_rows],0,8*sizeof(float));
for( k=0; k<rows;k++){
new_image_col[i*new_rows+8+k] = tempim[k*cols+i];
//new_image_col[i*new_rows+8+k] = imagetest1[k*cols+i];
}
memset(&new_image_col[i*new_rows+8+rows],0,8*sizeof(float));
}
Basekernel = 0.0;
for( a=8; a<=16; a++){
Basekernel += new_kernel[a];
}
for(m=0; m<cols; m++){// it was rows at br
for( n=0; n<rows; n++){
temp_sum = vdupq_n_f32(0);
if(n==0){
kernelSum = Basekernel;
}
else if(n <=8){
kernelSum += new_kernel[8-n];
}
else if(n>=rows-8){
kernelSum -=new_kernel[rows-n+8];
}
for( j=0; j<4; j++)
{
int kk=0;
if(j>=2)
{
kk=1;
}
neon_input = vld1q_f32(&new_image_col[m*new_rows+n+j*4+kk]);
neon_filter = vld1q_f32(&new_kernel[j*4+kk]);
temp_sum = vmlaq_f32(temp_sum,neon_input,neon_filter);
}
unsigned int t;
for( t=0; t<=3; t++){
temp_output += vgetq_lane_f32(temp_sum,t );
}
temp_output += new_image_col[m*new_rows+n+8] * new_kernel[8];
temp_output = (temp_output * BOOSTBLURFACTOR) / kernelSum + 0.5;
smoothedim[n*cols+m] = (short int )temp_output;
temp_output=0;
}
}
stopTimer(&totalTime);
printTimer(&totalTime);
free(tempim);
free(kernel);
return smoothedim;
}
inline v_float32x4 v_sqr_magnitude(const v_float32x4& a, const v_float32x4& b)
{
return v_float32x4(vmlaq_f32(vmulq_f32(a.val, a.val), b.val, b.val));
}
void meanStdDev(const Size2D &size,
const u16 * srcBase, ptrdiff_t srcStride,
f32 * pMean, f32 * pStdDev)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
size_t blockSize0 = 1 << 10, roiw4 = size.width & ~3;
f64 fsum = 0.0f, fsqsum = 0.0f;
f32 arsum[8];
uint32x4_t v_zero = vdupq_n_u32(0u), v_sum;
float32x4_t v_zero_f = vdupq_n_f32(0.0f), v_sqsum;
for (size_t i = 0; i < size.height; ++i)
{
const u16 * src = internal::getRowPtr(srcBase, srcStride, i);
size_t j = 0u;
while (j < roiw4)
{
size_t blockSize = std::min(roiw4 - j, blockSize0) + j;
v_sum = v_zero;
v_sqsum = v_zero_f;
for ( ; j + 16 < blockSize ; j += 16)
{
internal::prefetch(src + j);
uint16x8_t v_src0 = vld1q_u16(src + j), v_src1 = vld1q_u16(src + j + 8);
// 0
uint32x4_t v_srclo = vmovl_u16(vget_low_u16(v_src0));
uint32x4_t v_srchi = vmovl_u16(vget_high_u16(v_src0));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
float32x4_t v_srclo_f = vcvtq_f32_u32(v_srclo);
float32x4_t v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
// 1
v_srclo = vmovl_u16(vget_low_u16(v_src1));
v_srchi = vmovl_u16(vget_high_u16(v_src1));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
v_srclo_f = vcvtq_f32_u32(v_srclo);
v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
}
for ( ; j < blockSize; j += 4)
{
uint32x4_t v_src = vmovl_u16(vld1_u16(src + j));
float32x4_t v_src_f = vcvtq_f32_u32(v_src);
v_sum = vaddq_u32(v_sum, v_src);
v_sqsum = vmlaq_f32(v_sqsum, v_src_f, v_src_f);
}
vst1q_f32(arsum, vcvtq_f32_u32(v_sum));
vst1q_f32(arsum + 4, v_sqsum);
fsum += (f64)arsum[0] + arsum[1] + arsum[2] + arsum[3];
fsqsum += (f64)arsum[4] + arsum[5] + arsum[6] + arsum[7];
}
// collect a few last elements in the current row
for ( ; j < size.width; ++j)
{
f32 srcval = src[j];
fsum += srcval;
fsqsum += srcval * srcval;
}
}
// calc mean and stddev
f64 itotal = 1.0 / size.total();
f64 mean = fsum * itotal;
f64 stddev = sqrt(std::max(fsqsum * itotal - mean * mean, 0.0));
if (pMean)
*pMean = mean;
if (pStdDev)
*pStdDev = stddev;
#else
(void)size;
(void)srcBase;
(void)srcStride;
(void)pMean;
(void)pStdDev;
#endif
}
inline v_float32x4 v_muladd(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
{
return v_float32x4(vmlaq_f32(c.val, a.val, b.val));
}
static void SubbandCoherenceNEON(AecCore* aec,
float efw[2][PART_LEN1],
float xfw[2][PART_LEN1],
float* fft,
float* cohde,
float* cohxd) {
float dfw[2][PART_LEN1];
int i;
if (aec->delayEstCtr == 0)
aec->delayIdx = PartitionDelay(aec);
// Use delayed far.
memcpy(xfw,
aec->xfwBuf + aec->delayIdx * PART_LEN1,
sizeof(xfw[0][0]) * 2 * PART_LEN1);
// Windowed near fft
WindowData(fft, aec->dBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, dfw);
// Windowed error fft
WindowData(fft, aec->eBuf);
aec_rdft_forward_128(fft);
StoreAsComplex(fft, efw);
SmoothedPSD(aec, efw, dfw, xfw);
{
const float32x4_t vec_1eminus10 = vdupq_n_f32(1e-10f);
// Subband coherence
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_sd = vld1q_f32(&aec->sd[i]);
const float32x4_t vec_se = vld1q_f32(&aec->se[i]);
const float32x4_t vec_sx = vld1q_f32(&aec->sx[i]);
const float32x4_t vec_sdse = vmlaq_f32(vec_1eminus10, vec_sd, vec_se);
const float32x4_t vec_sdsx = vmlaq_f32(vec_1eminus10, vec_sd, vec_sx);
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_cohde = vmulq_f32(vec_sde.val[0], vec_sde.val[0]);
float32x4_t vec_cohxd = vmulq_f32(vec_sxd.val[0], vec_sxd.val[0]);
vec_cohde = vmlaq_f32(vec_cohde, vec_sde.val[1], vec_sde.val[1]);
vec_cohde = vdivq_f32(vec_cohde, vec_sdse);
vec_cohxd = vmlaq_f32(vec_cohxd, vec_sxd.val[1], vec_sxd.val[1]);
vec_cohxd = vdivq_f32(vec_cohxd, vec_sdsx);
vst1q_f32(&cohde[i], vec_cohde);
vst1q_f32(&cohxd[i], vec_cohxd);
}
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
cohde[i] =
(aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) /
(aec->sd[i] * aec->se[i] + 1e-10f);
cohxd[i] =
(aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) /
(aec->sx[i] * aec->sd[i] + 1e-10f);
}
}
void daxpy(int n, REAL da, REAL dx[], int incx, REAL dy[], int incy)
/*
constant times a vector plus a vector.
jack dongarra, linpack, 3/11/78.
*/
{
int i,ix,iy,m,mp1;
mp1 = 0;
m = 0;
if(n <= 0) return;
if (da == ZERO) return;
if(incx != 1 || incy != 1) {
/* code for unequal increments or equal increments
not equal to 1 */
ix = 0;
iy = 0;
if(incx < 0) ix = (-n+1)*incx;
if(incy < 0)iy = (-n+1)*incy;
for (i = 0;i < n; i++) {
dy[iy] = dy[iy] + da*dx[ix];
ix = ix + incx;
iy = iy + incy;
}
return;
}
/* code for both increments equal to 1 */
#ifdef ROLL
for (i = 0;i < n; i++) {
dy[i] = dy[i] + da*dx[i];
}
#endif
#ifdef UNROLL
m = n % 4;
if ( m != 0)
{
for (i = 0; i < m; i++)
dy[i] = dy[i] + da*dx[i];
if (n < 4) return;
}
for (i = m; i < n; i = i + 4)
{
dy[i] = dy[i] + da*dx[i];
dy[i+1] = dy[i+1] + da*dx[i+1];
dy[i+2] = dy[i+2] + da*dx[i+2];
dy[i+3] = dy[i+3] + da*dx[i+3];
}
#endif
#ifdef NEON
float cf[4];
float32x4_t x41, y41, c41, r41;
float32_t *ptrx1 = (float32_t *)dx;
float32_t *ptry1 = (float32_t *)dy;
float32_t *ptrc1 = (float32_t *)cf;
for (i=0; i<4; i++)
{
cf[i] = da;
}
m = n % 4;
if ( m != 0)
{
for (i = 0; i < m; i++)
dy[i] = dy[i] + da*dx[i];
if (n < 4) return;
}
ptrx1 = ptrx1 + m;
ptry1 = ptry1 + m;
c41 = vld1q_f32(ptrc1);
for (i = m; i < n; i=i+4)
{
x41 = vld1q_f32(ptrx1);
y41 = vld1q_f32(ptry1);
r41 = vmlaq_f32(y41, x41, c41);
vst1q_f32(ptry1, r41);
ptrx1 = ptrx1 + 4;
ptry1 = ptry1 + 4;
}
#endif
return;
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1],
int* extreme_filter_divergence) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const float32x4_t vec_15 = vdupq_n_f32(WebRtcAec_kMinFarendPSD);
float32x4_t vec_sdSum = vdupq_n_f32(0.0f);
float32x4_t vec_seSum = vdupq_n_f32(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_dfw0 = vld1q_f32(&dfw[0][i]);
const float32x4_t vec_dfw1 = vld1q_f32(&dfw[1][i]);
const float32x4_t vec_efw0 = vld1q_f32(&efw[0][i]);
const float32x4_t vec_efw1 = vld1q_f32(&efw[1][i]);
const float32x4_t vec_xfw0 = vld1q_f32(&xfw[0][i]);
const float32x4_t vec_xfw1 = vld1q_f32(&xfw[1][i]);
float32x4_t vec_sd = vmulq_n_f32(vld1q_f32(&aec->sd[i]), ptrGCoh[0]);
float32x4_t vec_se = vmulq_n_f32(vld1q_f32(&aec->se[i]), ptrGCoh[0]);
float32x4_t vec_sx = vmulq_n_f32(vld1q_f32(&aec->sx[i]), ptrGCoh[0]);
float32x4_t vec_dfw_sumsq = vmulq_f32(vec_dfw0, vec_dfw0);
float32x4_t vec_efw_sumsq = vmulq_f32(vec_efw0, vec_efw0);
float32x4_t vec_xfw_sumsq = vmulq_f32(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = vmlaq_f32(vec_dfw_sumsq, vec_dfw1, vec_dfw1);
vec_efw_sumsq = vmlaq_f32(vec_efw_sumsq, vec_efw1, vec_efw1);
vec_xfw_sumsq = vmlaq_f32(vec_xfw_sumsq, vec_xfw1, vec_xfw1);
vec_xfw_sumsq = vmaxq_f32(vec_xfw_sumsq, vec_15);
vec_sd = vmlaq_n_f32(vec_sd, vec_dfw_sumsq, ptrGCoh[1]);
vec_se = vmlaq_n_f32(vec_se, vec_efw_sumsq, ptrGCoh[1]);
vec_sx = vmlaq_n_f32(vec_sx, vec_xfw_sumsq, ptrGCoh[1]);
vst1q_f32(&aec->sd[i], vec_sd);
vst1q_f32(&aec->se[i], vec_se);
vst1q_f32(&aec->sx[i], vec_sx);
{
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4_t vec_dfwefw0011 = vmulq_f32(vec_dfw0, vec_efw0);
float32x4_t vec_dfwefw0110 = vmulq_f32(vec_dfw0, vec_efw1);
vec_sde.val[0] = vmulq_n_f32(vec_sde.val[0], ptrGCoh[0]);
vec_sde.val[1] = vmulq_n_f32(vec_sde.val[1], ptrGCoh[0]);
vec_dfwefw0011 = vmlaq_f32(vec_dfwefw0011, vec_dfw1, vec_efw1);
vec_dfwefw0110 = vmlsq_f32(vec_dfwefw0110, vec_dfw1, vec_efw0);
vec_sde.val[0] = vmlaq_n_f32(vec_sde.val[0], vec_dfwefw0011, ptrGCoh[1]);
vec_sde.val[1] = vmlaq_n_f32(vec_sde.val[1], vec_dfwefw0110, ptrGCoh[1]);
vst2q_f32(&aec->sde[i][0], vec_sde);
}
{
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_dfwxfw0011 = vmulq_f32(vec_dfw0, vec_xfw0);
float32x4_t vec_dfwxfw0110 = vmulq_f32(vec_dfw0, vec_xfw1);
vec_sxd.val[0] = vmulq_n_f32(vec_sxd.val[0], ptrGCoh[0]);
vec_sxd.val[1] = vmulq_n_f32(vec_sxd.val[1], ptrGCoh[0]);
vec_dfwxfw0011 = vmlaq_f32(vec_dfwxfw0011, vec_dfw1, vec_xfw1);
vec_dfwxfw0110 = vmlsq_f32(vec_dfwxfw0110, vec_dfw1, vec_xfw0);
vec_sxd.val[0] = vmlaq_n_f32(vec_sxd.val[0], vec_dfwxfw0011, ptrGCoh[1]);
vec_sxd.val[1] = vmlaq_n_f32(vec_sxd.val[1], vec_dfwxfw0110, ptrGCoh[1]);
vst2q_f32(&aec->sxd[i][0], vec_sxd);
}
vec_sdSum = vaddq_f32(vec_sdSum, vec_sd);
vec_seSum = vaddq_f32(vec_seSum, vec_se);
}
{
float32x2_t vec_sdSum_total;
float32x2_t vec_seSum_total;
// A B C D
vec_sdSum_total = vpadd_f32(vget_low_f32(vec_sdSum),
vget_high_f32(vec_sdSum));
vec_seSum_total = vpadd_f32(vget_low_f32(vec_seSum),
vget_high_f32(vec_seSum));
// A+B C+D
vec_sdSum_total = vpadd_f32(vec_sdSum_total, vec_sdSum_total);
vec_seSum_total = vpadd_f32(vec_seSum_total, vec_seSum_total);
// A+B+C+D A+B+C+D
sdSum = vget_lane_f32(vec_sdSum_total, 0);
seSum = vget_lane_f32(vec_seSum_total, 0);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard update.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
// Signal extreme filter divergence if the error is significantly larger
// than the nearend (13 dB).
*extreme_filter_divergence = (seSum > (19.95f * sdSum));
}
static float32x4_t vpowq_f32(float32x4_t a, float32x4_t b) {
// a^b = exp2(b * log2(a))
// exp2(x) and log2(x) are calculated using polynomial approximations.
float32x4_t log2_a, b_log2_a, a_exp_b;
// Calculate log2(x), x = a.
{
// To calculate log2(x), we decompose x like this:
// x = y * 2^n
// n is an integer
// y is in the [1.0, 2.0) range
//
// log2(x) = log2(y) + n
// n can be evaluated by playing with float representation.
// log2(y) in a small range can be approximated, this code uses an order
// five polynomial approximation. The coefficients have been
// estimated with the Remez algorithm and the resulting
// polynomial has a maximum relative error of 0.00086%.
// Compute n.
// This is done by masking the exponent, shifting it into the top bit of
// the mantissa, putting eight into the biased exponent (to shift/
// compensate the fact that the exponent has been shifted in the top/
// fractional part and finally getting rid of the implicit leading one
// from the mantissa by substracting it out.
const uint32x4_t vec_float_exponent_mask = vdupq_n_u32(0x7F800000);
const uint32x4_t vec_eight_biased_exponent = vdupq_n_u32(0x43800000);
const uint32x4_t vec_implicit_leading_one = vdupq_n_u32(0x43BF8000);
const uint32x4_t two_n = vandq_u32(vreinterpretq_u32_f32(a),
vec_float_exponent_mask);
const uint32x4_t n_1 = vshrq_n_u32(two_n, kShiftExponentIntoTopMantissa);
const uint32x4_t n_0 = vorrq_u32(n_1, vec_eight_biased_exponent);
const float32x4_t n =
vsubq_f32(vreinterpretq_f32_u32(n_0),
vreinterpretq_f32_u32(vec_implicit_leading_one));
// Compute y.
const uint32x4_t vec_mantissa_mask = vdupq_n_u32(0x007FFFFF);
const uint32x4_t vec_zero_biased_exponent_is_one = vdupq_n_u32(0x3F800000);
const uint32x4_t mantissa = vandq_u32(vreinterpretq_u32_f32(a),
vec_mantissa_mask);
const float32x4_t y =
vreinterpretq_f32_u32(vorrq_u32(mantissa,
vec_zero_biased_exponent_is_one));
// Approximate log2(y) ~= (y - 1) * pol5(y).
// pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
const float32x4_t C5 = vdupq_n_f32(-3.4436006e-2f);
const float32x4_t C4 = vdupq_n_f32(3.1821337e-1f);
const float32x4_t C3 = vdupq_n_f32(-1.2315303f);
const float32x4_t C2 = vdupq_n_f32(2.5988452f);
const float32x4_t C1 = vdupq_n_f32(-3.3241990f);
const float32x4_t C0 = vdupq_n_f32(3.1157899f);
float32x4_t pol5_y = C5;
pol5_y = vmlaq_f32(C4, y, pol5_y);
pol5_y = vmlaq_f32(C3, y, pol5_y);
pol5_y = vmlaq_f32(C2, y, pol5_y);
pol5_y = vmlaq_f32(C1, y, pol5_y);
pol5_y = vmlaq_f32(C0, y, pol5_y);
const float32x4_t y_minus_one =
vsubq_f32(y, vreinterpretq_f32_u32(vec_zero_biased_exponent_is_one));
const float32x4_t log2_y = vmulq_f32(y_minus_one, pol5_y);
// Combine parts.
log2_a = vaddq_f32(n, log2_y);
}
// b * log2(a)
b_log2_a = vmulq_f32(b, log2_a);
// Calculate exp2(x), x = b * log2(a).
{
// To calculate 2^x, we decompose x like this:
// x = n + y
// n is an integer, the value of x - 0.5 rounded down, therefore
// y is in the [0.5, 1.5) range
//
// 2^x = 2^n * 2^y
// 2^n can be evaluated by playing with float representation.
// 2^y in a small range can be approximated, this code uses an order two
// polynomial approximation. The coefficients have been estimated
// with the Remez algorithm and the resulting polynomial has a
// maximum relative error of 0.17%.
// To avoid over/underflow, we reduce the range of input to ]-127, 129].
const float32x4_t max_input = vdupq_n_f32(129.f);
const float32x4_t min_input = vdupq_n_f32(-126.99999f);
const float32x4_t x_min = vminq_f32(b_log2_a, max_input);
const float32x4_t x_max = vmaxq_f32(x_min, min_input);
// Compute n.
const float32x4_t half = vdupq_n_f32(0.5f);
const float32x4_t x_minus_half = vsubq_f32(x_max, half);
const int32x4_t x_minus_half_floor = vcvtq_s32_f32(x_minus_half);
// Compute 2^n.
const int32x4_t float_exponent_bias = vdupq_n_s32(127);
const int32x4_t two_n_exponent =
vaddq_s32(x_minus_half_floor, float_exponent_bias);
const float32x4_t two_n =
vreinterpretq_f32_s32(vshlq_n_s32(two_n_exponent, kFloatExponentShift));
// Compute y.
const float32x4_t y = vsubq_f32(x_max, vcvtq_f32_s32(x_minus_half_floor));
// Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
const float32x4_t C2 = vdupq_n_f32(3.3718944e-1f);
const float32x4_t C1 = vdupq_n_f32(6.5763628e-1f);
const float32x4_t C0 = vdupq_n_f32(1.0017247f);
float32x4_t exp2_y = C2;
exp2_y = vmlaq_f32(C1, y, exp2_y);
exp2_y = vmlaq_f32(C0, y, exp2_y);
// Combine parts.
a_exp_b = vmulq_f32(exp2_y, two_n);
}
return a_exp_b;
}
static void FilterAdaptationNEON(
int num_partitions,
int x_fft_buf_block_pos,
float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
float e_fft[2][PART_LEN1],
float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1]) {
float fft[PART_LEN2];
int i;
for (i = 0; i < num_partitions; i++) {
int xPos = (i + x_fft_buf_block_pos) * PART_LEN1;
int pos = i * PART_LEN1;
int j;
// Check for wrap
if (i + x_fft_buf_block_pos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// Process the whole array...
for (j = 0; j < PART_LEN; j += 4) {
// Load x_fft_buf and e_fft.
const float32x4_t x_fft_buf_re = vld1q_f32(&x_fft_buf[0][xPos + j]);
const float32x4_t x_fft_buf_im = vld1q_f32(&x_fft_buf[1][xPos + j]);
const float32x4_t e_fft_re = vld1q_f32(&e_fft[0][j]);
const float32x4_t e_fft_im = vld1q_f32(&e_fft[1][j]);
// Calculate the product of conjugate(x_fft_buf) by e_fft.
// re(conjugate(a) * b) = aRe * bRe + aIm * bIm
// im(conjugate(a) * b)= aRe * bIm - aIm * bRe
const float32x4_t a = vmulq_f32(x_fft_buf_re, e_fft_re);
const float32x4_t e = vmlaq_f32(a, x_fft_buf_im, e_fft_im);
const float32x4_t c = vmulq_f32(x_fft_buf_re, e_fft_im);
const float32x4_t f = vmlsq_f32(c, x_fft_buf_im, e_fft_re);
// Interleave real and imaginary parts.
const float32x4x2_t g_n_h = vzipq_f32(e, f);
// Store
vst1q_f32(&fft[2 * j + 0], g_n_h.val[0]);
vst1q_f32(&fft[2 * j + 4], g_n_h.val[1]);
}
// ... and fixup the first imaginary entry.
fft[1] = MulRe(x_fft_buf[0][xPos + PART_LEN],
-x_fft_buf[1][xPos + PART_LEN],
e_fft[0][PART_LEN],
e_fft[1][PART_LEN]);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
const float scale = 2.0f / PART_LEN2;
const float32x4_t scale_ps = vmovq_n_f32(scale);
for (j = 0; j < PART_LEN; j += 4) {
const float32x4_t fft_ps = vld1q_f32(&fft[j]);
const float32x4_t fft_scale = vmulq_f32(fft_ps, scale_ps);
vst1q_f32(&fft[j], fft_scale);
}
}
aec_rdft_forward_128(fft);
{
const float wt1 = h_fft_buf[1][pos];
h_fft_buf[0][pos + PART_LEN] += fft[1];
for (j = 0; j < PART_LEN; j += 4) {
float32x4_t wtBuf_re = vld1q_f32(&h_fft_buf[0][pos + j]);
float32x4_t wtBuf_im = vld1q_f32(&h_fft_buf[1][pos + j]);
const float32x4_t fft0 = vld1q_f32(&fft[2 * j + 0]);
const float32x4_t fft4 = vld1q_f32(&fft[2 * j + 4]);
const float32x4x2_t fft_re_im = vuzpq_f32(fft0, fft4);
wtBuf_re = vaddq_f32(wtBuf_re, fft_re_im.val[0]);
wtBuf_im = vaddq_f32(wtBuf_im, fft_re_im.val[1]);
vst1q_f32(&h_fft_buf[0][pos + j], wtBuf_re);
vst1q_f32(&h_fft_buf[1][pos + j], wtBuf_im);
}
h_fft_buf[1][pos] = wt1;
}
}
}
