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/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qs8_vaddc_minmax_ukernel__sse41_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__sse41_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->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
27 _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
28 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
29 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
34
35 for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
36 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
38 const __m128i vaGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
39 input_a += 24;
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
49 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
50 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
51 vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
52
53 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
54 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
55 vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
56
57 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
58 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
59 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
60 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
61 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
62 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
63
64 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
65 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
66 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
67 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
68 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
69 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
70
71 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
72 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
73 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
74
75
76 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
77 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
78
79 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
80 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
81
82 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
83 voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
84
85 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
86 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
87 output += 24;
88 }
89 if XNN_UNLIKELY(n != 0) {
90 do {
91 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
92 input_a += 8;
93
94
95 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
96 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
97
98 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
99
100 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
101
102 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
103 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
104
105 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
106 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
107
108 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
109
110 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
111 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
112 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
113
114 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
115 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
116 output += 8;
117 n -= 8 * sizeof(int8_t);
118 } else {
119 if (n & (4 * sizeof(int8_t))) {
120 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
121 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
122 output += 4;
123 }
124 if (n & (2 * sizeof(int8_t))) {
125 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
126 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
127 output += 2;
128 }
129 if (n & (1 * sizeof(int8_t))) {
130 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
131 }
132 n = 0;
133 }
134 } while (n != 0);
135 }
136 }
137