xref: /aosp_15_r20/external/XNNPACK/src/s16-window/gen/neon-shift12-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/s16-window/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/window.h>
18 
19 
xnn_s16_window_shift12_ukernel__neon_x8(size_t rows,size_t batch_size,const int16_t * input,const int16_t * weights,int16_t * output,uint32_t shift)20 void xnn_s16_window_shift12_ukernel__neon_x8(
21     size_t rows,
22     size_t batch_size,
23     const int16_t* input,
24     const int16_t* weights,
25     int16_t* output,
26     uint32_t shift)
27 {
28   assert(rows != 0);
29   assert(batch_size != 0);
30   assert(input != NULL);
31   assert(weights != NULL);
32   assert(output != NULL);
33   assert(shift == 12);
34 
35 
36   do {
37     const int16_t* w = weights;
38     size_t n = batch_size * sizeof(int16_t);
39 
40     // Remainder of full vectors
41     for (; n >= 8 * sizeof(int16_t); n -= 8 * sizeof(int16_t)) {
42       const int16x8_t vi = vld1q_s16(input); input += 8;
43       const int16x8_t vw = vld1q_s16(w); w += 8;
44       int32x4_t vacc_lo = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
45       int32x4_t vacc_hi = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
46       const int16x4_t vshift_lo = vqshrn_n_s32(vacc_lo, 12);
47       const int16x4_t vshift_hi = vqshrn_n_s32(vacc_hi, 12);
48       const int16x8_t vout = vcombine_s16(vshift_lo, vshift_hi);
49       vst1q_s16(output, vout); output += 8;
50     }
51 
52     assert(n % 2 == 0);
53     // Remainder of 1 to 7 batch_size
54     if XNN_UNLIKELY(n != 0) {
55       const int16x8_t vi = vld1q_s16(input); input = (const int16_t*) ((uintptr_t) input + n);
56       const int16x8_t vw = vld1q_s16(w);
57       int32x4_t vacc = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
58       int16x4_t vout = vqshrn_n_s32(vacc, 12);
59       if (n & (4 * sizeof(int16_t))) {
60         vst1_s16(output, vout); output += 4;
61         vacc = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
62         vout = vqshrn_n_s32(vacc, 12);
63       }
64       if (n & (2 * sizeof(int16_t))) {
65         vst1_lane_u32((void*) output, vreinterpret_u32_s16(vout), 0); output += 2;
66         vout = vext_s16(vout, vout, 2);
67       }
68       if (n & (1 * sizeof(int16_t))) {
69         vst1_lane_s16(output, vout, 0); output += 1;
70       }
71     }
72 
73   } while (--rows != 0);
74 }
75