Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-neon-ld64-u8.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/vbinary.h"
15
16
17 void xnn_qu8_vaddc_minmax_ukernel__neon_ld64_u8(
18 size_t batch,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(batch != 0);
25 assert(batch % sizeof(uint8_t) == 0);
26 assert(input_a != NULL);
27 assert(input_b != NULL);
28 assert(output != NULL);
29
30 const uint8x8_t va_zero_point = vld1_dup_u8(&params->scalar.a_zero_point);
31 const int32x4_t va_multiplier = vld1q_dup_s32(&params->scalar.a_multiplier);
32 const int32x4_t vright_shift = vdupq_n_s32(-params->scalar.shift);
33 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->scalar.output_zero_point);
34 const uint8x8_t voutput_min = vld1_dup_u8(&params->scalar.output_min);
35 const uint8x8_t voutput_max = vld1_dup_u8(&params->scalar.output_max);
36
37 const int32_t vxb = (int32_t) *input_b - (int32_t) params->scalar.b_zero_point;
38 const int32_t vb = params->scalar.b_multiplier;
39 const int32x4_t vbias = vdupq_n_s32(vxb * vb);
40
41 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
42 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
43
44 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
45
46 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
47 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
48
49 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
50 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
51
52 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
53
54 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
55
56 vout01234567 = vmax_u8(vout01234567, voutput_min);
57
58 vout01234567 = vmin_u8(vout01234567, voutput_max);
59
60 vst1_u8(output, vout01234567); output += 8;
61 }
62 if XNN_UNLIKELY(batch != 0) {
63 {
64 const uint8x8_t va01234567 = vld1_u8(input_a);
65
66 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
67
68 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
69 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
70
71 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
72 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
73
74 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
75
76 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
77 vout01234567 = vmax_u8(vout01234567, voutput_min);
78 vout01234567 = vmin_u8(vout01234567, voutput_max);
79
80 if (batch & (4 * sizeof(uint8_t))) {
81 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
82 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
83 }
84 if (batch & (2 * sizeof(uint8_t))) {
85 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
86 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
87 }
88 if (batch & (1 * sizeof(uint8_t))) {
89 vst1_lane_u8(output, vout01234567, 0);
90 }
91 }
92 }
93 }