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/vadd.h>
15
16
xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16(
18 size_t n,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 #if XNN_ARCH_ARM64
25 const uint8x16_t va_zero_point = vld1q_dup_u8(¶ms->neon.a_zero_point);
26 #else
27 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->neon.a_zero_point);
28 #endif
29 const int32x4_t va_multiplier = vld1q_dup_s32(¶ms->neon.a_multiplier);
30 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
31 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
32 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->neon.output_min);
33 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->neon.output_max);
34
35 const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point;
36 const int32_t vb = params->neon.b_multiplier;
37 const int32x4_t vbias = vdupq_n_s32(vxb * vb);
38
39 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
40 const uint8x16_t va0123456789ABCDEF = vld1q_u8(input_a); input_a += 16;
41
42 #if XNN_ARCH_ARM64
43 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget_low_u8(va_zero_point)));
44 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(va0123456789ABCDEF, va_zero_point));
45 #else // !XNN_ARCH_ARM64
46 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_zero_point));
47 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va0123456789ABCDEF), va_zero_point));
48 #endif // XNN_ARCH_ARM64
49
50 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
51 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
52 int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
53 int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
54
55 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
56 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
57 vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
58 vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
59
60 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
61 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
62
63 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
64
65 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
66
67 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
68
69 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 do {
73 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
74
75 #if XNN_ARCH_ARM64
76 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point)));
77 #else // !XNN_ARCH_ARM64
78 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
79 #endif
80
81 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
82 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
83
84 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
85 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
86
87 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
88
89 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
90 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
91 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
92
93 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
94 vst1_u8(output, vout01234567); output += 8;
95 n -= 8 * sizeof(uint8_t);
96 } else {
97 if (n & (4 * sizeof(uint8_t))) {
98 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
99 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
100 }
101 if (n & (2 * sizeof(uint8_t))) {
102 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
103 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
104 }
105 if (n & (1 * sizeof(uint8_t))) {
106 vst1_lane_u8(output, vout01234567, 0);
107 }
108 n = 0;
109 }
110 } while (n != 0);
111 }
112 }
113