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_x12(size_t batch,const int16_t * input,uint32_t * output)19 void xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x12(
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 >= 12; batch -= 12) {
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
33 int32x4_t vacc0 = vmull_s16(vi0.val[0], vi0.val[0]);
34 vacc0 = vmlal_s16(vacc0, vi0.val[1], vi0.val[1]);
35 int32x4_t vacc1 = vmull_s16(vi1.val[0], vi1.val[0]);
36 vacc1 = vmlal_s16(vacc1, vi1.val[1], vi1.val[1]);
37 int32x4_t vacc2 = vmull_s16(vi2.val[0], vi2.val[0]);
38 vacc2 = vmlal_s16(vacc2, vi2.val[1], vi2.val[1]);
39
40 vst1q_u32(output, vreinterpretq_u32_s32(vacc0)); output += 4;
41 vst1q_u32(output, vreinterpretq_u32_s32(vacc1)); output += 4;
42 vst1q_u32(output, vreinterpretq_u32_s32(vacc2)); output += 4;
43 }
44
45 // Remainder of full vectors
46 for (; batch >= 4; batch -= 4) {
47 const int16x4x2_t vi = vld2_s16(input); input += 8;
48
49 int32x4_t vacc = vmull_s16(vi.val[0], vi.val[0]);
50
51 vacc = vmlal_s16(vacc, vi.val[1], vi.val[1]);
52
53 vst1q_u32(output, vreinterpretq_u32_s32(vacc)); output += 4;
54 }
55
56 // Remainder of 1 to 3 elements
57 if XNN_UNLIKELY(batch != 0) {
58 const int16x4x2_t vi = vld2_s16(input);
59
60 int32x4_t vacc = vmull_s16(vi.val[0], vi.val[0]);
61 vacc = vmlal_s16(vacc, vi.val[1], vi.val[1]);
62
63 uint32x2_t vacc_lo = vreinterpret_u32_s32(vget_low_s32(vacc));
64 if (batch & 2) {
65 vst1_u32(output, vacc_lo); output += 2;
66 vacc_lo = vreinterpret_u32_s32(vget_high_s32(vacc));
67 }
68 if (batch & 1) {
69 vst1_lane_u32(output, vacc_lo, 0);
70 }
71 }
72 }
73