static void FilterAdaptation(aec_t *aec, float *fft, float ef[2][PART_LEN1],
int ip[IP_LEN], float wfft[W_LEN]) {
int i, j;
for (i = 0; i < NR_PART; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos;
// Check for wrap
if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= NR_PART * PART_LEN1;
}
pos = i * PART_LEN1;
#ifdef UNCONSTR
for (j = 0; j < PART_LEN1; j++) {
aec->wfBuf[pos + j][0] += MulRe(aec->xfBuf[xPos + j][0],
-aec->xfBuf[xPos + j][1],
ef[j][0], ef[j][1]);
aec->wfBuf[pos + j][1] += MulIm(aec->xfBuf[xPos + j][0],
-aec->xfBuf[xPos + j][1],
ef[j][0], ef[j][1]);
}
#else
for (j = 0; j < PART_LEN; j++) {
fft[2 * j] = MulRe(aec->xfBuf[0][xPos + j],
-aec->xfBuf[1][xPos + j],
ef[0][j], ef[1][j]);
fft[2 * j + 1] = MulIm(aec->xfBuf[0][xPos + j],
-aec->xfBuf[1][xPos + j],
ef[0][j], ef[1][j]);
}
fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-aec->xfBuf[1][xPos + PART_LEN],
ef[0][PART_LEN], ef[1][PART_LEN]);
aec_rdft_128(-1, fft, ip, wfft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
float scale = 2.0f / PART_LEN2;
for (j = 0; j < PART_LEN; j++) {
fft[j] *= scale;
}
}
aec_rdft_128(1, fft, ip, wfft);
aec->wfBuf[0][pos] += fft[0];
aec->wfBuf[0][pos + PART_LEN] += fft[1];
for (j = 1; j < PART_LEN; j++) {
aec->wfBuf[0][pos + j] += fft[2 * j];
aec->wfBuf[1][pos + j] += fft[2 * j + 1];
}
#endif // UNCONSTR
}
}
static void FilterAdaptation_background(aec_t *aec, float *fft, float ef[2][PART_LEN1]) {
int i, j;
for (i = 0; i < NR_PART; i++) {
int xPos = (i + aec->xfBufBlockPos_background)*(PART_LEN1);
int pos;
// Check for wrap
if (i + aec->xfBufBlockPos_background >= NR_PART) {
xPos -= NR_PART * PART_LEN1;
}
pos = i * PART_LEN1;
for (j = 0; j < PART_LEN1; j++) {
aec->wfBuf_background[0][pos + j] += MulRe(aec->xfBuf_background[0][xPos + j],
-aec->xfBuf_background[1][xPos + j],
ef[0][j], ef[1][j]);
aec->wfBuf_background[1][pos + j] += MulIm(aec->xfBuf_background[0][xPos + j],
-aec->xfBuf_background[1][xPos + j],
ef[0][j], ef[1][j]);
}
}
}
static void FilterAdaptation(aec_t *aec, float *fft, float ef[2][PART_LEN1]) {
int i, j;
for (i = 0; i < NR_PART; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos;
// Check for wrap
if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= NR_PART * PART_LEN1;
}
pos = i * PART_LEN1;
for (j = 0; j < PART_LEN; j++) {
fft[2 * j] = MulRe(aec->xfBuf[0][xPos + j],
-aec->xfBuf[1][xPos + j],
ef[0][j], ef[1][j]);
fft[2 * j + 1] = MulIm(aec->xfBuf[0][xPos + j],
-aec->xfBuf[1][xPos + j],
ef[0][j], ef[1][j]);
}
fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-aec->xfBuf[1][xPos + PART_LEN],
ef[0][PART_LEN], ef[1][PART_LEN]);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
// fft scaling
{
float scale = 2.0f / PART_LEN2;
for (j = 0; j < PART_LEN; j++) {
fft[j] *= scale;
}
}
aec_rdft_forward_128(fft);
aec->wfBuf[0][pos] += fft[0];
aec->wfBuf[0][pos + PART_LEN] += fft[1];
for (j = 1; j < PART_LEN; j++) {
aec->wfBuf[0][pos + j] += fft[2 * j];
aec->wfBuf[1][pos + j] += fft[2 * j + 1];
}
}
}
static void FilterFarSSE2(
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 __m128 x_fft_buf_re = _mm_loadu_ps(&x_fft_buf[0][xPos + j]);
const __m128 x_fft_buf_im = _mm_loadu_ps(&x_fft_buf[1][xPos + j]);
const __m128 h_fft_buf_re = _mm_loadu_ps(&h_fft_buf[0][pos + j]);
const __m128 h_fft_buf_im = _mm_loadu_ps(&h_fft_buf[1][pos + j]);
const __m128 y_fft_re = _mm_loadu_ps(&y_fft[0][j]);
const __m128 y_fft_im = _mm_loadu_ps(&y_fft[1][j]);
const __m128 a = _mm_mul_ps(x_fft_buf_re, h_fft_buf_re);
const __m128 b = _mm_mul_ps(x_fft_buf_im, h_fft_buf_im);
const __m128 c = _mm_mul_ps(x_fft_buf_re, h_fft_buf_im);
const __m128 d = _mm_mul_ps(x_fft_buf_im, h_fft_buf_re);
const __m128 e = _mm_sub_ps(a, b);
const __m128 f = _mm_add_ps(c, d);
const __m128 g = _mm_add_ps(y_fft_re, e);
const __m128 h = _mm_add_ps(y_fft_im, f);
_mm_storeu_ps(&y_fft[0][j], g);
_mm_storeu_ps(&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(
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 FilterFarSSE2(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 __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
const __m128 wfBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
const __m128 wfBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
const __m128 yf_re = _mm_loadu_ps(&yf[0][j]);
const __m128 yf_im = _mm_loadu_ps(&yf[1][j]);
const __m128 a = _mm_mul_ps(xfBuf_re, wfBuf_re);
const __m128 b = _mm_mul_ps(xfBuf_im, wfBuf_im);
const __m128 c = _mm_mul_ps(xfBuf_re, wfBuf_im);
const __m128 d = _mm_mul_ps(xfBuf_im, wfBuf_re);
const __m128 e = _mm_sub_ps(a, b);
const __m128 f = _mm_add_ps(c, d);
const __m128 g = _mm_add_ps(yf_re, e);
const __m128 h = _mm_add_ps(yf_im, f);
_mm_storeu_ps(&yf[0][j], g);
_mm_storeu_ps(&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]);
}
}
}
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]);
}
}
}
static void FilterFar(aec_t *aec, float yf[2][PART_LEN1])
{
int i;
for (i = 0; i < NR_PART; i++) {
int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= NR_PART*(PART_LEN1);
}
for (j = 0; 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]);
}
}
}
static void FilterAdaptationSSE2(aec_t *aec, float *fft, float ef[2][PART_LEN1]) {
int i, j;
for (i = 0; i < NR_PART; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= NR_PART * PART_LEN1;
}
#ifdef UNCONSTR
for (j = 0; j < PART_LEN1; j++) {
aec->wfBuf[pos + j][0] += MulRe(aec->xfBuf[xPos + j][0],
-aec->xfBuf[xPos + j][1],
ef[j][0], ef[j][1]);
aec->wfBuf[pos + j][1] += MulIm(aec->xfBuf[xPos + j][0],
-aec->xfBuf[xPos + j][1],
ef[j][0], ef[j][1]);
}
#else
// Process the whole array...
for (j = 0; j < PART_LEN; j+= 4) {
// Load xfBuf and ef.
const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
const __m128 ef_re = _mm_loadu_ps(&ef[0][j]);
const __m128 ef_im = _mm_loadu_ps(&ef[1][j]);
// Calculate the product of conjugate(xfBuf) by ef.
// re(conjugate(a) * b) = aRe * bRe + aIm * bIm
// im(conjugate(a) * b)= aRe * bIm - aIm * bRe
const __m128 a = _mm_mul_ps(xfBuf_re, ef_re);
const __m128 b = _mm_mul_ps(xfBuf_im, ef_im);
const __m128 c = _mm_mul_ps(xfBuf_re, ef_im);
const __m128 d = _mm_mul_ps(xfBuf_im, ef_re);
const __m128 e = _mm_add_ps(a, b);
const __m128 f = _mm_sub_ps(c, d);
// Interleave real and imaginary parts.
const __m128 g = _mm_unpacklo_ps(e, f);
const __m128 h = _mm_unpackhi_ps(e, f);
// Store
_mm_storeu_ps(&fft[2*j + 0], g);
_mm_storeu_ps(&fft[2*j + 4], h);
}
// ... and fixup the first imaginary entry.
fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-aec->xfBuf[1][xPos + PART_LEN],
ef[0][PART_LEN], ef[1][PART_LEN]);
aec_rdft_inverse_128(fft);
memset(fft + PART_LEN, 0, sizeof(float)*PART_LEN);
// fft scaling
{
float scale = 2.0f / PART_LEN2;
const __m128 scale_ps = _mm_load_ps1(&scale);
for (j = 0; j < PART_LEN; j+=4) {
const __m128 fft_ps = _mm_loadu_ps(&fft[j]);
const __m128 fft_scale = _mm_mul_ps(fft_ps, scale_ps);
_mm_storeu_ps(&fft[j], fft_scale);
}
}
aec_rdft_forward_128(fft);
{
float wt1 = aec->wfBuf[1][pos];
aec->wfBuf[0][pos + PART_LEN] += fft[1];
for (j = 0; j < PART_LEN; j+= 4) {
__m128 wtBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
__m128 wtBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
const __m128 fft0 = _mm_loadu_ps(&fft[2 * j + 0]);
const __m128 fft4 = _mm_loadu_ps(&fft[2 * j + 4]);
const __m128 fft_re = _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2 ,0));
const __m128 fft_im = _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3 ,1));
wtBuf_re = _mm_add_ps(wtBuf_re, fft_re);
wtBuf_im = _mm_add_ps(wtBuf_im, fft_im);
_mm_storeu_ps(&aec->wfBuf[0][pos + j], wtBuf_re);
_mm_storeu_ps(&aec->wfBuf[1][pos + j], wtBuf_im);
}
aec->wfBuf[1][pos] = wt1;
}
#endif // UNCONSTR
}
}
static void FilterAdaptationSSE2(
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, j;
for (i = 0; i < num_partitions; i++) {
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;
}
// Process the whole array...
for (j = 0; j < PART_LEN; j += 4) {
// Load x_fft_buf and e_fft.
const __m128 x_fft_buf_re = _mm_loadu_ps(&x_fft_buf[0][xPos + j]);
const __m128 x_fft_buf_im = _mm_loadu_ps(&x_fft_buf[1][xPos + j]);
const __m128 e_fft_re = _mm_loadu_ps(&e_fft[0][j]);
const __m128 e_fft_im = _mm_loadu_ps(&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 __m128 a = _mm_mul_ps(x_fft_buf_re, e_fft_re);
const __m128 b = _mm_mul_ps(x_fft_buf_im, e_fft_im);
const __m128 c = _mm_mul_ps(x_fft_buf_re, e_fft_im);
const __m128 d = _mm_mul_ps(x_fft_buf_im, e_fft_re);
const __m128 e = _mm_add_ps(a, b);
const __m128 f = _mm_sub_ps(c, d);
// Interleave real and imaginary parts.
const __m128 g = _mm_unpacklo_ps(e, f);
const __m128 h = _mm_unpackhi_ps(e, f);
// Store
_mm_storeu_ps(&fft[2 * j + 0], g);
_mm_storeu_ps(&fft[2 * j + 4], h);
}
// ... 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
{
float scale = 2.0f / PART_LEN2;
const __m128 scale_ps = _mm_load_ps1(&scale);
for (j = 0; j < PART_LEN; j += 4) {
const __m128 fft_ps = _mm_loadu_ps(&fft[j]);
const __m128 fft_scale = _mm_mul_ps(fft_ps, scale_ps);
_mm_storeu_ps(&fft[j], fft_scale);
}
}
aec_rdft_forward_128(fft);
{
float wt1 = h_fft_buf[1][pos];
h_fft_buf[0][pos + PART_LEN] += fft[1];
for (j = 0; j < PART_LEN; j += 4) {
__m128 wtBuf_re = _mm_loadu_ps(&h_fft_buf[0][pos + j]);
__m128 wtBuf_im = _mm_loadu_ps(&h_fft_buf[1][pos + j]);
const __m128 fft0 = _mm_loadu_ps(&fft[2 * j + 0]);
const __m128 fft4 = _mm_loadu_ps(&fft[2 * j + 4]);
const __m128 fft_re =
_mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2, 0));
const __m128 fft_im =
_mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3, 1));
wtBuf_re = _mm_add_ps(wtBuf_re, fft_re);
wtBuf_im = _mm_add_ps(wtBuf_im, fft_im);
_mm_storeu_ps(&h_fft_buf[0][pos + j], wtBuf_re);
_mm_storeu_ps(&h_fft_buf[1][pos + j], wtBuf_im);
}
h_fft_buf[1][pos] = wt1;
}
}
}
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;
}
}
}