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_qs8_vaddc_minmax_ukernel__neon_ld64_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16(
18 size_t n,
19 const int8_t* input_a,
20 const int8_t* input_b,
21 int8_t* output,
22 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const int8x8_t va_zero_point = vld1_dup_s8(¶ms->neon.a_zero_point);
25 const int32x4_t va_multiplier = vld1q_dup_s32(¶ms->neon.a_multiplier);
26 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
27 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
28 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
29 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
30
31 const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point;
32 const int32_t vb = params->neon.b_multiplier;
33 const int32x4_t vbias = vdupq_n_s32(vxb * vb);
34
35 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
36 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
37 const int8x8_t va89ABCDEF = vld1_s8(input_a); input_a += 8;
38
39 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
40 const int16x8_t vxa89ABCDEF = vsubl_s8(va89ABCDEF, va_zero_point);
41
42 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
43 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
44 int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
45 int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
46
47 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
48 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
49 vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
50 vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
51
52 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
53 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
54
55 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
56
57 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
58
59 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
60
61 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 do {
65 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
66
67 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
68
69 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
70 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
71
72 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
73 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
74
75 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
76
77 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
78 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
79 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
80
81 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
82 vst1_s8(output, vout01234567); output += 8;
83 n -= 8 * sizeof(int8_t);
84 } else {
85 if (n & (4 * sizeof(int8_t))) {
86 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
87 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
88 }
89 if (n & (2 * sizeof(int8_t))) {
90 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
91 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
92 }
93 if (n & (1 * sizeof(int8_t))) {
94 vst1_lane_s8(output, vout01234567, 0);
95 }
96 n = 0;
97 }
98 } while (n != 0);
99 }
100 }
101