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 <smmintrin.h>
13
14 #include <xnnpack/vaddsub.h>
15
16
xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x32(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_vaddc_minmax_ukernel__sse41_mul16_ld64_x32(
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_add_epi32(
25 _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
26 _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
27 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
28 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_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 >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 const __m128i vaGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
38 const __m128i vaOPQRSTUV = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 24)));
39 input_a += 32;
40
41
42 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
43 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
44 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
45 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
46 __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
47 const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
48 __m128i vaprodOPQRSTUVhi = _mm_mulhi_epu16(vaOPQRSTUV, va_multiplier_lo);
49 const __m128i vaprodOPQRSTUVlo = _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_lo);
50
51 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
52 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
53 vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
54 vaprodOPQRSTUVhi = _mm_add_epi16(vaprodOPQRSTUVhi, _mm_mullo_epi16(vaOPQRSTUV, 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 vaprodOPQRSTUVhi = _mm_sub_epi16(vaprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vaOPQRSTUV, 15), va_multiplier_lo));
60
61 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
62 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
63 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
64 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
65 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
66 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
67 __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
68 __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
69
70 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
71 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
72 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
73 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
74 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
75 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
76 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
77 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
78
79 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
80 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
81 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
82 __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
83
84
85 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
86 __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
87
88 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
89 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
90
91 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
92 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
93
94 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
95 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
96 output += 32;
97 }
98 if XNN_UNLIKELY(n != 0) {
99 do {
100 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
101 input_a += 8;
102
103
104 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
105 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
106
107 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
108
109 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
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_sra_epi32(vacc0123, vshift);
115 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
116
117 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
118
119 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
120 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
121 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
122
123 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
124 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
125 output += 8;
126 n -= 8 * sizeof(int8_t);
127 } else {
128 if (n & (4 * sizeof(int8_t))) {
129 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
130 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
131 output += 4;
132 }
133 if (n & (2 * sizeof(int8_t))) {
134 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
135 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
136 output += 2;
137 }
138 if (n & (1 * sizeof(int8_t))) {
139 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
140 }
141 n = 0;
142 }
143 } while (n != 0);
144 }
145 }
146