Debian Codegraph @

src/f16-f32acc-rsum/gen/f16-f32acc-rsum-neonfp16arith-u32-acc4.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32acc-rsum/neonfp16arith.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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
12 #include <arm_neon.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/reduce.h"
16
17
18 void xnn_f16_f32acc_rsum_ukernel__neonfp16arith_u32_acc4(
19 size_t batch,
20 const xnn_float16* input,
21 float* output,
22 const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(batch != 0);
25 assert(batch % sizeof(uint16_t) == 0);
26 assert(input != NULL);
27 assert(output != NULL);
28
29 const uint16_t* i = (const uint16_t*) input;
30 float32x4_t vacc0 = vmovq_n_f32(0.0f);
31 float32x4_t vacc1 = vmovq_n_f32(0.0f);
32 float32x4_t vacc2 = vmovq_n_f32(0.0f);
33 float32x4_t vacc3 = vmovq_n_f32(0.0f);
34 for (; batch >= 32 * sizeof(uint16_t); batch -= 32 * sizeof(uint16_t)) {
35 const float16x8_t vh01 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
36 const float16x8_t vh23 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
37 const float16x8_t vh45 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
38 const float16x8_t vh67 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
39
40 const float32x4_t vt0 = vcvt_f32_f16(vget_low_f16(vh01));
41 const float32x4_t vt1 = vcvt_f32_f16(vget_high_f16(vh01));
42 const float32x4_t vt2 = vcvt_f32_f16(vget_low_f16(vh23));
43 const float32x4_t vt3 = vcvt_f32_f16(vget_high_f16(vh23));
44 const float32x4_t vt4 = vcvt_f32_f16(vget_low_f16(vh45));
45 const float32x4_t vt5 = vcvt_f32_f16(vget_high_f16(vh45));
46 const float32x4_t vt6 = vcvt_f32_f16(vget_low_f16(vh67));
47 const float32x4_t vt7 = vcvt_f32_f16(vget_high_f16(vh67));
48
49 vacc0 = vaddq_f32(vacc0, vt0);
50 vacc1 = vaddq_f32(vacc1, vt1);
51 vacc2 = vaddq_f32(vacc2, vt2);
52 vacc3 = vaddq_f32(vacc3, vt3);
53 vacc0 = vaddq_f32(vacc0, vt4);
54 vacc1 = vaddq_f32(vacc1, vt5);
55 vacc2 = vaddq_f32(vacc2, vt6);
56 vacc3 = vaddq_f32(vacc3, vt7);
57 }
58 vacc0 = vaddq_f32(vacc0, vacc1);
59 vacc2 = vaddq_f32(vacc2, vacc3);
60 vacc0 = vaddq_f32(vacc0, vacc2);
61 for (; batch >= 4 * sizeof(uint16_t); batch -= 4 * sizeof(uint16_t)) {
62 const float16x4_t vh = vreinterpret_f16_u16(vld1_u16(i)); i += 4;
63 const float32x4_t vt = vcvt_f32_f16(vh);
64 vacc0 = vaddq_f32(vacc0, vt);
65 }
66 const float32x2_t vscale = vld1_dup_f32(&params->scalar.scale);
67 float32x2_t vacc = vadd_f32(vget_low_f32(vacc0), vget_high_f32(vacc0));
68 if XNN_UNLIKELY(batch & (2 * sizeof(uint16_t))) {
69 const float16x4_t vh = vreinterpret_f16_u32(vld1_dup_u32((const void*) i)); i += 2;
70 const float32x4_t vt = vcvt_f32_f16(vh);
71 vacc = vadd_f32(vacc, vget_low_f32(vt));
72 }
73 vacc = vpadd_f32(vacc, vacc);
74 if XNN_UNLIKELY(batch & (1 * sizeof(uint16_t))) {
75 const float16x4_t vh = vreinterpret_f16_u16(vld1_dup_u16(i));
76 const float32x4_t vt = vcvt_f32_f16(vh);
77 vacc = vadd_f32(vacc, vget_low_f32(vt));
78 }
79 vacc = vmul_f32(vacc, vscale);
80
81 float vout = vget_lane_f32(vacc, 0);
82 *output += vout;
83 }