1 // Auto-generated file. Do not edit!
2 // Template: src/s16-vlshift/neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11 #include <stddef.h>
12 #include <stdint.h>
13
14 #include <arm_neon.h>
15
16 #include <xnnpack/math.h>
17 #include <xnnpack/vlshift.h>
18
19
xnn_s16_vlshift_ukernel__neon_x32(size_t batch,const int16_t * input,int16_t * output,uint32_t shift)20 void xnn_s16_vlshift_ukernel__neon_x32(
21 size_t batch,
22 const int16_t* input,
23 int16_t* output,
24 uint32_t shift)
25 {
26 assert(batch > 0);
27 assert(input != NULL);
28 assert(output != NULL);
29 assert(shift < 16);
30
31 const int16x8_t vshift = vdupq_n_s16((int16_t) shift);
32
33 for (; batch >= 32; batch -= 32) {
34 const int16x8_t vi0 = vld1q_s16(input); input += 8;
35 const int16x8_t vi1 = vld1q_s16(input); input += 8;
36 const int16x8_t vi2 = vld1q_s16(input); input += 8;
37 const int16x8_t vi3 = vld1q_s16(input); input += 8;
38
39 const int16x8_t vout0 = vshlq_s16(vi0, vshift);
40 const int16x8_t vout1 = vshlq_s16(vi1, vshift);
41 const int16x8_t vout2 = vshlq_s16(vi2, vshift);
42 const int16x8_t vout3 = vshlq_s16(vi3, vshift);
43
44 vst1q_s16(output, vout0); output += 8;
45 vst1q_s16(output, vout1); output += 8;
46 vst1q_s16(output, vout2); output += 8;
47 vst1q_s16(output, vout3); output += 8;
48 }
49
50 // Remainder of full vectors
51 for (; batch >= 8; batch -= 8) {
52 const int16x8_t vi = vld1q_s16(input); input += 8;
53
54 const int16x8_t vout = vshlq_s16(vi, vshift);
55
56 vst1q_s16(output, vout); output += 8;
57 }
58
59 // Remainder of 1 to 7 batch
60 if XNN_UNLIKELY(batch != 0) {
61 const int16x8_t vi = vld1q_s16(input);
62
63 const int16x8_t vout = vshlq_s16(vi, vshift);
64 int16x4_t vout_lo = vget_low_s16(vout);
65
66 if (batch & 4) {
67 vst1_s16(output, vout_lo); output += 4;
68 vout_lo = vget_high_s16(vout);
69 }
70 if (batch & 2) {
71 vst1_lane_u32((void*) output, vreinterpret_u32_s16(vout_lo), 0); output += 2;
72 vout_lo = vext_s16(vout_lo, vout_lo, 2);
73 }
74 if (batch & 1){
75 vst1_lane_s16(output, vout_lo, 0);
76 }
77 }
78 }
79