1 // Copyright 2022 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7 #include <stddef.h>
8 #include <stdint.h>
9
10 #include <xnnpack/math.h>
11 #include <xnnpack/fft.h>
12
13 #include <arm_neon.h>
14
15
xnn_cs16_bfly4_samples1_ukernel__neon(size_t samples,int16_t * data,const size_t stride,const int16_t * twiddle)16 void xnn_cs16_bfly4_samples1_ukernel__neon(
17 size_t samples,
18 int16_t* data,
19 const size_t stride,
20 const int16_t* twiddle)
21 {
22 assert(samples == 1);
23 assert(data != NULL);
24 assert(stride != 0);
25 assert(twiddle != NULL);
26
27 const int16x8_t vi = vld1q_s16(data);
28 const int16x8_t vdiv4 = vdupq_n_s16(8191);
29 const int16x8_t vout = vqrdmulhq_s16(vi, vdiv4);
30
31 const int16x4_t vtmp5 = vsub_s16(vget_low_s16(vout), vget_high_s16(vout));
32 int16x4_t vout0 = vadd_s16(vget_low_s16(vout), vget_high_s16(vout));
33
34 const int16x4_t vtmp3 = vadd_s16(vget_low_s16(vout), vget_high_s16(vout));
35 const int16x4_t vtmp4 = vsub_s16(vget_low_s16(vout), vget_high_s16(vout));
36
37 const int16x4_t vtmp3hi = vext_s16(vtmp3, vtmp3, 2);
38 const int16x4_t vout2 = vsub_s16(vout0, vtmp3hi);
39 vout0 = vadd_s16(vout0, vtmp3hi);
40 const int16x4_t vtmp4rev = vrev64_s16(vtmp4);
41 const int16x4_t vout1r3i = vadd_s16(vtmp5, vtmp4rev);
42 const int16x4_t vout3r1i = vsub_s16(vtmp5, vtmp4rev);
43
44 vst1_lane_u32((void*) data, vreinterpret_u32_s16(vout0), 0); data += 2;
45 vst1_lane_s16(data, vout1r3i, 0); data += 1;
46 vst1_lane_s16(data, vout3r1i, 1); data += 1;
47 vst1_lane_u32((void*) data, vreinterpret_u32_s16(vout2), 0); data += 2;
48 vst1_lane_s16(data, vout3r1i, 0); data += 1;
49 vst1_lane_s16(data, vout1r3i, 1);
50 }
51