1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x24(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)])18 void xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x24(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m128i vbias = _mm_add_epi32(
26 _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
27 _mm_load_si128((const __m128i*) params->sse2.bias));
28 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
29 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
34
35 for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
36 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
37 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
38 __m128i vaGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input_a + 16));
39 input_a += 24;
40
41 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
42 va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
43 vaGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vaGHIJKLMN, vaGHIJKLMN), 8);
44
45 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
46 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
47 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
48 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
49 __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
50 const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
51
52 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
53 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
54 vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
55
56 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
57 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
58 vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
59
60 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
61 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
62 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
63 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
64 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
65 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
66
67 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
68 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
69 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
70 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
71 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
72 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
73
74 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
75 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
76 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
77
78 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
79 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
80 voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
81
82 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
83 vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
84 voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
85
86 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
87 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
88
89
90 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
91 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
92 output += 24;
93 }
94 if XNN_UNLIKELY(n != 0) {
95 do {
96 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
97 input_a += 8;
98
99 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
100
101 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
102 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
103
104 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
105
106 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
107
108 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
109 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
110
111 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
112 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
113
114 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
115 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
116 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
117
118 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
119
120 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
121 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
122 output += 8;
123 n -= 8 * sizeof(int8_t);
124 } else {
125 if (n & (4 * sizeof(int8_t))) {
126 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
127 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
128 output += 4;
129 }
130 if (n & (2 * sizeof(int8_t))) {
131 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
132 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
133 output += 2;
134 }
135 if (n & (1 * sizeof(int8_t))) {
136 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
137 }
138 n = 0;
139 }
140 } while (n != 0);
141 }
142 }
143