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_x8(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_x8(
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 >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
35 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
36 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
37 input_a += 8;
38 input_b += 8;
39
40 const __m128i vzero = _mm_setzero_si128();
41 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
42 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
43
44 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
45 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
46 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
47 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
48
49 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
50 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
51
52
53 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
54 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
55
56 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
57 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
58
59 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
60 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
61
62 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
63
64
65 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
66
67 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
68
69 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
70
71 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
72 output += 8;
73 }
74 if XNN_UNLIKELY(n != 0) {
75 {
76 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
77 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
78
79 const __m128i vzero = _mm_setzero_si128();
80 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
81 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
82
83 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
84 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
85 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
86 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
87
88 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
89 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
90
91
92 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
93 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
94
95 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
96 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
97
98 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
99 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
100
101 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
102
103 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
104 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
105 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
106
107 if (n & (4 * sizeof(uint8_t))) {
108 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
109 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
110 output += 4;
111 }
112 if (n & (2 * sizeof(uint8_t))) {
113 *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
114 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
115 output += 2;
116 }
117 if (n & (1 * sizeof(uint8_t))) {
118 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
119 }
120 }
121 }
122 }
123