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/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qu8_vadd_minmax_ukernel__sse41_mul16_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vadd_minmax_ukernel__sse41_mul16_ld64_x16(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
26 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
27 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
28 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo);
29 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_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 >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
36 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
38 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
39 const __m128i vb89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
40 input_a += 16;
41 input_b += 16;
42
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 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
49 __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
50 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
51 const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
52
53 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
54 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
55 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
56 vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
57
58
59 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
60 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
61 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
62 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
63
64 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
65 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
66 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
67 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
68
69 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
70 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
71 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
72 vaccCDEF = _mm_sra_epi32(vaccCDEF, 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
77
78 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
79
80 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
81
82 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
83
84 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
85 output += 16;
86 }
87 if XNN_UNLIKELY(n != 0) {
88 do {
89 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
90 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
91 input_a += 8;
92 input_b += 8;
93
94
95 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
96 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
97 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
98 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
99
100 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
101 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
102
103
104 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
105 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
106
107 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
109
110 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
111 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
112
113 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
114
115 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
116 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
117 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
118
119 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
120 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
121 output += 8;
122 n -= 8 * sizeof(uint8_t);
123 } else {
124 if (n & (4 * sizeof(uint8_t))) {
125 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
126 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
127 output += 4;
128 }
129 if (n & (2 * sizeof(uint8_t))) {
130 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
131 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
132 output += 2;
133 }
134 if (n & (1 * sizeof(uint8_t))) {
135 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
136 }
137 n = 0;
138 }
139 } while (n != 0);
140 }
141 }
142