1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/sse-mul16-ld64.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 <emmintrin.h>
13
14 #include <xnnpack/vaddsub.h>
15
16
xnn_qu8_vadd_minmax_ukernel__sse2_mul16_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vadd_minmax_ukernel__sse2_mul16_ld64_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_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
25 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
26 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
27 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo);
28 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_hi);
29 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
33
34 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
35 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
36 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
37 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
38 __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
39 input_a += 16;
40 input_b += 16;
41
42 const __m128i vzero = _mm_setzero_si128();
43 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
44 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
45 va89ABCDEF = _mm_unpacklo_epi8(va89ABCDEF, vzero);
46 vb89ABCDEF = _mm_unpacklo_epi8(vb89ABCDEF, vzero);
47
48 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
49 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
50 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
51 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
52 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
53 __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
54 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
55 const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
56
57 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
58 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
59 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
60 vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
61
62
63 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
64 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
65 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
66 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
67
68 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
69 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
70 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
71 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
72
73 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
74 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
75 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
76 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
77
78 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
79 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
80
81
82 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
83
84 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
85
86 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
87
88 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
89 output += 16;
90 }
91 if XNN_UNLIKELY(n != 0) {
92 do {
93 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
94 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
95 input_a += 8;
96 input_b += 8;
97
98 const __m128i vzero = _mm_setzero_si128();
99 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
100 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
101
102 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
103 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
104 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
105 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
106
107 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
108 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
109
110
111 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
112 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
113
114 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
115 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
116
117 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
118 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
119
120 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
121
122 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
123 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
124 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
125
126 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
127 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
128 output += 8;
129 n -= 8 * sizeof(uint8_t);
130 } else {
131 if (n & (4 * sizeof(uint8_t))) {
132 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
133 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
134 output += 4;
135 }
136 if (n & (2 * sizeof(uint8_t))) {
137 *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
138 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
139 output += 2;
140 }
141 if (n & (1 * sizeof(uint8_t))) {
142 *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
143 }
144 n = 0;
145 }
146 } while (n != 0);
147 }
148 }
149