xref: /aosp_15_r20/external/XNNPACK/src/s16-window/gen/neon-shift15-x24.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_shift15_ukernel__neon_x24(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_shift15_ukernel__neon_x24(
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 == 15);
34 
35 
36   do {
37     const int16_t* w = weights;
38     size_t n = batch_size * sizeof(int16_t);
39     for (; n >= 24 * sizeof(int16_t); n -= 24 * sizeof(int16_t)) {
40       const int16x8_t vi0 = vld1q_s16(input); input += 8;
41       const int16x8_t vi1 = vld1q_s16(input); input += 8;
42       const int16x8_t vi2 = vld1q_s16(input); input += 8;
43 
44       const int16x8_t vw0 = vld1q_s16(w); w += 8;
45       const int16x8_t vw1 = vld1q_s16(w); w += 8;
46       const int16x8_t vw2 = vld1q_s16(w); w += 8;
47 
48       const int16x8_t vout0 = vqdmulhq_s16(vi0, vw0);
49       const int16x8_t vout1 = vqdmulhq_s16(vi1, vw1);
50       const int16x8_t vout2 = vqdmulhq_s16(vi2, vw2);
51 
52       vst1q_s16(output, vout0); output += 8;
53       vst1q_s16(output, vout1); output += 8;
54       vst1q_s16(output, vout2); output += 8;
55     }
56 
57     // Remainder of full vectors
58     for (; n >= 8 * sizeof(int16_t); n -= 8 * sizeof(int16_t)) {
59       const int16x8_t vi = vld1q_s16(input); input += 8;
60       const int16x8_t vw = vld1q_s16(w); w += 8;
61       const int16x8_t vout = vqdmulhq_s16(vi, vw);
62       vst1q_s16(output, vout); output += 8;
63     }
64 
65     assert(n % 2 == 0);
66     // Remainder of 1 to 7 batch_size
67     if XNN_UNLIKELY(n != 0) {
68       const int16x8_t vi = vld1q_s16(input); input = (const int16_t*) ((uintptr_t) input + n);
69       const int16x8_t vw = vld1q_s16(w);
70       int16x4_t vout = vqdmulh_s16(vget_low_s16(vi), vget_low_s16(vw));
71       if (n & (4 * sizeof(int16_t))) {
72         vst1_s16(output, vout); output += 4;
73         vout = vqdmulh_s16(vget_high_s16(vi), vget_high_s16(vw));
74       }
75       if (n & (2 * sizeof(int16_t))) {
76         vst1_lane_u32((void*) output, vreinterpret_u32_s16(vout), 0); output += 2;
77         vout = vext_s16(vout, vout, 2);
78       }
79       if (n & (1 * sizeof(int16_t))) {
80         vst1_lane_s16(output, vout, 0); output += 1;
81       }
82     }
83 
84   } while (--rows != 0);
85 }
86