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_ld128_x32(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_ld128_x32(
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 #if XNN_ARCH_ARM64
25 const int8x16_t va_zero_point = vld1q_dup_s8(¶ms->neon.a_zero_point);
26 #else
27 const int8x8_t va_zero_point = vld1_dup_s8(¶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 int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
33 const int8x16_t voutput_max = vld1q_dup_s8(¶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 >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
40 const int8x16_t va0123456789ABCDEF = vld1q_s8(input_a); input_a += 16;
41 const int8x16_t vaGHIJKLMNOPQRSTUV = vld1q_s8(input_a); input_a += 16;
42
43 #if XNN_ARCH_ARM64
44 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point));
45 const int16x8_t vxa89ABCDEF = vsubl_high_s8(va0123456789ABCDEF, va_zero_point);
46 const int16x8_t vxaGHIJKLMN = vsubl_s8(vget_low_s8(vaGHIJKLMNOPQRSTUV), vget_low_s8(va_zero_point));
47 const int16x8_t vxaOPQRSTUV = vsubl_high_s8(vaGHIJKLMNOPQRSTUV, va_zero_point);
48 #else // !XNN_ARCH_ARM64
49 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point);
50 const int16x8_t vxa89ABCDEF = vsubl_s8(vget_high_s8(va0123456789ABCDEF), va_zero_point);
51 const int16x8_t vxaGHIJKLMN = vsubl_s8(vget_low_s8(vaGHIJKLMNOPQRSTUV), va_zero_point);
52 const int16x8_t vxaOPQRSTUV = vsubl_s8(vget_high_s8(vaGHIJKLMNOPQRSTUV), va_zero_point);
53 #endif // XNN_ARCH_ARM64
54
55 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
56 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
57 int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
58 int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
59 int32x4_t vaccGHIJ = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
60 int32x4_t vaccKLMN = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
61 int32x4_t vaccOPQR = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxaOPQRSTUV)), va_multiplier);
62 int32x4_t vaccSTUV = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxaOPQRSTUV)), va_multiplier);
63
64 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
65 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
66 vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
67 vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
68 vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
69 vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
70 vaccOPQR = vrshlq_s32(vaccOPQR, vright_shift);
71 vaccSTUV = vrshlq_s32(vaccSTUV, vright_shift);
72
73 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
74 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
75 const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
76 const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV)), voutput_zero_point);
77
78 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
79 int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV));
80
81 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
82 voutGHIJKLMNOPQRSTUV = vmaxq_s8(voutGHIJKLMNOPQRSTUV, voutput_min);
83
84 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
85 voutGHIJKLMNOPQRSTUV = vminq_s8(voutGHIJKLMNOPQRSTUV, voutput_max);
86
87 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
88 vst1q_s8(output, voutGHIJKLMNOPQRSTUV); output += 16;
89 }
90 if XNN_UNLIKELY(n != 0) {
91 do {
92 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
93
94 #if XNN_ARCH_ARM64
95 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point));
96 #else // !XNN_ARCH_ARM64
97 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
98 #endif
99
100 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
101 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
102
103 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
104 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
105
106 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
107
108 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
109 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
110 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
111
112 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
113 vst1_s8(output, vout01234567); output += 8;
114 n -= 8 * sizeof(int8_t);
115 } else {
116 if (n & (4 * sizeof(int8_t))) {
117 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
118 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
119 }
120 if (n & (2 * sizeof(int8_t))) {
121 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
122 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
123 }
124 if (n & (1 * sizeof(int8_t))) {
125 vst1_lane_s8(output, vout01234567, 0);
126 }
127 n = 0;
128 }
129 } while (n != 0);
130 }
131 }
132