1 // Auto-generated file. Do not edit!
2 // Template: src/cs16-vsquareabs/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/vsquareabs.h>
17
18
xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x16(size_t batch,const int16_t * input,uint32_t * output)19 void xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x16(
20 size_t batch,
21 const int16_t* input,
22 uint32_t* output) {
23
24 assert(batch != 0);
25 assert(input != NULL);
26 assert(output != NULL);
27
28 for (; batch >= 16; batch -= 16) {
29 const int16x4x2_t vi0 = vld2_s16(input); input += 8;
30 const int16x4x2_t vi1 = vld2_s16(input); input += 8;
31 const int16x4x2_t vi2 = vld2_s16(input); input += 8;
32 const int16x4x2_t vi3 = vld2_s16(input); input += 8;
33
34 int32x4_t vacc0 = vmull_s16(vi0.val[0], vi0.val[0]);
35 vacc0 = vmlal_s16(vacc0, vi0.val[1], vi0.val[1]);
36 int32x4_t vacc1 = vmull_s16(vi1.val[0], vi1.val[0]);
37 vacc1 = vmlal_s16(vacc1, vi1.val[1], vi1.val[1]);
38 int32x4_t vacc2 = vmull_s16(vi2.val[0], vi2.val[0]);
39 vacc2 = vmlal_s16(vacc2, vi2.val[1], vi2.val[1]);
40 int32x4_t vacc3 = vmull_s16(vi3.val[0], vi3.val[0]);
41 vacc3 = vmlal_s16(vacc3, vi3.val[1], vi3.val[1]);
42
43 vst1q_u32(output, vreinterpretq_u32_s32(vacc0)); output += 4;
44 vst1q_u32(output, vreinterpretq_u32_s32(vacc1)); output += 4;
45 vst1q_u32(output, vreinterpretq_u32_s32(vacc2)); output += 4;
46 vst1q_u32(output, vreinterpretq_u32_s32(vacc3)); output += 4;
47 }
48
49 // Remainder of full vectors
50 for (; batch >= 4; batch -= 4) {
51 const int16x4x2_t vi = vld2_s16(input); input += 8;
52
53 int32x4_t vacc = vmull_s16(vi.val[0], vi.val[0]);
54
55 vacc = vmlal_s16(vacc, vi.val[1], vi.val[1]);
56
57 vst1q_u32(output, vreinterpretq_u32_s32(vacc)); output += 4;
58 }
59
60 // Remainder of 1 to 3 elements
61 if XNN_UNLIKELY(batch != 0) {
62 const int16x4x2_t vi = vld2_s16(input);
63
64 int32x4_t vacc = vmull_s16(vi.val[0], vi.val[0]);
65 vacc = vmlal_s16(vacc, vi.val[1], vi.val[1]);
66
67 uint32x2_t vacc_lo = vreinterpret_u32_s32(vget_low_s32(vacc));
68 if (batch & 2) {
69 vst1_u32(output, vacc_lo); output += 2;
70 vacc_lo = vreinterpret_u32_s32(vget_high_s32(vacc));
71 }
72 if (batch & 1) {
73 vst1_lane_u32(output, vacc_lo, 0);
74 }
75 }
76 }
77