xref: /aosp_15_r20/external/XNNPACK/src/s16-window/gen/neon-shift12-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_shift12_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_shift12_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 == 12);
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       int32x4_t vacc0_lo = vmull_s16(vget_low_s16(vi0), vget_low_s16(vw0));
49       int32x4_t vacc0_hi = vmull_s16(vget_high_s16(vi0), vget_high_s16(vw0));
50       int32x4_t vacc1_lo = vmull_s16(vget_low_s16(vi1), vget_low_s16(vw1));
51       int32x4_t vacc1_hi = vmull_s16(vget_high_s16(vi1), vget_high_s16(vw1));
52       int32x4_t vacc2_lo = vmull_s16(vget_low_s16(vi2), vget_low_s16(vw2));
53       int32x4_t vacc2_hi = vmull_s16(vget_high_s16(vi2), vget_high_s16(vw2));
54 
55       const int16x4_t vshift0_lo = vqshrn_n_s32(vacc0_lo, 12);
56       const int16x4_t vshift0_hi = vqshrn_n_s32(vacc0_hi, 12);
57       const int16x4_t vshift1_lo = vqshrn_n_s32(vacc1_lo, 12);
58       const int16x4_t vshift1_hi = vqshrn_n_s32(vacc1_hi, 12);
59       const int16x4_t vshift2_lo = vqshrn_n_s32(vacc2_lo, 12);
60       const int16x4_t vshift2_hi = vqshrn_n_s32(vacc2_hi, 12);
61 
62       const int16x8_t vout0 = vcombine_s16(vshift0_lo, vshift0_hi);
63       const int16x8_t vout1 = vcombine_s16(vshift1_lo, vshift1_hi);
64       const int16x8_t vout2 = vcombine_s16(vshift2_lo, vshift2_hi);
65 
66       vst1q_s16(output, vout0); output += 8;
67       vst1q_s16(output, vout1); output += 8;
68       vst1q_s16(output, vout2); output += 8;
69     }
70 
71     // Remainder of full vectors
72     for (; n >= 8 * sizeof(int16_t); n -= 8 * sizeof(int16_t)) {
73       const int16x8_t vi = vld1q_s16(input); input += 8;
74       const int16x8_t vw = vld1q_s16(w); w += 8;
75       int32x4_t vacc_lo = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
76       int32x4_t vacc_hi = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
77       const int16x4_t vshift_lo = vqshrn_n_s32(vacc_lo, 12);
78       const int16x4_t vshift_hi = vqshrn_n_s32(vacc_hi, 12);
79       const int16x8_t vout = vcombine_s16(vshift_lo, vshift_hi);
80       vst1q_s16(output, vout); output += 8;
81     }
82 
83     assert(n % 2 == 0);
84     // Remainder of 1 to 7 batch_size
85     if XNN_UNLIKELY(n != 0) {
86       const int16x8_t vi = vld1q_s16(input); input = (const int16_t*) ((uintptr_t) input + n);
87       const int16x8_t vw = vld1q_s16(w);
88       int32x4_t vacc = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
89       int16x4_t vout = vqshrn_n_s32(vacc, 12);
90       if (n & (4 * sizeof(int16_t))) {
91         vst1_s16(output, vout); output += 4;
92         vacc = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
93         vout = vqshrn_n_s32(vacc, 12);
94       }
95       if (n & (2 * sizeof(int16_t))) {
96         vst1_lane_u32((void*) output, vreinterpret_u32_s16(vout), 0); output += 2;
97         vout = vext_s16(vout, vout, 2);
98       }
99       if (n & (1 * sizeof(int16_t))) {
100         vst1_lane_s16(output, vout, 0); output += 1;
101       }
102     }
103 
104   } while (--rows != 0);
105 }
106