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_qu8_vaddc_minmax_ukernel__avx_mul16_ld64_x8(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_vaddc_minmax_ukernel__avx_mul16_ld64_x8(
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_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 >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
36 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 input_a += 8;
38
39
40 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
41 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
42
43 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
44
45
46 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
47 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
48
49 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
50 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
51
52 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
53
54
55 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
56
57 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
58
59 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
60
61 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
62 output += 8;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 {
66 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
67
68
69 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
70 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
71
72 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
73
74
75 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
76 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
77
78 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
79 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
80
81 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
82
83 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
84 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
85 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
86
87 if (n & (4 * sizeof(uint8_t))) {
88 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
89 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
90 output += 4;
91 }
92 if (n & (2 * sizeof(uint8_t))) {
93 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
94 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
95 output += 2;
96 }
97 if (n & (1 * sizeof(uint8_t))) {
98 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
99 }
100 }
101 }
102 }
103