Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-neon-ld64-u32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/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_vadd_minmax_ukernel__neon_ld64_u32(
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 uint8x8_t vb_zero_point = vld1_dup_u8(&params->scalar.b_zero_point);
32 const int32x4_t va_multiplier = vld1q_dup_s32(&params->scalar.a_multiplier);
33 const int32x4_t vb_multiplier = vld1q_dup_s32(&params->scalar.b_multiplier);
34 const int32x4_t vright_shift = vdupq_n_s32(-params->scalar.shift);
35 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->scalar.output_zero_point);
36 const uint8x16_t voutput_min = vld1q_dup_u8(&params->scalar.output_min);
37 const uint8x16_t voutput_max = vld1q_dup_u8(&params->scalar.output_max);
38
39 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
40 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
41 const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
42 const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8;
43 const uint8x8_t vb89ABCDEF = vld1_u8(input_b); input_b += 8;
44 const uint8x8_t vaGHIJKLMN = vld1_u8(input_a); input_a += 8;
45 const uint8x8_t vbGHIJKLMN = vld1_u8(input_b); input_b += 8;
46 const uint8x8_t vaOPQRSTUV = vld1_u8(input_a); input_a += 8;
47 const uint8x8_t vbOPQRSTUV = vld1_u8(input_b); input_b += 8;
48
49 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
50 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
51 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point));
52 const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEF, vb_zero_point));
53 const int16x8_t vxaGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vaGHIJKLMN, va_zero_point));
54 const int16x8_t vxbGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vbGHIJKLMN, vb_zero_point));
55 const int16x8_t vxaOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vaOPQRSTUV, va_zero_point));
56 const int16x8_t vxbOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vbOPQRSTUV, vb_zero_point));
57
58 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
59 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
60 int32x4_t vacc89AB = vmulq_s32(vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
61 int32x4_t vaccCDEF = vmulq_s32(vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
62 int32x4_t vaccGHIJ = vmulq_s32(vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
63 int32x4_t vaccKLMN = vmulq_s32(vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
64 int32x4_t vaccOPQR = vmulq_s32(vmovl_s16(vget_low_s16(vxaOPQRSTUV)), va_multiplier);
65 int32x4_t vaccSTUV = vmulq_s32(vmovl_s16(vget_high_s16(vxaOPQRSTUV)), va_multiplier);
66
67 vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
68 vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
69 vacc89AB = vmlaq_s32(vacc89AB, vmovl_s16(vget_low_s16(vxb89ABCDEF)), vb_multiplier);
70 vaccCDEF = vmlaq_s32(vaccCDEF, vmovl_s16(vget_high_s16(vxb89ABCDEF)), vb_multiplier);
71 vaccGHIJ = vmlaq_s32(vaccGHIJ, vmovl_s16(vget_low_s16(vxbGHIJKLMN)), vb_multiplier);
72 vaccKLMN = vmlaq_s32(vaccKLMN, vmovl_s16(vget_high_s16(vxbGHIJKLMN)), vb_multiplier);
73 vaccOPQR = vmlaq_s32(vaccOPQR, vmovl_s16(vget_low_s16(vxbOPQRSTUV)), vb_multiplier);
74 vaccSTUV = vmlaq_s32(vaccSTUV, vmovl_s16(vget_high_s16(vxbOPQRSTUV)), vb_multiplier);
75
76 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
77 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
78 vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
79 vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
80 vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
81 vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
82 vaccOPQR = vrshlq_s32(vaccOPQR, vright_shift);
83 vaccSTUV = vrshlq_s32(vaccSTUV, vright_shift);
84
85 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
86 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
87 const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
88 const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV)), voutput_zero_point);
89
90 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
91 uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
92
93 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
94 voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
95
96 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
97 voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
98
99 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
100 vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
101 }
102 if XNN_UNLIKELY(batch != 0) {
103 do {
104 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
105 const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
106
107 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
108 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
109
110 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
111 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
112
113 vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
114 vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
115
116 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
117 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
118
119 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
120
121 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
122 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
123 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
124
125 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
126 vst1_u8(output, vout01234567); output += 8;
127 batch -= 8 * sizeof(uint8_t);
128 } else {
129 if (batch & (4 * sizeof(uint8_t))) {
130 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
131 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
132 }
133 if (batch & (2 * sizeof(uint8_t))) {
134 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
135 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
136 }
137 if (batch & (1 * sizeof(uint8_t))) {
138 vst1_lane_u8(output, vout01234567, 0);
139 }
140 batch = 0;
141 }
142 } while (batch != 0);
143 }
144 }