Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-neon-ld64-u16.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_u16(
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 uint8x16_t voutput_min = vld1q_dup_u8(&params->scalar.output_min);
35 const uint8x16_t voutput_max = vld1q_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 >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
42 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
43 const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8;
44
45 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
46 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point));
47
48 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
49 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
50 int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
51 int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
52
53 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
54 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
55 vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
56 vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
57
58 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
59 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
60
61 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
62
63 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
64
65 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
66
67 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
68 }
69 if XNN_UNLIKELY(batch != 0) {
70 do {
71 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
72
73 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
74
75 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
76 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
77
78 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
79 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
80
81 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
82
83 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
84 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
85 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
86
87 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
88 vst1_u8(output, vout01234567); output += 8;
89 batch -= 8 * sizeof(uint8_t);
90 } else {
91 if (batch & (4 * sizeof(uint8_t))) {
92 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
93 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
94 }
95 if (batch & (2 * sizeof(uint8_t))) {
96 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
97 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
98 }
99 if (batch & (1 * sizeof(uint8_t))) {
100 vst1_lane_u8(output, vout01234567, 0);
101 }
102 batch = 0;
103 }
104 } while (batch != 0);
105 }
106 }