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 <smmintrin.h>
13
14 #include <xnnpack/vaddsub.h>
15
16
xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vadd_minmax_ukernel__sse41_mul16_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_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul16.bias);
25 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
26 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
27 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_lo);
28 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_hi);
29 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
33
34 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
35 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
37 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
38 const __m128i vb89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
39 input_a += 16;
40 input_b += 16;
41
42
43 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
44 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
45 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
46 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
47 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
48 __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
49 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
50 const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
51
52 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
53 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
54 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
55 vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
56
57 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
58 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
59 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
60 vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
61
62 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
63 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
64 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
65 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
66
67 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
68 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
69 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
70 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
71
72 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
73 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
74 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
75 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
76
77 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
78 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
79
80
81 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
82
83 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
84
85 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
86
87 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
88 output += 16;
89 }
90 if XNN_UNLIKELY(n != 0) {
91 do {
92 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
93 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
94 input_a += 8;
95 input_b += 8;
96
97
98 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
99 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
100 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
101 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
102
103 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
104 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
105
106 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
107 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
108
109 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
110 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
111
112 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
113 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
114
115 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
116 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
117
118 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
119
120 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
121 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
122 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
123
124 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
125 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
126 output += 8;
127 n -= 8 * sizeof(int8_t);
128 } else {
129 if (n & (4 * sizeof(int8_t))) {
130 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
131 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
132 output += 4;
133 }
134 if (n & (2 * sizeof(int8_t))) {
135 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
136 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
137 output += 2;
138 }
139 if (n & (1 * sizeof(int8_t))) {
140 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
141 }
142 n = 0;
143 }
144 } while (n != 0);
145 }
146 }
147