float32x4x2_t test_vzipq_f32(float32x4_t a, float32x4_t b) {
// CHECK-LABEL: test_vzipq_f32
return vzipq_f32(a, b);
// CHECK: zip1 {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
// CHECK: zip2 {{v[0-9]+}}.4s, {{v[0-9]+}}.4s
}
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;
}
}
}
static void rftbsub_128_neon(float* a) {
const float* c = rdft_w + 32;
int j1, j2;
const float32x4_t mm_half = vdupq_n_f32(0.5f);
a[1] = -a[1];
// Vectorized code (four at once).
// Note: commented number are indexes for the first iteration of the loop.
for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) {
// Load 'wk'.
const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4,
const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31,
const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31,
const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28,
const float32x4_t wki_ = c_j1; // 1, 2, 3, 4,
// Load and shuffle 'a'.
// 2, 4, 6, 8, 3, 5, 7, 9
float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]);
// 120, 122, 124, 126, 121, 123, 125, 127,
const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]);
// 126, 124, 122, 120
const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]);
// 127, 125, 123, 121
const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]);
// Calculate 'x'.
const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0);
// 2-126, 4-124, 6-122, 8-120,
const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1);
// 3-127, 5-125, 7-123, 9-121,
// Calculate product into 'y'.
// yr = wkr * xr - wki * xi;
// yi = wkr * xi + wki * xr;
const float32x4_t a_ = vmulq_f32(wkr_, xr_);
const float32x4_t b_ = vmulq_f32(wki_, xi_);
const float32x4_t c_ = vmulq_f32(wkr_, xi_);
const float32x4_t d_ = vmulq_f32(wki_, xr_);
const float32x4_t yr_ = vaddq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120,
const float32x4_t yi_ = vsubq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121,
// Update 'a'.
// a[j2 + 0] -= yr;
// a[j2 + 1] -= yi;
// a[k2 + 0] += yr;
// a[k2 + 1] -= yi;
// 126, 124, 122, 120,
const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_);
// 127, 125, 123, 121,
const float32x4_t a_k2_p1n = vsubq_f32(yi_, a_k2_p1);
// Shuffle in right order and store.
// 2, 3, 4, 5, 6, 7, 8, 9,
const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n);
const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n);
// 124, 125, 126, 127, 120, 121, 122, 123
const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr);
// 2, 4, 6, 8,
a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_);
// 3, 5, 7, 9,
a_j2_p.val[1] = vsubq_f32(yi_, a_j2_p.val[1]);
// 2, 3, 4, 5, 6, 7, 8, 9,
vst2q_f32(&a[0 + j2], a_j2_p);
vst1q_f32(&a[122 - j2], a_k2_n.val[1]);
vst1q_f32(&a[126 - j2], a_k2_n.val[0]);
}
// Scalar code for the remaining items.
for (; j2 < 64; j1 += 1, j2 += 2) {
const int k2 = 128 - j2;
const int k1 = 32 - j1;
const float wkr = 0.5f - c[k1];
const float wki = c[j1];
const float xr = a[j2 + 0] - a[k2 + 0];
const float xi = a[j2 + 1] + a[k2 + 1];
const float yr = wkr * xr + wki * xi;
const float yi = wkr * xi - wki * xr;
a[j2 + 0] = a[j2 + 0] - yr;
a[j2 + 1] = yi - a[j2 + 1];
a[k2 + 0] = yr + a[k2 + 0];
a[k2 + 1] = yi - a[k2 + 1];
}
a[65] = -a[65];
}
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);
}